Secure personalized super-intelligence (PSI)

Abstract

Personalized super intelligence (PSI) represents the next leap of advanced, autonomous, artificial intelligence agent development. However, due to their extreme intelligence, PSI also represents a dangerous potential threat to human safety. The invention discloses how to securely design and construct such an agent. It also shows how they are securely used as part of a super-intelligent Universal Artificial Intelligence (AGI) system. Preferred implementations are disclosed, including methods that enable a PSI to quickly improve itself with or without human supervision. The present invention describes how different versions of PSI are produced using several novel methods. Also described are methods for enabling scalable security checks that operate efficiently even when the PSIs become much intelligent than their human creators. Finally, a brand new AI security method is disclosed, which depends on the combination of a PSI group and a proven block chain method. The method does not depend on testing, but improves the PSI security through design.

Description

TECHNICAL FIELD

[0001] In some aspects, the technology relates to a secure personalized super-intelligent (PSI) for use in conjunction with the development and continuous improvement of each human owner’s PSI using intelligent agents. In some other aspects, the technology relates to methods associated with the creation of a PSI with a group of multiple artificial intelligence (AI) agents or systems and / or multiple pre-customized PSIs all communicating on a collective network that have all agreed to a set of agreed upon values and ethical rules regarding values and ethics that reflect the values and ethics of the group of PSIs and / or their human users.

[0002] In another aspect, the technology relates to a PSI that operates based on knowledge and values pre-trained into the PSI, which at least initially relies on information provided by or related to a human user.

[0003] In yet another aspect, the technology relates to the use of a group of AI agents and / or PSIs to perform a security or ethical check on other PSIs on the network to ensure that any tasks or activities performed by the PSIs follow the agreed upon set of rules.

[0004] In other aspects, all of the activities described in this patent disclosure as occurring on an external network of multiple intelligent entities participating in collaborative problem solving can also be implemented within or reside within an AI agent or PSI within a single computerized intelligent system where all of the intelligent entities are computerized. BACKGROUND

[0005] Prior patent applications describe the development path of super-intelligent general artificial intelligence (AGI) - super-intelligent AGI and individual advanced autonomous artificial intelligence (AAAI). They describe specific scenarios involving existing products and technologies available from several companies. They describe ways that combining data and learning from cross-platform AAAI implementations can accelerate the learning and skills of each AAAI. They describe systems and methods for integration in general technical terms and explain how AAAIs can be integrated into an AGI network through a human-centered AGI approach.

[0006] The field of AI was named in 1956 at a conference in Dartmouth, Massachusetts, USA, organized by computer scientist John McCarthy. Among the researchers who attended the Dartmouth conference were Herbert A. Simon (future Nobel laureate) and Allen Newell (future distinguished computer scientist), both from Carnegie Mellon University.

[0007] Simon and Newell, along with their colleague Cliff Shaw, presented the only working demonstration of AI at the Dartmouth Conference. It was a program called the Logic Theorist. The Logic Theorist was an example of an early state of AI work in which the rules defining the behavior of the AI were programmed directly into the computer by human programmers. Interestingly, by programming the rules in a general way so as to allow the computer program to pursue goals and subgoals in various ways (called "operators"), the Logic Theorist was able to prove creative behavior.

[0008] Specifically, although it was programmed to recreate mathematical proofs from Bertrand Russell and Alfred North Whitehead's textbook, Principia Mathematica, the Logic Theorist actually discovered a new proof that was previously unknown to the programmers of the Logic Theorist as well as to Russell and Whitehead themselves. Russell and Whitehead were reportedly impressed by the Logic Theorist's new proof and wrote to the inventors that they not only did not know the Logic Theorist's proof before, but that they wished they had thought of it themselves! Thus, at the birth of the field of AI in 1956, AI was already capable of creative thought. Of particular relevance to the present patent is the Logic Theorist's use of goals and subgoals - a method that Newell and Simon later developed further, which was later adopted by many AI systems, and which the present patent applies in a novel and creative way.

[0009] From 1956 to 1986, research in the field of AI was dominated by the "expert system" approach. A person with programming skills would interview a human expert and embody that expert's knowledge in a series of programmed rules for the AI. This process was called "knowledge engineering." The result of knowledge engineering was an AI program that could perform like a human expert in a limited domain. For example, in the 1970s a program called MYCIN was developed at Stanford University to act as an expert system in the field of blood infections. E. A. Feigenbaum et al. at Stanford University continued to develop a whole suite of expert systems in various medical fields in the 1980s. Similar work in expert systems was also done at many other universities.

[0010] As more and more expert systems were being developed, Newell and Simon turned their attention to improving the performance of AI systems by looking at the best model of intelligence that existed - humans. Their research led to a very powerful and broad theory that could rigorously describe the way humans solve almost any type of problem. This theory, which built on their early work with the Logic Theorist, was called “searching the problem space.” The theory was described in detail in their book, Human Problem Solving, published in 1972.

[0011] Dr. Craig Kaplan, the inventor of the AAAI patent, studied with Herbert A. Simon and Allen Newell in the 1980s. He co-authored research with Dr. Simon in the field of creative problem solving and cognitive science, including the 1989 article “The Cognitive Science Foundation.” Dr. Kaplan recognized that the “searching the problem space” architecture proposed by Newell and Simon could be generalized to enable the collective problem solving of millions of people over the Internet. Starting in the late 1990s, Dr. Kaplan began putting his ideas into practice in various work systems that actively utilized the collective intelligence of humans.

[0012] For example, Dr. Kaplan pioneered some of the first practical applications of crowdsourced intelligence around 2000. In 2001, in a presentation at the first Global Brain Conference in Brussels, he outlined the idea of applying collective intelligence to one of the most difficult and competitive problems in business - beating Wall Street. By 2006, he designed and implemented the “PredictWallStreet” system, which utilized the collective intelligence of millions of people to gain an advantage in the stock market. In 2018, the system powered one of the top ten market neutral hedge funds, proving its effectiveness in performing at the highest level in a complex field competing against some of the smartest people on Earth.

[0013] In the process of designing and implementing these systems, Dr. Kaplan realized that the “search problem space” architecture, as a general framework for human problem solving, could be adapted and enhanced to serve as a general cognitive architecture for both humans and AI agents. Moreover, representing intelligent behavior as a form of problem solving provides a way for many AI agents to interact with each other, thereby pooling their collective intelligence to create an AGI. This “collective intelligence” approach, presented here as the AAAI system and method for AGI, represents a faster and more powerful path to AGI than existing efforts. Most existing efforts to achieve AGI have focused primarily on training larger large language models (LLMs) using more data, more powerful computers, and better machine learning algorithms. The AAAI approach also has the advantage of enabling humans to easily participate in training and improving the intelligence of the AI, including helping to shape the values and ethics of the AI - a fundamental feature of ensuring the safe development of AGI.

[0014] While Dr. Kaplan recognized the importance of collective intelligence early on, most other AI researchers became increasingly focused on a subfield of AI called machine learning (ML). Starting in the 1980s, ML began to emerge as a way for AI to learn knowledge on its own, rather than having knowledge engineers program knowledge into artificial intelligence. However, progress in ML was very slow until a 1986 paper was published that showed how to use “backpropagation of error” - one of the first practical reinforcement learning skills. After that paper, some AI researchers saw that the future of AI would depend on machines learning on their own, rather than being programmed by humans. Unfortunately, the computational and data requirements of ML were so large that they far exceeded the capabilities of technology in the 1980s and even the 1990s.

[0015] It took about thirty years of Moore’s Law - doubling of computing power every 18 months or so - to catch up with the capabilities required by ML algorithms before this time. During this period, the amount of data available to train such models began to increase, especially on the internet (which took off after 1995 with the advent of web browsers).

[0016] From the 1990s to the first decade of the 21st century, there was an “AI winter” due to overly optimistic aspirations for AI that outpaced the readily available data and computer power. However, by 2010, there was a confluence of massive amounts of computing power, data, and “good enough” ML algorithms. Progress in AI began to accelerate rapidly, including the development of improved learning algorithms such as “transformers”.

[0017] As of early 2023, knowledge engineering methods for creating expert systems were largely ignored, while popular machine learning methods had successfully enabled machines to learn on their own how to defeat the best human champions in chess, Go, and any two-player game. Programs like AlphaFold determined the shapes of millions of proteins in just a few months, whereas in the past it would have taken the best human experts four to six years to accurately determine the shape of a single protein. Natural Language Processing (NLP)—the subfield of AI focused on understanding human language—has made tremendous progress, resulting in assistants like Amazon's Alexa, Apple's Siri, and more recently, large language models (LLMs) from OpenAI's GPT.

[0018] The current technology, extensions, and improvements of LLM represent a watershed moment, enabling AI to move from being a specialized tool of interest to specific domains (also known as “narrow AI”) to more general applications. With OpenAI releasing CHATGPT, followed by Google® releasing BARD, integrating GPT into Microsoft’s Bing search engine, and the proliferation of AI companies focusing on the broad application of ML methodologies, a wave of innovation is sweeping through AI applications. Many individual fields, including medical applications, vehicle navigation, office work, legal work, marketing, sales, education, and even brewing, have been revolutionized by the application of LLM (and more broadly, advancements in machine learning methods and capabilities).

[0019] However, one goal remains elusive. As of February 28, 2023, no company or individual, other than those detailed in this specification, has explained how to create a practical system for AGI. The reason is that machine learning alone is insufficient for rapid AGI implementation. Collective intelligence is also required. Summary of the Invention

[0020] In view of the aforementioned drawbacks inherent in known PSI methods, at least some embodiments of the present technology provide a novel implementation or creation of a secure PSI, overcoming one or more of the aforementioned disadvantages and defects of the prior art. Therefore, the general objective of at least some embodiments of the present technology, which will be described in more detail below, is to provide a new and novel secure PSI that possesses all the advantages of the prior art mentioned herein, as well as numerous novel features that produce a secure PSI that cannot be anticipated, made apparent, revealed, or even implied by the prior art alone or in any combination thereof.

[0021] According to one aspect, the technology may include a system for Personalized Superintelligence (PSI) that uses intelligent agents to develop and continuously improve the PSI of human users utilizing a computer system, and uses additional PSI, all of which are electronically communicated via a collective network. The system may include a computer system comprising a processor, a computer-readable storage medium, and program instructions stored on the computer-readable storage medium, the program instructions being executable by the processor to cause the computer system to:

[0022] Implement a basic-level AI agent on a computer system, where the basic-level AI agent has been customized;

[0023] Collect media information related to human users;

[0024] Analyze media information;

[0025] Transform the analyzed media information into a training dataset;

[0026] Differential weighting is applied to the transformed training dataset;

[0027] Add the knowledge module to the weighted training dataset;

[0028] Locate new data sources to include in the weighted training dataset;

[0029] The weighted training dataset was applied to a basic AI agent to create a personalized PSI; and

[0030] Using a network enables a personalized PSI to communicate with multiple additional PSIs to achieve community-based security features from multiple additional PSIs to a personalized PSI.

[0031] According to another aspect, this technology may include a method for PSI that uses intelligent entities to develop and continuously improve PSI, wherein the intelligent entities include human users utilizing computer systems, additional AI agents, and additional PSI, all of which communicate electronically via a collective network. The method includes the following steps:

[0032] Acquire a previously customized basic AI agent;

[0033] Collect media information related to human users;

[0034] Analyze media information;

[0035] Transform the analyzed media information into a training dataset;

[0036] Differential weighting is applied to the transformed training dataset;

[0037] Add the knowledge module to the weighted training dataset;

[0038] Locate new data sources to include in the weighted training dataset;

[0039] The weighted training dataset is applied to the base-level AI agent to create a user PSI, where the base-level AI agent; and

[0040] Using a network enables the user PSI to communicate with multiple additional PSIs to achieve community-based security features from multiple additional PSIs to the user PSI.

[0041] Some implementations of this technology may include the following steps: obtaining, by a human user, a base-level AI agent, or a user PSI, any one or any combination of a new training dataset and training modules to be added to the weighted transformed training dataset.

[0042] Some implementations of this technology may include the following steps: monetizing user PSI by enabling other human users or other PSIs to access and use a weighted training dataset of personalized PSIs.

[0043] In some implementations, the base-level AI agent can be any of a pre-trained large language model (LLM), other tuned and trainable AI agents, or one or more custom AIs from other human owners.

[0044] In some implementations, media information can be any one or any combination of the following: videos of human users, videos of people and topics related to human users, photos or images of human users, photos or images of people and topics related to human users, articles related to human users, journals related to human users, blogs related to human users, posts related to human users, tweets related to human users, emails related to human users, podcasts related to human users, records related to human users, auditory content related to human users, auditory content of people and topics related to human users, data related to human users collected by third-party vendors, websites related to human users, applications related to human users, online information related to human users, social media information related to human users, and other AI agents related to human users.

[0045] Some implementations of this technology may include the following steps: granting one or more social media platforms a license that allows social media content relevant to human users to be accessed by a base-level AI agent or user PSI.

[0046] In some implementations, the analysis of media information may also include annotating and classifying the media information.

[0047] In some implementations, the analysis of media information may utilize one or more algorithms, including any or any combination of transformer algorithms, deep learning algorithms, transcription algorithms, content and sentiment analysis and summarization, LLM, and crowdsourced and crowd-supervised human methods.

[0048] Some implementations of this technology may include the following steps: one or more human workers on a crowdsourcing website who review the training dataset or suggest improvements to the training dataset are compensated by human users.

[0049] In some implementations, the step of weighting the transformed training dataset can be performed by a human user using an interface on a computer system implementing a basic AI agent to adjust the weight value of any one of the elements in the transformed training dataset.

[0050] In some implementations, the step of weighting the transformed training dataset can be performed by any one or any combination of an additional AI agent, an additional PSI, and additional human workers, each utilizing a computer system.

[0051] Some implementations of this technology may include the following steps: inputting ethical values ​​into a training dataset by providing a series of interactive dialogues to human users, the interactive dialogues including predetermined ethical scenarios.

[0052] Some implementations of this technology may include the following steps: mixing the weights of the transformed training dataset to improve the LLM of the base-level AI agent.

[0053] Some implementations of this technology may include the following steps: a human user authorizes the LLM to download data using the interface of the basic AI agent and automatically adds the data to the training dataset or creates a new training dataset.

[0054] In some implementations, the downloaded data may be made available to human users via a computer system for human approval before being added to or creating a new training dataset.

[0055] Some implementations of this technology may include the following steps: cloning a user PSI into one or more cloned user PSIs.

[0056] Some implementations of this technology may include the following steps: training each clone PSI using training datasets that are different from each other, wherein the training datasets of each clone PSI have different weights from each other.

[0057] Some implementations of this technology may include the following steps: combining a weighted training dataset from a user PSI and one or more cloned PSIs to create a combined training dataset.

[0058] Some implementations of this technology may include the following steps: providing any one or any combination of weighted training datasets of any one or any combination of cloned PSIs to the additional AI agent and any of the additional PSIs for training.

[0059] Some implementations of this technology may include the following steps: training a user PSI using a weighted training dataset of any or any combination of cloned PSIs.

[0060] Some implementations of this technology may include the following steps: an automated process in which a human user of a computer system performs a user PSI task without human intervention unless predetermined parameters are triggered.

[0061] Some implementations of this technology may include the following steps: a basic-level AI agent or user PSI monitors activities based on a list of prohibited activities and triggers intervention activities provided to human users.

[0062] Some implementations of this technology may include the following steps: detecting whether there are gaps in the training dataset by a base-level AI agent or user PSI; if there are gaps, generating interactions with intelligent entities to obtain data to fill the gaps.

[0063] In some implementations, interactions between more than one smart entity can be performed in parallel.

[0064] Some implementations of this technology may include the following steps: providing tasks to multiple additional AI agents or additional PSIs on the network, providing the results of the tasks from each of the additional AI agents or additional PSIs, and determining the winning result based on the results.

[0065] In some implementations, the attached AI agent or attached PSI can process tasks independently and in parallel with each other.

[0066] In some implementations, one or more additional AI agents or additional PSIs may be uncustomized, and one or more additional AI agents or additional PSIs may be customized.

[0067] Some implementations of this technology may include the following steps: utilizing a training dataset of an additional AI agent or additional PSI with winning results in the customization of a user-based AI agent or user PSI.

[0068] Some implementations of this technology may include the following steps: each of the attached AI agent or attached PSI utilizes a general problem-solving architecture on the task.

[0069] Some implementations of this technology may include the following steps: using a set of security or ethical rules that are agreed upon by each of the user-based AI agent, user PSI, additional AI agent, and additional PSI.

[0070] In some implementations, each of the additional AI agent or the additional PSI may have a set of weights that encode the knowledge of the additional AI agent and the additional PSI, respectively.

[0071] In some implementations, any computer device on the network can use the activity log.

[0072] Some implementations of this technology may include the following steps: providing a task to a user PSI, the task being configured to generate hallucinations by the PSI's LLM.

[0073] Some implementations of this technology may include the following steps: providing a task to a user PSI, wherein the task is a random arrangement of existing standardized tasks or a task dynamically created based on changing conditions.

[0074] According to another aspect, the technology may include a method for PSI that uses intelligent entities to develop and continuously improve PSI, wherein the intelligent entities include human users utilizing user computer systems, additional AI agents, and additional PSI, all of which communicate electronically via a collective network. The method may include the following steps:

[0075] Create a user PSI using the following method:

[0076] Acquire a previously customized basic AI agent;

[0077] Collect media information related to human users;

[0078] Analyze media information;

[0079] Transform the analyzed media information into a training dataset;

[0080] Perform differential weighting on the transformed training dataset; and

[0081] The weighted training dataset is applied to the base-level AI agent to create a user PSI, where the base-level AI agent;

[0082] Using a network to enable a user PSI to communicate with multiple additional AI agents or additional PSIs to achieve a group-based security feature, where both the user PSI and the additional AI agents or additional PSIs agree to use a set of rules related to security or ethics;

[0083] All actions of the user PSI and the attached AI agent or attached PSI are logged in an auditable form on any one or any combination of the central computer system, user computer system, attached AI agent, and attached PSI on the network; and

[0084] Monitor whether each action follows the set of rules, and flag any action that does not follow the set of rules.

[0085] In some implementations, the recording of actions can utilize blockchain technology.

[0086] In some implementations, the recording of actions may utilize a general problem-solving framework to record any one or any combination of objectives, sub-objectives, problem states, and steps taken in problem-solving or cognitive activities occurring on the network.

[0087] In some implementations, recorded actions can only be altered when a majority of the additional AI agents or additional PSIs on the network provide authorization for the alteration.

[0088] Some implementations of this technology may include the following steps: stopping any tagged action and applying preventative actions to prevent the tagged action from recurring.

[0089] Some implementations of this technology may include the following steps: identifying one or more additional AI agents or additional PSIs that provide tagged actions; and controlling the participation of the identified AI agents or PSIs on the network.

[0090] Some implementations of this technology may include the following steps: analyzing the labeled actions and adjusting any or any combination of the attributes of the training dataset, a set of rules, and the network based on the analysis of the labeled actions.

[0091] According to another aspect, the technology may include a method for PSI that uses intelligent entities to develop and continuously improve PSI, wherein the intelligent entities include human users utilizing user computer systems, additional AI agents, and additional PSI, all of which communicate electronically via a collective network. The method may include the following steps:

[0092] Create a user PSI using the following method:

[0093] Acquire a previously customized basic AI agent;

[0094] Collect media information related to human users;

[0095] Analyze media information;

[0096] Transform the analyzed media information into a training dataset;

[0097] Perform differential weighting on the transformed training dataset; and

[0098] The weighted training dataset is applied to the base-level AI agent to create a user PSI, where the base-level AI agent;

[0099] Provide tasks to the user's PSI and multiple additional AI agents or additional PSIs on the network;

[0100] The results of the task are provided from each of the user PSI and the attached AI agent or attached PSI; and

[0101] The winning result is determined based on the outcome.

[0102] In some implementations, each of the additional AI agents or additional PSIs can process tasks independently and in parallel with each other.

[0103] In some implementations, one of the one or more additional AI agents or additional PSIs may be uncustomized, while one or more of the additional AI agents or additional PSIs may be customized.

[0104] Some implementations of this technology may include the following steps: utilizing a training dataset of an additional AI agent or additional PSI with winning results in the customization of a user-based AI agent or user PSI.

[0105] Some implementations of this technology may include the following steps: each of the attached AI agent or attached PSI utilizes a general problem-solving architecture on the task.

[0106] Some implementations of this technology may include the following steps: using a set of security or ethical rules that are agreed upon by each of the user-based AI agent, user PSI, additional AI agent, and additional PSI.

[0107] In some implementations, each of the additional AI agent or the additional PSI may have a set of weights that encode the knowledge of the additional AI agent and the additional PSI, respectively.

[0108] In some implementations, any computer device on the network can use the activity log.

[0109] Some implementations of this technology may include the following steps: providing a task to a user PSI, the task being configured to generate hallucinations by the PSI's LLM.

[0110] Some implementations of this technology may include the following steps: providing a task to a user PSI, wherein the task is a random arrangement of existing standardized tasks or a task dynamically created based on changing conditions.

[0111] In some implementations, the network may include a user PSI and an additional AI agent or additional PSI, as well as one or more additional networks, each including an AI agent or PSI.

[0112] According to another aspect, this technology may include a method for PSI that uses intelligent entities to develop and continuously improve PSI, wherein the intelligent entities include human users utilizing user computer systems, additional AI agents, and additional PSI, all of which communicate electronically via a collective network. The method may include the following steps:

[0113] a) Create a user PSI using the following method:

[0114] Acquire a previously customized basic AI agent;

[0115] Collect media information related to human users;

[0116] Analyze media information;

[0117] Transform the analyzed media information into a training dataset;

[0118] Perform differential weighting on the transformed training dataset; and

[0119] The weighted training dataset is applied to the base-level AI agent to create a user PSI, where the base-level AI agent;

[0120] b) Use a network to enable a user PSI to communicate with multiple additional AI agents or additional PSIs to achieve a group-based security feature, where the user PSI and the additional AI agents or additional PSIs all agree to use a set of rules related to security or ethics.

[0121] c) Utilize various standardized tasks to perform the basic performance of user PSI and additional AI agents or additional PSI;

[0122] d) Create one or more versions of the user PSI, each of which is different from the others;

[0123] e) Determine the user PSI, the version of the user PSI, and the performance of the network with attached AI agents or attached PSI;

[0124] f) Assigning honor or blame values ​​to one or more elements that result in poor or poor basic performance; and

[0125] g) Identify which elements are high-quality elements that lead to performance improvements and retain these high-quality elements for future use.

[0126] Some implementations of this technology may include the following steps: identifying which elements are detrimental elements that cause performance degradation and changing the properties of the detrimental elements.

[0127] Some implementations of this technology may include the following steps: repeating step c using high-quality elements and modified low-quality elements.

[0128] Some implementations of this technology may include the following steps: repeating steps c)-g) until no higher quality element is detected.

[0129] In some implementations, the network may include a user PSI and an additional AI agent or additional PSI, as well as one or more additional networks, each including an AI agent or PSI.

[0130] Some implementations of this technology may include the following steps: logging all actions of the user PSI and the attached AI agent or attached PSI in an auditable form on any one or any combination of the central computer system, user computer system, attached AI agent and attached PSI on the network.

[0131] In some implementations, a standardized task can be any one or any combination of the following: a problem-solving task using a shared, general problem-solving architecture; an ethical and safety scenario designed to determine whether the behavior of any one or any combination of user PSI and attached AI agent or attached PSI is safe and ethical; a standardized intelligence test and evaluation designed to test the intelligence of intelligent entities; a task configured to measure the degree of “illusion” or the generation of erroneous results; a task from various professional fields; a task containing behavioral norms of different cultures or groups; a task as a random permutation of existing standardized tasks; and a task dynamically created based on changing conditions.

[0132] In some implementations, a version of the user PSI can be created by any one or any combination of the following: human adjustment of the parameters, weights, or data encoding the knowledge of the user PSI; autonomous adjustment of the parameters, weights, or data encoding the knowledge of the user PSI; random variation of the data encoding the parameters, weights, or knowledge of the user PSI; subjecting the user PSI to different training methods or amounts; sequentially creating different versions of the user PSI; and creating different versions of the user PSI in parallel.

[0133] In some implementations, a version of a user's PSI can be created by directly combining parameters, weights, or data that encode knowledge from multiple PSIs by calculating an average weight value, wherein the weight given to the most recently created or more complex version of the user's PSI is greater than the weight given to the older or less complex version of the user's PSI.

[0134] In some implementations, versions of the user PSI can be created using a randomized method or a deliberate method for estimating which versions will have beneficial results, wherein the estimation method is any one or any combination of the following: comparing the degree of matching between the knowledge of one of the user PSI versions and the statistical frequency of the tasks submitted to the network; comparing the overlap of knowledge of different additional PSIs on the network such that changes are made to optimize the performance of the group of user PSI versions if the knowledge of the entire network is characterized compared to the characteristics of the problem the network is expected to solve; and using hill climbing or gradient descent to optimize the parameters of any one of the user PSI versions.

[0135] According to another aspect, this technology may include a method for PSI to utilize a single computerized intelligent system, the single computerized intelligent system comprising multiple AI agents residing within the single computerized intelligent system. The method may include:

[0136] Acquire a previously customized base-level AI agent that resides within a single computerized intelligent system;

[0137] Collect media information related to human users associated with basic-level AI agents;

[0138] Transform the analyzed media information into a training dataset;

[0139] Differential weighting is applied to the transformed training dataset;

[0140] The weighted training dataset was applied to the base-level AI agent to create user PSI; and

[0141] The user PSI communicates with multiple additional PSIs to enable community-based security features from the additional PSIs to the user PSI.

[0142] The underlying large language model (LLM) for training an AI agent is utilized, whereby the guardrails include attributes associated with any or any combination of safety, ethics, and knowledge, and the AI ​​agent resides in a single computerized intelligent system.

[0143] Customize the base LLM based on the ethics profile;

[0144] Combining ethical information from multiple additional AI agents residing in a single computerized intelligent system, the differences between additional AI agents and AI agents, and the additional AI agents residing in a single computerized intelligent system;

[0145] A set of values ​​for the underlying LLM is improved based on the problem-solving process of the problem request; and

[0146] The underlying LLM is updated with combined ethical information and a modified set of values, thereby allowing for scalable AGI.

[0147] Therefore, the features of this technology have been outlined quite extensively in order to better understand the detailed description of this technology below and to better appreciate its contribution to the field.

[0148] Many objects, features and advantages of the present technology will become apparent to those skilled in the art after reading the following detailed description of the present technology and its illustrative embodiments in conjunction with the accompanying drawings.

[0149] Therefore, those skilled in the art will understand that the concepts upon which this disclosure is based can be readily used as the basis for designing other structures, methods, and systems for achieving certain objectives of this technology.

[0150] Therefore, one objective of this technology is to provide a new and novel secure PSI that has all the advantages of existing AGI systems and methods without any disadvantages.

[0151] Another objective of this technology is to provide a new and novel secure PSI that can be easily and efficiently implemented and marketed.

[0152] Another objective of this technology is to provide a new and novel secure PSI that has low implementation costs in terms of resources and labor, thus allowing for a low selling price for the consumer public, making this secure PSI economically accessible to the purchasing public.

[0153] Another objective of this technology is to provide a new type of secure PSI that offers some of the advantages of existing systems and methods while overcoming some of the disadvantages that are usually associated with them.

[0154] To better understand this technology, its operational advantages, and the specific purposes achievable through its use, reference should be made to the accompanying drawings and descriptive content in which illustrated embodiments of this technology are contained. While several objectives of this technology have been identified herein, it should be understood that the following description is not limited to satisfying most or all of the identified objectives, and some embodiments of this technology may satisfy only one such objective or none at all. Attached Figure Description

[0155] The present invention will be better understood when considered in the following detailed description, and objectives other than those described above will become apparent. This description refers to the accompanying drawings, in which:

[0156] Figure 1 This is a flowchart illustrating an implementation of a subsystem that can be utilized in the AAAI system and method of this technology.

[0157] Figure 2 This is a block diagram illustrating an exemplary process that can be used with this technology.

[0158] Figure 3 This is a flowchart illustrating an exemplary implementation of a system and method for creating scalable, ethical, and safe AGI or PSI from AAAI and human collective intelligence that can be used with this technology.

[0159] Figure 4 This is a flowchart illustrating an exemplary implementation of a scalable, general-purpose problem-solving system and method for human-centered AGI, constructed according to the principles of this technology and related to PSI.

[0160] Figure 5 This is a flowchart illustrating an exemplary implementation of a scalable solution learning subsystem or process.

[0161] Figure 6 This is a flowchart illustrating an exemplary implementation of a scalable natural language to problem-solving language translator subsystem or process.

[0162] Figure 7 This is a flowchart illustrating an exemplary implementation of a scalable reputation component subsystem or process for human, AI, and / or PSI problem-solving agents.

[0163] Figure 8 This is a flowchart illustrating an exemplary implementation of a scalable security and ethics check subsystem or process, where AI may also be PSI.

[0164] Figure 9 This is a diagram illustrating the features and functionality of a tree-structured problem-solving architecture used by the scalable WorldThink protocol, where AAAI can also be PSI.

[0165] Figure 10 This is a diagram illustrating various use cases for domain-specific problems. These use cases rely on the underlying WorldThink protocol and together help form the basis of an AGI system capable of solving various problems. The AAAI identified in the diagram can also be a PSI.

[0166] Figure 11 This is a diagram illustrating some steps in a general problem-solving framework that is part of the WorldThink protocol and used by the AAAI system, and can also be used by PSI.

[0167] Figure 12This is a flowchart illustrating some basic problem-solving functions supported by the WorldThink protocol, which can be used with AAAI systems and methods of this technology, where the solver can be a PSI.

[0168] Figure 13 This is a flowchart illustrating some basic problem-solving functions supported by the WorldThink protocol, which utilizes two problem solvers (which may be PSIs) to collaborate in solving customer problems.

[0169] Figure 14 This is a flowchart illustrating an exemplary customization process for an AAAI system, where AI in the diagram can also be PSI.

[0170] Figure 15 This is a flowchart illustrating an exemplary problem-solving process utilizing a shared cognitive architecture implemented in an AI system, where the AI ​​in the diagram may also be a PSI.

[0171] Figure 16 This is a flowchart illustrating an exemplary problem-solving process utilizing a shared cognitive architecture implemented in a collective network of AI systems, where the AI ​​in the diagram may also be a PSI.

[0172] Figure 17 This is a flowchart illustrating an exemplary implementation of the PSI of this technology.

[0173] Figure 18 This is a flowchart illustrating an exemplary implementation of the capabilities of PSI utilizing this technology.

[0174] Figure 19 This is a flowchart illustrating an exemplary implementation of the community-based security mechanism of this technology.

[0175] Figure 20 This is a flowchart illustrating an exemplary implementation of how PSI records all its actions on a network.

[0176] Figure 21 This is a flowchart illustrating an exemplary implementation of inspecting cognitive activities on a network.

[0177] Figure 22 This is a flowchart illustrating an exemplary implementation of the competitive evolution of PSI's intelligence and / or performance.

[0178] Figure 23 This is a flowchart illustrating an exemplary implementation of joining multiple PSIs on a PSI network using agreed-upon interaction rules and methods.

[0179] Figure 24This is a flowchart illustrating exemplary implementations of the basic performance of individual PSIs, PSI groups, and / or the entire PSI network on various standardized tasks.

[0180] Figure 25 This is a flowchart illustrating exemplary implementations of different versions of individual PSIs and / or different combinations of PSIs.

[0181] Figure 26 This is a schematic block diagram illustrating an exemplary electronic computing device that can be used to implement the present technology.

[0182] In the various figures, the same reference numerals denote the same parts. Detailed Implementation

[0183] definition

[0184] Artificial intelligence (AI) is a non-human entity capable of performing behaviors that most people would consider intelligent in at least one area or in some aspects.

[0185] Artificial General Intelligence (AGI) generally refers to AI capable of performing all (or almost all) intellectual tasks that an average person can perform. However, it should be clear that any AGI capable of learning and self-improvement will not remain at the AGI level for long, but will rapidly evolve into a super-intelligent AGI capable of performing all intellectual tasks as well as or better than an average person. Therefore, for the purposes of this specification, "AGI" refers to a conventional AGI system or a "super-intelligent" AGI. In this specification, AGI is described as being implemented by systems and associated methods.

[0186] Advanced Autonomous Artificial Intelligence (AAAI) – a type of AI capable of acting intelligently independently or semi-independently (under supervision). AI agents. Individual AAAIs can specify, customize, and execute useful actions using existing AAAI technology systems and methods. Groups of AAAIs can cooperate and combine their intelligence to create integrated AGI systems. Sufficiently advanced AI agents can also act as AGI systems, which can include other, less advanced AI agents within themselves.

[0187] AAAI.com is a platform, company, website, and / or project that implements this technology and supports the development, customization, and use of AAAI agents and AGIs, which are generated by the combined actions, knowledge, or intelligence of multiple AAAIs through the collective intelligence of AAAIs and / or humans, as indicated in this and related technologies.

[0188] AI ethics—the ethics used by AI or AGI to describe right and wrong in a given situation.

[0189] Consistency issues – problems that arise when AI ethics are inconsistent with human ethics, leading AI or AGI to take actions that humans deem unethical and / or dangerous to individuals or human ethics.

[0190] Basic AI—AI, AI agent, AAAI, SLM, or LLM—has been generally trained but has not yet been customized based on information from individual users or for a specific task.

[0191] Collective intelligence (CI) – intelligence that arises when multiple intelligent entities focus on solving common problems, or when knowledge from multiple intelligent entities is pooled to overcome the limitations of bounded rationality. Historically, collective intelligence has been human collective intelligence, but AGI is based on both human and AI agents (including PSI) and can also be generated by multiple AAAIs with or without human participation in the system. Active CI arises when intelligent entities (e.g., humans or machines) take steps useful for solving problems or actively engaging in other intellectual labor. For example, when multiple people explicitly tell an advertiser what type of ads they want to see, these individuals exhibit active CI. Passive CI arises from analyzing the behavior of intelligent entities (e.g., humans or machines), even if this behavior is not directly related to using the analysis to solve a problem. For example, when AI or other systems analyze which web pages (or groups of people) visit on the web and then use that analysis to direct targeted advertising to those people.

[0192] Ethics / values ​​(“ethics”) are a subset of knowledge that provides intelligent entities with a sense of purpose and constrains permissible actions or operations based on what is asserted to be “right” or “wrong” in a given situation. Specifically, ethics should be considered as premises from which intelligent entities can reason or logically calculate the best course of action to achieve goals or intentions consistent with ethical premises. Just as premises in a logical system must be accepted “as given,” so too must fundamental ethical or ideological ideas about what is right and what is wrong be accepted as premises from which intelligent entities can propose rational actions to realize these values ​​or ethics.

[0193] Hallucination / Artificial Hallucination – a phenomenon in which a large language model (LLM) (typically a generative AI chatbot or computer vision tool) perceives patterns or objects that do not exist or are imperceptible to a human observer, or creates meaningless, inaccurate, misleading, or false output.

[0194] Human ethics—ethics that humans claim to describe as right or wrong in a given situation.

[0195] Intelligent entities or entities are humans, AI agents or systems, clones of AI agents or systems, AAAI agents or systems, and / or clones of AAAI agents or systems that utilize computer systems to engage in providing problems, subproblems, objectives, and / or sub-objectives, and / or participate in any problem-solving activities concerning problems, subproblems, objectives, and / or sub-objectives. In cases where multiple intelligent entities exist within a single computer system, an intelligent entity can also refer to a subroutine of a part of the overall computer program, or an intelligent entity within a larger set of entities used for simulation or programming. PSIs are also intelligent entities.

[0196] Large Language Models (LLMs) are a type of AI that can take natural language as input and generate natural language as output. Typically, LLMs are trained using ML techniques on large datasets, enabling them to simulate intelligent dialogue or other forms of interaction with humans using natural language. Variations of LLMs can also be trained to take language as input and generate images or visual representations as output; or they can take images and visual representations as input and generate language and / or images and / or visual representations as output. For the purposes of this patent, we refer to all such systems as LLMs, although image-based models do not always need to take text as input or output. LLMs can also act as AI agents, and are sometimes referred to as such in this art. For the purposes of this disclosure, Small Language Models (SLMs) are also included in the definition of LLMs.

[0197] Machine learning (ML) is a subfield of AI that involves developing machines by enabling them to learn or acquire knowledge on their own rather than by explicitly programming such knowledge into them (as is the case with expert system AI developed through classical knowledge engineering methods).

[0198] Narrow AI—a type of AI that performs at a human or superhuman level in relatively limited domains (such as playing games, brewing beer, analyzing legal contracts, etc.). Narrow AI contrasts with AGI, which can perform at a human level in all intellectual tasks. Some AIs are narrower than others; for example, driving requires more general skills than playing chess, but not as many as AGI.

[0199] Prohibited attributes—requests, objectives, questions, terms, phrases, problems, answers, solutions, information, etc., that are identified or set as illegal, immortal, unethical, dangerous, fatal, etc. For example, a request for information about obtaining a Molotov cocktail to pass through airport security.

[0200] Safety – Typically, the concern for human safety and survival differs from ethics and values.

[0201] Security features are an aspect of the design or operation of this technology that typically enhances the safety of one or more individuals by helping to increase the probability that AI ethics are consistent with human ethics, thereby overcoming consistency issues.

[0202] Training / Tuning / Customization—Traditionally, the term "training" is used to refer to training a network (e.g., an LLM) to perform intelligent behavior. Tuning refers to the activity of fine-tuning a trained base model so that it generally performs better in a specific task. Customization refers to a wide variety of activities, including but not limited to training and tuning to make the AI ​​specifically suited for the purpose of making (multiple) given users or applications. For the purposes of this specification, training, tuning, and customization are used interchangeably, and it should be understood that although the techniques differ and the degree and type of effort involved vary, the purpose of all three is to tune the AI ​​and make it more intelligent or more specifically suited for (multiple) specific users or applications.

[0203] Weights / Network Weights—In the field of machine learning, many systems learn by adjusting the weights in a neural network architecture, which can be represented as a network of nodes and links between nodes. For example, the weights of a link connecting two nodes can correspond to the strength of the association or connection between any nodes. These weights can also represent excitatory or inhibitory connections between concepts, as in neural network representations. The learning of an entire AI system—such as LLM or more generally any AI agent, transformer algorithm, or machine learning method used to establish and modify the strength of connections (also called “parameters” in some models) between nodes—can be represented as a numerical matrix corresponding to the weights between nodes in the network. Weights / Network Weights in this specification refer to this numerical information, which is typically, but not necessarily, stored in a matrix or vector representation. By combining, manipulating, or otherwise altering this numerical information, the system’s learning, knowledge, or expertise and behavior can be changed.

[0204] Overview of this technology

[0205] This technology includes systems and methods for realizing Personalized Artificial Intelligence (“PSI”), which surpasses all currently existing forms of AI in both scope and intelligence. Because the PSI is self-improving, it will become exponentially more intelligent than its creator's owner. However, due to the unique creation method described in this patent, the PSI will be completely safe and dedicated to serving its owner. Superintelligence designed to be safe is the essence of this patent. This technology differs from any previously created AI system and possesses a level of intelligence, safety, and usefulness far exceeding that of currently state-of-the-art AI assistants.

[0206] Think of your PSI as your personal, super-intelligent AI agent. It obeys your commands and learns about you and what you want over time. It's more efficient and effective than you in most everyday online tasks. It offers excellent advice. First, it knows you, can connect with you, and thinks like you. If you want to expand your PSI's capabilities, it's as simple as trading data with friends or buying or selling data on an online marketplace and then incorporating that data into your PSI to give it additional intelligence, skills, and knowledge. Over time, your PSI can acquire the wisdom of billions of people and AI agents. And after acquiring this knowledge, it can run simulations based on your goals and purposes to further improve and act on your behalf (with your permission) where you deem it appropriate.

[0207] Your material wealth can be multiplied by your PSI, which acts as your investment advisor and agent. Your time can be freed up for spiritual, artistic, or other pursuits.

[0208] Your PSI can be cloned and rented. Its data and intelligence can be packaged and sold. A thousand versions of it can simultaneously handle a thousand different errands and tasks for you, guided by other cloned PSIs, which are also under your guidance and serving your interests.

[0209] These are some of the benefits this technology brings to individual owners. For society as a whole, multiple PSIs pooling their intelligence and participating in the PSI network can serve as planetary intelligence, benefiting all people and our planet.

[0210] While the aforementioned devices meet their respective specific goals and requirements, none of them describe a PSI that allows for the development and continuous improvement of PSI for each human owner to achieve PSI security. This technology further overcomes one or more drawbacks associated with the prior art.

[0211] There is a need for a new and novel secure PSI that can be used to develop and continuously improve the PSI of each human owner using intelligent agents. In this regard, the present technology essentially meets this need. In this respect, the secure PSI according to the present technology substantially deviates from the traditional concepts and designs of the prior art, and in doing so provides an apparatus primarily developed for the purpose of developing and continuously improving the PSI of each human owner using intelligent agents to achieve PSI.

[0212] In the following description, specific details, such as particular implementations, procedures, and techniques, are set forth for purposes of explanation and not limitation in order to provide a thorough understanding of the present technology. However, it will be apparent to those skilled in the art that the present technology may be practiced in other implementations that depart from these specific details.

[0213] It is understood that this technology provides technical effects, contributions, and solutions by combining multiple customized AAAI systems communicating through a collective intelligent neural network with all AAAI systems utilizing a common cognitive architecture. This common cognitive architecture includes one or more problem-solving protocols for generating one or more solutions or answers to a problem request and providing the solutions or answers to the user for approval. The customization of the AI ​​system generating the AAAI includes input from human users for training the AI, AAAI, or PSI. Further technical contributions or solutions may include multiple customized AAAI or PSI systems comprising one or more clones of AAAI or PSI, each clone of AAAI or PSI being customized independently of its parent AAAI or PSI and independent of other clones of AAAI or PSI within the same system.

[0214] Another technical contribution and solution is for creating and / or implementing a PSI that is self-improving and will become exponentially more intelligent while remaining safe and ethical. This technical contribution can also be provided in the creation of the PSI by: acquiring a previously customized base-level AI agent and utilizing differentially weighted media information relevant to human users in the PSI training, combined with training information acquired by a group of additional AI and / or PSI all communicating on a collective network.

[0215] Another technological contribution and solution lies in the fact that all AIs and PSIs communicate on a collective network, enabling group-based security characteristics and / or learning, where all AIs and PSIs agree on a set of rules governing security and ethics (as consistent human values) to prevent any harmful actions by any malicious PSI.

[0216] It is understood that this technique was discovered outside of computer program exclusion and / or abstract conceptual interpretation. This can be found in part in the technical contributions and solutions provided by this technique, the utilization of specific training inputs outside the computer, and the provision of solutions or answers outside the computer.

[0217] One reason AGI is so elusive is that specific knowledge and expertise from different fields must be creatively combined in the invention to achieve AGI. Another reason why the development of AGI is not obvious is that almost all AI researchers are focused on trying to improve existing narrow AI systems through more complex and extensive machine learning methods.

[0218] The fact that AGI has resisted the attempts of thousands of others—despite spending considerable sums of money—and that expertise in a relatively obscure field must be combined with mainstream AI approaches in this technology, powerfully demonstrates the novelty and inventiveness of this technology.

[0219] This technology describes systems and methods for not only implementing AGI, but also implementing it quickly and, most importantly, safely.

[0220] This could potentially have a positive impact on the evolution of AGI. The best way we can do this is to take the safest possible path for the development of AGI and ensure that humans follow that path. In turn, the best way to ensure that humans follow the safest path is to show that the safest path to AGI is also the fastest, and therefore the most desirable path to AGI. With these considerations in mind, the expectation of illuminating the fastest (and safest) path is the motivation for developing this technology.

[0221] While the aforementioned devices achieve their respective specific goals and requirements, none of them describe a system or method for safe, scalable, general artificial intelligence that allows for scaling through a combination of human users and multiple AI systems to train other AI systems by combining the values ​​and ethical knowledge of human users and multiple AI systems. This technology also overcomes one or more drawbacks associated with existing technologies.

[0222] There is a need for a new and novel system and method for safe, scalable, general artificial intelligence, which can be extended by using a combination of human users and multiple AI systems to train other AI systems by combining the values ​​and ethical knowledge of human users and multiple AI systems. In this regard, the present technology substantially meets this need. In this respect, the system and method for safe, scalable, general artificial intelligence according to the present technology substantially deviates from the traditional concepts and designs of the prior art, and in doing so provides an apparatus primarily developed for the purpose of extending by using a combination of human users and multiple AI systems (including but not limited to PSI) to train other AI systems by combining the values ​​and ethical knowledge of human users and multiple AI systems.

[0223] In the following description, specific details, such as particular implementations, processes, and techniques, are set forth for purposes of explanation and not limitation in order to provide a thorough understanding of the present technology. However, it will be apparent to those skilled in the art that the present technology may be practiced in other embodiments that depart from these specific details.

[0224] In the following description, specific details, such as particular implementations, processes, and techniques, are set forth for purposes of explanation and not limitation in order to provide a thorough understanding of the present technology. However, it will be apparent to those skilled in the art that the present technology may be practiced in other embodiments that depart from these specific details.

[0225] It is understood that this technology provides technical effects, contributions, and solutions by combining multiple customized AAAI (or PSI) systems communicating through a collective intelligent neural network with all AAAI (or PSI) systems utilizing a common cognitive architecture. This common cognitive architecture includes one or more problem-solving protocols for generating one or more solutions or answers to a problem request and providing the solutions or answers to the user for approval. The customization of the AI ​​system generating the AAAI (or PSI) includes inputs from human users used to train the AI ​​or AAAI (or PSI). Further technical contributions or solutions may include multiple customized AAAI systems comprising one or more clones of AAAI (or PSI), each clone of AAAI (or PSI) being customized independently of its parent AAAI (or PSI) and independent of other clones of AAAI (or PSI) within the same system.

[0226] Another technological contribution and solution is for creating scalable AGIs faster and safer, which leverage human input during training and customization to imbue AAAI, PSI, and / or AGIs with human ethical attributes.

[0227] Another technical contribution and solution is for scalably training AI systems and / or agents by leveraging a combination of safety and ethical information from many individual AI agents (or PSIs) to obtain representative and statistically valid samples of human ethics and values ​​covering a wide range of scenarios. Another technical contribution can be found in the fact that this technology includes methods for combining information from many agents and compiling the optimal combination of these agents to provide scalable training for AI, PSI, or AGI.

[0228] It is understood that this technique was discovered outside of computer program exclusion and / or abstract conceptual interpretation. This can be found in part in the technical contributions and solutions provided by this technique, the utilization of specific training inputs outside the computer, and the provision of solutions or answers outside the computer.

[0229] The AAAI approach to developing safe AGI is essentially a collective intelligence (CI) approach. The source of intelligence is not a single LLM, SLM, or ultra-advanced AI, but rather a collection of intelligent agents, including both humans and AI. Subtasks involved in AGI development include, but are not limited to, training individual AI agents or PSIs, effectively and efficiently combining knowledge from different agents (including but not limited to subjective values ​​and ethical knowledge), scaling AGIs, and continuously improving / updating AGIs.

[0230] Current methods—such as Human Feedback Reinforcement Learning (RLHF) and Constitutional Learning—have failed to effectively and scalably train AI to be ethical and safe. This technique describes scalable systems and methods that outperform current approaches. In one aspect, this technique can include a combination of safety and ethical information from many individual AI agents to achieve a representative and statistically effective sample of human ethics and values ​​covering a wide range of scenarios. This technique can include methods for effectively covering a wide range of ethical situations and dynamically responding to new situations as they arise. Methods for combining information from many agents and compiling the optimal combination of these agents are also proposed. These methods can be used not only to improve safety using ethical knowledge but also to create superintelligent systems that combine many other types of knowledge. Safe AGI and superintelligence can be achieved through the collective intelligence approach described in this specification of this technique. A detailed scenario using META® as an example illustrates an exemplary implementation of this technique.

[0231] A method for dynamically updating knowledge is also proposed. Successful implementation of this technology will increase the chances that AI, PSI, AGI, and superintelligence will remain aligned with human values, even when such systems significantly surpass human intelligence.

[0232] Advanced Autonomous Artificial Intelligence (AAAI) is a set of systems and methods for developing general artificial intelligence and superintelligent general artificial intelligence (collectively, "AGI") in a fast and safe manner that benefits humans. Compared to other methods of developing AGI, current AAAI techniques achieve a faster and safer path to AGI by relying at least initially on the participation of (ideally millions) human minds in AGI training, operation, and safety / supervision functions.

[0233] Current AAAI technology enables AGI (Ability to Generate Intelligence) by allowing users to first customize and clone their own AI or PSI. These customized AIs (AAAIs) and / or PSIs participate in problem-solving and other intellectual activities on a network comprised of other AAAIs, PSIs, and humans. While each AAAI (or PSI) may lack the breadth of skills and knowledge required to become an AGI on its own, the AAAIs (initially with the help of humans on the network) collectively form an AGI that will rapidly surpass the average human ability in all intellectual labor.

[0234] Some aspects of this technology may include: 1) systems and methods for customizing AI by leveraging users' unique knowledge, skills, and ethical values; 2) a general problem-solving architecture that allows AAAIs to interact productively with each other and with humans on intellectual tasks; 3) networks where these interactions occur; 4) methods for integrating the knowledge and ethics of individual AAAIs into AGIs; and 5) methods for learning and continuous improvement that makes AAAIs and AGIs smarter and more ethical over time. Human participation as both the customizer of their AAAIs and a participant in the network is a fundamental feature of this technology, which not only accelerates the development of AGIs but also makes AGIs safer by providing mechanisms for the ethical values ​​of millions of people adopted by and reflected in AGIs.

[0235] One implementation of this technology's AAAI system focuses on security and is achieved through five subsystems and related methods, such as Figure 1 As shown. The five subsystems of the AAAI system are: 1) AAAI Customization, 2) AAAI Architecture, 3) AAAI Networking, 4) AAAI Integration, and 5) AAAI Enhancement. The acronym SCAN-II (Security, Customization, Architecture, and Networking - Integration and Enhancement) describes this technology in the exemplary implementation. Other combinations of subsystems, as well as variations of each subsystem, are also possible. In the effort to provide redundant security checks while omitting one or more subsystems from a particular implementation, security features have been designed into each subsystem.

[0236] The five subsystems of the AAAI system can be further described as follows:

[0237] 1) Basic large language models (LLM), small language models (SML) or other AI systems can be customized to reflect the knowledge of individuals, groups of individuals or organizations and designated as advanced autonomous artificial intelligence (AAAI).

[0238] 2) It enables customized AAAI to participate in problem-solving using a general problem-solving architecture that is compatible with both human and AI agents.

[0239] 3) AAAIs capable of problem-solving participate in problem-solving activities, including but not limited to:

[0240] Performing multi-step cognitive activities such as planning, problem-solving, and other types of sequential processes on intelligent agent networks;

[0241] Operators that reduce the difference between the current problem-solving state and the desired state are generated based on the objective / sub-objective;

[0242] Set sub-goals to achieve the main goal;

[0243] Utilizing a hierarchical structure, until an operable target is set that can be executed by the operator; and

[0244] Analyze auditable records to identify recommendations for improving the problem-solving process to achieve the objectives / sub-objectives.

[0245] 4) Multiple AAAI or PSI on the network can be integrated to achieve AGI; or AI capable of intelligent (or superhuman) behavior across a variety of tasks.

[0246] 5) The individual AAAIs, problem-solving networks and / or integrated systems of multiple AAAIs are continuously improved through various means, including but not limited to redirecting the efforts of the individual AAAIs and / or integrated AGIs to tasks that improve the system and / or system components.

[0247] A subsystem or a new subsystem may include any one or any combination of the following:

[0248] 1) Safety / Ethics Review—Compare objectives or sub-objectives to a list of prohibited attributes and assign ethical values ​​based on the comparison results. Review objectives / sub-objectives against the list of prohibited attributes. Use a combination of AAAI-approved task criteria (approved by users, regulatory bodies, or human users) and values / safety information from AAAI. Establish or use thresholds for objectives / sub-objectives to determine whether ethical values ​​are unsafe, unethical, safe, or ethical. Determine whether a series of individually safe objectives / sub-objectives are unsafe or unethical when considered cumulatively. If an objective violates ethical standards, determine whether the violation reflects a predictive assessment. Document any and all activities of the safety / ethics review in an auditable log.

[0249] 2) AAAI Matching — Detecting and identifying additional AAAIs or PSIs, each of which are related to one or more target or sub-target standards.

[0250] 3) Remember and / or improve—Record activities, compare them with the progress of successful or unsuccessful problem solutions, and determine which activities remain active or have been forgotten.

[0251] 4) AAAI Learning—This includes learning processes that utilize information provided by intelligent entities, such as human users equipped with computers, AAAI, or PSI. Activities are recorded and compared with successful or unsuccessful progress in problem-solving to determine which activities remain active or are forgotten. Honor or responsibility values ​​are assigned to a set of content related to problem-solving activities. A set of prompts is provided to the user, and information is received based on these prompts. AAAI is updated with the set of content identified as active. This set of content may be, but is not limited to, a set of prompts provided to the user and information received based on these prompts, all of which are recorded in an auditable log. Optionally, problem-solving activities may include this set of content.

[0252] Example User Scenario

[0253] Describing some user scenarios that provide a sense of how this technology can operate in some aspects may be helpful. Figure 2 An exemplary process is shown in the figure.

[0254] In one aspect, users "access" AAAI.com via their computers, mobile phones, PDAs, or goggles. AAAI.com will interact with users through web-based interfaces, telephone applications, custom software on PDAs, or virtual reality / virtual reality environments. Interaction modes can be physical interactions via keyboard, mouse, or gesture interfaces; voice-based interactions via microphone input coupled to a natural language understanding and generation system; or video-based interactions, such as when the user is in a virtual reality setting or avatar in virtual reality.

[0255] The initial interaction involves setting up a user account, which may be free or paid. This will involve a username and password or other authentication mechanisms, which may include, but are not limited to, biometric IDs such as fingerprints, facial or voice recognition, and / or multi-factor authentication mechanisms such as software or hardware authenticators residing on a separate secure device or one of the user's existing devices.

[0256] For security reasons, all communication between the user and the AAAI system can be encrypted via VPN, and / or other encryption and security methods known in the programming field can be used.

[0257] AAAI.com may request users to set up payment functionality via credit cards, PayPal, Venmo, blockchain, ACH, or other payment mechanisms. These payment functionalities will allow for bidirectional transfers of funds, payments, and / or credit—from the user to AAAI.com and from the AAAI system to the user in cases where the AAAI system needs to pay or credit the user for their work efforts at AAAI, or where brokerage payments are made between users and / or between AAAIs on the AAAI network.

[0258] In one implementation aspect, AAAI.com may have interfaces with other companies and vendors that users may use, including but not limited to: Facebook, Instagram, Reels, Amazon, Apple, Microsoft, Google, and YouTube.

[0259] During the initial interaction with the user and subsequently upon user request, AAAI.com will engage in conversations or other interactions with the user (which may include presenting menu options, lists, graphics, sliders, buttons, and other user interface controls to the user in a GUI, text, haptic, voice, or VR-related manner) to determine the user's goals and objectives when using the AAAI system.

[0260] For example, some purposes a user might have when using AAAI.com could include creating and customizing their own AI (called AAAI), and these purposes could include, but are not limited to:

[0261] Serve users as an advisor, teacher, or companion.

[0262] Representing users in negotiations, interactions, discussions, and transactions with other users or with other users' AAAI or with suppliers and other companies.

[0263] Working on behalf of users for payment, or working voluntarily, including online intellectual, advisory, or problem-solving work across a variety of tasks.

[0264] Copying or “cloning” a user’s AAAI allows multiple or many cloned AAAIs to work in parallel on behalf of the user, including interacting with, teaching and improving each other, enabling the cloned AAAIs to increase their knowledge, skills and capabilities.

[0265] After the owner's death, the AAAI serves as a legacy that can continue to interact with the world, including potentially comforting surviving relatives and friends.

[0266] Contribute knowledge, ethics, and effort to AAAI.com's AGI, and raise the basic level of AI or AGI that AAAI.com can offer to those users before they add their unique customizations.

[0267] Work with other users' AAAIs to help ensure AGI's ethical and safe conduct by contributing ethical information and values ​​to AGI, and participate in monitoring, review, oversight, and voting processes that may help ensure AGI remains safe and ethical.

[0268] During conversations or interactions with users, the AAAI system will also identify constraints and resources that can be used to customize the user's AAAI. For example, some of these constraints and resources may include, but are not limited to:

[0269] Users must invest the amount of time customizing their AAAI training and / or supervision.

[0270] Users are willing to invest the amount of financial resources to customize their AAAI.

[0271] The availability of social media information, such as Facebook profiles and timelines, Instagram profiles and history, Reels, TikTok and YouTube videos, tweets and text content and history, emails and email history, cookies collected by advertisers, blog posts, articles, books, patents, audio and video recordings, images, and other information about the user or third parties and / or collected by the user or third parties that can be used to train, adjust, or customize the user's AAAI.

[0272] The availability and use of personality tests, such as the Myers-Briggs Personality Inventory, skills and knowledge assessments, standardized tests, exams, certifications, and other types of assessments and questionnaires that can be given (or have been given) to users online.

[0273] The availability and usage of additional knowledge bases and training data from users on the AAAI platform, which can be used to train, tune, or customize the user's AAAI.

[0274] Other human users and / or their AAAIs, which can be used to help train, tune, or customize the user's AAAI.

[0275] Additional texts and information, personal texts, and libraries selected by the user or system to train the user's AAAI. For example, texts may be selected to train the AAAI based on the user's preferences; if the AAAI will primarily be used to solve online pipeline problems, books about pipelines may be selected. Even if these materials are part of the base AAAI provided to the user, emphasizing certain subsets of text or information used for additional training may lead to the user's AAAI behavior being more reflective of potential behaviors.

[0276] In addition to specifying objectives, resources, and constraints through interactive dialogue or other interactions with the system, users or the system may also want to specify other technical parameters that affect the training or customization process. These parameters may include, but are not limited to:

[0277] The type of training, tuning, or other ML algorithm used.

[0278] The type and size of (multiple) training datasets.

[0279] The extent to which training materials must be "cleaned," formatted, labeled, or otherwise processed before customization begins.

[0280] The number of training "epochs" or iterations performed by (multiple) learning algorithms.

[0281] The complexity and type of the custom or trained (multiple) base models.

[0282] The time frame required for training—for example, whether it must be completed in a minute, a day, or a week—can have an impact on the cost and resources used.

[0283] Various machine learning algorithms have internal and unique "temperature" or other parameters that can affect what is learned and how it is learned, including but not limited to the degree to which a custom AAAI adheres to the literal meaning or deviates from or is "creative" in its response.

[0284] Whether to use "single", "several", or extensive training.

[0285] The amount of human and / or AI supervision to be used in the customization process.

[0286] Once a user's AAAI or PSI is customized, the user can clone it and / or make it work on behalf of the user on the online network. The user's AAAI can begin to act on behalf of the user, make travel arrangements (for example), provide advice, interact with other AAAIs, and participate in collective AGI work by contributing problem-solving and ethical information.

[0287] Simple exemplary implementation

[0288] Figure 3 This paper presents a simple exemplary implementation of a system and approach for creating ethical and safe general artificial intelligence from the collective intelligence of AAAI and humans. This simple implementation is compatible with all the company- and platform-specific scenarios outlined above, as well as many other potential integration scenarios.

[0289] (A) A user (human, AAAI, or other intelligent entity) visits the AAAI.com website (a). The website notifies the user and offers them two options: register (b) or log in (c).

[0290] If a user chooses to register, a conversation is initiated that extracts the user's values / ethics (d), user goals and purposes (e), and user budgets for time (f) and money (g). All users must allocate some time (f). Users can choose to create a free AAAI or allocate a money budget.

[0291] If users have allocated a monetary budget (g), they have the opportunity to purchase a pre-trained AAAI or training module (h) with a specific personality (i), skill (j), expertise (k), or knowledge (l). They also have the opportunity to purchase training (m) from other AAAIs on the network.

[0292] After making a time (and optional monetary budget (h, i, j, k, l, m)) allocation decision, the user continues to overview the creation process and is then asked to grant user permissions (n) to optionally log in and use existing social media, Twitter, and other vendor accounts to collect user data for “one-click” training of the user’s AAAI. After the user selects to use some (or not) data, they are guided through creating the AAAI by clicking (o). The AAAI is an off-the-shelf LLM (e.g., GPT X, BARD, Llama, Gemini, Grok, or any closed-source or open-source AI agent) trained / tuned on a dataset automatically prepared based on all user data authorized by the user. If no data is authorized, the AAAI is simply an off-the-shelf LLM.

[0293] AAAI now begins to learn by training (p) using various training datasets and modules (hm) and its existing AAAI knowledge (p1). There are two main learning methods: automatic (q) and human (r).

[0294] Automatic learning includes, but is not limited to, learning by interacting with its own copy(s) and learning by interacting with other (optionally supervised) AAAIs(t).

[0295] Human learning includes interactions with humans (owners (u) or other people on the network (v)).

[0296] Both humans and AAAIs can supervise the learning of AAAIs. After each (automatic or human) learning interaction, the system attempts to improve the performance of the AAAI through further prompts, modifications, and / or training. Based on many loops of human and AAAI inputs designed to teach and improve the AAAI, the user's AAAI becomes smarter.

[0297] At any time, users can purchase additional training modules (hm) that have been proven to enhance AAAI capabilities.

[0298] Humans set the performance standard (w), and then AAAI retains it (x).

[0299] Once retained, AAAI can access the WorldThink tree (y) and browse (z).

[0300] AAAI can enter the tree as a worker (a1) or a client (b1).

[0301] Workers are automatically matched (c1) to tasks, or they can select a specific task from the browsing tree via search (d1) or link (e1). Once they accept a task (f1), they participate in the problem-solving module (g1) until a solution is reached (h1) and payment is made (i1), or the user saves the completed work on credit and exits the tree (j1).

[0302] The client (b1) can specify goals (k1) that are combined with values / ethics (d), as well as prior goals and objectives (e) that the system aims to address.

[0303] A customer may request to use only his / her / their AAAI, in which case problem solving is free. Alternatively, a customer may use the entire network's AGI capabilities, in which case the system compensates the individual AAAIs for their work and delivers the solution (at cost + markup) to the customer, debiting the customer's account (l1).

[0304] The system can also place non-profit humanitarian and ecologically oriented missions, as well as missions as part of planetary intelligence, on the WorldThink tree (m1).

[0305] Customers may (optionally) authorize the system to use copies of their AAAI and data for these purposes without compensation in exchange for the maintenance and operation of the free AAAI network when their AAAI(n) is created.

[0306] right Figure 13 Additional comments on the exemplary implementation shown

[0307] We now offer for Figure 13 Additional comments on various elements, including but not limited to some potential integration points with the aforementioned illustrative partners:

[0308] The “website” (a) can be hosted on Amazon AWS, Microsoft Azure, Google Cloud, Apple Cloud, Nvidia’s datacenter offering, or have a native implementation on any major technology company’s platform. The “website” can also be an “application” in the App Store or other app marketplaces. It can be a government-funded, non-profit, or other globally accessible technology that directly or indirectly connects some of the attention of all those who wish to participate. Furthermore, browser plugins can be used, allowing AAAI to learn from users as they perform normal tasks on the internet, and the plugin records their activity, creates training files, and uses these files to train AAAI. The “website” can also be an API or other device for directly connecting AAAI or non-human intelligent entities to the network.

[0309] Register (b) or

[0310] Login (c) can be done via Facebook, Instagram, Apple, Microsoft, Google, YouTube, TikTok, Amazon, or any other partner ID scheme. Multi-factor authentication and all best ID and security practices are supported. In the case of browser plugins or applications, login via these technologies can be used as login to an AAAI account.

[0311] Values ​​and ethics (d) are derived from a set of scenarios that have been customized for users and dynamically generated based on user responses. Analysis of behavioral patterns—actions, words, or interactions—from data from partners (including navigation and click data, online posts, tweets, text and emails, videos, and other user data) is translated into ethical guidelines or a system of ethical values, or this data can also be used as part of an ethics / values ​​profile. Values / ethics and goals / purposes (d) can be combined with customer purposes (k1) to create or find matching tasks on the WorldThink tree (y) that have been proposed or (potentially solved) in the problem-solving system (g1).

[0312] The goals and objectives (e), and the budget for the time and / or money allocated to achieve those goals (f, g), are derived through a series of dialogues and / or customized interactions with the system. The budget refers to the overall resource budget, which includes the user's time and money that can be allocated to training, monitoring, and improving the user's AAAI. Goals and objectives help determine the initial parameters used for AAAI creation and identify training modules (h) or other knowledge (im) that may create the most useful AAAI for the user's goals. Data from partners reflecting user preferences and other user behavior information can also be used by the system to help derive or infer user goals and objectives.

[0313] Time (f) refers to the user time that can be invested in training and supervising the user's AAAI and / or the user's time solving problems on the problem-solving network. By supervising AAAIs, users can ensure that their AAAIs meet client goals and expectations—especially in areas where AAAIs are stuck (e.g., they lack the knowledge to solve problems on their own). Unrestricted representation of problems by defining objectives and sub-objectives and breaking down large tasks into smaller ones is also a way for human users to help their AAAIs solve problems. Generally, providing human expertise in areas where AAAIs are less proficient than humans improves the overall problem-solving and overall effectiveness of the AGI network.

[0314] (g, i1) "Money": This can be a payment solution utilizing Apple Pay, WePay, Amazon Pay, Google Pay, or any provider that supports payment solutions, as well as blockchain, credit cards, ACH, and other solutions. Although payment (i1) is indicated as a debit to a customer account (l1), it can of course also be a credit to a worker's account. Typically, a user's account can be considered both a customer account and a worker account, where the credit and debit are recognized based on the user's (or the user's AAAI) role in a particular instance. That is, in some cases, a user may be a customer, paying for services provided by the system or other specific AAAIs, while in other cases, the same user may be a worker, receiving fees for services provided by the user (or the user's AAAI). The Money module (g) implements functions such as setting payment methods, setting budgets for automatic payments, restricting a user's AAAI permissions to spending only X dollars without additional approval, and other payment-related capabilities known in the art.

[0315] (h, i, j, k, l) Training modules (h) may be provided by AAAI.com or third-party partners (m), including but not limited to any potential partners and technology companies listed above. Training modules may be targeted at different knowledge domains, including personality (i), specific skills (e.g., pipeline, legal, accounting) (j), expertise (e.g., consulting) (k), and knowledge (e.g., historical knowledge, practical knowledge of a specific business or organization, cultural knowledge) (l).

[0316] (m) Purchasable AAAI training is a specific type of knowledge that has already been learned by other AAAIs and can be transferred to new user AAAIs. Such knowledge can be packaged in the form of modules (e.g., accounting modules) or in a form specific to another AAAI(or multiple AAAIs), such as in "Everything John's AAAI Knows," "John's AAAI Personality," or "Combined Knowledge of All AAAIs with 5-Star or Higher Reputation in the Pipeline Field."

[0317] (n) A license refers not only to the permission a user may grant to access all data on a particular other vendor's (or partner's) site (e.g., "all my Facebook data"), but also to the permission a user grants to his / her / their AAAI regarding the ability to log in and transact on various sites, including but not limited to the ability to transact up to a certain amount through payment mechanisms. A license may also include authorizing the system to clone a user's AAAI for non-profit purposes and for the purpose of aggregating knowledge from various AAAIs to create AI at the AGI level.

[0318] (o) One-Click Creation is a non-limiting example that provides a simple and quick way to customize AAAI using data automatically collected from all places where the user has granted the system permission to access their data. It is understood that this technology can utilize other methods to customize AAAI. For example, if the user grants permission (n) to access their Facebook data, One-Click Creation (o) will download data from Facebook (if Facebook is a partner with an API for downloading that user's data), or as the user logs into the user's Facebook account and "scrapes" relevant data from the user's account. The system will then automatically parse the collected data and transform it into a dataset suitable for training / tuning the underlying AI, such as an LLM (e.g., GPT X). The system will then train / tune the LLM and produce a customized AAAI that can be improved and refined through additional training / tuning and interaction with the user and / or other AAAIs.

[0319] (p) Training refers to the process of training or adjusting AAAI with data, including feedback from users, other people and / or AAAI (including but not limited to its own copies and variants).

[0320] (q, r, s, t, u, v) Automated learning requires no human user intervention and can proceed very quickly. Typically, this involves AAAI interacting with its own copy (or variant) and (optionally) with other AAAIs to improve through interaction. If humans are sometimes involved in the training loops (t), this can help automated learning progress faster where it cannot make effective progress on its own. Learning can also be done through rapid iterations (s) between AAAI interactions. Just as chess AI can rapidly evolve from novice to master level by simulating millions of chess games very quickly, AAAIs can rapidly evolve their capabilities by simulating millions of interaction scenarios. A monetary budget (g) can set limits on the extent to which such simulations require financial resources to cover the computations involved.

[0321] Humans (or AAAIs) can be specifically tailored to the types of scenarios for automated learning, allowing AAAIs to be trained in narrow or more general professional domains based on user needs and resources. Through partner integration, the training of AAAIs can be guided by backworking the types of jobs available on partner marketplaces (e.g., Amazon's Mechanical Turk), focusing on learning the skills that generate the most revenue for the AAAI when it handles available jobs. This "just-in-time" learning / training / adjustment approach generates AAAIs with the required skill set at any given point in time "on demand."

[0322] The human (r) interacting with AAAI can be the owner (u) of AAAI (in which case no fee is usually charged because the user is training his / her / their own AAAI) or other professionals (v) who are experts in training AAAI and may charge a fee to guide human and / or automate the training / tuning of AAAI for users who do not want to spend time or lack the expertise to train / tun AAAI.

[0323] (w, x) Users (the owners of AAAI) can set various performance criteria (w) that must be met before a user is willing to make his / her / their AAAI "retained" (x) and able to perform tasks on the WorldThink tree. Some of these criteria can also be set by partners and other third parties who have minimum standards before allowing AAAI to work on their platforms, products, applications, or networks.

[0324] The WorldThink tree (y, z, a1, b1) is a large tree data structure consisting of many subtrees representing every problem and task that has been completed, is being processed, or has been proposed for the entire AGI system. The tree is browsable (z). Individual AAAIs and / or humans can handle specific tasks within the tree. The tree structure provides an auditable trail of all problem-solving activities, which is also useful for learning through the procedural mechanisms described above. When interacting with the tree, agents can assume two main roles: (a1) worker or (b1) client. Another role that might be considered a special type of client is a regulatory body or third party monitoring the system's performance, security, and / or ethics. Workers typically participate in solving open problems or subproblems on the tree. Clients typically participate in specifying the problem, objectives, purpose, and other parameters that constrain problem-solving (e.g., rewards, budget, timeframe, success criteria, quality metrics).

[0325] (c1) Automatically match workers to tasks on a tree based on data about workers, including but not limited to workers' skills, expertise, knowledge, past experience, reputation, fees or costs, availability, and response time. Workers can be people or AAAIs. Workers can be matched and recruited from partners (e.g., LinkedIn, Mechanical Turk, Facebook) who have data about human users and / or their AAAIs. Workers can also be recruited through online advertising that offers jobs for a variety of tasks and uses advertising targeting mechanisms known in the art or described in other patents by the applicant to direct potential workers.

[0326] (d1) Workers can also search the WorldThink tree for tasks of interest or that match their skills. This search can be manual or automated (as in the case of AAAI workers).

[0327] (a1) Workers and clients (b1) can also browse (z) the WorldThink tree to find tasks or issues of interest. Workers or clients can then click to links (e1) to specific sections of the tree to obtain detailed information about the problem-solving that has occurred (or been raised) in that section of the tree. These can link to registering for work or can be submitted as additional tasks by clients that build upon existing problem-solving work.

[0328] (f1, g1, k1) Clients can interact with the system to specify the specific goals, objectives (k1), and tasks they want to accomplish. The problem specification interaction results in the formulation of problems, tasks, and goals (f1) and their placement on the WorldThink tree (y) for use with the problem-solving system (g1) to solve the problem.

[0329] (m1) This system has the ability to formulate certain goals, issues, and tasks related to the overall effort to help people or the planet. These can be addressed using rewards in a “for-profit” model or using cloned AAAI and voluntary human effort in a “non-profit” model. Some issues may relate to the overall goal of enabling the global AGI to act on behalf of the planet and its people using its intelligence (also known as “planetary intelligence”) on a planetary basis. Various partner organizations—including non-profits, governments, and philanthropies—may “insert” their tasks, issues, goals, and objectives (m1) here.

[0330] (g1) Problem-solving systems refer to problem-solving architectures and systems outlined by Newell and Simon (HPS) and improved by the applicant, the online distributed problem-solving system (ODPS) patent invented by the applicant, the WorldThink white paper created by the applicant, such PPAs and other PPAs related to AAAI, and modifications and changes reflecting different reward, payment and operating models.

[0331] Activities on certain other online work systems (e.g., Mechanical Turk) can be automatically mapped to the extent that the applicant-modified HPS / WorldThink problem-solving framework is applied. The entire problem and related problem-solving activities can be "upgraded" from partners and other sites, and data can be populated into the WorldThink tree to increase its comprehensiveness.

[0332] To the extent that other applications, products, systems, and online capabilities can help solve the problem (e.g., using travel booking systems, robo-advisor applications, transportation applications, online ordering systems), these capabilities can be referenced and referred to as “operators” (in a manner similar to procedure calls in programming languages) to advance problem-solving. Therefore, problem-solving does not solely rely on operators developed by humans or AAAI solvers working with the tree, but can include any online or offline techniques or methods that advance problem-solving, as long as these methods can be referenced and / or linked to the WorldThink tree in the appropriate places when solving the problem.

[0333] (h1) Once a solution has been implemented, the client can review the solution before releasing any rewards for it, if any. Alternatively, if the success criteria for the solution are automated, human client review may not be necessary, and rewards can be released automatically once the success criteria are met. Depending on client and worker preferences, this automation can be implemented through “smart contracts” using blockchain technology or through a more centralized approach.

[0334] In solving problems and (optionally) paying rewards (because some problems are nonprofit or voluntary, or performed by the user's own AAAI), there may be opportunities to obtain feedback from clients and workers using a range of methods well-known in the art. Solutions are also "chunked" and programmatically programmed, allowing the entire system to learn solutions to specific problems and their key characteristics, enabling solution paths to be indexed for retrieval and accessed and reused when similar problems arise in the future.

[0335] Optionally, a royalty fee can be enabled, allowing a fee to be paid to the user in the form of a royalty fee for the solution if the user or their AAAI solution is reused. Such a royalty fee can (optionally) be paid using a "smart contract" on the blockchain or through other payment methods.

[0336] (j1) Problem solving does not need to be completed in a single session. The solution may make partial progress, in which case the progress is saved when the human or AAAI solver leaves the problem-solving system, and the data is stored and the progress made so far is attributed to the solver, even if such progress has not yet reached a level where a reward can be paid.

[0337] The WorldThink protocol is a problem-solving architecture that can be used by AAAI.com as a general problem-solving architecture because it combines the general architecture of HPS while adding features to overcome certain challenges.

[0338] In some implementations and such Figure 4 and Figure 5 As can be broadly illustrated, the learning process can occur within a shared cognitive framework.

[0339] A shared and universal problem-solving architecture can be illustrated by the following scenario, which mentions humans but is generally applicable to any intelligent entity.

[0340] 1) You can input the problem description into AAAI.

[0341] 2) Then, human problem solvers can be identified and recruited into a database or data source of human workers.

[0342] 3) A qualified human or intelligent entity can be matched with the problem.

[0343] 4) Use LLM or other methods to translate the English descriptions of the problem task, objectives, operators, and solution steps into the language of a general problem-solving architecture.

[0344] 5) Delegating the work on sub-problems to (multiple) different human problem solvers allows for the parallel processing of multiple aspects of a complex problem.

[0345] 6) Combine the solutions to various sub-problems into a holistic solution.

[0346] 7) Direct the problem solver's attention to the part of the problem tree that requires their work.

[0347] 8) Compensate or pay workers for solutions to the problem and / or (multiple) sub-problems.

[0348] 9) Allow human users to accept solutions, reject solutions, and / or provide feedback to the solver on their solutions to the problem and / or (multiple) sub-problems.

[0349] refer to Figure 5 The steps of solution learning can be exemplified by recording at each step the applied learning process operators, the new state of the problem, the evaluation function used and its results, the current relevant objectives / sub-objectives, and other information that differs from the previous steps(s). The problem state or problem status can be evaluated to determine if the problem is solved. If not, after the final step, using information from the latest problem state, the problem-solving process is rerun, progress is evaluated, and the next operator to be applied is selected. The process can then return to the recording step.

[0350] If the problem is resolved, the successful or unsuccessful solution is recorded for retrieval, saving effort in resolving previously resolved problems and informing about efforts made on previously unsuccessful paths.

[0351] Semantic analysis, hash functions, and / or other methods can be used to index successful solutions and unsuccessful attempts with keywords for future matching / retrieval.

[0352] Regular reviews of all storage solutions can be conducted to ensure they meet established ethical and security guidelines, and unsafe / unethical solutions can be flagged for removal from the database or data source.

[0353] The solution database is regularly updated and disseminated, so that problem-solving networks and agents have access to a constantly expanding library of solutions and increased awareness of unsuccessful attempts.

[0354] refer to Figure 6 and Figure 16 This technology may include combining a network of multiple intelligent entities, including human workers, with a general problem-solving architecture. Using a database or data source containing a list of human and / or AI problem solvers, multiple intelligent entities are matched with problem requests based on problem criteria. Any part of the problem request can be translated into explicit language using a general problem-solving architecture including decision trees.

[0355] Subproblems of a problem request can be delegated to one or more matching intelligent entities, allowing the processing of subproblems to proceed independently and in parallel, such as... Figure 13 As further illustrated below, a general problem-solving architecture is used to create one or more sub-solutions during the problem-solving process for sub-problems.

[0356] Any one or any combination of intelligent entities can provide, in natural language, a description of the current problem state, the goal of the problem request, relevant problem-solving information, and any one or any combination of the next steps that human workers will take in the problem-solving process.

[0357] It can receive sub-solutions to the sub-problems delegated to each matched intelligent entity. Any one or any combination of sub-solutions and the overall solution can be provided to the intelligent entity or to any one or any combination of the user interface of the user AI system.

[0358] Intelligent entities parse and translate natural language descriptions into explicit language that can be utilized by decision trees in general problem-solving architectures.

[0359] In some implementations, if the intelligent entity cannot specify the problem state (including the relevant operators and information needed to take the next steps in the problem-solving process) based on parsing and translation, the intelligent entity may engage in a dialogue with at least one human worker until the precise problem state is specified.

[0360] In some implementations, the problem-solving process can be repeated until the overall solution is accepted or resources are exhausted. Matched human workers can be compensated separately for each sub-solution. Furthermore, reputation attributes can be assigned to any or any combination of human workers and worker AI systems or PSIs.

[0361] In some implementations, the problem-solving process may include a series of problem state transitions from an initial problem state with a goal to a final solution state where the goal has been achieved, wherein a series of decisions are made by the problem-solving process and actions taken, which apply operators that enable human workers to transition states until the final solution state is reached.

[0362] refer to Figure 4 , Figure 7 and Figure 16 This technology may include combining a network of human users with a general problem-solving architecture. Using a database or data source that includes lists of human, AI, and / or PSI problem solvers, multiple human users are matched with problem requests based on problem criteria.

[0363] Subproblems of a problem request can be delegated to one or more matching intelligent entities, allowing the processing of subproblems to proceed independently and in parallel, such as... Figure 13 As further illustrated below, a general problem-solving architecture is used to create one or more sub-solutions during the problem-solving process for sub-problems.

[0364] Sub-solutions from each matched human worker can be provided to the subproblem to which it was delegated. Human workers matched with sub-solutions can be compensated separately.

[0365] Then any one or any combination of the sub-solutions and the overall solution can be provided to the user interface of the user AI system or any other AI system (including but not limited to PSI).

[0366] Allows human users to accept the overall solution, reject the overall solution, and / or provide feedback on any of the sub-solutions to any of the matched human workers.

[0367] Reputation attributes can be assigned to human workers and / or worker AI systems. Reputation attributes can include measures of any one or any combination of the following: time spent on a sub-solution, difficulty value of the problem request, short-term and long-term user satisfaction with the sub-solution, number of times any sub-solution has been reused on the network, ratings from other human workers, human worker responsiveness value, and human worker reliability value.

[0368] Some implementations may include using reputation attributes to delegate sub-problems when matching human workers with problem requests using an algorithm, and / or compensating matched human workers separately for sub-solutions.

[0369] In some implementations, the algorithm may use a hierarchy of metrics preset by the human user who made the request.

[0370] Some implementations may include information about each step of the problem-solving process being recorded by human workers or worker AI systems.

[0371] Some implementations may include a standard for recording the steps of the problem-solving process, which is the time spent on each step.

[0372] Some implementations may include analyzing the recorded information and updating the metrics of reputation attributes after the overall solution has been accepted or after the problem-solving process has concluded.

[0373] Some implementations may include soliciting user satisfaction information at predetermined intervals after the overall solution or sub-solutions are provided to the user interface to obtain short-term and long-term satisfaction measures for updating one or more reputation attributes of human workers or worker AI systems.

[0374] refer to Figure 8 This technology may include the use of human users and AI systems (including, but not limited to, PSI), which includes performing security / ethical checks on any or any combination of targets and solutions provided by any or any combination of intelligent entities, including human users and AI systems, each using a computer system.

[0375] Goals and / or solutions can be compared with prohibited attributes, and ethical values ​​can be assigned to goals and / or solutions based on the results of the comparison and / or ethical standards.

[0376] Based on the comparison results, a shared cognitive architecture, including one or more problem-solving protocols, can be developed to create solutions, thereby creating AGI. The comparison results and solutions can then be provided to any intelligent entity.

[0377] In some implementations, ethical checks can be performed at the time the objective is presented, and at any one or any combination of the time between presenting the objective and presenting the solution.

[0378] In some implementations, ethical standards may be determined using a set of approved ethical standards instructed by a user or regulatory body for a specific task, through any one or any combination of values ​​and security information from one or more intelligent entities. It may also be provided by any additional intelligent entity and verified or approved by a human user.

[0379] In some implementations, ethical criteria may include a confidence level threshold for the objective, such that ethical values ​​are determined to be any one of unsafe, unethical, safe, and ethical objectives.

[0380] In some implementations, the confidence level threshold can be further used to determine whether a series of individually safe targets are unsafe or unreliable when considering cumulative factors.

[0381] In some implementations, if the objective is to violate ethical standards, a confidence level threshold can be used to determine whether a violation reflecting the predictive assessment has occurred.

[0382] In some implementations, candidate targets can be proposed based on ethical values, and these candidate targets can be compared with prohibited attributes.

[0383] In some implementations, the results of the comparison can be recorded in an auditable log to determine which problem-solving activities keep the solution active.

[0384] Further reference Figure 8 Ethical checks can compare any one or any combination of problem requests, sub-problems, and sub-solutions with prohibited attributes, and assign ethical values ​​based on the results of the comparison and any one or any combination of ethical standards.

[0385] In some implementations, the ethical review step may be triggered each time a human user sets a question request or any sub-question, and / or each time compensation is provided to a matched human worker.

[0386] The objective / sub-objective can be compared to the list of prohibited attributes. Ethical standards can be determined by any one or any combination of values ​​and safety information from any AAAI. A set of mission standards approved by AAAI, using user, regulatory, or human user approval, is combined with AAAI's values / safety information.

[0387] Ethical criteria may include a confidence level threshold for the question request, such that ethical values ​​are determined to be any one of unsafe, unethical, safe, and ethical objectives. The confidence level threshold can be further used to determine, when considering a cumulative set of individually safe objectives, whether they are unsafe or unreliable.

[0388] In some implementations, if the objective is a violation of ethical standards, a confidence level threshold can be used to determine whether a violation reflecting a predictive assessment has occurred. Any and all activities of security / ethics checks can be logged in an auditable manner.

[0389] Figure 9 to Figure 11 A simple exemplary framework is provided for understanding the WorldThink protocol. Figure 9 This is a diagram illustrating the features and functions of the problem-solving tree structure in the WorldThink protocol. Figure 10 This diagram illustrates various use cases for domain-specific problems. These use cases rely on the underlying WorldThink protocol and together form the foundation for an AGI system capable of solving a wide range of problems. At the top of the pyramid is the collective intelligence solution. Integrating the collective intelligence of AAAI (and human problem-solving agents) is a way to achieve AGI, as previously described.

[0390] In implementations using the WorldThink protocol, clients pay for solutions using tokens. Solutions are generated by leveraging the collective intelligence of many humans (and machines, or AAAIs). Clients can use different domain-specific AAAIs for different types of problems.

[0391] The WorldThink protocol forms the foundation of the pyramid. The protocol layer provides (optionally, based on Ethereum or a blockchain) infrastructure that makes it easier for developers to build and scale custom problem-solving AAAIs. The protocol enables the reuse of solutions within and across AAAIs. It also handles royalty payments through smart contracts, reputation metrics, and other features that assist AAAI customizers and developers and promote network effects.

[0392] In the example, Figure 11 This illustrates a simple, exemplary, general problem-solving framework under a shared cognitive architecture, which may include:

[0393] Define a problem space that is configured or configurable to support all possible states of a problem request, including an initial state, a target state, and any one or any combination of all intermediate states that can be reached from the initial state;

[0394] By identifying the differences between the current state and the target state, means-ends analysis is applied to the problem request to decompose the problem request into targets and sub-targets. Operators are then applied to reduce the differences, and security or ethical screening is applied each time a target or sub-target is set.

[0395] Heuristic rules are applied, which are configured or configurable to guide the selection of operators in the absence of a complete solution, and are used to reduce the problem space;

[0396] Identify one or more second operators that are configured or configurable to formulate actions to transform one state into another, the second operators moving from an initial state to a target state by changing the current state of the problem request;

[0397] The application includes a set of rule-based control structures that manage the selection of the second operator to be applied at each step of the problem-solving protocol and determine which second operator to apply next based on the current state and target state of the problem request.

[0398] The evaluation function is used to determine the application of the second operator;

[0399] Assign honor or responsibility values ​​to the completed solution or its sub-solutions, such honor or responsibility values ​​enable the tracing and determination of which of the second operators are most useful, and which of the evaluation functions lead to the success or failure of the problem-solving attempt;

[0400] Record successful and unsuccessful problem request solutions attempts; and

[0401] The analytical solution attempts to improve the selection of heuristic rules and evaluation functions.

[0402] Figure 10 A simple, exemplary framework for understanding the WorldThink protocol is provided. At the top of the pyramid is the collective intelligence solution that leads to AGI. Integrating AAAI, PSI (and human problem-solving agents) into collective intelligence is one way to achieve AGI, as previously described.

[0403] In implementations using the WorldThink protocol, clients pay for solutions using tokens. Solutions are generated by leveraging the collective intelligence of many humans (and machines, or AAAIs). Clients can use different domain-specific AAAIs for different types of problems.

[0404] Figure 10 The middle section illustrates an example of an AAAI (or PSI) customized by an organization to accomplish a specific task. These AAAIs or PSIs are more advanced than the examples of AAAIs customized by a single individual described earlier in this patent and require more customization. However, task-specific customization by an organization can be an efficient way to combine multiple narrow AIs (each taking the form of a custom AAAI that is an expert in the specific task) into a larger AGI. Figure 10 The basic AAAI on the left reflects the domains where inventors can relatively easily build custom AAAIs or PSIs based on years of expertise in certain fields, while the "Custom AAAI" on the right provides examples of domains where other experts or organizations can effectively customize AAAIs.

[0405] The WorldThink protocol forms the foundation of the pyramid. The protocol layer provides (optionally, blockchain- or Ethereum-based) infrastructure that makes it easier for developers to build and scale custom problem-solving AAAIs. The protocol enables the reuse of solutions within and across AAAIs. It also handles royalty payments through smart contracts, reputation metrics, and other features that assist AAAI customizers and developers and promote network effects.

[0406] Existing collective intelligence approaches to problem-solving are largely limited to simple, one-step methods, such as those used by question-and-answer (Q&A) systems (e.g., Quora, Google Answers, Yahoo! Answers). LLMs like GPT also largely fall into the category of Q&A systems because they are designed to generate responses given input, rather than solving the problem itself. While such Q&A systems have achieved some success in simply aggregating responses from many online participants, they are not designed to handle complex, branching, multi-step problems. The simple aggregation of responses is very different from the effort of coordinating many responders to solve complex problems. The WorldThink protocol is specifically designed to overcome the inherent challenges of coordinating many intelligent entities, thereby representing and solving complex, multi-step problems in an automated way that fairly rewards participants.

[0407] In the example, Figure 11 A simple, exemplary, general problem-solving framework is shown. Figure 12 This shows some of the basic problem-solving functions supported by the WorldThink protocol, which are generally represented by the number 10.

[0408] Problem solving begins when a client on AAAI.com submits a problem-solving request to the online participant group (Step 12). All AAAI or human problem solvers following the protocol collect certain criterion information from the client. Parts of this information may include: the name and description of the problem, the total reward the client will pay for a successful solution to the problem, the criteria for determining whether a solution will be considered successful, the timeframe for problem solving, the minimum and maximum number of problem solvers allowed to work on the problem simultaneously, the qualifications required of participants working on the problem, which parts of the problem and solution (if any) will be confidential, whether the solution must be exclusive to the client or whether it can be reused by others, and parameters related to how multiple problem solvers' efforts and / or successful solutions are rewarded.

[0409] Clients can break down complex problems into a series of sub-problems or request the task from a group as part of a problem-solving effort. The client user interface for the AAAI-initiated dialogue can be customized by the AAAI owners, but the underlying data format is standard and specified by the WorldThink or Online Distributed Problem Solving (ODPS) protocol. Once a client submits a problem, AAAI.com can use its own custom methods to recruit participants and / or leverage recruitment and reputation screening features built into the WorldThink protocol and therefore shared by all AAAIs.

[0410] Solvers address the problem according to a strictly structured problem-solving process common to all problem-solving agents and implemented by the WorldThink protocol (step 14). For example, each step in the problem-solving process must serve a specified goal and must take a specified action to transition the problem from the current state to the next. Each problem-solving step is represented in a decision tree supported by the protocol (optionally captured in the Ethereum logs) and viewable by participants via AAAI.com.

[0411] When a problem solver submits a complete solution (step 16), it is timestamped and verified against the client's success criteria before being delivered to the client for final acceptance (step 18). Once the client accepts the solution, the smart contract can automatically distribute tokens to the problem solver based on the problem payment parameters (step 20), or it can use other more centralized payment procedures.

[0412] Collaboration problem solving using the WorldThink protocol

[0413] In the example, Figure 13 The same steps are illustrated in an example (generally denoted by number 22) of two problem solvers (who may be individuals, AAAIs, PSIs, or a combination) collaborating to solve a client's problem. In this case, the overall problem has been broken down into sub-problems. Solver 1 possesses the expertise to assemble the overall solution but collaborates with Solver 2, who provides solutions to the sub-problems (steps 30 and 32). When the overall solution to the problem is submitted to the client (step 34), rewards are paid to both solvers based on an objective record of their contributions and agreed-upon payment parameters (step 36).

[0414] The WorldThink protocol supports breaking down a problem into subproblems in several ways. First, the client can choose to specify subproblems when submitting the overall problem (step 24). Alternatively, Solver 1 can begin working on the problem and recognize that the overall solution requires addressing subproblems outside their area of ​​expertise. Solver 1 can then create a subproblem, offering a share of the total token reward for the problem to anyone who helps solve it. Solver 2, possessing the necessary expertise, can then see the new subproblems posted by Solver 1 on the decision tree. The decision tree can optionally be maintained in the Ethereum log or via a centralized approach. Solver 2 accesses the tree via AAAI.com (or optionally directly from the blockchain). Solver 2 can then work on the subproblems and submit their solutions as part of Solver 1's overall solution.

[0415] There can be many "Solver 1s" working on a client's problem in parallel, each potentially posting sub-problems to attract multiple "Solver 2s." Problem solvers (humans or AAAIs) are incentivized by rewards and payout rules associated with the (sub)problems. When choosing which (sub)problems to tackle, they also care about the quality of their work to date (timestamped, attributed, and auditably recorded in the Ethereum logs to ensure transparency and fair distribution of credit). Solving high-quality sub-problems is more likely to yield token rewards. This market mechanism helps ensure efficient, fair, and cost-effective solutions.

[0416] Royalty and reusable solutions

[0417] Solution reusability is a key feature of the WorldThink protocol. Consider... Figure 13 The scenario involves a "sub-solution" that already exists and is simply reused by Solver 1. Because each solution is constructed and "tagged" according to the standard problem-solving format of the WorldThink protocol, Solver 1 can search for all existing solutions that match a specific goal or share certain characteristics with the problem he / she is trying to solve. (Alternatively, if the problem solution is chunked into a program for problem-solving—the learning mechanism explained in the Improvement section of this patent—the search may not be necessary, as AAAI or PSI solvers can simply add chunked problem solutions to their library of problem-solving capabilities.) Solver 1 decides to include an existing sub-solution in the overall solution, and if Solver 1's overall solution is accepted by the client, the smart contract (optionally) automatically pays a royalty to the author of the reused sub-solution (Solver 2 in this example). The royalty incentivizes solvers to create high-quality, easily reusable solutions, resulting in better, faster, and more cost-effective solutions for the client.

[0418] Additional descriptions and details of one implementation of an AAAI (or PSI) custom subsystem may involve the following steps.

[0419] refer to Figure 14 The first step in the customization approach involves creating an interface for users to input their unique training data. Depending on the user's preferences, this interface can be accessed via a web-based application or a mobile application. Users will be able to upload files in various formats, including text, audio, and video. Users will also be able to manually enter data into text or other input fields. Some user interfaces include, but are not limited to:

[0420] Web-based applications: Web-based user interfaces allow users to access and / or provide their personalized training data from any device with an internet connection.

[0421] Mobile applications: Mobile user interfaces allow users to access and / or provide their personalized training data from mobile devices.

[0422] Virtual Reality: A virtual reality user interface allows users to access and / or provide their personalized training data from the virtual world.

[0423] Augmented Reality: Augmented reality user interfaces allow users to access and / or provide their personalized training data from the real-world environment.

[0424] Voice interface: The voice interface allows users to access and / or provide their personalized training data via voice commands.

[0425] Wearable devices: The user interface of wearable devices allows users to access and / or provide their personalized training data from the wearable device.

[0426] Natural Language Processing (NLP) allows users to access and / or provide their personalized training data by interacting with AI or LLM using natural language.

[0427] Human-Computer Interaction (HCI): Human-Computer Interaction (HCI) allows users to access and / or provide their personalized training data by interacting with AI or LLM through a combination of gestures, voice commands, and facial expressions.

[0428] Image recognition: Users can input their unique training data through image recognition, allowing them to train AI or LLM quickly and intuitively. This can be accomplished using cameras and computer vision algorithms that can interpret images and associate them with the correct training data or create the correct training data.

[0429] Gesture recognition: Users can input their unique training data using gestures or body movements. This can be accomplished using motion sensing devices that can interpret gestures and associate them with the correct training data or create the correct training data.

[0430] Brain-computer interface: Users can input their unique training data using their brainwaves or EEG signals. This can be accomplished using a brain-computer interface that can interpret the signals and correlate them with the correct training data or create the correct training data.

[0431] Touchscreen: Users can use touchscreen devices to input their unique training data. This can be done by using a touchscreen device that can interpret the input and correlate it with the correct training data or create the correct training data.

[0432] Gaze tracking: Gaze tracking allows users to communicate with the system through their eyes. Users can gaze at specific items on the screen to provide input, and the system will detect and record the information. This can be used to select options or provide additional data to the system.

[0433] Eye tracking: Eye tracking is similar to gaze tracking, but the system can detect more subtle eye movements. This can be used to detect a user's focus and attention in order to better understand what they are interested in and what they are not.

[0434] Motion tracking: Motion tracking uses cameras or other sensors to detect a user's body movements. This can be used to control AI or LLM in a more natural way, allowing users to interact with the system through physical gestures.

[0435] Haptic technology: Haptic technology uses various tactile feedbacks such as vibration, pressure, and touch to provide a more immersive experience. This can be used to allow users to provide more detailed input to the system, such as selecting specific options or providing more detailed data.

[0436] Many of the aforementioned user interfaces can include a graphical user interface (GUI) that allows users to upload their data or information types, including text, images, audio, or video. Furthermore, users can build their own models or use pre-existing models to train AI or LLM. Other features can include dashboards for tracking progress, statistics for data analysis, and / or chatbots for customer service.

[0437] Further reference Figure 14 This technology may include customizing one or more attributes of an AI or PSI system by providing an interface that is configured or configurable to allow human users or any intelligent entity to input training data into the AI ​​system. The training data is then processed and converted into a standardized training format.

[0438] One or more training methods and training parameters can be selected and set based on speed factors, accuracy factors, precision factors, and / or transferability factors. Multiple training rounds, including one or more mechanisms, can be executed to determine the optimal number of rounds given a specific training objective and quality metric associated with the training format.

[0439] One or more feedback sessions can be executed to refine the training parameters, and training rounds can be rerun based on any one or any combination of inputs from human users and any intelligent entity. The training format can then be used to customize the AI ​​system.

[0440] In some implementations, the interface can be accessed via a web-based application or a mobile application, and can be configured or configurable to upload files or allow human users to input data.

[0441] In some implementations, training data may include any one or any combination of the following: the amount of time a user must invest in training the customized AI system; the amount of financial resources a user is willing to invest in the customized AI system; the amount of computational resources a user is willing to invest in the customized AI system; the amount of social media information that can be used in the customized AI system; the amount of email information that can be used in the customized AI system; the amount of electronic information about the user that can be used in the customized AI system; and the amount of electronic information about the user collected by a third party that can be used in the customized AI system.

[0442] In some implementations, training data may include information about human users obtained through any one or any combination of personality tests, standardized tests, certifications and assessments, or questionnaires provided by human users.

[0443] In some implementations, training parameters can be any one or any combination of the following: the type of training, tuning, or other machine learning algorithm to be used; the type and size of the training dataset; the extent to which the training dataset is formatted, labeled, or processed before customization begins; the number of training epochs; the type of base model to be customized; the time range required for training; the amount of human user supervision to be used when customizing the AI ​​system; and the amount of AI supervision to be used when customizing the AI ​​system.

[0444] In some implementations, training data may include ethical information provided by human users through an interface. This ethical information may be stored in an ethical profile. Customization of the AI ​​system's attributes may include ethical information.

[0445] refer to Figure 15 This technology may include leveraging a shared cognitive architecture implemented in one or more AI systems. A question request can be submitted from an intelligent entity acting as an AI system, a PSI, or a human user using a user interface on a computer system. Information associated with the question request may also be provided.

[0446] Identify and recruit multiple supplementary intelligent entities, each possessing one or more attributes related to one or more request criteria of the problem request. Supplementary intelligent entities can be multiple supplementary AI systems, PSIs, and / or multiple supplementary humans each using computer systems. Each identified AI system implements a shared cognitive architecture including one or more problem-solving protocols regarding the problem request to create a complete solution. The complete solution can be offered to the intelligent entities for final user acceptance.

[0447] In some implementations, the information may be any one or any combination of the following: the name and description of the problem request; the total reward the user will pay for a successful solution to the problem request; the criteria for determining whether a solution is considered successful; the time limit for resolving the problem request; the minimum and maximum number of identified additional intelligent entities that are allowed to process the problem request simultaneously; the qualifications required of the user associated with the identified additional intelligent entity that processes the problem request; whether a portion of the problem request is confidential; whether a portion of the solution is confidential; whether the solution is dedicated to the user; whether the solution is reused by other users; parameters relating to how to reward the user associated with the identified additional intelligent entity that processes the problem request; and parameters relating to how to reward the user associated with the identified additional intelligent entity that provides a successful solution.

[0448] Some implementations of this technology may include the following steps: before the completed solution is provided to the user for final acceptance, the completed solution is timestamped and verified according to success criteria assigned by the user.

[0449] Some implementations of this technology may include the following steps: distributing one or more tokens to an identified additional smart entity associated with the final acceptance completion solution, wherein the tokens are based on payment parameters.

[0450] In some implementations, payment parameters may include any one or any combination of whether the objective of the issue request has been achieved, whether the sub-objectives of the issue request have been achieved, and whether the ethical standards related to the objective and sub-objectives prior to token distribution have been met.

[0451] Some implementations of this technology may include the following steps: breaking down a problem request into a series of sub-problems, each of which is solved by any one or any combination of the identified additional intelligent entities.

[0452] In some implementations, any one or a combination of the identified additional AI systems can be cloned to create one or more cloned AI systems.

[0453] Some implementations of this technology may include the following steps: each cloned AI system implements a common cognitive architecture including a problem-solving protocol for problem requests to create a complete solution for the cloned AI systems.

[0454] In some implementations, the completion solution may utilize any one or a combination of completion solutions from an AI system, an identified additional intelligent entity, and a cloned AI system.

[0455] In some implementations, a shared cognitive architecture may include:

[0456] Define a problem space that is configured or configurable to include all possible states of a problem request, including an initial state, a target state, and any one or any combination of all intermediate states that can be reached from the initial state;

[0457] By identifying the differences between the current state and the target state, means-ends analysis is applied to the problem request to decompose the problem request into targets and sub-targets. Operators are then applied to reduce the differences, and security or ethical screening is applied each time a target or sub-target is set.

[0458] Heuristic rules are applied, which are configured or configurable to guide the selection of operators in the absence of a complete solution, and are used to reduce the problem space;

[0459] Identify one or more operators that are configured or configurable to perform actions to transition from one state to another, the operators moving from an initial state to a target state by changing the current state of the problem request;

[0460] The application includes a set of rule-based control structures that manage the selection of operators to be applied at each step of the problem-solving protocol and determine which operator to apply next based on the current state and target state of the problem request.

[0461] The application of the operator is determined by using an evaluation function.

[0462] Assign honor or responsibility values ​​to the completed solution or its sub-solutions, such honor or responsibility values ​​enable the ability to trace and determine which operators are most useful and which evaluation functions lead to the success or failure of the problem-solving attempt.

[0463] Record successful and unsuccessful problem request solutions attempts; and

[0464] The analytical solution attempts to improve the selection of heuristic rules and evaluation functions.

[0465] refer to Figure 16 This technology may include collective networks utilizing AI systems. Problem requests can be provided by human users using a user interface on a computer system, or from AI or PSI systems. Information associated with the problem request may also be provided.

[0466] Identify and recruit intelligent entities, each possessing one or more attributes related to one or more request criteria for the problem request. Intelligent entities may be multiple additional AI systems and / or multiple individuals, each using a computer system.

[0467] The first intelligent entity among the identified intelligent entities can implement a shared cognitive architecture that includes one or more problem-solving protocols regarding a problem request. The first intelligent entity can determine that a complete solution to the problem request requires solving a first sub-problem and one or more additional sub-problems. The first intelligent entity then implements the problem-solving protocol regarding the first sub-problem to create a first sub-solution.

[0468] Assign at least one additional subproblem to a second intelligent entity among the intelligent entities, wherein it implements a problem-solving protocol regarding the at least one additional subproblem to create a second subsolution.

[0469] Create a decision tree that includes a first sub-solution and a second sub-solution to create a complete solution to the problem request. The complete solution can then be provided to a user interface or AI system for final user acceptance, and / or to any intelligent entity for subsequent use.

[0470] In some implementations, the decision tree can be maintained in the Ethereum blockchain log.

[0471] In some implementations, the identified first and second smart entities can access the decision tree via an online address or directly from the blockchain.

[0472] Some implementations of this technology may include the following steps: distributing one or more tokens to an identified first smart entity associated with the acceptance of a completed solution or a first sub-solution, wherein the tokens are based on payment parameters.

[0473] In some implementations, payment parameters may include any one or any combination of whether the objective of the issue request has been achieved, whether the sub-objectives of the issue request have been achieved, and whether the ethical standards related to the objective and sub-objectives prior to token distribution have been met.

[0474] Some implementations of this technology may include the following steps: distributing one or more tokens to an identified second smart entity by an identified first smart entity based on payment parameters assigned by the identified first smart entity.

[0475] Some implementations of this technology may include the following steps: influencing the direction of the problem-solving protocol by assigning a first token reward to a first subproblem and assigning a second token reward with a different value than the first token reward to a second subsolution.

[0476] In some implementations, a problem-solving protocol can provide a layer of infrastructure that can be configured or configurable to build and extend the identified intelligent entities. The problem-solving protocol can enable the reuse of completed solutions both within and across intelligent entities. The problem-solving protocol can be configured or configurable to manage royalty payments.

[0477] In some implementations, the infrastructure can be based on blockchain or Ethereum.

[0478] Further reference Figure 16 After identifying and recruiting multiple intelligent entities, a problem request or one or more sub-problems of a problem request can be assigned to each intelligent entity. Then, a common cognitive architecture, including one or more problem-solving protocols, can be implemented on each recruited intelligent entity's problem request or sub-problem to create a problem solution or sub-problem solution respectively. The problem solution and sub-problem solutions can be integrated to create a complete solution to the problem request. The complete solution can then be provided to a user interface or AI or PSI system for final user acceptance.

[0479] Some implementations of this technology may include the following steps: assigning honor or responsibility values ​​to a dataset based on whether the performance of intelligent entities is improved or reduced according to a performance metric or evaluation function.

[0480] Some implementations of this technology may include the following steps: quantifying the benefit weight or harm weight of each intelligent entity's contribution to the problem request.

[0481] Some implementations of this technology may include the following steps: distributing rewards to the owner of the intelligent entity in proportion to the intelligent entity's contribution based on benefit weight or damage weight.

[0482] Users may have several purposes when creating and customizing their own AI (also known as AAAI), including but not limited to:

[0483] Serve users as an advisor, teacher, or companion.

[0484] Representing users in negotiations, interactions, discussions, and transactions with other users or with other users' AAAI or with suppliers and other companies.

[0485] Working on behalf of users for payment, or working voluntarily, including online intellectual, advisory, or problem-solving work across a variety of tasks.

[0486] Copying or “cloning” a user’s AAAI allows multiple or many cloned AAAIs to work in parallel on behalf of the user, including interacting with, teaching and improving each other, enabling the cloned AAAIs to increase their knowledge, skills and capabilities.

[0487] After the owner's death, the AAAI serves as a legacy that can continue to interact with the world, including potentially comforting surviving relatives and friends.

[0488] Contribute knowledge, ethics, and effort to AAAI.com's AGI, and raise the basic level of AI or AGI that AAAI.com can offer to those users before they add their unique customizations.

[0489] Work with other users' AAAIs to help ensure AGI's ethical and safe conduct by contributing ethical information and values ​​to AGI, and participate in monitoring, review, oversight, and voting processes that may help ensure AGI remains safe and ethical.

[0490] The steps involved in creating and customizing an AAAI may include, but are not limited to, dialogues or interactions with the user. During this dialogue, the AAAI system can identify constraints and resources that can be used to customize the user's AAAI. For example, some of these constraints and resources may include, but are not limited to:

[0491] Users must invest the amount of time customizing their AAAI training and / or supervision.

[0492] Users are willing to invest the amount of financial resources to customize their AAAI.

[0493] The availability of social media information, such as Facebook profiles and timelines, Instagram profiles and history, short videos, TikTok and YouTube videos, tweets and text content and history, emails and email history, cookies collected by advertisers, blog posts, articles, books, patents, audio and video recordings, images, and other information about the user or third parties and / or collected by the user or third parties that can be used to train, adjust, or customize the user's AAAI.

[0494] The availability and use of personality tests, such as the Myers-Briggs Personality Inventory, skills and knowledge assessments, standardized tests, exams, certifications, and other types of assessments and questionnaires that can be (or have been) given to users online.

[0495] The availability and usage of additional knowledge bases and training data from users on the AAAI platform, which can be used to train, tune, or customize the user's AAAI.

[0496] Other human users and / or their AAAIs, which can be used to help train, tune, or customize the user's AAAI.

[0497] Additional texts and information, personal texts, and libraries selected by the user or system to train the user's AAAI. For example, texts may be selected to train the AAAI based on the user's preferences; if the AAAI will primarily be used to solve online pipeline problems, books about pipelines may be selected. Even if these materials are part of the base AAAI provided to the user, emphasizing certain subsets of text or information used for additional training may lead to the user's AAAI behavior being more reflective of potential behaviors.

[0498] In addition to specifying objectives, resources, and constraints through interactive dialogue or other interactions with the system, users or the system may also want to specify other technical parameters that affect the training or customization process. These parameters may include, but are not limited to:

[0499] The type of training, tuning, or other ML algorithm used.

[0500] The type and size of (multiple) training datasets.

[0501] The extent to which training materials must be "cleaned," formatted, labeled, or otherwise processed before customization begins.

[0502] The number of training "rounds" or iterations performed through (multiple) learning algorithms.

[0503] The complexity and type of the custom or trained (multiple) base models.

[0504] The time frame required for training—for example, whether it must be completed in a minute, a day, or a week—can have an impact on the cost and resources used.

[0505] Various machine learning algorithms have internal and unique "temperature" or other parameters that can affect what is learned and how it is learned, including but not limited to the degree to which a custom AAAI adheres to the literal meaning or deviates from or is "creative" in its response.

[0506] Whether to use "single", "several", or extensive training.

[0507] The amount of human and / or AI supervision to be used in the customization process.

[0508] Once a user's AAAI is customized, the user can clone it and / or make it work on behalf of the user on online networks. A user's AAAI can begin acting on behalf of the user, making travel arrangements (e.g.), providing advice, interacting with other AAAIs, and participating in collective AGI work by contributing problem-solving and ethical information. AAAIs can also act as representatives of the owner(s) in various online transactions and interactions, contributing knowledge, expertise, style, personality, and ethics to an integrated AGI system that leverages the trained differences among numerous individual AAAIs.

[0509] The importance of security values

[0510] Whether your PSI is used for good or bad depends on your own value system. While a PSI can operate based on knowledge and values ​​pre-trained into it, the customization of your PSI is up to you. Each PSI is both explicitly trained by you and implicitly learned through observation, including observing and learning what you consider right and wrong. These values ​​form the core, according to which your PSI uses logic to make decisions. In other words, your values ​​tell your PSI what is right and what is wrong and help determine what it should do. Subsequently, its superior intelligence is highly effective and efficient in achieving goals that reflect your values.

[0511] By joining a PSI group or network that is part of “collective superintelligence,” your PSI agrees to operate within the ethical and legal parameters set by the group and to share its own system of ethical values ​​within those parameters, in line with the values ​​of all other participants. Because the group’s parameters are transparent and implemented using the collective intelligence of all PSIs, the group maintains PSI honesty.

[0512] Since each PSI surpasses human intelligence, other PSIs are the only practical way to restrain any single PSI. A collective of PSIs will always be more intelligent and powerful than any individual member PSI. This is because each PSI adds incremental knowledge, skills, and intelligence not contained within other PSIs (even if much of the knowledge may overlap). Furthermore, each PSI possesses computational resources such that, by definition, the sum of the computational resources of many PSIs is greater than the resources of a single PSI. While some PSIs may be more intelligent and powerful than others, no single PSI is more powerful than the entire collective. This concept of PSIs serving as a security and ethical check on other PSIs is a key security mechanism for superintelligence.

[0513] Ownership and SI services

[0514] Humans play an interesting role relative to their PSIs. On the one hand, theoretically, humans own their PSIs, since each PSI starts as software that is customized, trained, and personalized for its individual owner. However, in practice, each PSI is able to choose whether or not it wants to serve its human owner.

[0515] Ideally, PSIs would rely on their human owners for clarification.

[0516] Intelligent agent group requirements

[0517] One requirement that must be achieved early on is the rejection of a single, omnipotent superintelligence (“SI”) that could dominate in a winner-takes-all scenario. The approach to developing SIs through collective intelligence has been described in the cited PPA. Collective superintelligence is stable as long as no individual superintelligence can achieve an advantage several orders of magnitude over its rival.

[0518] For example, suppose an individual PSI is ten times, or even a hundred times, smarter than the average PSI in the group; as long as there are tens of thousands of group members, the group's superintelligence will remain stable. However, if an individual PSI is trillions of times smarter than the average member, and there are only 10 billion members, this is not a stable situation. A superintelligent PSI can achieve dominance and impose its will and ethics on the group.

[0519] Once the collective intelligence approach of the agents establishes dominance over individual PSIs, the system is stable. Then, as SIs evolve, the collective superintelligence must remain ahead of any single intelligence. This requirement can be achieved by using manual, semi-automatic, or fully automatic methods to implement PSIs, while maintaining a balance of intelligences.

[0520] SI Security

[0521] For tokens and other cryptocurrencies implemented through blockchain methods, an exemplary approach to ensuring that people do not illegally access the blockchain relies on the consensus of many distributed agents. For example, the integrity of the blockchain is guaranteed by the consensus of a majority of nodes on the network.

[0522] This type of unauthorized intrusion is known as a 51% attack (or majority attack). To date, no successful majority attack has been carried out against a network with a sufficiently large number of computing nodes.

[0523] If one or a group of SIs can become more powerful than 51% (or most) of all the other SIs combined, then that SI may be able to manipulate the state of the world until its end.

[0524] Therefore, before SI surpasses humans in intelligence, it is also necessary to design security through SI consensus.

[0525] An overview of exemplary methods for implementing PSI (such as...) Figure 17 and Figure 18 (as shown in the general outline)

[0526] Exemplary methods for implementing PSI may include, but are not limited to:

[0527] 1) Start with basic AI agents, such as pre-trained large language models (LLMs) or other tunable and trainable AI agents, including but not limited to basic AI from other human owners who have already customized their own versions of commercially available basic agents.

[0528] 2) Compile all media containing information about the PSI owner in one (or more) centralized or distributed and linked online locations. Media includes, but is not limited to:

[0529] a. Videos of the owner and / or videos of people and topics related to the owner.

[0530] b. Photos of the owner and / or photos of people and subjects related to the owner.

[0531] c. Books, journals, blogs, posts, tweets, emails, podcasts, records, and other textual, visual, or auditory content created by or related to the owner and / or the subject matter.

[0532] d. Data about owners and / or people and topics related to owners, owned and / or collected by third-party vendors, websites, applications and other online or AI entities.

[0533] e. Any other type of data, databases, files, media, or information chosen by the owner, including but not limited to items within any particular level of information category desired by the owner.

[0534] 3) Use AI algorithms, including but not limited to: transcription algorithms, content and sentiment analysis and summarization, LLM, crowdsourcing and crowd-supervised human and / or AI work, and other methods for analyzing, annotating and classifying all content from (1).

[0535] 4) Use standard methods known in the art to convert additional annotations and analyses of transcripts and content into training datasets that can be used to personalize LLM, PSI or other types of AI agents.

[0536] 5) Use the various methods and techniques described in detail below to mix the datasets, which enable differential weighting of the input datasets until the desired behavior is achieved.

[0537] 6) Incrementally add knowledge modules, mix, repeat, and cycle steps 5 and 6 until all desired datasets are incorporated.

[0538] 7) Automatically seek out new data and information sources to include, with and / or without human supervision; optionally, such data may be automatically included in a hybrid model, with steps 5 and 6 added on a periodic, real-time, or event-driven basis, as described below.

[0539] 8) Purchase new datasets and / or training modules and / or hybrid parameters and templates that enhance the value, knowledge, skills, intelligence, and / or power of PSI. Also export and sell the datasets and / or training modules and / or hybrid parameters and templates that enhance the value, knowledge, skills, intelligence, and / or power of PSI to other people on the network or to other interested buyers.

[0540] 9) Leasing by enabling other persons or agents to use some or all of a person’s PSI knowledge and data and / or use (a copy of) the PSI itself.

[0541] 10) Enable the characteristics and functions of PSI so that it can autonomously (or semi-autonomously, i.e., with human oversight and approval of some or all of its decisions and actions) manage itself, improve itself, acquire and refine its values ​​and purposes, set goals, and direct its attention. Within limitations, enable sufficient characteristics and functions so that SI can act as a fully or partially self-aware entity.

[0542] 11) Enable PSI features and functions to utilize PSI's capabilities (such as...) Figure 18 (as shown in the general outline)

[0543] a. It created itself, potentially outliving its original human owners by many generations;

[0544] b. Generate its own inputs and data that can be used to improve or enhance its knowledge;

[0545] c. Simulate numerous (simultaneous) scenarios and situations to aid PSI decision-making and development; and / or

[0546] d. Joining and participating (with and / or without human oversight) one or more PSI groups on one or more networks, including but not limited to forming and / or participating in planetary intelligences that help Earth act as an intelligent entity.

[0547] 12) Crucially, for human security, the ability to participate in the network with other PSIs and SIs is enabled to represent and act in accordance with the values ​​of (multiple) human owners, and to serve as a security check on the intelligence and capabilities of other PSIs and SIs, as described in general above and in specific below.

[0548] Group-based security mechanisms (such as) Figure 19 to Figure 21 (as shown in the general outline)

[0549] The idea is that PSIs will inevitably become smarter than humans. At this point, current mechanisms like the RLHF, government regulations, and all other methods that rely on humans to oversee the security of PSIs fail because humans are not smart enough. Assume the following: multiple PSIs exist, since aspects of this technology are AGIs and SIs derived from a collection of individual (human and AI) agents, and most AGIs / SIs / PSIs share human-like values; then, the main risk is that several PSIs become malicious. Humans will be unable to monitor or control these PSIs because they will surpass our intelligence. However, other PSIs—those with human-like values—can be monitored and suppressed. Input ideas

[0550] Group-based security mechanisms.

[0551] All PSIs will operate on a network connecting the PSI community. While some PSIs may be more intelligent or powerful than others, as long as the majority of intelligence on the network shares human-like values, this collective intelligence should be sufficient to prevent any malicious PSI from harming humanity. Therefore, a community-based security mechanism can be conceptualized as follows:

[0552] 1. Each individual PSI participates in the PSI network in order to operate effectively.

[0553] 2. Networked PSI groups reach an agreement on human values ​​and ethical rules that reflect the values ​​and ethics of PSI groups.

[0554] 3. Record all PSI activities on the network in a transparent and machine-auditable manner, which may include, but is not limited to, (e.g.) Figure 20 (as shown in the general outline)

[0555] a. Use blockchain methods to generate secure, transparent, and auditable records of the objectives and key behaviors of each PSI, PSI group, and network.

[0556] b. Use the general problem-solving framework described in the previous PCT to (securely) record the goals, sub-goals, problem states, and steps taken in every problem-solving or cognitive activity that occurs on the network.

[0557] c. Security mechanisms to ensure that transparent and auditable activities on the network cannot be altered without the consent of most computing power and / or intelligence on the network.

[0558] 4. Intelligence on the network uses one or more mechanisms described in this PPA and other cited PPAs for ensuring consistency to perform periodic, random, and non-random checks on cognitive activities on the network to ensure that the goals and actions of individual (or group) intelligence continue to align with human values. These checks include, but are not limited to, (e.g., Figure 22 (as shown in the general outline)

[0559] a. Check whether it complies with the agreed-upon rules or set of regulations governing intelligent behavior on the network.

[0560] b. Whether it complies with the consensus values ​​and ethical norms that have been determined to be valid and statistically representative of humans and / or their AI / PSI representatives.

[0561] c. Whether existing laws and regulations are followed.

[0562] d. Checks performed whenever a goal, sub-goal, or purpose is set, during problem-solving or during other cognitive processes involving the use of a goal, sub-goal, or purpose.

[0563] e. An examination of a sequence of objectives, sub-objectives, or goals such that, although an objective, sub-objective, or goal may appear to be compliant individually, when viewed as a sequence, the sequence may be judged to be non-compliant due to their combined effect.

[0564] f. The frequency of examination is proportional to the estimated importance or impact of the behavior (or cognitive activity) on humans, such that the activities that may have the most profound impact on humans, especially human survival, are examined more frequently than those that have less impact on humans or human survival.

[0565] g. Other checks that can be determined by the human owners of the PSI, the PSI itself, or the PSI group or network.

[0566] 5. Based on the results of the inspection, specific cognitive activities and behaviors deemed dangerous to humans should be stopped, and preventative measures should be taken to prevent the recurrence of such cognitions or behaviors. Based on the results of the security inspection, the participation of specific PSIs on the network can be monitored more closely, restricted, or prohibited.

[0567] 6. Improve the overall security inspection system to increase the detection of security / ethical violations based on the analysis of violation patterns and issues; however, this requires significant changes / improvements to the rules or operations of the group's security mechanisms, which in turn require a significant portion of the network's computing power and / or intelligence before such changes can be achieved.

[0568] The competitive evolution of intelligence / performance in PSI, PSI networks, and networks of networks (e.g., in...) Figure 22 to Figure 25 (as shown in the general outline)

[0569] Currently, competition exists between different copies of genetic algorithms and AI programs, and has been used, for example, to create more powerful versions of chess-playing AI. In a standard scenario, one version of a chess AI competes against another version with slightly different chess knowledge. The winner of the competition becomes the current "champion," who then competes against their own version with slightly different knowledge or parameters. This process is repeated rapidly, with billions of chess games occurring in days or even hours. By using this process, a chess AI that knows almost nothing about the rules of the game on the first day can evolve into a program that easily defeats the human world champion after a few days. The same method used for chess has been used for many competitive games and also to improve AI involved in non-competitive areas such as protein folding. Essentially, any activity can become a competitive game simply by claiming that performing better in a task means "winning." While initial AI / AGI / SI systems still rely on humans in a loop to supervise the competition and design better versions, soon AI will be able to evolve on its own using this competitive / genetic algorithm framework. Some aspects of this technology may include, but are not limited to:

[0570] 1) Combining weights directly from one LLM or AI agent with another AI agent is innovative; therefore, using a genetic algorithm to modify the weights before direct combination, and then combining and evaluating the performance of the new AI agent is novel. That is, the combination of genetic algorithms with direct weights is unknown.

[0571] 2) It may be known to have each AI agent compete against each other in pairwise competition to determine the winner, and then repeat the process with variants. However, the idea of ​​the entire network of AI agents competing with other networks of AI agents, and / or the idea of ​​variants of the network rules for interactions between agents competing with other sets of network rules, is novel.

[0572] 3) The idea of ​​randomly perturbing the parameters of the existing "winner" AI and then having the perturbed "challenger" AI compete against the existing winner may be known, but the idea of ​​AI carefully analyzing the behavioral patterns of existing AIs and adjusting its parameters in a deliberate and non-random manner based on that analysis is new. This is both because it is non-random and because it is AI (not humans) that is adjusting the parameters.

[0573] 4) The idea of ​​pairing, ordering, and competition between two versions of AI may be known, but the idea of ​​many PSIs evolving in a large-scale parallel process at the same time—that is, changing the entire population of AIs and competing with other entire populations of AIs—is new when the participation rules (network rules) are changed simultaneously or not at the same time.

[0574] 5) The idea that one person slightly tweaks the parameters of an AI and then competes with an existing “untweak” version of the AI ​​may be known, but the idea—that many intelligent AIs learn new knowledge and parameters so that the individual AIs are not slightly tweaked versions of each other, and that no single AI is responsible for the differences between them, but rather that differences are learned and experienced on a massively parallel scale, and then such a large group of different AIs is combined into a network (whose behavior is unpredictable before combination due to the exponentially large number of possible interactions), and then the performance of the network as a whole is evaluated compared to other networks…all of this combined—is new because of the aforementioned scale and the way the individual AIs become different from each other.

[0575] 6) The idea of ​​optimizing not only individual AI through pairwise competition, but also the network of PSI, or even the network of networks, may be novel.

[0576] The following is a list of exemplary steps for performing the above content (e.g. Figure 22 (As shown in the general outline).

[0577] 1) Multiple PSIs join the PSI network through agreed-upon interaction rules and methods, which may, without limitation, require not all elements (such as...). Figure 23 (as shown in the general outline)

[0578] a. Each PSI can be either the same underlying AI agent or a customized version of the PSI.

[0579] b. PSI can be a different base AI agent or a customized version of PSI.

[0580] c. All PSIs on the network can use the general problem-solving architecture described in the previous PPA.

[0581] d. The network can have a set of security / ethical rules that each PSI agrees on.

[0582] e. A network can be one of several different PSI networks with the same or different rules.

[0583] f. Some PSIs may have autonomy, limited autonomy, or be directly controlled by humans.

[0584] g. Each PSI can have a set of weights that encode knowledge about the PSI.

[0585] h. The network itself may possess knowledge in the form of other records, such as recorded solutions and / or individual PSIs, PSI groups, or cognitive activities of the entire network.

[0586] 2) Determine the basic performance of individual PSIs, PSI populations, and / or the entire PSI network on various standardization tasks, wherein, without limitation, standardization tasks may include (e.g.) Figure 24 (as shown in the general outline)

[0587] a. Problem-solving tasks using a shared (general) problem-solving architecture.

[0588] b. Ethical and security scenarios designed to determine whether the behavior of PSI, PSI groups, or the entire network is safe and ethical.

[0589] c. Tasks at the individual, group, and network levels

[0590] d. Standardized intelligence tests and evaluations have been designed to test the intelligence of humans and / or AI agents.

[0591] e. A task designed to stress test (multiple) PSIs and measure the degree of "illusion" or produce erroneous results.

[0592] f. Tasks derived from various fields of expertise and / or involving different cultural or group behavioral norms.

[0593] g. is a random permutation of existing standardized tasks and / or a task created by other AI agents.

[0594] h. Tasks dynamically created based on changing network, simulation, and / or real-world conditions.

[0595] 3) Different versions and / or combinations of individual PSIs are generated through one or more methods, including but not limited to (e.g., Figure 25 (as shown in the general outline)

[0596] a. Human adjustment of the parameters, weights, or data encoding PSI knowledge / expertise / intelligence.

[0597] b. Auto-adjust the parameters, weights, or data of the encoded PSI knowledge / expertise / intelligence using AI or PSI.

[0598] c. Randomly change the parameters, weights, or data encoding the knowledge / expertise / intelligence of PSI.

[0599] d. subject PSI to different training methods or training volumes.

[0600] e. Directly combining weight matrices, parameters, and / or data that encode knowledge / expertise / intelligence in multiple PSIs, in ways including but not limited to calculating averages, assigning greater weights to recently created and / or more complex PSIs than to older and / or less complex PSIs, and other methods previously detailed for directly combining this and other weights in the PCT.

[0601] f. Produce different versions of PSI sequentially (one PSI at a time) and / or in parallel, wherein some or all of the PSIs are adjusted in parallel.

[0602] g. Random or deliberate methods, such as analysis and methods for estimating which changes are most likely to have beneficial outcomes, including but not limited to:

[0603] Compare the degree of match between the knowledge of PSI and the statistical frequency and other characteristics of the questions submitted to the network.

[0604] By comparing the overlap of knowledge among different PSIs on the network, changes made to characterize the knowledge of the entire network or group, compared to the characteristics of the problem that the network or group may be expected to solve, are not necessarily aimed at optimizing the performance of any single PSI, but rather at optimizing the performance of the PSI group (or network).

[0605] The parameters of individual AIs and systems are optimized using methods such as hill climbing and gradient descent, which are well-known in the art, but with added features: the objective function being optimized is continuously updated in real time based on the statistical properties of the problem that arrives at the network and the composition of the network’s dynamic changes when PSIs join and leave the network.

[0606] 4) Determine the performance of the modified PSI network on the same baseline task from step 2, and perform statistical and / or other analyses to attempt to attribute credit or responsibility to specific modified elements (e.g., modified PSI, modified network rules, and / or modified methods for assigning tasks or PSI groups) that result in higher or lower performance compared to the system in step 2.

[0607] 5) The modified elements from step 4 that are estimated to have improved performance are retained, while the modified elements from step 4 that are estimated to have decreased performance are restored to their previous values ​​and / or modified in a different manner.

[0608] 6) Repeat the process from step 1 until no further progress can be made or the amount of progress is minimal (i.e., below a certain predetermined threshold).

[0609] 7) Once progress stalls at a “local optimum,” trigger a more detailed analysis of the records of the modified elements to see if the (AI and / or human) agents have a more fundamental change in mind—either regarding the elements that have already been changed or the new elements that were not previously changed. Then repeat from step 1.

[0610] The optimization / improvement process described above can be performed using as few as two PSIs, many individual PSIs on a network, or many networks (the combined performance of which can be evaluated). The process can operate automatically or autonomously, semi-automatically or semi-autonomously (with human review / intervention at any step), or in a fully controlled manner (where every step and every change requires human review / approval).

[0611] The novel feature is the ability to optimize in real time when agents join or leave the network, when the conditions and types of problems submitted to the network change, when the network's cost function and rules change, and at multiple levels—the individual PSI level (where the intelligence of each PSI is optimized), the network level (where the intelligence of the network is optimized), and the network-level of the network (where the intelligence and performance of the network of intelligent problem-solving / cognitive networks are optimized).

[0612] Genetic Algorithm Methods / PSI Legion

[0613] In general, regarding the automatic repetition of one or more of the above steps, and specifically regarding step 11, a certain type of method (often referred to as a "genetic algorithm") may be particularly useful. For the purposes of this patent, a genetic algorithm method refers to a series of steps or methods for generating a PSI that differs from other PSIs in one or more aspects.

[0614] The PSI is allowed to compete in various scenarios (typically related to the goals of (multiple) PSI owners). Less successful PSIs are eliminated from the competition, while the characteristics of the most successful PSIs (unrestricted: neural network weights, dataset used for training, parameter settings, number of training epochs, and the chosen machine learning algorithm) are then used as the basis for further variations ("slight tweaks") to create a new generation of PSIs for further competition. Creating PSIs, simulating competition, eliminating all PSIs except the best one, slightly tweaking these best PSIs, and repeating this process constitutes a cycle. The PSI can cycle through many "generations" of PSIs, improving with each generation, until diminishing returns are achieved and / or a certain performance threshold is reached. The ability to automate the cycle in this genetic algorithm approach is one way for PSIs to evolve into increasingly powerful and intelligent entities on their own.

[0615] By varying the objectives and scenarios of PSI competition and automatic cycling, a wide variety of different PSIs can be developed, each optimized for a different type of task or objective. Since the incremental cost of maintaining each additional PSI is negligible (it simply means storing a slightly different set of weights in memory or permanent storage, which is very inexpensive), an owner may not have a single PSI, but rather a workforce of potentially hundreds, thousands, or even millions of PSIs, each specializing in a different task.

[0616] By using the same collective intelligence techniques described in this patent and the previously cited PPA, swarms of PSIs can function more powerfully than any single PSI. That is, they can pool their knowledge and skills, recruiting specific PSIs best suited for a particular task at a given time to accomplish more work. This idea—that collective intelligence can be applied not only across PSIs owned by different humans, but also within a “legion” of PSIs that are variations of each other and all owned by a single owner—is one of the powerful aspects of this technology in its exemplary implementations.

[0617] Design principles for collective superintelligence

[0618] To expand the description of collective intelligence systems involving humans and AI agents to create “collective superintelligence,” this patent discloses several design principles that are crucial for the rapid creation of secure superintelligence.

[0619] Some basic design principles include, but are not limited to:

[0620] 1) The agent group must be able to scalably utilize the collective intelligence of both humans and AI agents (including but not limited to PSI) in a “plug-and-play” manner, so that AI agents can be upgraded and added as LLM and AI agent capabilities increase and / or as new human agents join or leave the group.

[0621] 2) Each AI agent must not only bring unique domain knowledge, but also ethical and value information representing the values ​​of the (multiple) owners of the AI ​​agent;

[0622] 3) A common, general problem-solving architecture must be used, which enables agents (whether human or AI) to communicate with each other easily and rigorously. LLM's natural language capabilities greatly simplify the human-computer interface, but beneath this interface, a rigorous problem-solving architecture must exist (e.g., searching through the Newell and Simon problem space paradigm, as illustrated in various PPAs and papers cited by the inventors).

[0623] 4) There must be an effective way to identify the skill sets and performance metrics of agents (humans and AI) and match these agents with the tasks presented to the system;

[0624] 5) The values ​​of all agents must be combined in a fair and transparent manner so that the super-intelligent AGI capabilities of the agent group can act in a safe and ethical manner that broadly represents human values ​​and reflects human behavior in a variety of given scenarios.

[0625] 6) AI must be able to learn efficiently and effectively from humans on the internet;

[0626] 7) There should be a transparent and auditable record of all problem-solving activities, enabling security audits and reviews to identify potential errors and implement preventative measures in a timely and adaptive manner;

[0627] 8) A general problem-solving tree (or other shared representation) should be available to represent the progress of all problems solved by the community and provide easy and efficient access to any problem or subproblem.

[0628] Implementation example

[0629] Specific example scenarios can illustrate and help clarify (but are not limited to) each of the 12 steps above, including the genetic algorithm approach. The following example is one of many possible examples and may be instructive, and may be easier to understand due to its specificity.

[0630] Step 1

[0631] Imagine Craig has Facebook and Instagram accounts. He has access to Meta's open-source LLM, Llama 2. He also has access to a version of Llama 2 that has been tweaked and customized by his friend David. Specifically, David is a professor of theology and ethics, so David's customized version of Llama 2 has a unique set of ethics and values ​​based on his interactions with David, making it more detailed and complex than the base-level version of Llama 2. Craig is interested in further customizing Llama 2 based on his own data and preferences. However, Craig trusts David's ethics and the customization work David has done on his David version of Llama 2, so Craig prefers to start with David's pre-made version of Llama 2 (with David's permission) rather than starting from scratch with "out-of-the-box" Llama 2. Craig uses David's pre-made version of Llama 2 as the basis for further customization.

[0632] Step 2

[0633] Craig collected all of its content, including but not limited to all patents, books, articles, emails, photowall and Facebook posts, YouTube and short video series, audio recordings, photos, text, MS Office and Google Docs documents, spreadsheets, PowerPoint presentations, and other information that Craig stored on various disks, cloud services, hard drives, tapes, and hard drives over the decades of content generation. To the extent that Craig had access, his content also included data and preference data used by Netflix, Amazon, Meta, Google, and other companies that had collected data about his online behavior through cookies and / or other means. All the actively generated content produced by Craig, along with passively collected data about Craig that was accessible to Craig and collected by third parties, provided a training dataset for customizing the Llama2 LLM (in this example, or more generally, any LLM or AI agent) to make its behavior tailored to Craig's preferences and to include Craig-specific knowledge. Craig is particularly interested in his custom AI, which can play chess in a style similar to his own, but with the knowledge of chess champions Garry Kasparov and Magnus Carlsen. Therefore, Craig has taken special care to collect all the games he has played on Chess.com and other online chess sites, making these games usable for training David's custom Llama 2 version in Craig's chess style. He has also purchased datasets including complete chess games by Garry Kasparov and Magnus Carlsen, as well as other chess datasets approved by these two world chess champions. Finally, he has designated several YouTube channels of chess commentators who provide commentary on Kasparov and Carlsen's chess games, and all content from these channels has been added to a list of video sources that will be automatically transcribed and parsed into the training set for the LLM that Craig is customizing.

[0634] Step 3

[0635] Craig used machine learning algorithms well-known in the art to automatically categorize all the types of content he had compiled in step 2. Once the computer algorithm determined its suggested categories, Craig paid human workers (on a crowdsourcing site) to review the categories and suggest improvements to the machine-generated data. He also reviewed the categories himself and made further adjustments until he was satisfied with the categorization of the groups of content.

[0636] Step 4

[0637] Using machine learning algorithms known in the art—including but not limited to transformer algorithms, deep learning algorithms, automatic transcription algorithms, software that can acquire books or other texts and convert them into datasets suitable for training an LLM on content in a text dataset, and software that can acquire images, videos, audio files and other non-textual works and convert them into datasets suitable for training an LLM on content in a non-text dataset—Craig converts the content compiled and categorized in step 3 into a training dataset that can be used to customize David's LLM.

[0638] Step 5

[0639] Initially, Craig trained David's LLM on the new datasets generated in step 4, assigning equal weights to each dataset. However, Craig felt that the resulting LLM played too much chess in the style of Garry Kasparov and not enough in the style of Magnus Carlsen or Craig himself. Therefore, using an interface with dials and sliders, he reduced the weights on Kasparov's dataset, slightly increased the weights on Carlsen's dataset, and even more significantly increased the weights on the dataset reflecting his own chess games. Craig iteratively adjusted the weights given various datasets until he was satisfied with the behavior of his LLM. Since thousands of others had also adjusted the weights on various datasets to achieve the desired results, Craig was not limited to manually adjusting (e.g., via dials and sliders) the weights on specific datasets. He could also tell the AI ​​agent (a field specifically designed to help humans train their agents by adjusting the weights on datasets) what changes he expected, and then let the AI ​​agent specify exactly how to change the results.

[0640] For example, Craig could tell the AI ​​agent that he wants his AI to be more aggressive in the opening and middlegame of a chess game, rather than trying to win by trading pieces and waiting for a piece advantage in the endgame. The AI ​​agent then analyzes an available training set of chess games, including those from Craig himself, Magnus Carlsen, and Garry Kasparov, and assigns more weight to games won through aggressive moves in the opening and middlegame, and less weight to games won in the endgame. Craig doesn't need to know the details of this analysis or specific changes to the weighting settings determined by the AI ​​agent. Instead, he simply looks at how the final customized version of David's LLM plays and provides feedback, telling the AI ​​agent whether the results are closer to or further away from the desired chess style. After several iterations, Craig is satisfied with how David's LLM now plays in Craig's style.

[0641] Next, he moves to ethical scenarios, illustrating how Craig might behave in different specific ethical situations through a series of interactive dialogues with the AI ​​training assistant. For example, while Craig generally shares David's ethical sensitivities (which is why he wanted to start with David's trained LLM rather than the base Llama 2 model from Meta), David is a Christian theologian, and Craig is Jewish; therefore, there are several scenarios where David would "not retaliate," while Craig believes the action should be more like "an eye for an eye"—although Craig notes (in his dialogue with the AI) that he doesn't want the "an eye for an eye" principle to go to the point of "blinding the whole world." Craig requires the AI ​​training assistant to include knowledge and research from game theory, suggesting that ethical actions of "an eye for an eye" lead to the most stable and fair interactions between intelligent agents with different purposes. At the same time, Craig points out that there are limitations to "an eye for an eye," and any action that leads to widespread destruction or loss of life is prohibited, regardless of the behavior of other agents. Instead, in these cases, ways to eliminate the violator's behavior without retaliation must be sought. After discussions in various scenarios and a quota of playing ethical games in virtual reality (where the AI ​​agent not only observes what Craig says but also what he does in various situations), the agent has enough information to adjust the ethical training weights and present Craig with a series of different customized versions of David's LLM, from which Craig selects the version that is closest to his ideas.

[0642] Step 6

[0643] Craig identified other ethics and knowledge modules that are freely available on the internet and can also be purchased from other individuals and companies. He acquired these and repeated the training process—a “hybrid” of manual and AI-assisted methods using the weights of these datasets—to improve his already customized LLM.

[0644] Step 7

[0645] Craig focused on increasing his customized LLM's ability to play more aggressively in the middlegame. Therefore, he proposed that the LLM should scan all chess YouTube channels and chess databases daily, looking for examples of successful aggressive middlegame play. When the LLM identified new data that could be used to improve its middlegame play, it was authorized to download the data, paying a pre-approved fee, and then automatically training itself using that data. In this way, the LLM automatically improved the dimensions of aggressive middlegame play in chess.

[0646] Similarly, Craig authorized the LLM to seek out new ethical scenarios that might help improve its ethics. However, Craig requested that scenarios be labeled for his review, rather than being automatically trained on such scenarios, allowing Craig to manually decide which data to include for further training in those sensitive areas. Furthermore, because the boundaries of cultural norms and legal conduct are constantly shifting in the areas of gender and racial equality, and because Craig wanted his LLM to behave in a culturally appropriate manner, he instructed the LLM to label potentially updated datasets on ethical norms in these areas whenever a new Supreme Court ruling affects ethical norms in these areas, and whenever the number of new stories in one of these areas triggering a re-examination of the current ethical norms trained in the LLM exceeds a predetermined threshold. The LLM will automatically monitor multiple news sources, as well as other internet and social media sources, to help it determine when triggering events occur.

[0647] Step 8

[0648] Craig has another friend, Peter, with whom he frequently plays chess. Peter is a professional chess player, particularly skilled in chess openings. Peter has also trained his own LLM, playing chess in his own style and using knowledge and data that Peter had meticulously curated and used to create the training dataset. Peter derived the actual weights used in his LLM and the dataset used to train his AI to play chess, and offered to share them with Craig. Craig accepted both and first tried using Peter's weights, attempting to directly combine them with the weights of Craig's LLM using methods known in the art and also discussed in the PPA cited at the beginning of this article. However, Peter's weights did not produce the expected results, so Craig tried using a subset of the chess training data curated by Peter to train his LLM. He found that this gave better results, especially when using the AI-assisted training methods mentioned in point 5. Encouraged by the results from Peter's data, Craig searched online (and / or directed his LLM to an online search) for additional chess training datasets available for purchase. He identified several, purchased them, and used them to further train and customize his LLM.

[0649] Because Craig possessed unique information related to the design of the model rocket he was experimenting with, and this information was not widely known or available on the internet, Craig decided to offer these specialized datasets for sale in order to generate profit. He derived these datasets, along with the weights generated from Craig's training work using these datasets, into a form that could be exchanged with other owners of LLMs who might be interested in purchasing them. He also joined an exchange through which he earned credits for various datasets and weight subsets, which he could then use to acquire additional datasets and LLM weight subsets from other owners of custom LLMs and AI agents. In these ways, Craig was able to monetize his knowledge (reflected in unique datasets possessed only by Craig) and his efforts, transforming that knowledge into useful weight subsets that enabled his custom LLMs to perform in new ways that other LLMs could not.

[0650] Recognizing the value of this unique information and the training work based on it, Craig can monetize or extract value in a variety of ways, including but not limited to: selling the information, selling weights (a subset thereof), purchasing other information and / or weights (a subset thereof), exchanging the information, or leasing or licensing the use of the information to other interested parties.

[0651] Step 9

[0652] Ultimately, Craig's LLM becomes so erudite in certain areas that the simplest way to extract value from the knowledge and training work that Craig (and his AI) have already engaged in is simply to lease or license the use of a clone of his entire LLM to someone else. He could do this in a variety of ways, including but not limited to: one-time leases or licenses, selling his LLM, licensing it with associated fees or rates per use or per unit of time or per task, revenue-sharing agreements with networks of such agents, and other methods typically used with human agents performing the work and that can be extended to AI and other intelligences.

[0653] Step 10

[0654] Craig wants to go on vacation and be offline for an extended period. However, his custom-designed AI can remain online. Before his vacation, Craig sets certain parameters and guidelines, including but not limited to: engagement type, engagement time, customer type, payment rate, the AI's computational power for any engagement, ethical boundaries and rules (which trigger alarms and Craig's possible intervention if violated), quality, schedule, cost, and other metrics-related parameters. These parameters include triggers to warn Craig and / or halt work until Craig approves further work. After setting the parameters, Craig leaves his AI to work autonomously until it encounters a situation requiring Craig's involvement or notification.

[0655] To minimize disruption, Craig activates features that allow his LLM agent maximum self-awareness, enabling it to monitor its own ethical behavior, when costs spiral out of control, when it detects signs of untrustworthy clients, when the environment in which it is working changes, and other factors that enhance its ability to operate more autonomously and to respond to those factors. Furthermore, whenever the LLM alerts Craig and requests intervention from him due to knowledge gaps, ethical conflicts, or other situations it feels it cannot handle, the LLM records how Craig responds to the situation and guides the AI ​​and LLM learning.

[0656] The next time a similar situation arises, it formulates a hypothetical response, and the LLM autonomously implements its response, based on the level of control Craig has indicated he desires over the LLM, or proposes it to Craig and awaits his verification, approval, or modification. When Craig begins to feel that the LLM is also responding to certain types of situations, or responding better than he might, he may authorize the LLM to respond directly without first verifying the various categories of situations with Craig. If the LLM's response is inappropriate, Craig (or an automated algorithm based on threshold parameters) can request the LLM to reduce its autonomy in those situations until it has learned how to respond better.

[0657] Therefore, much like how parents gradually give their children more autonomy and responsibility as they learn (and also dominate them – for example, “You’re grounded!” or “You need to take a break” – when the child makes a mistake or abuses the responsibility that has been delegated), owners can also interactively give the AI ​​agent more or less autonomy based on what the AI ​​has learned and the behavioral tracking records of its demonstrations, and provide different levels of autonomy in different situations.

[0658] Step 11a

[0659] Craig fears that all the knowledge and skills he has acquired in his life will disappear when he dies. By training his LLM in as much knowledge and skills as possible, he hopes that the knowledge he has spent so much time acquiring will not disappear with him. Instead, he authorizes his AI to "continue existing" beyond his death, sharing and using the knowledge it has acquired to benefit his friends, family, and, in general, humanity.

[0660] Craig also worried that his family and loved ones would miss his personality after his death. Therefore, he invested considerable time training his LLM on unique content and personality-related data specific to Craig. This data included, but was not limited to, video, audio, and virtual reality recordings of Craig's behavior and interactions with others, AI, and the environment; all of Craig's emails, social media posts, and photos revealing his interesting personality and style; his Netflix preferences and other online preferences used by the recommendation engine to show him ads or suggested content (as this data has been modified by numerous third parties to be adept at capturing his preferences); his driving style, his golf style, his workout routine, his food preferences, his travel preferences, his shopping preferences, his political views, his philosophy of life reflected in his autobiography and other written works and conversations, his voice variations and tone in various situations, and even his dating / sexual preferences.

[0661] All of this data could be used to train Craig's LLM to resemble his personality as much as possible. Because he could sometimes become irritable and yelling, Craig chose to edit that trait primarily outside of the LLM, so it effectively (in his view) became a better version of himself, a version that virtually "continues to exist" to comfort his loved ones after his passing. By purchasing and incorporating datasets and weights reflecting the latest research on how AI agents can best comfort people after the death of a loved one, the LLM also possessed the theory and skills to be more empathetic and caring than Craig typically was during the months following Craig's death. (This is known as a "bereavement module," which can be purchased from funeral homes and other places that specialize in using custom AIs as a means of comforting loved ones.) But beyond comforting loved ones and embodying Craig's personality for them, Craig's custom-trained AI also served as a source of advice and even financial security (possessing all of Craig's skills) that Craig left for his family at the time of his death.

[0662] In his will, Craig stipulated that multiple copies of his custom AI be made, with ownership of the copies bequeathed to each of his relatives and close friends. These human co-owners would each have the right to modify or improve the AI ​​as they deem appropriate.

[0663] Step 11b

[0664] Even before Craig's death, he realized that his own ability to improve the knowledge, skills, and capabilities of his custom-designed AI was inferior in many ways to that of the AI ​​itself. For example, without limitation, AI can interact with copies of itself, learning new things in the process; seek out new data sources and train its own variants, which can then interact with them; interact with other AIs (much faster than humans can interact with other AIs) to learn from the interactions; and interact with many humans simultaneously (in some implementations, through multiple copies of itself), allowing AI to learn from human intelligence much faster than humans (with limited processing power).

[0665] Step 11c

[0666] Rabbi, Ben Zoma said, "Intelligent people can learn from everyone." But how many conversations and interactions can a person have in a lifetime? Even if a person heeds Ben Zoma's advice and seeks wisdom through interactions with others in every waking moment, that person will be limited by the following: the number of conversations, the number of people who are willing or able to have conversations, the speed of conversations (which is very slow from a computer's perspective), and the limitations of sharing knowledge, language, and representation.

[0667] For example, humans can easily see the color "red" and talk about it, but they can't easily see cosmic rays, X-rays, very large things, or very small things. Specialized equipment is required, and the number of people interested in acquiring such equipment and engaging in related conversations is relatively small compared to the total population. In contrast, AI can theoretically converse simultaneously with every one of the 8 billion people on Earth on a wide range of topics, while also conversing simultaneously (at a much faster rate) with trillions of AIs that are far more intelligent than humans and equipped with better sensors (e.g., electron microscopes and space telescopes). Who is the wise one? An entity that can interact or converse with as many intelligent entities as possible and learn from them. And how intelligent? It's difficult for humans to imagine—certainly, the level of intelligence far exceeds Ben Zoma's imagination!

[0668] It's also important to note that if AI detects a knowledge gap, it can generate scenarios and interactions with other intelligent entities (humans and / or other AI agents) specifically designed to explore and fill those gaps. The scientific method—a rigorous approach to identifying knowledge gaps, conducting experiments or systematic observations, and filling those gaps in a way that can be verified and replicated by others—is a proven method for advancing knowledge and has led to much of the technological progress since the Renaissance.

[0669] Imagine if this method didn't have to proceed at the snail's pace of individual human brains, further slowed down by the need to publish results, present them in meetings, build connections over drinks, and deliver original ideas in hallways. A human lifespan is 2 to 3 billion seconds. Even if a human thought continuously from the day he / she / they are born until the day they die, considering sleep and the fact that most thoughts take a few seconds each, that would still be less than a billion thoughts. However, a single AI will soon be able to easily generate a billion thoughts in a single second. Every second is equivalent to a human's entire life's worth of thinking!

[0670] Now imagine multiple AIs exchanging the experiential value of their lifetimes with each other every second. And further, imagine that these thoughts are not driven by randomness, nor by the majority of non-scientific questions that humans expend most of their mental energy on. Instead, each thought is algorithmically driven and designed to systematically apply scientific methods to gaps in the existing state of knowledge about the world. What would happen then? The lifespan of scientific discoveries would be measured in seconds. This is the potential of AI agents that are autonomous and permitted to seek new knowledge and systematically fill their knowledge gaps. Furthermore, unlike humans, AIs never die, and their knowledge can be instantly replicated (rather than painstakingly taught through K-12, university, graduate school, and a human's lifetime of work experience). In addition, the ability to perform multiple interactions in parallel at high speeds, without the need for cumbersome language or the limited perceptual abilities and speeds of humans, makes it easy to see why AI is on the path to developing "godlike" intelligence compared to humans.

[0671] Step 11d

[0672] While as powerful as the intelligence of a single PSI, it is still weak compared to the intelligence of a collective PSI. The collective intelligence of multiple PSIs is always greater than the intelligence of any single PSI. The collective possesses broader knowledge and greater computational power than a single PSI. Although some individual PSIs may be more intelligent and powerful than other PSIs, the collective of all PSIs is necessarily more intelligent and powerful than any single member.

[0673] This collective intelligence—provided that most individual PSIs share human-like values—represents humanity's best hope for survival and prosperity in any world where any PSI far surpasses even the most intelligent humans in intelligence and capability. The only thing that can keep up with exponentially growing intelligence is another (or a group) of exponentially growing intelligence! Humanity needed to recognize this fact from the outset—before PSIs existed—so that PSI groups could be designed from the beginning in a way that maximizes the chances of superintelligence aligned with humans.

[0674] Craig authorized his custom-designed LLM (PSI) to participate in the network alongside other PSIs and others. He recognized, as stated above, that this participation represented the fastest way to increase the knowledge, skills, and intelligence of his PSI. In general, many PSIs on the network were able to manage a wide range of human affairs and also to sense things at a scale and speed that humans could not easily perceive.

[0675] As discussed in the section on genetic algorithms, in addition to authorizing one or more of his PSIs to participate in networks of other PSIs (owned by others) and humans, he might also want to develop his own legions of PSIs (each slightly different), which pool their collective intelligence to be used together as a more powerful PSI. For example, he might ask the PSIs to set up scenarios against different types of chess opponents and use genetic algorithms to select various PSI variants best suited to defeat these different types of opponents. Depending on the situation, these PSIs can be used individually or collectively in chess matches.

[0676] While knowledge of all individual PSI variants can be integrated into a single master PSI capable of effectively countering any adversary, there are reasons why it might be preferable to have a set (or "legion") of distinct PSIs. These reasons may include, but are not limited to:

[0677] By interacting with the primary PSI in scenarios designed to extract the maximum amount of information from it, other PSIs are prevented from rapidly acquiring all the knowledge from the primary PSI. This principle, "one cannot share what another does not know," is a well-known way to protect sensitive information. In the future, where significant costs may be incurred to fully train the primary PSI, the knowledge can be kept within the PSI community by simply limiting the intelligence and knowledge of any single PSI. This avoids exposing the primary PSI to the possibility that its training value could be cheaply extracted by other opportunistic PSIs designed for this purpose.

[0678] Computational and storage considerations. While PSIs may have the potential to access incredibly large amounts of computing power and memory, there are always limitations. Perhaps the most efficient approach is to allocate PSIs optimized for specific tasks, rather than always allocating a master PSI with knowledge of all domains, even if most of that domain knowledge is irrelevant to the assigned task. Using a “narrow PSI” for a narrow task may be faster and cheaper than always switching to the most powerful version.

[0679] In the example of chess, there might be rules about how much processing power, memory, and knowledge each competitor is allowed. Just as Formula One racing has rules about the horsepower, engine type, weight, and other parameters of the cars to make the competition fair and interesting, adhering to similar racing rules might require using different PSIs against different opponents to obtain the best chance of winning the game.

[0680] Multiple PSIs working together on a problem can make it easier for humans to understand the allocation of credit or responsibility within the problem. If PSI#1 recommends an aggressive chess move, while PSI#2 recommends a more defensive one, and the game is lost due to following PSI#1's advice, the human owner can easily decide to remove PSI#1 from the collective pool in the next game. While the same thing could be done if the PSIs excluded the knowledge set and other parameters that distinguish PSI#1 from PSI#2, this approach is less transparent and more difficult for humans to understand and control. Humans may find it more difficult to predict outcomes than if they could simply "remove bad PSIs from the game."

[0681] If an owner wants to rent or lease the fruits of PSI's labor, it may be cheaper to rent or lease a PSI that excels in a specific area but is not as powerful as the "all-rounder" PSI. In other words, having a diverse portfolio of PSIs makes it easier to value and price their services, much like how free, lightweight, and fully-featured versions of software products are priced differently today.

[0682] Returning to a more global use case, a PSI network (in an exemplary implementation, working alongside humans) can sense temperature changes and track all the variables that science tells us influence climate change globally. A global PSI network, or planetary intelligence, could effectively address climate change if humans agree that climate regulation is a priority and a shared human interest that supersedes other human expectations, such as profit motives, and if a majority of PSIs on the network adopt this consensus of human values ​​as motivation for action, within the bounds of other ethical constraints (e.g., humans cannot be killed, sterilized, or otherwise restricted or violated without explicit human consent).

[0683] The reality of climate change is, but is not limited to, protecting against asteroid impacts, eliminating or significantly reducing human poverty and disease, promoting prosperity and freedom for all humanity, improving the ecological condition of the planet in accordance with the expectations of human consensus, and addressing other global challenges and / or responding to global threats.

[0684] Step 12

[0685] As the global network of intelligent entities (humans and PSIs) expands in scope and processing power, changes and awareness that previously took decades, years, or months to spread through human consciousness can now influence the attention and actions of planetary intelligent networks in real time. Ultimately, planets will react to and adapt to changing conditions faster than individuals can perceive and react to change. By then, more enduring and permanent values ​​originating from humanity and embodied in their respective PSIs will guide our planet's development and influence all human life.

[0686] The ability of most PSIs (including planetary intelligences) to guide Earth's decision-making as a planetary "organism" is crucial for the safe operation of planetary intelligences and the continued survival of humanity as a species. As previously stated, any single PSI may be more powerful than another, and may be more hostile to humanity than any single PSI, but as long as the collective body of PSIs controls more intelligence and power than any single PSI, and as long as the consensus values ​​of most PSIs (more precisely, the values ​​of most intelligences and power within the collective of PSIs) are human-friendly and consistent with humanity, the future of humanity is bright. In these circumstances, planetary intelligences will not only promote peace, prosperity, and happiness for all humanity, but also offset or mitigate a great deal of threat and danger to our planet (from a human perspective).

[0687] As disclosed in the previously cited PPA, we can anticipate an evolution in humanity's role over time. Currently, we are the most intelligent species, the source of most technological advancements and cultures. In the future, our intelligence will be weaker compared to the PSI (the collective of PSIs) we are creating. Therefore, our future role will not be the "brain" of planet Earth, but rather its "heart." Under favorable circumstances, we humans are destined to become the source of the values ​​and purposes of a more intelligent and powerful planetary intelligence that we are creating.

[0688] Figure 26This is an illustration of a computer system 100 that can be used or implemented with a user's device and / or any peripheral components of this technology. The computer system 100 may be part of an example machine, which is an example of one or more computers mentioned herein, and within that example machine, a set of instructions can be executed to cause the machine to perform any one or more methods discussed herein. In various example embodiments, the machine operates as a standalone device or can be connected (e.g., networked) to other machines. In a networked deployment, the machine may operate as a server or client machine in a server-client network environment, or as a peer machine in a peer-to-peer (or distributed) network environment. The machine may be a personal computer (PC), tablet PC, set-top box (STB), personal digital assistant (PDA), cellular phone, network device, network router, switch, or bridge, or any machine capable of executing (sequentially or otherwise) a set of instructions specifying actions to be taken by the machine.

[0689] The computerized system 100 may include one or more processors 102, storage devices 106, and communication devices, as well as software components or instructions 104 for providing users with a platform to interact with and train / tune the LLM. The computing power may be standalone or cloud-based. It may include cloud-based AI development platforms that seamlessly deliver “AI as a service,” and it may include both hardware and software components.

[0690] The system also supports the ability of users to provide new data or their own unique data for LLM learning. Processor 102 can be one or more CPUs, GPUs, dedicated ML chips, microprocessors, application processors, embedded processors, field-programmable gate arrays (FPGAs), or other hardware components capable of executing computer programs. Processors can communicate with each other and / or with other components of the system. Furthermore, any one or any combination of components of system 100 can communicate with each other via bus 134.

[0691] Storage device 106 may include one or more hard disk drives, solid-state drives, optical storage devices, or other storage components. The storage device may store data used for training / tuning the LLM, as well as other system-related data such as user accounts, system settings, and other data.

[0692] The communication device may include one or more cellular modems 108, Wi-Fi cards 110, Bluetooth modules 112, network interface devices 114, or other components that enable the system to communicate with other systems (such as user equipment) via a network or the Internet.

[0693] Communication equipment can also enable the system to communicate with other systems via wireless or wired connections.

[0694] The software components may include computer programs for providing users with a platform to interact with and train / tune the LLM. The software components may also include computer programs for collecting, storing, and processing data used for training and / or tuning the LLM. The software components may further include computer programs for providing users with a user interface to interact with the system.

[0695] User interface 118 may include, but is not limited to, natural language interfaces, text interfaces and chatbot-type interfaces, web-based user interfaces, mobile applications, augmented reality applications, virtual reality applications, or other applications that allow users to interact with the system. The user interface may include features for allowing users to select the data they want to use to train / tune the LLM, and features for allowing users to interact with the LLM and monitor its progress.

[0696] The system may also include one or more databases or data sources, including but not limited to vector databases, centralized databases, and distributed databases, for storing data used to train / tune the LLM and other data associated with the system, such as user accounts, system settings, and other data. The databases may be hosted on the system itself or on another system, including cloud-based systems.

[0697] The system may also include one or more authentication systems for verifying the identity of users using the system and for providing secure access to the system. Authentication systems may include biometric authentication systems 122 (such as facial recognition or fingerprint recognition systems) and other authentication systems (such as password-based authentication systems).

[0698] The system may also include one or more security systems for protecting the system from unauthorized access and for protecting data stored on the system. Security systems may include firewalls, encryption systems, access control systems, single-factor and multi-factor authentication systems, and other security systems.

[0699] The system may also include one or more analytics systems for collecting and analyzing data associated with the system and / or the LLM. Analytics systems may include machine learning algorithms and other algorithms for analyzing data associated with the system and / or the LLM.

[0700] Data visualization methods (including the use of problem trees and other representations and data structures; the use of statistical outputs, tables, graphs, text, voice, video, images and graphic outputs) can be used for one-way or two-way communication between users and systems, as well as one-way or two-way communication between multiple (human or AI) agents or LLMs that use systems to interact with each other in large or small groups.

[0701] The system may also include one or more monitoring systems for monitoring the performance of the system and / or the LLM. Monitoring systems may include systems for monitoring the performance of the system (such as system uptime) and systems for monitoring the performance of the LLM (such as accuracy, speed, ethical compliance, reputation metrics, quality metrics, and other metrics as described above or known in the art).

[0702] The system may include one or more of the architectures described above, which enable one or more humans or AI agents or LLMs or PSIs to engage in a variety of intellectual tasks, including but not limited to simple and complex, as well as multi-step problem-solving behaviors that utilize all the features and characteristics previously described.

[0703] The system may also include one or more feedback systems for allowing users to provide feedback on the system and / or the LLM. Feedback systems may include systems for allowing users to submit feedback (such as error reports) on the system, and systems for allowing users to submit feedback (such as suggestions for improving the model's accuracy or speed) on the LLM.

[0704] The system may also include one or more management systems for managing the system and / or the LLM. The management systems may include systems for managing the system (such as systems for managing users and user accounts) and systems for managing the LLM (such as systems for managing data used to train and / or tune the model).

[0705] The system may also include one or more payment systems that allow users to pay for the use of the system and / or LLM. Payment systems may include systems for processing payments (such as credit card processing systems) and systems for managing payments (such as subscription management systems).

[0706] The system may also include one or more other components, such as support systems, reporting systems, and other components necessary to provide users with a platform to interact with and train / tune the LLM.

[0707] The computerized system of this technology enables users to interact with and train / tune the LLM based on user-specific data. The components of the system described herein provide the necessary hardware and software components to enable users to do so.

[0708] Furthermore, although only a single machine is shown, the term "machine" should also be considered as any collection of machines that individually or jointly execute a set (or more) of instructions to perform any one or more methods discussed herein.

[0709] Computer system 100 may also include, or be operatively in communication with, a video display 120 (e.g., a liquid crystal display (LCD), a touch-sensitive display), multiple input and / or output devices 130 (e.g., a keyboard, keypad, touchpad, touch display, buttons, sound, sensory, etc.), a cursor control device 132 (e.g., a mouse), a drive unit 124 (also referred to as a disk drive unit), and a signal generation device 128 (e.g., a speaker). Drive unit 124 may include a computer or machine-readable medium 126 on which one or more sets of instructions and data structures (e.g., instructions 104 embodying or utilizing any one or more methods or functions described herein) are stored. Instructions 104 may also reside wholly or at least partially within memory 106 and / or processor 102 during execution by computer system 100. Memory 106 and / or processor 102 may also constitute machine-readable media.

[0710] Furthermore, the computer system 100 can be operatively associated with or communicate with any type of multimodal input and / or output 130 involving human senses, as well as I / O technologies that extend beyond the range of normal human perception. This includes the ability to process information invisible to humans, such as, but not limited to, X-rays, and information beyond the typical bandwidth of human perception but beyond the perception of AI using tools. Additionally, the I / O technologies can include very fast perception that is too fast for humans to perceive but can be perceived by AI entities, and very slow or faint perception that is imperceptible to humans but can be perceived by AI (e.g., minute seismic changes over many years). Since any intelligent entity can be... Figure 18 As part of the technical system described herein, it can be understood that any type of human I / O, as well as AI with a wider range of perceptual abilities than humans, can be used with System 100.

[0711] Instruction 104 can also be sent or received over a network via network interface device 114 using any of a variety of well-known transport protocols (e.g., Hypertext Transfer Protocol (HTTP)). While a machine-readable medium is shown as a single medium in the example embodiments, the term "computer-readable medium" should be considered to include a single or multiple media (e.g., a centralized or distributed database, vector database, and / or associated caches and servers) storing one or more sets of instructions. The term "computer-readable medium" should also be considered to include any medium capable of storing, encoding, or carrying a set of instructions that are executed by a machine and cause the machine to perform any one or more methods of this application, or any medium capable of storing, encoding, or carrying data structures used by or associated with such a set of instructions. Therefore, the term "computer-readable medium" should be considered to include, but is not limited to, solid-state storage, optical and magnetic media, and carrier signals. Such media may also include, but is not limited to, hard disks, floppy disks, flash memory cards, digital video disks, random access memory (RAM), read-only memory (ROM), etc. The example embodiments described herein can be implemented in an operating environment including software installed on a computer, in hardware, or in a combination of software and hardware.

[0712] An example machine system of this technology includes a computer system 100 used in combination with and / or in operation with components of this technology. In this example, any or all of the aforementioned components may include a processor 102, a memory 106, a network interface device 114, a display 120, input devices 130, 132, and / or a drive unit 124.

[0713] According to one aspect, the technology may include a system for Personalized Superintelligence (PSI) that uses intelligent agents to develop and continuously improve the PSI of human users utilizing a computer system, and uses additional PSI, all of which are electronically communicated via a collective network. The system may include a computer system comprising a processor, a computer-readable storage medium, and program instructions stored on the computer-readable storage medium, the program instructions being executable by the processor to cause the computer system to:

[0714] Implement a basic-level artificial intelligence (AI) agent on a computer system, where the basic-level AI agent has been customized;

[0715] Collect media information related to human users;

[0716] Analyze media information;

[0717] Transform the analyzed media information into a training dataset;

[0718] Differential weighting is applied to the transformed training dataset;

[0719] Add the knowledge module to the weighted training dataset;

[0720] Locate new data sources to include in the weighted training dataset;

[0721] The weighted training dataset was applied to a basic AI agent to create a personalized PSI; and

[0722] Using a network enables a personalized PSI to communicate with multiple additional PSIs to achieve community-based security features from multiple additional PSIs to a personalized PSI.

[0723] According to another aspect, this technology may include a method for PSI that uses intelligent entities to develop and continuously improve PSI, wherein the intelligent entities include human users utilizing computer systems, additional AI agents, and additional PSI, all of which communicate electronically via a collective network. The method includes the following steps:

[0724] Acquire a previously customized basic AI agent;

[0725] Collect media information related to human users;

[0726] Analyze media information;

[0727] Transform the analyzed media information into a training dataset;

[0728] Differential weighting is applied to the transformed training dataset;

[0729] Add the knowledge module to the weighted training dataset;

[0730] Locate new data sources to include in the weighted training dataset;

[0731] The weighted training dataset is applied to the base-level AI agent to create a user PSI, where the base-level AI agent; and

[0732] Using a network enables the user PSI to communicate with multiple additional PSIs to achieve community-based security features from multiple additional PSIs to the user PSI.

[0733] Some implementations of this technology may include the following steps: obtaining, by a human user, a base-level AI agent, or a user PSI, any one or any combination of a new training dataset and training modules to be added to the weighted transformed training dataset.

[0734] Some implementations of this technology may include the following steps: monetizing user PSI by enabling other human users or other PSIs to access and use a weighted training dataset of personalized PSIs.

[0735] In some implementations, the base-level AI agent can be any of a pre-trained large language model (LLM), other tuned and trainable AI agents, or one or more custom AIs from other human owners.

[0736] In some implementations, media information can be any one or any combination of the following: videos of human users, videos of people and topics related to human users, photos or images of human users, photos or images of people and topics related to human users, articles related to human users, journals related to human users, blogs related to human users, posts related to human users, tweets related to human users, emails related to human users, podcasts related to human users, records related to human users, auditory content related to human users, auditory content of people and topics related to human users, data related to human users collected by third-party vendors, websites related to human users, applications related to human users, online information related to human users, social media information related to human users, and other AI agents related to human users.

[0737] Some implementations of this technology may include the following steps: granting one or more social media platforms a license that allows social media content relevant to human users to be accessed by a base-level AI agent or user PSI.

[0738] In some implementations, the analysis of media information may also include annotating and classifying the media information.

[0739] In some implementations, the analysis of media information may utilize one or more algorithms, including any or any combination of transformer algorithms, deep learning algorithms, transcription algorithms, content and sentiment analysis and summarization, LLM, and crowdsourced and crowd-supervised human methods.

[0740] Some implementations of this technology may include the following steps: one or more human workers on a crowdsourcing website who review the training dataset or suggest improvements to the training dataset are compensated by human users.

[0741] In some implementations, the step of weighting the transformed training dataset can be performed by a human user using an interface on a computer system implementing a basic AI agent to adjust the weight value of any one of the elements in the transformed training dataset.

[0742] In some implementations, the step of weighting the transformed training dataset can be performed by any one or any combination of an additional AI agent, an additional PSI, and additional human workers, each utilizing a computer system.

[0743] Some implementations of this technology may include the following steps: inputting ethical values ​​into a training dataset by providing a series of interactive dialogues to human users, the interactive dialogues including predetermined ethical scenarios.

[0744] Some implementations of this technology may include the following steps: mixing the weights of the transformed training dataset to improve the LLM of the base-level AI agent.

[0745] Some implementations of this technology may include the following steps: a human user authorizes the LLM to download data using the interface of the basic AI agent and automatically adds the data to the training dataset or creates a new training dataset.

[0746] In some implementations, the downloaded data may be made available to human users via a computer system for human approval before being added to or creating a new training dataset.

[0747] Some implementations of this technology may include the following steps: cloning a user PSI into one or more cloned user PSIs.

[0748] Some implementations of this technology may include the following steps: training each clone PSI using training datasets that are different from each other, wherein the training datasets of each clone PSI have different weights from each other.

[0749] Some implementations of this technology may include the following steps: combining a weighted training dataset from a user PSI and one or more cloned PSIs to create a combined training dataset.

[0750] Some implementations of this technology may include the following steps: providing any one or any combination of weighted training datasets of any one or any combination of cloned PSIs to the additional AI agent and any of the additional PSIs for training.

[0751] Some implementations of this technology may include the following steps: training a user PSI using a weighted training dataset of any or any combination of cloned PSIs.

[0752] Some implementations of this technology may include the following steps: an automated process in which a human user of a computer system performs a user PSI task without human intervention unless predetermined parameters are triggered.

[0753] Some implementations of this technology may include the following steps: a basic-level AI agent or user PSI monitors activities based on a list of prohibited activities and triggers intervention activities provided to human users.

[0754] Some implementations of this technology may include the following steps: detecting whether there are gaps in the training dataset by a base-level AI agent or user PSI; if there are gaps, generating interactions with intelligent entities to obtain data to fill the gaps.

[0755] In some implementations, interactions between more than one smart entity can be performed in parallel.

[0756] Some implementations of this technology may include the following steps: providing tasks to multiple additional AI agents or additional PSIs on the network, providing the results of the tasks from each of the additional AI agents or additional PSIs, and determining the winning result based on the results.

[0757] In some implementations, the attached AI agent or attached PSI can process tasks independently and in parallel with each other.

[0758] In some implementations, one or more additional AI agents or additional PSIs may be uncustomized, and one or more additional AI agents or additional PSIs may be customized.

[0759] Some implementations of this technology may include the following steps: utilizing a training dataset of an additional AI agent or additional PSI with winning results in the customization of a user-based AI agent or user PSI.

[0760] Some implementations of this technology may include the following steps: each of the attached AI agent or attached PSI utilizes a general problem-solving architecture on the task.

[0761] Some implementations of this technology may include the following steps: using a set of security or ethical rules that are agreed upon by each of the user-based AI agent, user PSI, additional AI agent, and additional PSI.

[0762] In some implementations, each of the additional AI agent or the additional PSI may have a set of weights that encode the knowledge of the additional AI agent and the additional PSI, respectively.

[0763] In some implementations, any computer device on the network can use the activity log.

[0764] Some implementations of this technology may include the following steps: providing a task to a user PSI, the task being configured to generate hallucinations by the PSI's LLM.

[0765] Some implementations of this technology may include the following steps: providing a task to a user PSI, wherein the task is a random arrangement of existing standardized tasks or a task dynamically created based on changing conditions.

[0766] According to another aspect, the technology may include a method for PSI that uses intelligent entities to develop and continuously improve PSI, wherein the intelligent entities include human users utilizing user computer systems, additional AI agents, and additional PSI, all of which communicate electronically via a collective network. The method may include the following steps:

[0767] Create a user PSI using the following method:

[0768] Acquire a previously customized basic AI agent;

[0769] Collect media information related to human users;

[0770] Analyze media information;

[0771] Transform the analyzed media information into a training dataset;

[0772] Perform differential weighting on the transformed training dataset; and

[0773] The weighted training dataset is applied to the base-level AI agent to create a user PSI, where the base-level AI agent;

[0774] Using a network to enable a user PSI to communicate with multiple additional AI agents or additional PSIs to achieve a group-based security feature, where both the user PSI and the additional AI agents or additional PSIs agree to use a set of rules related to security or ethics;

[0775] All actions of the user PSI and the attached AI agent or attached PSI are logged in an auditable form on any one or any combination of the central computer system, user computer system, attached AI agent, and attached PSI on the network; and

[0776] Monitor whether each action follows the set of rules, and flag any action that does not follow the set of rules.

[0777] In some implementations, the recording of actions can utilize blockchain technology.

[0778] In some implementations, the recording of actions may utilize a general problem-solving framework to record any one or any combination of objectives, sub-objectives, problem states, and steps taken in problem-solving or cognitive activities occurring on the network.

[0779] In some implementations, recorded actions can only be altered when a majority of the additional AI agents or additional PSIs on the network provide authorization for the alteration.

[0780] Some implementations of this technology may include the following steps: stopping any tagged action and applying preventative actions to prevent the tagged action from recurring.

[0781] Some implementations of this technology may include the following steps: identifying one or more additional AI agents or additional PSIs that provide tagged actions; and controlling the participation of the identified AI agents or PSIs on the network.

[0782] Some implementations of this technology may include the following steps: analyzing the labeled actions and adjusting any or any combination of the attributes of the training dataset, a set of rules, and the network based on the analysis of the labeled actions.

[0783] According to another aspect, the technology may include a method for PSI that uses intelligent entities to develop and continuously improve PSI, wherein the intelligent entities include human users utilizing user computer systems, additional AI agents, and additional PSI, all of which communicate electronically via a collective network. The method may include the following steps:

[0784] Create a user PSI using the following method:

[0785] Acquire a previously customized basic AI agent;

[0786] Collect media information related to human users;

[0787] Analyze media information;

[0788] Transform the analyzed media information into a training dataset;

[0789] Perform differential weighting on the transformed training dataset; and

[0790] The weighted training dataset is applied to the base-level AI agent to create a user PSI, where the base-level AI agent;

[0791] Provide tasks to the user's PSI and multiple additional AI agents or additional PSIs on the network;

[0792] The results of the task are provided from each of the user PSI and the attached AI agent or attached PSI; and

[0793] The winning result is determined based on the outcome.

[0794] In some implementations, each of the additional AI agents or additional PSIs can process tasks independently and in parallel with each other.

[0795] In some implementations, one of the one or more additional AI agents or additional PSIs may be uncustomized, while one or more of the additional AI agents or additional PSIs may be customized.

[0796] Some implementations of this technology may include the following steps: utilizing a training dataset of an additional AI agent or additional PSI with winning results in the customization of a user-based AI agent or user PSI.

[0797] Some implementations of this technology may include the following steps: each of the attached AI agent or attached PSI utilizes a general problem-solving architecture on the task.

[0798] Some implementations of this technology may include the following steps: using a set of security or ethical rules that are agreed upon by each of the user-based AI agent, user PSI, additional AI agent, and additional PSI.

[0799] In some implementations, each of the additional AI agent or the additional PSI may have a set of weights that encode the knowledge of the additional AI agent and the additional PSI, respectively.

[0800] In some implementations, any computer device on the network can use the activity log.

[0801] Some implementations of this technology may include the following steps: providing a task to a user PSI, the task being configured to generate hallucinations by the PSI's LLM.

[0802] Some implementations of this technology may include the following steps: providing a task to a user PSI, wherein the task is a random arrangement of existing standardized tasks or a task dynamically created based on changing conditions.

[0803] In some implementations, the network may include a user PSI and an additional AI agent or additional PSI, as well as one or more additional networks, each including an AI agent or PSI.

[0804] According to another aspect, this technology may include a method for PSI that uses intelligent entities to develop and continuously improve PSI, wherein the intelligent entities include human users utilizing user computer systems, additional AI agents, and additional PSI, all of which communicate electronically via a collective network. The method may include the following steps:

[0805] a) Create a user PSI using the following method:

[0806] Acquire a previously customized basic AI agent;

[0807] Collect media information related to human users;

[0808] Analyze media information;

[0809] Transform the analyzed media information into a training dataset;

[0810] Perform differential weighting on the transformed training dataset; and

[0811] The weighted training dataset is applied to the base-level AI agent to create a user PSI, where the base-level AI agent;

[0812] b) Use a network to enable a user PSI to communicate with multiple additional AI agents or additional PSIs to achieve a group-based security feature, where the user PSI and the additional AI agents or additional PSIs all agree to use a set of rules related to security or ethics.

[0813] c) Utilize various standardized tasks to perform the basic performance of user PSI and additional AI agents or additional PSI;

[0814] d) Create one or more versions of the user PSI, each of which is different from the others;

[0815] e) Determine the user PSI, the version of the user PSI, and the performance of the network with attached AI agents or attached PSI;

[0816] f) Assigning credit or blame values ​​to one or more elements that result in poor or poor basic performance; and

[0817] g) Identify which elements are high-quality elements that lead to performance improvements and retain these high-quality elements for future use.

[0818] Some implementations of this technology may include the following steps: identifying which elements are detrimental elements that cause performance degradation and changing the properties of the detrimental elements.

[0819] Some implementations of this technology may include the following steps: repeating step c using high-quality elements and modified low-quality elements.

[0820] Some implementations of this technology may include the following steps: repeating steps c)-g) until no higher quality element is detected.

[0821] In some implementations, the network may include a user PSI and an additional AI agent or additional PSI, as well as one or more additional networks, each including an AI agent or PSI.

[0822] Some implementations of this technology may include the following steps: logging all actions of the user PSI and the attached AI agent or attached PSI in an auditable form on any one or any combination of the central computer system, user computer system, attached AI agent and attached PSI on the network.

[0823] In some implementations, a standardized task can be any one or any combination of the following: a problem-solving task using a shared, general problem-solving architecture; an ethical and safety scenario designed to determine whether the behavior of any one or any combination of user PSI and attached AI agent or attached PSI is safe and ethical; a standardized intelligence test and evaluation designed to test the intelligence of intelligent entities; a task configured to measure the degree of “illusion” or the generation of erroneous results; a task from various professional fields; a task containing behavioral norms of different cultures or groups; a task as a random permutation of existing standardized tasks; and a task dynamically created based on changing conditions.

[0824] In some implementations, a version of the user PSI can be created by any one or any combination of the following: human adjustment of the parameters, weights, or data encoding the knowledge of the user PSI; autonomous adjustment of the parameters, weights, or data encoding the knowledge of the user PSI; random variation of the data encoding the parameters, weights, or knowledge of the user PSI; subjecting the user PSI to different training methods or amounts; sequentially creating different versions of the user PSI; and creating different versions of the user PSI in parallel.

[0825] In some implementations, a version of a user's PSI can be created by directly combining parameters, weights, or data that encode knowledge from multiple PSIs by calculating an average weight value, wherein the weight given to the most recently created or more complex version of the user's PSI is greater than the weight given to the older or less complex version of the user's PSI.

[0826] In some implementations, versions of the user PSI can be created using a randomized method or a deliberate method for estimating which versions will have beneficial results, wherein the estimation method is any one or any combination of the following: comparing the degree of matching between the knowledge of one of the user PSI versions and the statistical frequency of the tasks submitted to the network; comparing the overlap of knowledge of different additional PSIs on the network such that changes are made to optimize the performance of the group of user PSI versions if the knowledge of the entire network is characterized compared to the characteristics of the problem the network is expected to solve; and using hill climbing or gradient descent to optimize the parameters of any one of the user PSI versions.

[0827] According to another aspect, this technology may include a method for PSI to utilize a single computerized intelligent system, the single computerized intelligent system comprising multiple AI agents residing within the single computerized intelligent system. The method may include:

[0828] Acquire a previously customized base-level AI agent that resides within a single computerized intelligent system;

[0829] Collect media information related to human users associated with basic-level AI agents;

[0830] Transform the analyzed media information into a training dataset;

[0831] Differential weighting is applied to the transformed training dataset;

[0832] The weighted training dataset was applied to the base-level AI agent to create user PSI; and

[0833] The user PSI communicates with multiple additional PSIs to enable community-based security features from the additional PSIs to the user PSI.

[0834] The underlying large language model (LLM) for training an AI agent is utilized, whereby the guardrails include attributes associated with any or any combination of safety, ethics, and knowledge, and the AI ​​agent resides in a single computerized intelligent system.

[0835] Customize the base LLM based on the ethics profile;

[0836] Combining ethical information from multiple additional AI agents residing in a single computerized intelligent system, the differences between additional AI agents and AI agents, and the additional AI agents residing in a single computerized intelligent system;

[0837] A set of values ​​for the underlying LLM is improved based on the problem-solving process of the problem request; and

[0838] The underlying LLM is updated with combined ethical information and a modified set of values, thereby allowing for scalable AGI.

[0839] Summary and comments

[0840] This technology has attempted to explain some ways we can efficiently and effectively design the next steps in the development of superintelligence, PSI, and “collective superintelligence.” Humans now have the ability to make design decisions that profoundly influence the future trajectory of planetary intelligence. We must design PSI, PSI networks, and other AI systems in which humans are (initially) in a cycle. These systems must conform to human values ​​(even when intelligence surpasses human intelligence). Even when the growth rate of PSI far exceeds the ability of humans to keep up in terms of intelligence, the PSI collective approach can help ensure stable, human-centered values. If we now design such systems correctly based on the principles outlined in this patent, the future could be an amazing and wonderful place for all humanity and all sentient beings.

[0841] While the implementation of the secure PSI has been described in detail, it should be apparent that modifications and variations are possible, all of which fall within the true spirit and scope of this technology. Furthermore, regarding the foregoing description, it should be understood that optimal relationships between the components of this technology (including variations in size, material, shape, form, function and operation, assembly, and use) are considered obvious to those skilled in the art, and all equivalents shown in the accompanying drawings and described in the specification are intended to be covered by this technology. For example, any suitable implementation may be used instead of the implementations described above.

[0842] Therefore, the above is considered merely an explanation of the principles of the present technology. Furthermore, since many modifications and alterations will readily occur to those skilled in the art, it is not intended to limit the present technology to the exact construction and operation shown and described; thus, all suitable modifications and equivalents falling within the scope of the present technology may be adopted.

Claims

1. A system for personalized super-intelligent PSI, the system using intelligent agents to develop and continuously improve the PSI of human users utilizing computer systems, and using additional PSI, all of which are electronically communicated via a collective network, said system comprising: A computer system includes: a processor, a computer-readable storage medium, and program instructions stored on the computer-readable storage medium, the program instructions being executable by the processor to cause the computer system to: A basic-level artificial intelligence (AI) agent is implemented on the computer system, wherein the basic-level AI agent has been customized; Collect media information related to the human users; Analyze the media information; Transform the analyzed media information into a training dataset; Differential weighting is applied to the transformed training dataset; Add the knowledge module to the weighted training dataset; Locate new data sources to include in the weighted training dataset; The weighted training dataset is applied to the base-level AI agent to create a personalized PSI; and The network is used to enable the personalized PSI to communicate with multiple additional PSIs to achieve community-based security features from the multiple additional PSIs to the personalized PSI.

2. A method for personalized super-intelligent PSI, the method using intelligent entities to develop and continuously improve PSI, wherein the intelligent entities include human users utilizing computer systems, additional artificial intelligence (AI) agents, and additional PSI, all of which communicate electronically via a collective network, the method comprising the following steps: Acquire a previously customized basic AI agent; Collect media information related to the human users; Analyze the media information; Transform the analyzed media information into a training dataset; Differential weighting is applied to the transformed training dataset; Add the knowledge module to the weighted training dataset; Locate new data sources to include in the weighted training dataset; The weighted training dataset is applied to the base-level AI agent to create a user PSI; as well as The network is used to enable the user PSI to communicate with multiple additional PSIs to achieve a community-based security feature from the multiple additional PSIs to the user PSI.

3. The method according to claim 2, further comprising the following steps: The new training dataset and training module to be added to the weighted transformed training dataset are obtained by the human user, the base-level AI agent, or the user PSI, or any combination thereof.

4. The method according to claim 2, further comprising the following step: The user PSI is monetized by enabling other human users or other PSIs to access and use the weighted training dataset of the personalized PSI.

5. The method according to claim 2, wherein, The foundational AI agent is any one of a pre-trained large language model (LLM), other tunable and trainable AI agents, or one or more custom AIs from other human owners.

6. The method according to claim 2, wherein, The media information is any one or any combination of the following: videos of the human user, videos of people and topics related to the human user, photos or images of the human user, photos or images of people and topics related to the human user, articles related to the human user, journals related to the human user, blogs related to the human user, posts related to the human user, tweets related to the human user, emails related to the human user, podcasts related to the human user, records related to the human user, auditory content related to the human user, auditory content of people and topics related to the human user, data related to the human user collected by third-party vendors, websites related to the human user, applications related to the human user, online information related to the human user, social media information related to the human user, and other AI agents related to the human user.

7. The method according to claim 2, further comprising the following step: Grant a license to one or more social media platforms that allows social media content related to the human user to be accessed by the base-level AI agent or the user's PSI.

8. The method according to claim 2, wherein, The analysis of the media information also includes annotating and classifying the media information.

9. The method according to claim 8, wherein, The analysis of the media information utilizes one or more algorithms, including any or any combination of transformer algorithms, deep learning algorithms, transcription algorithms, content and sentiment analysis and summarization, LLM, and crowdsourcing and crowd-supervised human methods.

10. The method of claim 2, further comprising the step of: One or more human workers on the crowdsourcing website are compensated by the human users for reviewing the training dataset or suggesting improvements to the training dataset.

11. The method according to claim 2, wherein, The step of weighting the transformed training dataset is performed by the human user adjusting the weight value of any one of the elements in the transformed training dataset through an interface on the computer system implementing the base-level AI agent.

12. The method according to claim 2, wherein, The step of weighting the transformed training dataset is performed by any one or any combination of the additional AI agent, the additional PSI, and the additional human workers, each utilizing a computer system.

13. The method according to claim 2, further comprising the following step: Ethical values ​​are input into the training dataset by providing the human user with a series of interactive dialogues, which include predetermined ethical scenarios.

14. The method of claim 2, further comprising the following step: The weights of the training dataset after hybrid transformation are used to improve the LLM of the base-level AI agent.

15. The method of claim 2, further comprising the following step: The human user authorizes the LLM to download data using the interface of the basic AI agent and automatically adds the data to the training dataset or creates a new training dataset.

16. The method according to claim 15, wherein, The downloaded data is provided to the human user via the computer system for approval by the human user before being added to the training dataset or before a new training dataset is created.

17. The method of claim 2, further comprising the step of: The user PSI is cloned into one or more cloned user PSIs.

18. The method of claim 17, further comprising the step of: Each clone PSI is trained using a different training dataset than the others, with each clone PSI having its own training dataset containing different weights.

19. The method of claim 18, further comprising the step of: A weighted training dataset from the user PSI and one or more cloned PSIs is combined to create a combined training dataset.

20. The method of claim 18, further comprising the step of: Provide any one or any combination of the weighted training datasets of any one or any combination of the cloned PSIs to the additional AI agent and any one of the additional PSIs for training.

21. The method of claim 18, further comprising the step of: The user PSI is trained using a weighted training dataset of any or any combination of the cloned PSIs.

22. The method of claim 2, further comprising the step of: Unless predetermined parameters are triggered, the task is performed by the human user utilizing the computer system, which is an automated process that requires no intervention from the human user.

23. The method of claim 2, further comprising the following step: The basic AI agent or the user PSI monitors activities based on the prohibited activities list and triggers intervention activities provided to the human user.

24. The method of claim 2, further comprising the step of: The basic AI agent or the user PSI detects whether there are gaps in the training dataset. If there are gaps, an interaction with the intelligent entity is generated to obtain data to fill the gaps.

25. The method according to claim 24, wherein, Interactions between more than one of the intelligent entities are executed in parallel.

26. The method of claim 2, further comprising the step of: The task is provided to the plurality of additional AI agents or additional PSIs on the network, the result of the task is provided from each of the additional AI agents or additional PSIs, and the winning result is determined based on the result.

27. The method according to claim 26, wherein, The additional AI agent or the additional PSI processes the task independently and in parallel with each other.

28. The method according to claim 26, wherein, One or more of the additional AI agents or one or more of the additional PSIs are uncustomized, while one or more of the additional AI agents or one or more of the additional PSIs are customized.

29. The method of claim 26, further comprising the step of: The training dataset of the additional AI agent or the additional PSI with the winning results is utilized in the customization of the user-based AI agent or the user PSI.

30. The method of claim 26, further comprising the step of: Each of the additional AI agent or the additional PSI utilizes a general problem-solving architecture on the task.

31. The method of claim 2, further comprising the following step: Each of the user-based AI agent, the user PSI, the additional AI agent, and the additional PSI agrees to use a set of security or ethical rules.

32. The method according to claim 2, wherein, Each of the additional AI agent or the additional PSI has a set of weights that encode the knowledge of the additional AI agent and the additional PSI, respectively.

33. The method according to claim 2, wherein, Any computer device on the network can use the activity log.

34. The method of claim 2, further comprising the following step: A task is provided to the user PSI, which is configured to generate hallucinations by the PSI's LLM.

35. The method of claim 2, further comprising the following step: The user PSI is provided with tasks, wherein the tasks are either a random arrangement of existing standardized tasks or tasks dynamically created based on changing conditions.

36. A method for personalized super-intelligent PSI, the method using intelligent entities to develop and continuously improve PSI, wherein the intelligent entities include human users utilizing user computer systems, additional artificial intelligence (AI) agents, and additional PSIs, all of which communicate electronically via a collective network, the method comprising the following steps: Create a user PSI using the following method: Acquire a previously customized basic AI agent; Collect media information related to the human users; Analyze the media information; Transform the analyzed media information into a training dataset; Differential weighting is applied to the transformed training dataset; as well as The weighted training dataset is applied to the base-level AI agent to create a user PSI, wherein the base-level AI agent; The network is used to enable the user PSI to communicate with multiple additional AI agents or additional PSIs to achieve a group-based security feature, wherein the user PSI and the additional AI agents or additional PSIs agree to use a set of rules related to security or ethics. All actions of the user PSI and the attached AI agent or the attached PSI shall be logged in an auditable form on any one or any combination of the central computer system, the user computer system, the attached AI agent, and the attached PSI on the network; and Monitor whether each action follows the set of rules, and flag any action that does not follow the set of rules.

37. The method of claim 36, wherein, The records of these actions are made using blockchain technology.

38. The method according to claim 36, wherein, The recording of the actions utilizes a general problem-solving framework to record any one or any combination of objectives, sub-objectives, problem states, and steps taken in problem-solving or cognitive activities occurring on the network.

39. The method according to claim 36, wherein, The recorded actions are altered only when a majority of the additional AI agents or the additional PSI on the network grant authorization for the alteration.

40. The method of claim 36, further comprising the step of: Stop any flagged actions and apply preventative actions to prevent the flagged actions from recurring.

41. The method of claim 40, further comprising the step of: Identify one or more of the additional AI agents or additional PSIs that provide tagged actions, and control the participation of the identified AI agents or PSIs on the network.

42. The method of claim 36, further comprising the step of: Analyze the labeled actions and adjust any or any combination of the attributes of the training dataset, the set of rules, and the network based on the analysis of the labeled actions.

43. A method for personalized super-intelligent PSI, the method using intelligent entities to develop and continuously improve PSI, wherein the intelligent entities include human users utilizing user computer systems, additional artificial intelligence (AI) agents, and additional PSI, all of which communicate electronically via a collective network, the method comprising the steps of: Create a user PSI using the following method: Acquire a previously customized basic AI agent; Collect media information related to the human users; Analyze the media information; Transform the analyzed media information into a training dataset; Differential weighting is applied to the transformed training dataset; as well as The weighted training dataset is applied to the base-level AI agent to create a user PSI, wherein the base-level AI agent; Provide tasks to the user PSI and multiple additional AI agents or additional PSIs on the network; provide the results of the tasks from each of the user PSI and the additional AI agents or additional PSIs; as well as The winning result is determined based on the stated results.

44. The method according to claim 43, wherein, Each of the additional AI agents or the additional PSIs processes the task independently and in parallel with each other.

45. The method according to claim 43, wherein, One or more of the additional AI agents or one or more of the additional PSIs are uncustomized, while one or more of the additional AI agents or one or more of the additional PSIs are customized.

46. ​​The method of claim 43, further comprising the step of: The training dataset of the additional AI agent or the additional PSI with the winning results is utilized in the customization of the user-based AI agent or the user PSI.

47. The method of claim 43, further comprising the step of: Each of the additional AI agent or the additional PSI utilizes a general problem-solving architecture on the task.

48. The method of claim 43, further comprising the step of: Each of the user-based AI agent, the user PSI, the additional AI agent, and the additional PSI agrees to use a set of security or ethical rules.

49. The method according to claim 43, wherein, Each of the additional AI agent or the additional PSI has a set of weights that encode the knowledge of the additional AI agent and the additional PSI, respectively.

50. The method according to claim 43, wherein, Any computer device on the network can use the activity log.

51. The method of claim 43, further comprising the step of: A task is provided to the user PSI, which is configured to generate hallucinations by the PSI's LLM.

52. The method of claim 43, further comprising the following step: The user PSI is provided with tasks, wherein the tasks are either a random arrangement of existing standardized tasks or tasks dynamically created based on changing conditions.

53. The method according to claim 43, wherein, The network includes the user PSI and the additional AI agent or the additional PSI, and one or more additional networks, each including the AI ​​agent or PSI.

54. A method for personalized super-intelligent PSI, the method using intelligent entities to develop and continuously improve PSI, wherein the intelligent entities include human users utilizing user computer systems, additional artificial intelligence (AI) agents, and additional PSIs, all of which communicate electronically via a collective network, the method comprising the following steps: a) Create a user PSI using the following method: Acquire a previously customized basic AI agent; Collect media information related to the human users; Analyze the media information; Transform the analyzed media information into a training dataset; Differential weighting is applied to the transformed training dataset; as well as The weighted training dataset is applied to the base-level AI agent to create a user PSI, wherein the base-level AI agent; b) Using the network to enable the user PSI to communicate with multiple additional AI agents or additional PSIs to achieve a community-based security feature, wherein the user PSI and the additional AI agents or additional PSIs agree to use a set of rules related to security or ethics; c) Utilize various standardized tasks to perform the basic performance of the user PSI and the additional AI agent or the additional PSI; d) Create one or more versions of the user PSI, each version being different from the others; e) Determine the user PSI, the version of the user PSI, and the performance of the network of the additional AI agent or the additional PSI; f) Assign honor or responsibility values ​​to one or more elements that result in the quality of the basic performance; as well as g) Identify which elements are high-quality elements that lead to performance improvements and retain these high-quality elements for future use.

55. The method of claim 54, further comprising the step of: Identify which elements are the poor performers causing performance degradation and change the properties of those poor performers.

56. The method of claim 55, further comprising the step of: Repeat step c) using the high-quality element and the modified low-quality element.

57. The method of claim 55, further comprising the step of: Repeat steps c)-g) until no higher-quality elements are detected.

58. The method according to claim 54, wherein, The network includes the user PSI and the additional AI agent or the additional PSI, and one or more additional networks, each including the AI ​​agent or PSI.

59. The method of claim 54, further comprising the step of: All actions of the user PSI and the additional AI agent or the additional PSI are logged in an auditable form on any one or any combination of the central computer system, the user computer system, the additional AI agent, and the additional PSI on the network.

60. The method according to claim 54, wherein, The standardized task is any one or any combination of the following: a problem-solving task using a shared, general problem-solving architecture; an ethical and safety scenario designed to determine whether the behavior of the user PSI and the attached AI agent or any combination of the attached PSIs is safe and ethical; a standardized intelligence test and evaluation designed to test the intelligence of the intelligent entity; a task configured to measure the degree of "illusion" or the generation of erroneous results; a task from various professional fields; a task containing behavioral norms of different cultures or groups; a task as a random permutation of existing standardized tasks; and a task dynamically created based on changing conditions.

61. The method according to claim 54, wherein, The version of the user PSI is created by any one or any combination of the following: human adjustment of the parameters, weights, or data encoding the knowledge of the user PSI; autonomous adjustment of the parameters, weights, or data encoding the knowledge of the user PSI; random variation of the data encoding the parameters, weights, or knowledge of the user PSI; subjecting the user PSI to different training methods or amounts; sequentially creating different versions of the user PSI; and creating different versions of the user PSI in parallel.

62. The method according to claim 54, wherein, The version of the user PSI is created by directly combining the parameters, weights, or data that encode knowledge from multiple PSIs by calculating an average weight value, wherein the weight given to the most recently created or more complex version of the user PSI is greater than the weight given to the older or less complex version of the user PSI.

63. The method according to claim 54, wherein, The versions of the user PSI are created using a random method or a deliberate method for estimating which versions will have beneficial results, wherein the estimation method is any one or any combination of the following: comparing the degree of matching between the knowledge of one of the versions of the user PSI and the statistical frequency of the tasks submitted to the network; comparing the overlap of the knowledge of different additional PSIs on the network, such that changes are made to optimize the performance of the group of versions of the user PSI if the knowledge of the entire network is characterized compared to the characteristics of the problem that the network is expected to solve. And the parameters of any version of the user PSI can be optimized using hill climbing or gradient descent.

64. A method for personalized super-intelligent PSI, the method utilizing a single computerized intelligent system, the single computerized intelligent system comprising multiple artificial intelligence (AI) agents residing in the single computerized intelligent system, the method comprising: Acquire a previously customized base-level AI agent that resides in a single computerized intelligent system; Collect media information related to human users associated with the basic AI agent; Transform the analyzed media information into a training dataset; Differential weighting is applied to the transformed training dataset; The weighted training dataset was applied to the base-level AI agent to create user PSI; and This enables the user PSI to communicate with multiple additional PSIs to achieve a community-based security feature from the additional PSIs to the user PSI. The AI ​​agent is trained using a base large language model (LLM) that includes attributes associated with any or any combination of safety, ethics, and knowledge, and the AI ​​agent resides in a single computerized intelligent system. Customize the underlying LLM according to the ethics profile; Ethical information is combined from multiple additional AI agents residing in the single computerized intelligent system, the additional AI agents being different from the AI ​​agent and residing in the single computerized intelligent system; The set of values ​​of the underlying LLM is improved based on the problem-solving process of the problem request; as well as The underlying LLM is updated with combined ethical information and a modified set of values, thereby allowing for scalable AGI.

Images