AI system using a single-personality transfer method

Abstract

In cross-device conversations, sharing only the conversation history is insufficient, as it cannot maintain the intermediate generation state during inference, resulting in a loss of conversational continuity. This could lead to delayed responses during device switching, reconnection, or the use of old tokens, potentially resulting in duplicate assignment of the same AI personality. Furthermore, if personality-related data remains on the device, there is a risk of leakage in the event of loss or theft. [Solution] A single AI personality entity is maintained on the cloud and, without duplicating the entity, is exclusively assigned to one terminal for operation. When switching terminals, the session token is invalidated, and the receipt of invalidation completion responses and connection disconnection responses serves as confirmation of blocking. Interlocked exclusive control is implemented, allowing assignment to the new terminal only after the time of successful blocking confirmation. The personality state, including conversation history, internal context, and inference state, is maintained in the state memory unit and applied to the temporary execution unit of the new terminal to realize a restart state that allows inference to continue. The terminal becomes an empty shell upon termination by being erased or having its key discarded, as it is stateless.

Description

Technical Field

[0001] The present invention relates to a technology for exclusively allocating an AI personality held on the cloud and operating it among a plurality of terminals.

[0002] The present invention relates to a technology that applies the state of an AI personality as a resumable state to a terminal after switching and does not leave the personality state on the terminal side.

Background Art

[0003] In the conventional method of using AI services on a plurality of terminals, it is easy to generate separate processing units for each terminal, and the consistency of the personality as seen by the user may be lost.

[0004] Furthermore, if a structure remains in which the same personality can operate on a plurality of terminals simultaneously during the process of switching terminals, problems may arise from the viewpoints of consistency, security, and responsibility tracking.

[0005] Also, although conversation history and settings can be synchronized, the framework for ensuring continuity including the internal state necessary for continuous inference is not clear.

[0006] On the other hand, there is also a conventional technique for continuously processing while regarding the conversation state as a single dialogue subject.

Prior Art Documents

Patent Documents

[0007]

Patent Document 1

Non-Patent Documents

[0008]

Non-Patent Document 1

[0009] In conventional methods of using AI services across multiple devices, separate processing units tend to be generated for each device, and the behavior of the AI ​​personality can differ from device to device even for the same user, potentially leading to a loss of consistency.

[0010] In conventional methods, while conversation history and settings information can be synchronized between devices, there is no clear framework to guarantee continuity, including the internal state necessary for continuing inference, and the flow of dialogue and processing is interrupted when switching devices.

[0011] If a structure remains that allows the same AI personality to be assigned to multiple terminals simultaneously, it may lead to inconsistent responses and processing results, posing problems from the standpoint of consistency, security, and accountability.

[0012] Furthermore, in configurations where conversation history and internal state are widely retained on the device, the loss or theft of the device could lead to the leakage of personal information, posing a significant information leakage risk for both users and service providers.

[0013] Furthermore, if information regarding the movement of AI personalities and their assignment to terminals is managed across multiple systems, it becomes difficult to verify the consistency of usage across terminals and to audit it, making it difficult to identify the cause of problems and assign responsibility when they occur. [Means for solving the problem]

[0015] To solve the above problems, the present invention provides an AI system that maintains a single AI personality entity on the cloud and operates by exclusively assigning that entity to one of multiple terminals without duplicating it.

[0016] The AI ​​system of the present invention comprises a personality management unit that manages the AI ​​personality entity, a session management unit that controls the assignment to and switching of the AI ​​to a terminal, and a state storage unit that stores the personality state.

[0017] When an AI personality entity moves between terminals, the session management unit invalidates the session token corresponding to the terminal to which it was previously assigned (hereinafter referred to as the "original terminal"), and confirms that it has been blocked upon receiving a confirmation of the completion of the invalidation and a confirmation of the connection being broken. It then implements interlocked exclusive control, which does not permit the start of assignment to the new terminal for requests that arrive before the time when the blocking confirmation is established.

[0018] The state memory unit stores personality states, including conversation history, internal context, and reasoning state, on the cloud.

[0019] The Personality Management Department maintains a single AI personality entity to which the personality state is applied, on the cloud.

[0020] When the assignment to the terminal begins, the AI ​​personality entity takes the form of an AI personality state package as a restart state that allows for continued inference, by applying the personality state stored in the state memory unit.

[0021] Also, when the AI personality state package is returned from the terminal to which it is assigned to the cloud, the state memory unit receives the personality state including the conversation history, internal context, and inference state generated or updated on the terminal, and updates the stored personality state.

[0022] The terminal is configured as a stateless terminal that does not permanently store personality-related data. Upon session termination or deassignment, it is either left in a reusable state or rendered undecryptable by discarding the encryption key. The state memory unit centrally manages operation logs, conversation history, personality transfer history, and terminal assignment history.

Advantages of the Invention

[0023] Since the AI personality is exclusively assigned as a single entity, personality fragmentation for each terminal is suppressed.

[0024] Structural prevention of double assignment is achieved by order constraint based on the blocking establishment time and assignment rejection processing when assignment fails.

[0025] With the AI personality state package, the state information necessary for continuous inference is applied as a resumable state, enabling continuous conversations across terminals.

[0026] Since the terminal is configured as a device and shelling out is irreversibly executed, the risk of information leakage in case of terminal loss or theft is reduced or eliminated.

[0027] Since the personality movement history and terminal assignment history are retained, auditability and accountability tracing ability are improved.

Brief Description of the Drawings

[0028] [Figure 1] Block diagram showing the overall configuration of the AI system (Figure 1), and sequence showing terminal switching processing (Figure 2) [Figure 2] Re-terminal switching sequence diagram to another terminal (Figure 3), and diagram showing state transition accompanying terminal switching (Figure 4) [Modes for carrying out the invention]

[0029] As shown in Figure 1, the AI ​​system of this embodiment includes a personality management unit 31, a session management unit 32, a state storage unit 33, and a terminal input / output unit 34 in the cloud 30. The terminal input / output unit 34 manages the inflow and outflow of the AI ​​personality state package 70 to the cloud 30.

[0030] The AI ​​personality entity managed by the personality management unit 31 takes the form of an AI personality state package when it is assigned to a terminal. An AI personality state package is a restart state to which the personality state stored in the state storage unit 33 is applied, allowing for continued reasoning. It is the form of the AI ​​personality entity when it moves, capable of separately holding the conversation history, internal context, and reasoning state that are generated or updated at the destination terminal. The reasoning state refers to the state in which the AI ​​personality state package continues to generate responses based on the previous conversation context when it is temporarily executed at the destination terminal.

[0031] The session management unit 32 controls the exclusive allocation to terminals. When moving between terminals, it invalidates the session token corresponding to the original terminal and confirms the disconnection when it receives the completion response and the connection disconnection response.

[0032] The terminal connection path switching unit 51 switches the connection path between the cloud 30 and each terminal based on instructions from the session management unit 32. The terminal connection path switching unit 51 establishes a connection path only to the terminal to which it is assigned, and blocks the connection paths to terminals other than the assigned terminal. Furthermore, when establishing a return connection path from a terminal to the cloud 30, the return connection path is established only after the blockage confirmation for the outbound connection path from the cloud 30 to the terminal has been completed.

[0033] The session management unit 32 only permits assignment to terminals after the time when the blocking confirmation is successful, and refuses assignment to terminals if the blocking confirmation is not successful. In addition, requests using old tokens or requests that arrive before the time when the blocking confirmation is successful are treated as invalid requests.

[0034] The state storage unit 33 holds at least the conversation history, internal context 42, and reasoning state 43 as personality states, and can hold the updated conversation history as the conversation progresses. When the assignment to the terminal begins, the personality states held in the state storage unit 33 are applied to the temporary execution unit 12 of the new terminal 10 as a resumeable state, and input and output using the input / output unit 11 becomes possible on the new terminal 10. The temporary execution unit 12 updates the personality states based on the input data from the terminal and saves the updated personality states to the state storage unit 33.

[0035] Among the terminals that are candidates for assignment and can be connected, the terminal to which the AI ​​personality state package 70 will be moved will be designated as the new terminal 10. The assigned terminal will be configured as a stateless terminal that does not permanently store personality-related data. At the end of the session, the temporary area within the terminal will be erased, or access to the temporary area will be disabled by discarding the encryption key to the temporary area, thereby performing a "shelling" 60 that prevents personality-related data from remaining in a reusable state. The state storage unit 33 will maintain the personality movement history and terminal assignment history, ensuring auditability and accountability. [Examples]

[0036] The AI ​​system shown in Figure 1 comprises a personality management unit 31 that holds a single AI personality entity on the cloud 30, a session management unit 32 that controls terminal assignment, a state storage unit 33 that stores the personality state, a terminal input / output unit 34 on the cloud side, and a terminal connection path switching unit 51, and is configured so that multiple terminals can connect to the cloud 30.

[0037] In the terminal switching process shown in Figure 2, the personality transfer request transmitted from the new terminal 10 is received by the session management unit 32 via the entry / exit point to the cloud 30. In response to this receipt, the session management unit 32 controls the terminal connection path switching unit 51 to establish a forward connection path between the cloud 30 and the new terminal 10.

[0038] Furthermore, the AI ​​personality entity managed by the personality management unit 31 that receives the personality transfer request is not duplicated but takes the form of an AI personality state package 70, which is exclusively assigned from the cloud 30 to the new terminal 10. On the new terminal 10, the input / output unit 11 is used to perform conversation processing using the conversation history A41, the internal context 42, and the inference state 43, which is then executed by the temporary execution unit 12.

[0039] When the AI ​​personality state package 70 in the new terminal 10 transitions to the allocation termination process, the session management unit 32 invalidates the session token corresponding to the new terminal 10, controls the terminal connection path switching unit 51 to close the forward connection path between the cloud 30 and the new terminal 10, and simultaneously establishes a return connection path from the new terminal 10 to the cloud 30.

[0040] The AI ​​personality state package 70 stores the conversation history A41, internal context 42, and inference state 43 information generated on the new terminal 10 as a personality state and returns it to the cloud 30.

[0041] The AI ​​personality state package 70, having returned to the cloud 30, sends the personality state generated at the new terminal 10, along with the personality movement history and terminal assignment history related to the transfer, to the state storage unit 33 via the terminal input / output unit 34, thereby ensuring auditability and accountability. After sending return completion information to the session management unit 32 indicating that the return to the cloud 30 is complete, it reverts to the form of the AI ​​personality entity managed by the personality management unit 31.

[0042] When the session management unit 32 receives information from the AI ​​personality state package 70 that the return process is complete, it controls the terminal connection path switching unit 51 to block the return connection path between the new terminal 10 and the cloud 30, and transitions the session to the terminated state.

[0043] The time at which session termination by the session management unit 32 is established is determined by the cloud-side reference time and does not depend on the terminal-side time or latency.

[0044] The empty shelling 60 shown in Figure 3 is executed when the session management unit 32 confirms that the session has ended or has been unassigned to the assigned terminal.

[0045] Decomposition 60 is a process in which, based on the control of the session management unit 32, personality-related data and work data held in the temporary execution unit 12 are erased, or the temporary key for decrypting or reusing said data is discarded.

[0046] Figure 3 shows the state in which the AI ​​system has received a new movement request from a terminal via the terminal input / output unit 34.

[0047] Upon receiving a movement request, the terminal input / output unit 34 transmits the movement request to the session management unit 32 and the state storage unit 33. The state storage unit 33 then outputs personality state information, including the conversation history A41, internal context 42, and inference state 43, to the personality management unit 31 via the terminal input / output unit 34.

[0048] The session management unit 32 notifies the personality management unit 31 of the move request via the terminal input / output unit 34, and controls the terminal connection path switching unit 51 to open a forward connection path to the new terminal 10. Based on the personality state information output from the state storage unit 33 and the move request, the personality management unit 31 starts moving the AI ​​personality entity to the new terminal 10 in the form of an AI personality state package 70.

[0049] The AI ​​personality state package 70, while retaining the personality state information applied by the state memory unit 33, uses the input / output unit 11 on the new terminal 10 to execute conversation processing using the conversation history B45, internal context 42, and inference state 43 in the temporary execution unit 12.

[0050] The inference state 43 may include computational states and intermediate generation states during the inference process.

[0051] The inference state 43 may include, for example, cache information, a generation in progress, an undetermined token sequence, and an intermediate computation state.

[0052] The inference state 43 may include, for example, cache information related to the attention mechanism and buffer states related to the stream processing of multimodal inputs.

[0053] This ensures that the continuity of dialogue and processing, which cannot be achieved by simply sharing conversation history, is maintained on the new device 10.

[0054] When the AI ​​personality state package 70 in the new terminal 10 transitions to the allocation termination process, the session management unit 32 invalidates the session token corresponding to the new terminal 10, controls the terminal connection path switching unit 51 to close the forward connection path between the cloud 30 and the new terminal 10, and simultaneously establishes a return connection path from the new terminal 10 to the cloud 30.

[0055] The AI ​​personality state package 70 stores the conversation history B45, internal context 42, and inference state 43 information generated on the new terminal 10 as a personality state and returns it to the cloud 30.

[0056] The AI ​​personality state package 70, having returned to the cloud 30, sends the personality state generated at the new terminal 10, along with the personality movement history and terminal assignment history related to the transfer, to the state storage unit 33 via the terminal input / output unit 34, thereby ensuring auditability and accountability. After sending return completion information to the session management unit 32 indicating that the return to the cloud 30 is complete, it reverts to the form of the AI ​​personality entity managed by the personality management unit 31.

[0057] As a result, the state memory unit 33 holds conversation history A41 and conversation history B45, as well as the internal context 42 and inference state 43 as continuous information of the assigned terminal and the new terminal, and manages these as a series of personality states originating from a single AI personality entity.

[0058] When the session management unit 32 receives information from the AI ​​personality state package 70 that the return process is complete, it controls the terminal connection path switching unit 51 to block the return connection path between the new terminal 10 and the cloud 30, and transitions the session to the terminated state.

[0059] As shown in Figure 4, the state transitions associated with terminal switching are represented, including the disconnected state, and terminals that have undergone decontamination are shown as decontamination 60.

[0060] As shown in Figure 4, the state storage unit 33 stores conversation history A41, conversation history B45, internal context 42, and inference state 43.

[0061] The following describes the common processing that occurs when the terminal connection path switching unit 51 fails to confirm the disconnection of the connection path under the control of the session management unit 32.

[0062] When returning the AI ​​personality state package 70 from the new terminal 10 to the cloud 30, if the terminal connection path switching unit 51 fails to confirm the disconnection of the outbound connection path, the session management unit 32 will not establish the return connection path by controlling the terminal connection path switching unit 51.

[0063] If the blockage confirmation on the forward connection path is not completed within a predetermined time, the session management unit 32 forcibly revokes the allocation of the new terminal 10 related to the forward connection path based on administrator authentication, and records the time of revocation as the time of blockage completion.

[0064] The session management unit 32 controls the terminal connection path switching unit 51 after the expiration time to establish a return connection path from the new terminal 10 to the cloud 30, and treats any delayed responses, reconnection requests, or requests using the old token that arrive from the new terminal 10 after the expiration time as invalid.

[0065] After the AI ​​personality state package 70 returns from the new terminal 10 to the cloud 30, if the terminal connection path switching unit 51 fails to confirm the connection on the return path, the session management unit 32 does not transition the session on the return path to the terminated state and performs interlocked exclusive control, which rejects the new assignment to the assigned terminal.

[0066] If the blockage confirmation on the return connection path is not completed within a predetermined time, the session management unit 32, based on administrator authentication, forcibly terminates the remaining sessions of the terminals to be unassigned on the return connection path and records the termination time as the blockage completion time.

[0067] The session management unit 32 permits new assignments to other terminals after the expiration time, and treats any delayed responses, reconnection requests, or requests using old tokens received from the terminals subject to the release after the expiration time as invalid.

[0068] Assignment to the designated terminal will only be permitted after the time when the blocking confirmation has been completed.

[0069] The time of blockage is determined by the cloud's reference time and does not depend on the device's time or latency.

[0070] · Modification The AI ​​personality state package 70 only needs to satisfy a format that allows inference to be resumed on the terminal after the switch, and is not limited to a specific model structure.

[0071] The terminal can be any computer device capable of communication, whether portable or stationary.

[0072] The cloud can be a single device or a distributed configuration, and the personality management unit 31, session management unit 32, and state storage unit 33 can be the same device or distributed devices.

[0073] The type, order, timeout value, and handling of failures used for blocking confirmation can be modified within the limits of satisfying double allocation prevention and order constraints.

[0074] The contents and representation format of the AI ​​personality state package 70 may be modified to the extent that inference continuity is maintained. [Explanation of Symbols]

[0075] 10 New Devices 11 Input / output section 12 Temporary Execution Unit 30 Cloud 31 Personality Management Department 32 Session Management Department 33 State memory unit 34 Terminal Input / Output Section 41 Conversation History A 42 Internal Context 43 Inference State 45 Conversation History B 51 Terminal connection path switching unit 60. Shed skin 70 AI Personality State Package

Claims

1. An AI system comprising a personality management unit that maintains a single AI personality entity on the cloud, a session management unit that controls terminal assignment, and a state storage unit that stores the personality state, wherein the AI ​​personality entity is exclusively assigned to the terminal without duplication and operated accordingly, the AI ​​personality entity is not reproduced as the same personality state by referencing or duplication regardless of the terminal's location, but only one complete personality state is always activated, the session management unit, when assigning to a terminal, permits the assignment to that terminal only after the time when a blocking confirmation for the terminal to be unassigned has been established, and does not permit the assignment to that terminal until the blocking confirmation has been established, thereby preventing duplicate assignment of the same personality state to other terminals or the terminal to be unassigned, and the terminal comprises an input / output unit and a temporary execution unit, and is configured as a stateless terminal that does not permanently store the personality state.

2. The AI ​​system according to claim 1, wherein the session management unit, upon transfer between terminals, invalidates the session token corresponding to the terminal to be unassigned, confirms the blockage upon receipt of a completion response for the invalidation and a connection disconnection response, permits assignment to the destination terminal only after the time the blockage confirmation is established, maintains a state of refusing assignment to the destination terminal until the blockage confirmation is established, forcibly revokes the assignment of the terminal to be unassigned based on administrator authentication if the blockage confirmation is not established within a predetermined time, records the time of revocation as the blockage establishment time, permits assignment to the destination terminal only after that time, and treats delayed responses or reconnection requests or requests using old tokens from the terminal to be unassigned as invalid.

3. The state memory unit stores conversation history, internal context, and reasoning state as personality states. It also maintains the inference state, and the inference state includes the calculation state and intermediate generation state during the inference process, and is allocated to the terminal. Claim to apply the aforementioned personality state to the temporary execution unit of the new terminal as a restartable state at the start of the application. The AI ​​system described in 1.

4. The AI ​​system according to claim 1, wherein the state storage unit centrally manages operation logs and conversation history, as well as personality transfer history and terminal assignment history, and the personality transfer history and terminal assignment history are recorded in a verifiable manner so as to whether the same personality state is activated on multiple terminals at a single time for the AI ​​personality entity, and whether the activated state on the destination terminal started after the activated state on the source terminal ended, thereby maintaining consistency and auditability across terminals.

5. The AI ​​system according to claim 1, wherein the stateless terminal is configured not to permanently store personality-related data, and at the end of a session or when the allocation is terminated, the temporary area within the terminal is erased, access to the temporary area is disabled, or the encryption key is discarded to make the remaining data undecryptable, thereby rendering the personality-related data an empty shell.

6. The AI ​​system according to Claim 1, wherein the personality state is treated as an AI personality state package when moved between terminals, the movement of the AI ​​personality state package between terminals is not performed directly between terminals but via the cloud, the personality state including conversation history, internal context and inference state is stored in the state storage unit on the cloud, the inference state includes a calculation state and an intermediate generation state during inference, such storage does not activate the AI ​​personality entity in the state storage unit, and when moving to a new terminal, the personality state is applied in a state that can be inherited, and processing based on the personality state is performed continuously.

7. A program for an information processing device to realize the functions of each part constituting the AI ​​system described in Claim 1.

8. A recording medium readable by a computer or other information processing device, on which the program described in claim 7 is recorded.

Images