Artificial intelligence (AI) explanations

Abstract

A method of training an LLM for an on-line dating service includes receiving a first explanation for a profile of a first seeker and a profile of a first target; training the LLM, at least in part based on the profile of the first seeker, the profile of the first target, and the first explanation, to produce a trained LLM; and producing a second explanation by the trained LLM, at least in part based on a profile of a second seeker and a profile of a second target.

Description

Attorney Docket8463-01021TITLE ARTIFICIAL INTELLIGENCE (Al) EXPLANATIONSBACKGROUND TECHNICAL FIELD

[0001] This disclosure relates to information and communication technology for implementation of business processes of social networking and, in particular, to generating an artificial intelligence (Al) explanation of the relevance of user profiles.RELATED ART

[0002] In a typical on-line dating service, users have profiles about themselves (i.e., as “targets”) to solicit the interest of other users (i.e. , “seekers”). In some dating services, users categorize themselves as only a target or as only a seeker. However, in many dating services, users operate intermittently as both targets and seekers.

[0003] Typically, users send indicators of interest (e.g., “likes”) when operating as a seeker. Thus, for a “match” to occur as a result of a mutual exchange of likes, each user in a pair momentarily operates as a seeker to send a respective like to the other user in the pair.

[0004] In a conventional on-line dating service, matching can be challenging. For example, in one on-line dating service, users may view several hundred profiles before they exchange contact information with another user and go on a date. Indeed, a further estimate is that a user declines the vast majority of recommendations by such a service.BRIEF SUMMARY

[0005] In a first implementation of the present disclosure, a method of training an LLM for an on-line dating service includes receiving a first explanation for a profile of a first seeker and a profile of a first target; training the LLM, at least in part based on the profile of the first seeker, the profile of the first target, and the firstAttorney Docket8463-01022explanation, to produce a trained LLM; and producing a second explanation by the trained LLM, at least in part based on a profile of a second seeker and a profile of a second target.

[0006] In a second implementation of the present disclosure, an apparatus includes a memory that stores a first explanation for a profile of a first seeker on an online dating service and that stores a profile of a first target on the on-line dating service; and a processor configured to train a large language model (LLM), at least in part based on the profile of the first seeker, the profile of the first target, and the first explanation, to produce a trained LLM, wherein the trained LLM produces a second explanation, at least in part based on a profile of a second seeker on the on-line dating service and a profile of a second target on the on-line dating service.

[0007] In a third implementation of the present disclosure, a computer-readable medium is encoded with instructions that, when executed by a processor of a computing device, cause the computing device to perform operations. The operations include receiving a first explanation for a profile of a first seeker of an on-line dating service and a profile of a first target of the on-line dating service; training the LLM, at least in part based on the profile of the first seeker, the profile of the first target, and the first explanation, to produce a trained LLM; and receiving a second explanation by the trained LLM, at least in part based on a profile of a second seeker of the on-line dating service and a profile of a second target of the on-line dating service.BRIEF DESCRIPTION OF DRAWINGS

[0008] Fig. 1 illustrates a conventional portion of a profile of a target shown to a seeker on an on-line dating service.

[0009] Fig. 2 illustrates a portion of a profile of a target shown to a seeker on an on-line dating service, according to an implementation of the present disclosure.

[0010] Fig. 3 illustrates a portion of a profile of a target shown to a seeker on an on-line dating service, according to an implementation of the present disclosure.Attorney Docket8463-01023[oon] Fig. 4 illustrates an example of key connection points, a short form explanation, and a long form explanation, based on a profile of a seeker and a profile of a target, according to an implementation of the present disclosure.

[0012] Fig. 5 illustrates an algorithm for a system for engineering a prompt on a large language model (LLM), according to an implementation of the present disclosure.

[0013] Fig. 6 illustrates an algorithm for training an LLM, according to an implementation of the present disclosure.

[0014] Fig. 7 illustrates an algorithm for fine-tuning an LLM, according to an implementation of the present disclosure.

[0015] Fig. 8 illustrates an algorithm for causing a display of a profile of a target with an explanation, according to an implementation of the present disclosure.

[0016] Fig. 9 illustrates an algorithm for creating a match on an on-line dating service, according to an implementation of the present disclosure.

[0017] Fig. 10 illustrates an example of a computing device, according to an implementation of the present disclosure.DETAILED DESCRIPTION

[0018] Although on-line dating services serve relevant recommendations to their users, there is a need to improve the ecosystem of online dating to provide more engaging ways for users to interact and, ultimately, obtain more matches. The present inventors believe that highlighting or contextualizing the relevance of these recommendations can lead to more matches.

[0019] Thus, to promote matches among users, implementations of the present disclosure can find key connection points between a profile of a seeker and a profile of a target. Representations of these key connection points can be displayed as explanations of relevance to the seeker to help them make a swiping decision regarding the target.

[0020] In various implementations, a computing device of the seeker can display these explanations on profiles of targets that the on-line dating service suggestsAttorney Docket8463-01024to the seeker. The explanations can help contextualize why a profile of a particular target is being shown to the seeker and what might make this target a good match for the seeker. The explanations can use the profile information of both the seeker and the target to come up with key connection points that connect these two people.

[0021] Various implementations of the present disclosure can incorporate, fine-tune, and apply machine-learning (for example, in the form of a large language model [LLM]). Some implementations in the context of an on-line dating service can take a seeker’s profile and a target’s profile as inputs. The implementation can then generate an explanation of why that target, who is shown to that seeker, was specifically selected for the seeker by the on-line dating service.

[0022] Fig. 1 illustrates a conventional portion of a profile of a target shown to a seeker on an on-line dating service. A server associated with the on-line dating service can transmit the profile of the target to a computing device of the seeker. The computing device can then display at least a portion of the profile of the target in the context of the on-line dating service. In many implementations, this context is an application program associated with (e.g., produced by) the on-line dating service. In other implementations, this context is a web site.

[0023] The profile of the target can include text and other media. The other media can be or include one or more images, one or more videos, or one or more audio files. The audio files can be audio clips of the target (e.g., a voice recording) or of a song indicated by the target (e.g., their favorite song on a streaming site).

[0024] The other media can additionally or alternatively be one or more links (e.g., uniform resource locators [URLs]) to such images, videos, or audio files. Thus, although the server might not transmit the images, videos, or audio files, the computing device of the seeker can retrieve the other media.

[0025] The profile of the target can also include tags, such as hash tags. In various implementations, these tags can be set by the server or by the target.

[0026] In the example of Fig. 1, the computing device of the seeker displays a portion of the profile of the target. As shown in Fig. 1, this portion can include a photograph of the target displayed in the background of the profile.Attorney Docket8463-01025

[0027] The computing device can overlay text of the profile onto the photograph. In the example of Fig. 1, this text indicates a name of the target (“Sarah”), an age of the target (“21”), a job of the target (“Welder”), and interests of the target (e.g., “Movies”). The target can indicate this information at the time of creation of their profile by providing their interests directly or inputting photos with their interests into an algorithm of the on-line dating service and the algorithm generating interest tags based on the photos, for example. In various implementations, the target can define their own job and interests; in some implementations, the target selects from among a predetermined number of jobs or interests.

[0028] Research indicates the seeker will decide whether to send a like to the target in less than two seconds. In that limited period of time, it might not be apparent to the seeker what the relevance of the target is. Therefore, the seeker might not recognize the potential in establishing a match with the target.

[0029] Thus, Fig. 2 illustrates a portion of a profile of a target shown to a seeker on an on-line dating service, according to an implementation of the present disclosure. As shown in Fig. 2, the profile has been supplemented by the inclusion of an explanation 240.

[0030] In many implementations of the present disclosure, the explanation 240 is generated by a machine learning (commonly called “artificial intelligence” [Al]) algorithm, at least in part based on the profile of the seeker and the profile of the target. In particular, the explanation 240 can be based on similarities between the profile of the seeker and the profile of the target.

[0031] For context, Fig. 2 also includes an indication 220 to the seeker that the explanation 240 is related to that particular seeker. That is, the indication 220 can prevent the seeker from believing the explanation 240 is simply a portion of the profile of the target intended by the target for a general audience.

[0032] The explanation 240 is an example of a “long form explanation” in which the explanation follows a grammatical structure of a few (e.g. one to two) sentences. In various implementations, the explanation 240 can include more sentences,Attorney Docket8463-01026particularly in implementations in which information can be communicated in a smaller space (e.g., Chinese).

[0033] As shown in Fig. 2, the computing device of the seeker can also display functional buttons 260 associated with the context of the on-line dating service, which are for illustration purposes only and are not to be construed as limiting. For example, each of the functional buttons 260 can indicate a particular operation to be performed relative to the profile of the target. For example, the seeker can select one of the functional buttons 260 to send a like to the target , or even a functional button to send a message with the like. The seeker can select a different one of the functional buttons 260 to dismiss the profile of the target without sending a like.

[0034] In select implementations, some operations relative to the profile of the target can be performed without selection of one of the functional buttons 260. For example, the seeker can send a like to the target by dragging right on the photograph of the target, for example. Relatedly, the seeker can dismiss the profile of the target without sending a like by dragging left on the photograph of the target.

[0035] The computing device can also display navigation icons 280. The seeker can use the navigation icons 280 to navigate within the application program or website of the on-line dating service, for example.

[0036] Fig. 3 illustrates a portion of a profile of a target shown to a seeker on an on-line dating service, according to an implementation of the present disclosure. As shown in Fig. 3, the profile has been supplemented by the inclusion of an explanation 340.

[0037] The explanation 340 differs from the explanation 240 in that the explanation 340 is a “short form explanation.” More specifically, the explanation 340 is a list of a predetermined number (e.g., four) of bullet points. In various implementations, the explanation 340 can include more or fewer bullet points. Like the explanation 240, the explanation 340 can be based on similarities between the profile of the seeker and the profile of the target.Attorney Docket8463-01027

[0038] Fig. 4 illustrates an example of key connection points, a short form explanation, and a long form explanation, based on a profile of a seeker and a profile of a target, according to an implementation of the present disclosure.

[0039] A key connection point is a human-intelligible similarity between a portion of the profile of the seeker and a portion of the profile of the target. For example, the text “country man” in the profile of the seeker forms a key connection point with the text “good ol tatted country boy.”

[0040] Similarly, the shared text “heavy metal” and “tattoos” in the profile of the seeker and the profile of the target forms key connection points. That is, the formation of the key connection points is not limited to the text entered by the seeker or by the target. For example, as shown in Fig. 4, the text “tattoos” does not appear in the text of the profile of the seeker; rather, the text “tattoos” appears only in the image tags of the profile of the seeker. Thus, the key connection points can be formed based on image tags generated by image recognition performed on images uploaded to the on-line dating service by the seeker or target.

[0041] In various implementations, a profile of a user can include an audio file, such as a song, or a link to such an audio file. In this situation, there might be tags associated with the audio file indicating, for example, a genre of the song. Therefore, in some implementations of the present disclosure, the key connection points can be based on a tag associated with the audio file.

[0042] A short form explanation can be based on the profile of the seeker and the profile of the target. As shown in Fig. 4, the short form explanation is not explicitly limited to the key connection points. For example, the short form explanation indicates the target is a motorcycle mechanic. However, because the profile of the seeker does not explicitly reference motorcycles or mechanics, neither does the key connection points. Thus, the Al possibly created this explanation based on the welder job of the seeker and the motorcycle mechanic job of the target.

[0043] Further, because the seeker understands their own identity, the short form explanation can focus on aspects of the profile of the target. For example,Attorney Docket8463-01028the short form explanation indicates the target is a motorcycle mechanic, rather than indicating the seeker is a welder.

[0044] A long form explanation can be based on the profile of the seeker and the profile of the target. As shown in Fig. 4, the long form explanation is not explicitly limited to the key connection points. For example, the long form explanation indicates the target has an interest in motorcycles. However, because the profile of the seeker does not explicitly reference motorcycles, neither does the key connection points. Thus, the Al possibly created this explanation based on some other aspect of the profile of the seeker.

[0045] Fig. 5 illustrates an algorithm 500 for a system for engineering a prompt on a large language model (LLM), according to an implementation of the present disclosure. The algorithm 500 begins at 510 and advances to 520.

[0046] In 520, the system receives a profile pair of the Ith seeker and the Ith target. Because LLMs are often trained iteratively, I is, for example, a natural number. That is, the system initially receives a profile of a first seeker and a profile of a first target. The algorithm 500 then advances to 525.

[0047] In 525, the system can apply a processing algorithm to the profile of the Ith seeker and the profile of the Ith target to produce a processed profile of the Ith seeker and a processed profile of the Ith target. For example, the processing algorithm can delete or modify information in the profile of the Ith seeker and the profile of the Ith target. In various implementations, the deleted information can include a blood type, a social media username, a zodiac sign, information on cannabis use, and so on. In several implementations, the deleted information can additionally or alternatively relate to sexual orientation, sexual content, or policy violating content. The algorithm 500 then advances to 530.

[0048] In 530, a prompting LLM receives a profile of the Ith seeker and a profile of the Ith target. In implementations in which the system applies a processing algorithm to the profile of the Ith seeker in 525, the prompting LLM can receive the processed profile of the Ith seeker. Similarly, in implementations in which the system applies a processing algorithm to the profile of the Ith target in 525, the prompting LLMAttorney Docket8463-01029can receive the processed profile of the Ith target. In implementations in which the system does not apply a processing algorithm to the profile of the Ith seeker or the profile of the Ith target in 525, then the prompting LLM receives the profile of the Ith seeker or the profile of the Ith target.

[0049] The prompting LLM can produce one or more outputs based on the profiles received in 530. The prompting LLM can be prompt-engineered to produce such outputs, based on pairs of profiles. These outputs can include a key connection point and a prompting explanation, as discussed in connection with Figs. 2-4, for example.

[0050] The prompting explanation can be a short form explanation or a long form explanation. In various implementations, the long form explanation can vary in tone, voice, or format. For example, the long form explanation can set forth its information in a formal, humorous, or over-the-top tone.

[0051] The algorithm 500 then advances to 540.

[0052] In 540, the output of the prompting LLM can be scored. The scoring can be based on a hallucination score, a coherence score, and a reasonability score. The hallucination score concerns whether the output of the prompting LLM is verifiably true, including whether the output is incorrect or misleading. For example, if the key connection point is “skiing,” but neither the profile of the Ith seeker or the profile of the Ith target mentions skiing, then that is a hallucination. Further, the key connection point might be “water sports,” which might be inferred from the profile of the Ith seeker mentioning “beach” and the profile of the Ith target mentioning “ocean,” but is not verifiable based on the profile of the Ith seeker and the profile of the Ith target: this key connection point, too, is a hallucination.

[0053] The coherence score concerns whether the output is grammatically incorrect or otherwise at odds with native speaker usage. For example, a key connection point using the phrase “weekend drinker,” although not grammatically incorrect, is not how a native English speaker would typically express that key connection point. Thus, such a phrase can be scored as incoherent.

[0054] The reasonability score concerns whether the output is reasonable. For example, a long form explanation that includes a phrase that violates a contentAttorney Docket8463-010210policy of the on-line dating service (e.g., because it is overly sexual) would be scored as unreasonable.

[0055] As discussed above, both the key connection point and the prompting explanation can be scored.

[0056] Each score can be as simple as a binary value. For example, the hallucination score can be 1, if the entirety of the explanation is factually accurate, based on the profile of the Ith target. If less than the entirety of the explanation is factually accurate, based on the profile of the Ith target, then the hallucination score can be o. The score can also be a percentage or on any other scale.

[0057] The algorithm 500 then advances to 550.

[0058] In 550, the system can determine whether the output is satisfactory. In some implementations, the system can determine whether the score of the output exceeds a predetermined threshold. In various implementations, the system can determine whether a history value of the score exceeds a predetermined threshold. In at least one implementation, the system can determine whether a most recent history value (e.g., last 50 scores) is within a predetermined value (e.g., 10%) of a preceding history value (e.g., the 50 scores preceding the last 50 scores). If the system determines in 550 that the output is satisfactory, the algorithm 500 advances to S570. If the system determines in 550 that the output is not satisfactory, then the algorithm 500 advances to 560.

[0059] In 560, the system can prompt-engineer the prompting LLM, at least in part based on the profile of the Ith seeker, the profile of the Ith target, the key connection point, and the prompting explanation. This new prompt can produce a further LLM output. For example, the prompting LLM can be improved with prodiving bad examples and desired examples.

[0060] The algorithm 500 can then advance to 565.

[0061] In 565, the system can modify the processing algorithm. For example, the processing algorithm can be modified to delete additional information or to normalize information. For example, if the prompting LLM did not produce a particular key connection point, the processing algorithm can be modified to promoteAttorney Docket8463-010211the production of the key connection point. As an example, the phrase “appreciates presents” in a profile can be normalized to “likes gifts.” In many implementations, the processing algorithm is modified by a human. The algorithm 500 then advances to 568.

[0062] In 568, the system increments to the next I. The algorithm 500 then advances to 520, in which a different pair of profiles of a seeker and a target are received. To be clear, the same profile of the previous seeker and a profile of a different target can be considered a different pair. Similarly, a profile of a different seeker and the profile of the previous target can likewise be considered a different input.

[0063] Briefly returning to 550, if the system determines the score is satisfactory, the algorithm 500 advances to S570. In 570, the algorithm 500 concludes.

[0064] With one implementation of the algorithm 500, within four iterations, the hallucinations dropped to 26%, incoherence to 44%, and reasonableness to 74%.

[0065] Fig. 6 illustrates an algorithm 600 for training an LLM, according to an implementation of the present disclosure. The algorithm 600 begins at 610 and advances to 620.

[0066] In 620, the system receives a Jth explanation for a profile of a Jth seeker and a profile of a Jth target. The Jth explanation can be written by a human or a teacher Al model, for example. The Jth seeker can be the last Ith seeker, a previous Ith seeker, or any other seeker. The Jth target can be the last Ith target, a previous Ith target, or any other target. The algorithm 600 then advances to 630.

[0067] In 630, the LLM is trained at least in part based on the profile of the Jth seeker, the profile of the Jth target, and the Jth explanation, to produce a trained LLM. In an implementation in which the Jth explanation received at 620 is received from a human, this training can be implemented with reinforcement learning from human feedback or supervised finetuning. The algorithm 600 then advances to 635.

[0068] In 635, the system receives a profile of a Kth seeker and a profile of a Kth target. The pair of the Kth seeker and the Kth target generally differs from the pair of the Jth seeker and the Jth target to avoid over-fitting the LLM. SomeAttorney Docket8463-010212implementations use the same seeker or the same target, and it is possible for the same users to be used in the opposite (seeker / target) role.

[0069] In addition, the system can apply the processing algorithm discussed in connection with 565 to obtain a processed profile of the Kth seeker and a processed profile of the Kth target.

[0070] The algorithm 600 then advances to 640.

[0071] In 640, the trained LLM produces a Kth explanation, at least in part based on a profile of the Kth seeker and a profile of the Kth target. The Kth explanation can be a short form explanation or a long form explanation.

[0072] The algorithm 600 then advances to 650.

[0073] In 650, the system scores the Kth explanation. This scoring can be the same or similar to the scoring performed in 540. For example, the scoring can be based on a hallucination score, a coherence score, and a reasonability score.

[0074] The algorithm 600 then advances to 660.

[0075] In 660, the system determines whether the Kth explanation is satisfactory. This determination can be based on the same or similar scoring as discussed in the context of 550. For example, the system can determine whether the score of the Kth explanation exceeds a predetermined threshold.

[0076] If the system determines in 660 that the Kth explanation is satisfactory, then the algorithm 600 advances to 670. If the system determines in 660 that the Kth explanation is not satisfactory, then the algorithm 600 advances to 663.

[0077] In 663, the system can receive a rewritten Kth explanation, at least in part based on the profile of the Kth seeker and the profile of the Kth target. The rewritten Kth explanation can be received from a human or from a teacher AL

[0078] The algorithm 600 then advances to 666.

[0079] In 666, the system sets J equal to K and iterates to the next K. The algorithm 600 then advances to 630 for another iteration.

[0080] Briefly returning to 660, if the system determines the Kth explanation is satisfactory, the algorithm 600 advances to 670. In 670, the algorithm 600 concludes.Attorney Docket8463-010213[oo8i] Various implementation of the algorithm 600 can typically use ~2ooo human-written explanations and ~sooo outputs from the Al teacher model from the algorithm 500.

[0082] Fig. 7 illustrates an algorithm 700 for fine-tuning an LLM, according to an implementation of the present disclosure. The algorithm 700 begins at 710 and advances to 715.

[0083] In 715, the system receives a profile of an Lth seeker and a profile of an Lth target. The pair of the Lth seeker and the Lth target generally differs from the pair of the last Kth seeker and the last Kth target to avoid over-fitting the LLM. Some implementations use the same seeker or the same target, and it is possible for the same users to be used in the opposite (seeker / target) role.

[0084] In addition, the system can apply the processing algorithm discussed in connection with 565 to obtain a processed profile of the Lth seeker and a processed profile of the Lth target.

[0085] The algorithm 700 then advances to 720.

[0086] In 720, the system receives an Lth explanation from the trained LLM, at least in part based on the processed profile of the Lth seeker and the processed profile of the Lth target. The explanation can be a short form explanation or a long form explanation. The algorithm 700 then advances to 730.

[0087] In 730, the system receives a revised explanation, at least in part based on the profile of the Lth seeker and the profile of the Lth target. The revised explanation can be received from a human or from a teacher Al. The algorithm 700 then advances to 740.

[0088] In 740, the trained LLM is trained, at least in part based on the profile of the Lth seeker, the profile of the Lth target, the Lth explanation, and the revised explanation, to produce an intermediate LLM.

[0089] Thus, reinforcement learning from human feedback (RLHF), and, more specifically, human preference alignment, can be implemented when the revised explanation is received from a human.

[0090] The algorithm 700 then advances to 745.Attorney Docket8463-010214

[0091] In 745, the system receives a profile of an Mth seeker and a profile of an Mth target. The pair of the Mth seeker and the Mth target generally differs from the pair of the Lth seeker and the Lth target to avoid over-fitting the LLM. Some implementations use the same seeker or the same target, and it is possible for the same users to be used in the opposite (seeker / target) role.

[0092] In addition, the system can apply the processing algorithm discussed in connection with 565 to obtain a processed profile of the Mth seeker and a processed profile of the Mth target.

[0093] The algorithm 700 then advances to 750.

[0094] In 750, the intermediate LLM produces an Mth explanation, at least in part based on the profile of the Mth seeker and the profile of the Mth target. The Mth explanation can be produced based on the processed profile of the Mth seeker and the processed profile of the Mth target, for example.

[0095] The algorithm 700 then advances to 760.

[0096] In 760, the system scores the Mth explanation. This scoring can be the same or similar to the scoring performed in 540. For example, the scoring can be based on a hallucination score, a coherence score, and a reasonability score.

[0097] The algorithm 700 then advances to 770.

[0098] In 770, the system determines whether the Mth explanation is satisfactory. This determination can be based on the same or similar scoring as discussed in the context of 550. For example, the system can determine whether the score of the Mth explanation exceeds a predetermined threshold.

[0099] If the system determines in 770 that the Mth explanation is satisfactory, then the algorithm 700 advances to 780. If the system determines in 770 that the Mth explanation is not satisfactory, then the algorithm 700 advances to 775.

[0100] In 775, the system sets L equal to M and iterates M. The algorithm 700 then advances to 730.

[0101] This iteration using both the explanation and the revised explanation can achieve a fine-tuned output from the LLM with about 1000 samples.Attorney Docket8463-010215

[0102] Other implementations can use supervised fine-tuning which might use approximately 5000 manually-written, diverse explanations.

[0103] Fig. 8 illustrates an algorithm 800 for causing a display of a profile of a target with an explanation, according to an implementation of the present disclosure. The algorithm 800 begins at 810 and advances to 820.

[0104] In 820, the system receives a profile of an Nth seeker. The profile can be received based on, for example, a login from a computing device of the Nth seeker. The algorithm 800 then advances to 830.

[0105] In 830, the system receives a profile of an Nth target. For example, the system can determine the profile of the Nth target, at least in part based on the login and a matching algorithm applied to the Nth seeker. In some implementations, the Nth target can be determined based on an alternative criterion. For example, the Nth seeker might have requested a profile boost. The algorithm 800 then advances to 835.

[0106] In 835, the system can apply the processing algorithm discussed in connection with 565 to obtain a processed profile of the Nth seeker and a processed profile of the Nth target. The algorithm 800 then advances to 840.

[0107] In 840, the system produces an Nth explanation, at least in part based on the trained LLM, the profile of the Nth seeker, and the profile of the Nth target. The Nth explanation can be based on the processed profile of the Nth seeker and the processed profile of the Nth target, for example. The Nth explanation can be a short form explanation or a long form explanation. The algorithm 800 then advances to 850.

[0108] In 850, the system causes a display of the profile of the Nth target and the Nth explanation on a computing device of the Nth seeker. For example, a server of the system can transmit the profile of the Nth target and the Nth explanation to the computing device. The display on the computing device can resemble Fig. 2 or Fig. 3, for example.

[0109] The algorithm 800 then advances to 860 and concludes.

[0110] Fig. 9 illustrates an algorithm 900 for creating a match on an on-line dating service, according to an implementation of the present disclosure.

[0111] The algorithm 900 begins at 910 and advances to 920.Attorney Docket8463-010216[oii2] In 920, the system determines whether a like was received from a seeker on a profile of a target. For example, if a computing device of the seeker has a display similar to Fig. 2 or Fig. 3, then the seeker can drag right on the photograph of the target to indicate the like. Alternatively, the seeker can tap on the heart icon or the star icon to indicate the like.

[0113] Additionally, the seeker can drag left on the photograph of the target to dismiss the profile of the target. This gesture can indicate the seeker does not like the target. Alternatively, the seeker can tap on the X icon to indicate that the seeker does not like the target.

[0114] If the system determines that a like of the target was not received from the seeker, then the system can advance to 960. If the system determines that a like of the target was received from the seeker, then the algorithm 900 advances to 930.

[0115] In 930, the system can display the profile of the seeker on a computing device of the target. The algorithm 900 then advances to 940.

[0116] In 940, the system determines whether a like was received from the target on the profile of the seeker. For example, if a computing device of the target has a display similar to Fig. 2 or Fig. 3, then the target can drag right on the photograph of the seeker to indicate the like. Alternatively, the target can tap on the heart icon or the star icon to indicate the like.

[0117] If the system determines that a like of the seeker was not received from the target, then the system can advance to 960. If the system determines that a like of the seeker was received from the target, then the algorithm 900 advances to 950.

[0118] In 950, the system determines that a match has been made between the target and the seeker. Thus, in some on-line dating services, the target and the seeker can freely communicate between themselves. In some on-line dating services, a particular user initiates the conversation, such as based on role (e.g., the target) or based on gender (e.g., female). In some on-line dating services, the duration of the free communication is limited to a predetermined number of messages or a predetermined time period (e.g., 24 hours). In various on-line dating services, a limitation on the communication canAttorney Docket8463-010217removed as a premium service, for example, upon payment of a fee or viewing a number of advertisements. The system then advances to 960.

[0119] In 960, the algorithm 900 concludes.

[0120] Implementations of the present disclosure can include any, all, or portions of the algorithms of Figs. 5-9. Thus, for example, an implementation can include only the algorithm of Fig. 6, only the algorithms of Figs. 6 and 8 (and not Figs. 5 and 7), only the algorithms of Figs. 7 and 8 (and not Figs. 5 and 6), and so on. The particular implementation can be selected, based on the adequacy of the training obtained by the particular algorithm(s) and the design constraints.

[0121] Further, the preceding discussion uses mathematical variables like I and J to set forth its iterative nature. Thus, one of ordinary skill in the art would understand that I, J, K, L, M, and N can be natural numbers or whole numbers. A person of ordinary skill in the art would understand that these mathematical sets can be communicated using different phrases, such as “non-negative integers.”

[0122] Further, these variables provide context to the algorithms of Figs. 5-8. Thus, it is to be understood that every Ith seeker profile is not mutually exclusive to, for example, every Jth seeker profile.

[0123] Further, the algorithms of Figs. 5-7, in particular, can be performed hundreds or thousands of times or more.

[0124] Further, in the preceding explanation, the processing algorithm is set forth as distinct from the LLM. In many implementations, the functionality of the processing algorithm can be incorporated into the training of the LLM. Thus, some implementations of the LLM can achieve a similar processing effect without a separate algorithm.

[0125] In addition, it is to be understood that the processed profile of a seeker is based on the (unprocessed) profile of the seeker. Thus, if an operation is performed based on the profile of the seeker, then the operation can alternatively be performed on the processed profile of the seeker in some implementations. Likewise, if an operation is performed based on a processed profile, the operation can alternatively be performed based on the (unprocessed) profile in some implementations. ForAttorney Docket8463-010218example, implementations in which the processing algorithm is incorporated into the LLM can use the (unprocessed) profiles rather than the processed profiles, although an implementation can involve additional iterations.

[0126] In various implementations, the LLMs of the present disclosure can be implemented on a processor of a server associated with the on-line dating service. In some implementations, the LLMs are implemented by a third party, and the processor of the server communicates with the LLMs via a network. In a few implementations, the LLMs are implemented on the respective computing devices of the users. In at least one implementation, the LLMs produced by at least one of the algorithm 500, the algorithm 600, or the algorithm 700 is burned into a memory to produce the LLM of the algorithm 800.

[0127] Fig. 10 illustrates a computing device 1000, according to an implementation of the present disclosure. The computing device of a user and the server can each be implemented with the computing device 1000.

[0128] The computing device 1000 can include a network interface 1010, a user input interface 1020, a memory 1030, a program 1035, a processor 1040, a user output interface 1050, and a bus 1055.

[0129] Although illustrated within a single housing, the computing device 1000 can be distributed across plural housings or sub-systems that cooperate in executing program instructions. In some implementations, the computing device 1000 can include one or more blade server devices, standalone server devices, personal computers (including laptop computers and tablet computers), routers, hubs, switches, bridges, firewall devices, intrusion detection devices, mainframe computers, network-attached storage devices, smartphones and other mobile telephones, and other computing devices. Although the computing device executes the Windows OS, macOS, or Linux in many implementations, the hardware can be configured according to a Symmetric Multiprocessing (SMP) architecture or a Non-Uniform Memory Access (NUMA) architecture.

[0130] The network interface 1010 provides one or more communication connections and / or one or more devices that allow for communication between the computing device 1000 and other computing systems (not shown) over a communicationAttorney Docket8463-010219network, collection of networks (not shown), or the air, to support the artificial intelligence explanations outlined herein. The network interface 1010 can communicate using various networks (including both internal and external networks) such as near-held communications (NFC), Wi-Fi™, Bluetooth, Ethernet, cellular (e.g., 3G, 4G, 5G), white space, 802.11X, satellite, LTE, GSM / HSPA, CDMA / EVDO, DSRC, CAN, GPS, facsimile, or any other wired or wireless interface. Other interfaces can include physical ports (e.g., Ethernet, USB, HDMI, etc.), interfaces for wired and wireless internal subsystems, and the like. Similarly, nodes and user equipment (e.g., mobile devices) of a computing system that includes the computing device 1000 can also include suitable interfaces for receiving, transmitting, and / or otherwise communicating data or information in a network environment.

[0131] The user input interface 1020 can receive one or more inputs from a human. The user input interface 1020 can be or include a mouse, a touchpad, a keyboard, a touchscreen, a trackball, a camera, a microphone, a joystick, a game controller, a scanner, and / or any other input device.

[0132] The memory 1030, also termed a “storage,” can include or be one or more computer-readable storage media readable by the processor 1040 and that store software. The memory 1030 can be implemented as one storage device or across multiple co-located or distributed storage devices or sub-systems. The memory 1030 can include additional elements, such as a controller, that communicate with the processor 1040. The memory 1030 can also include storage devices and / or sub-systems on which data and / or instructions are stored. The computing device 1000 can access one or more storage resources to access information to carry out any of the processes indicated in this disclosure and, in particular, Figs. 5-9. In various implementations, the memory 1030 stores the program 1035 to execute at least a portion of the algorithms illustrated in Figs.5-9. Further, the program 1035, when executed by the computing device 1000 generally and / or the processor 1040 specifically, can direct, among other functions, generation of artificial intelligence explanations, as described herein.

[0133] The memory 1030 can be or include a read-only memory (ROM), a programmable ROM (PROM), an erasable PROM (EPROM), an electrically erasableAttorney Docket8463-010220PROM (EEPROM), a random-access memory (RAM), a dynamic RAM (DRAM), a static RAM (SRAM), a field programmable gate array (FPGA), a hard drive, a cache memory, a flash memory, a removable disk, or a tape reel. The memory 1030 can be or include resistive RAM (RRAM) or a magneto-resistive RAM (MRAM). The information being tracked, sent, received, or stored in the computing device can be provided in any database, register, table, cache, queue, control list, or storage structure, based on particular implementations, all of which could be referenced in any suitable timeframe.

[0134] The processor 1040 (e.g., a processing unit) can be or include one or more hardware processors and / or other circuitry that retrieve and execute software, especially the program 1035, from the memory 1030. The processor 1040 can be implemented within one processing device, chip, or package and can also be distributed across multiple processing devices, chips, packages, or sub-systems that cooperate. In some implementations, the processor 1040 is or includes a Graphics Processing Unit (GPU) or neural processing unit (NPU).

[0135] The processor 1040 can have any register size, such as a 32-bit register or a 64-bit register, among others. The processor 1040 can include multiple cores. Implementations of the processor 1040 are not limited to any particular number of threads. The processor 1040 can be fabricated by any process technology, such as iqnm process technology.

[0136] The user output interface 1050 outputs information to a human user. The user output interface 1050 can be or include a display (e.g., a screen), a touchscreen, speakers, a printer, or a haptic feedback unit. In many implementations, the user output interface 1050 can be combined with the user input interface 1020. For example, some such implementations include a touchscreen, a headset including headphones and a microphone, or a joystick with haptic feedback.

[0137] In implementations including multiple computing devices, a server of a system or, in a serverless implementation, a peer can use one or more communications networks that facilitate communication among the computing devices to achieve the artificial intelligence explanations, as outlined herein. For example, the one or more communications networks can include or be a local area network (LAN) or wideAttorney Docket8463-010221area network (WAN) that facilitate communication among the computing devices. One or more direct communication links can be included between the computing devices. In addition, in some cases, the computing devices can be installed at geographically distributed locations. In other cases, the multiple computing devices can be installed at one geographic location, such as a server farm or an office.

[0138] As used herein, the terms “storage media” or “computer-readable storage media” can refer to non-transitory storage media, such as non-limiting examples of a hard drive, a memory chip, an ASIC, and cache memory, and to transitory storage media, such as carrier waves or propagating signals.

[0139] Aspects of the system can be implemented in various manners, e.g., as a method, a system, a computer program product, or one or more computer-readable storage media. Accordingly, aspects of the present disclosure can take the form of a hardware implementation, a software implementation (including firmware, resident software, or micro-code) or an implementation combining software and hardware aspects that can generally be referred to herein as a “module” or a “system.” Functions described in this disclosure can be implemented as an algorithm executed by one or more hardware processing units, e.g., the processor 1040. In various embodiments, different operations and portions of the operations of the algorithms described can be performed by different processing units. In some implementations, the operations can be achieved by reciprocating software in one or more computing devices. The program 1035 can be implemented using reciprocating software, for example. Furthermore, aspects of the present disclosure can take the form of a computer program product embodied in one or more computer-readable media having computer-readable program code embodied, e.g., encoded or stored, thereon. In various implementations, such a computer program can, for example, be downloaded (or updated) to existing devices and systems or be stored upon manufacture of these devices and systems.

[0140] Any suitable permutation can be applied to a physical implementation, including the design of the communications network in which the system is implemented. In one embodiment, the bus 1055 can share hardware resources with the memory 1030 and the processor 1040. In this alternative implementation, theAttorney Docket8463-010222computing device 1000 can be provided with separate hardware resources including one or more processors and memory elements.

[0141] In example implementations, various other components of the computing device 1000 can be installed in different physical areas or can be installed as single units.

[0142] The computing device 1000 can be configured to facilitate communication with machine devices (e.g., vehicle sensors, instruments, electronic control units (ECUs), embedded devices, actuators, displays, etc.) through the bus 1055. Other suitable communication interfaces can also be provided for an Internet Protocol (IP) network, a user datagram protocol (UDP) network, or any other suitable protocol or communication architecture enabling network communication with machine devices.

[0143] The innovations in this detailed description can be implemented in a multitude of different ways, for example, as defined and covered by the claims and / or select examples. In the description, reference is made to the drawings where like reference numerals can indicate identical or functionally similar elements. Elements illustrated in the drawings are not necessarily drawn to scale. Additionally, certain implementations can include more elements than illustrated in a drawing and / or a subset of the elements illustrated in a drawing. Further, some implementations can incorporate a suitable combination of features from two or more drawings.

[0144] The disclosure describes various illustrative implementations and examples for implementing the features and functionality of the present disclosure. The components, arrangements, and / or features are described in connection with various implementations and are merely examples to simplify the present disclosure and are not intended to be limiting. In the development of actual implementations, implementationspecific decisions can be made to achieve specific goals, including compliance with system, business, and / or legal constraints, which can vary from one implementation to another. Additionally, while such a development effort might be complex and timeconsuming, it would be a routine undertaking for those of ordinary skill in the art having the benefit of this disclosure.Attorney Docket8463-010223

[0145] The systems, methods and devices of this disclosure have several innovative aspects, no one of which is solely responsible for the attributes disclosed herein. Some objects or advantages might not be achieved by implementations described herein. Thus, for example, certain implementations can operate in a manner that achieves or optimizes one advantage or group of advantages as taught herein and not other objects or advantages as taught or suggested herein.

[0146] In one example implementation, electrical circuits of the drawings can be implemented on a board of an associated electronic device. The board can be a general circuit board that can hold various components of the internal electronic system of the electronic device and, further, provide connectors for other peripherals. More specifically, the board can provide the electrical connections by which other components of the system can communicate electrically. Any processors (inclusive of digital signal processors, microprocessors, supporting chipsets, etc.) and computer-readable, non-transitory memory elements can be coupled to the board based on configurations, processing demands, and computer designs. Other components such as external storage, additional sensors, controllers for audio / video display, and peripheral devices can be attached to the board as plug-in cards, via cables, or integrated into the board itself. In various implementations, the functionalities described herein can be implemented in emulation form as software or firmware running within one or more configurable (e.g., programmable) elements arranged in a structure that supports these functions. A non-transitory, computer-readable storage medium can include instructions to allow one or more processors to carry out the emulation.

[0147] In another example implementation, the electrical circuits of the drawings can be implemented as stand-alone modules (e.g., a device with associated components and circuitry configured to perform a specific application or function) or implemented as plug-in modules into application specific hardware of electronic devices. Implementations of the present disclosure can be readily included in a system-on-chip (SOC) package. An SOC represents an integrated circuit (IC) that integrates components of a computer or other electronic system into one chip. The SOC can contain digital, analog, mixed-signal, and often radio frequency functions on one chip substrate. OtherAttorney Docket8463-010224implementations can include a multi-chip-module (MCM), with a plurality of separate ICs located within one electronic package and that interact through the electronic package. In various other implementations, the processors can be implemented in one or more silicon cores in Application Specific Integrated Circuits (ASICs), Field Programmable Gate Arrays (FPGAs), programmable array logic (PAL), generic array logic (GAL), and other semiconductor chips.

[0148] The specifications, dimensions, and relationships outlined herein (e.g., the number of processors and logic operations) have been offered for non-limiting purposes of example and teaching. For example, various modifications and changes can be made to the arrangements of components. The description and drawings are, accordingly, to be regarded in an illustrative sense, not in a restrictive sense.

[0149] The numerous examples provided herein described interaction in terms of two, three, or more electrical components for purposes of clarity and example. The computing device or a system including the computing device can be consolidated in any manner. Along similar design alternatives, the illustrated components, modules, and elements of the drawings can be combined in various possible configurations within the scope of this disclosure. In certain cases, one or more of the functionalities of a given set of flows might be more clearly described by referencing a limited number of electrical elements. The electrical circuits of the drawings are readily scalable and can accommodate many components, as well as more complicated / sophisticated arrangements and configurations. Accordingly, the provided examples do not limit the scope or inhibit the teachings of the electrical circuits as potentially applied to a myriad of other architectures.

[0150] In this disclosure, references to various features (e.g., elements, structures, modules, components, steps, operations, characteristics, etc.) included in “one implementation,” “example implementation,” “an implementation,” “another implementation,” “some implementations,” “various implementations,” “other implementations,” “alternative implementation,” and the like are intended to mean that any such features can be included in one or more implementations of the present disclosure and might or might not necessarily be combined in the same implementations.Attorney Docket8463-010225Some operations can be deleted or omitted where appropriate, or these operations can be modified or changed considerably. In addition, the timing of these operations can be altered considerably. The preceding operational flows have been offered for purposes of example and discussion. Implementations described herein provide flexibility in that any suitable arrangements, chronologies, configurations, and timing mechanisms can be provided.Examples

[0151] In Example AMi, a method of training an LLM for an on-line dating service includes receiving a first explanation for a profile of a first seeker and a profile of a first target; training the LLM, at least in part based on the profile of the first seeker, the profile of the first target, and the first explanation, to produce a trained LLM; and producing a second explanation by the trained LLM, at least in part based on a profile of a second seeker and a profile of a second target.

[0152] Example AM2 is the method of Example AMi, further comprising: training the trained LLM, at least in part based on the profile of the second seeker, the profile of the second target, and the second explanation, if a determination determines the second explanation is not satisfactory.

[0153] Example AM3 is the method of Example AMi or AM2, further comprising: receiving a third explanation from the trained LLM, at least in part based on a profile of a third seeker and a profile of a third target; receiving a revised explanation, at least in part based on the profile of the third seeker and the profile of the third target; and training the trained LLM, at least in part based on the profile of the third seeker, the profile of the third target, the third explanation, and the revised explanation, to produce an intermediate LLM.

[0154] Example AM4 is the method of Example AM3, further comprising: producing a fourth explanation by the intermediate LLM, at least in part based on a profile of a fourth seeker and a profile of a fourth target; and training the intermediate LLM, at least in part based on the profile of the fourth seeker, the profile of the fourth target, the fourth explanation, and a revised explanation between the profile of theAttorney Docket8463-010226fourth seeker and the profile of the fourth target, if a determination determines the intermediate LLM is not satisfactory, at least in part based on the fourth explanation.

[0155] Example AM5 is the method of any of Examples AM1-AM4, further comprising: receiving a profile of a training seeker and a profile of a training target; producing a key connection point by a prompting LLM, at least in part based on the profile of the training seeker and the profile of the training target; and training the prompting LLM, at least in part based on the profile of the training seeker, the profile of the training target, and the key connection point, to provide the LLM.

[0156] Example AM6 is the method of Example AM5, further comprising: producing a prompting explanation by the prompting LLM, at least in part based on the profile of the training seeker and the profile of the training target; and training the prompting LLM, at least in part based on the prompting explanation, to provide the LLM.

[0157] Example AM7 is the method of any of Examples AM1-AM6, further comprising: receiving a profile of a fifth seeker; receiving a profile of a fifth target; producing a fifth explanation at least in part based on the trained LLM, the profile of the fifth seeker, and the profile of the fifth target; and causing a display of the profile of the fifth target and the fifth explanation on a computing device of the fifth seeker.

[0158] In Example AA1, an apparatus includes a memory that stores a first explanation for a profile of a first seeker on an on-line dating service and that stores a profile of a first target on the on-line dating service; and a processor configured to train a large language model (LLM), at least in part based on the profile of the first seeker, the profile of the first target, and the first explanation, to produce a trained LLM, wherein the trained LLM produces a second explanation, at least in part based on a profile of a second seeker on the on-line dating service and a profile of a second target on the online dating service.

[0159] Example AA2 is the apparatus of Example AA1, wherein the processor is further configured to train the trained LLM, at least in part based on the profile of the second seeker, the profile of the second target, and the second explanation, if a determination determines the second explanation is not satisfactory.Attorney Docket8463-010227[oi6o] Example AA3 is the apparatus of Example AA1 or Example AA2, wherein the trained LLM produces a third explanation, at least in part based on a profile of a third seeker and a profile of a third target, the processor receives a revised explanation, at least in part based on the profile of the third seeker and the profile of the third target, and the trained LLM is trained, at least in part based on the profile of the third seeker, the profile of the third target, the third explanation, and the revised explanation, to produce an intermediate LLM.

[0161] Example AA4 is the apparatus of Example AA3, wherein the intermediate LLM produces a fourth explanation, at least in part based on a profile of a fourth seeker and a profile of a fourth target, and the intermediate LLM is trained, at least in part based on the profile of the fourth seeker, the profile of the fourth target, the fourth explanation, and a revised explanation between the profile of the fourth seeker and the profile of the fourth target, if a determination determines the intermediate LLM is not satisfactory, at least in part based on the fourth explanation.

[0162] Example AA5 is the apparatus of any of Examples AA1-AA4, wherein the processor receives a profile of a training seeker and a profile of a training target, a prompting LLM produces a key connection point, at least in part based on the profile of the training seeker and the profile of the training target, and the prompting LLM is trained, at least in part based on the profile of the training seeker, the profile of the training target, and the key connection point, to provide the LLM.

[0163] Example AA6 is the apparatus of Example AA5, wherein the prompting LLM produces a prompting explanation, at least in part based on the profile of the training seeker and the profile of the training target, and the prompting LLM is trained, at least in part based on the prompting explanation, to provide the LLM.

[0164] Example AA7 is the apparatus of any of Examples AA1-A6, wherein the processor receives a profile of a fifth seeker and a profile of a fifth target, a fifth explanation is produced, at least in part based on the trained LLM, the profile of the fifth seeker, and the profile of the fifth target, and the processor causes a display of the profile of the fifth target and the fifth explanation to the fifth seeker.Attorney Docket8463-010228

[0165] In Example ACi, a computer-readable medium encoded with instructions that, when executed by a processor of a computing device, cause the computing device to perform operations comprising: receiving a first explanation for a profile of a first seeker of an on-line dating service and a profile of a first target of the on-line dating service; training the LLM, at least in part based on the profile of the first seeker, the profile of the first target, and the first explanation, to produce a trained LLM; and receiving a second explanation by the trained LLM, at least in part based on a profile of a second seeker of the on-line dating service and a profile of a second target of the on-line dating service.

[0166] Example AC2 is the medium of Example ACi, the operations further comprising: training the trained LLM, at least in part based on the profile of the second seeker, the profile of the second target, and the second explanation, if a determination determines the second explanation is not satisfactory.

[0167] Example AC3 is the medium of Example ACi or Example AC2, the operations further comprising: receiving a third explanation from the trained LLM, at least in part based on a profile of a third seeker of the on-line dating service and a profile of a third target of the on-line dating service; receiving a revised explanation, at least in part based on the profile of the third seeker and the profile of the third target; and training the trained LLM, at least in part based on the profile of the third seeker, the profile of the third target, the third explanation, and the revised explanation, to produce an intermediate LLM.

[0168] Example AC4 is the medium of Example AC3, the operations further comprising: receiving a fourth explanation by the intermediate LLM, at least in part based on a profile of a fourth seeker and a profile of a fourth target; and training the intermediate LLM, at least in part based on the profile of the fourth seeker, the profile of the fourth target, the fourth explanation, and a revised explanation between the profile of the fourth seeker and the profile of the fourth target, if a determination determines the intermediate LLM is not satisfactory, at least in part based on the fourthexplanation.Attorney Docket8463-010229

[0169] Example AC5 is the medium of any of Examples AC1-AC4, the operations further comprising: receiving a profile of a training seeker and a profile of a training target; producing a key connection point by a prompting LLM, at least in part based on the profile of the training seeker and the profile of the training target; producing a prompting explanation by the prompting LLM, at least in part based on the profile of the training seeker and the profile of the training target; training the prompting LLM, at least in part based on the profile of the training seeker, the profile of the training target, the key connection point, and the prompting explanation, to provide the LLM.

[0170] Example AC6 is the medium of any of Examples AC1-AC5, the operations further comprising: receiving a profile of a fifth seeker; receiving a profile of a fifth target; producing a fifth explanation at least in part based on the trained LLM, the profile of the fifth seeker, and the profile of the fifth target; and causing a display of the profile of the fifth target and the fifth explanation on a computing device of the fifth seeker.

[0171] Example BM1 is a method including receiving a profile of a first seeker and a profile of a first target; receiving a first explanation from a large language model (LLM), at least in part based on the profile of the first seeker and the profile of the first target; receiving a revised explanation, at least in part based on the profile of the first seeker and the profile of the first target; training the LLM, at least in part based on the profile of the first seeker, the profile of the first target, the first explanation, and the revised explanation, to produce a revised LLM; and receiving a second explanation from the revised LLM, at least in part based on a profile of a second seeker and a profile of a second target.

[0172] Example BA1 is an apparatus including a memory that stores a profile of a first seeker and a profile of a first target; and a processor configured to receive a first explanation from a large language model (LLM), at least in part based on the profile of the first seeker and the profile of the first target, to receive a revised explanation, at least in part based on the profile of the first seeker and the profile of the first target, and to train the LLM, at least in part based on the profile of the first seeker,Attorney Docket8463-010230the profile of the first target, the first explanation, and the revised explanation, to produce a revised LLM. A second explanation is received from the revised LLM, at least in part based on a profile of a second seeker and a profile of a second target.

[0173] Example BCi is a computer-readable medium encoded with instructions that, when executed by a processor of a computing device, cause the computing device to perform operations comprising: receiving a profile of a first seeker and a profile of a first target; receiving a first explanation from a large language model (LLM), at least in part based on the profile of the first seeker and the profile of the first target; receiving a revised explanation, at least in part based on the profile of the first seeker and the profile of the first target; training the LLM, at least in part based on the profile of the first seeker, the profile of the first target, the first explanation, and the revised explanation, to produce a revised LLM; and receiving a second explanation from the revised LLM, at least in part based on a profile of a second seeker and a profile of a second target.

Claims

Attorney Docket8463-010231ClaimsWe claim:

1. A method of training an LLM for an on-line dating service, the method comprising:receiving a first explanation for a profile of a first seeker and a profile of a first target;training the LLM, at least in part based on the profile of the first seeker, the profile of the first target, and the first explanation, to produce a trained LLM; and producing a second explanation by the trained LLM, at least in part based on a profile of a second seeker and a profile of a second target.

2. The method of claim 1, further comprising:training the trained LLM, at least in part based on the profile of the second seeker, the profile of the second target, and the second explanation, if a determination determines the second explanation is not satisfactory.

3. The method of claim 1 or claim 2, further comprising:receiving a third explanation from the trained LLM, at least in part based on a profile of a third seeker and a profile of a third target;receiving a revised explanation, at least in part based on the profile of the third seeker and the profile of the third target; andtraining the trained LLM, at least in part based on the profile of the third seeker, the profile of the third target, the third explanation, and the revised explanation, to produce an intermediate LLM.

4. The method of claim 3, further comprising:producing a fourth explanation by the intermediate LLM, at least in part based on a profile of a fourth seeker and a profile of a fourth target; andAttorney Docket8463-010232training the intermediate LLM, at least in part based on the profile of the fourth seeker, the profile of the fourth target, the fourth explanation, and a revised explanation between the profile of the fourth seeker and the profile of the fourth target, if a determination determines the intermediate LLM is not satisfactory, at least in part based on the fourth explanation.

5. The method of any preceding claim, further comprising:receiving a profile of a training seeker and a profile of a training target; producing a key connection point by a prompting LLM, at least in part based on the profile of the training seeker and the profile of the training target; andtraining the prompting LLM, at least in part based on the profile of the training seeker, the profile of the training target, and the key connection point, to provide the LLM.

6. The method of claim 5, further comprising:producing a prompting explanation by the prompting LLM, at least in part based on the profile of the training seeker and the profile of the training target; and training the prompting LLM, at least in part based on the prompting explanation, to provide the LLM.

7. The method of any preceding claim, further comprising:receiving a profile of a fifth seeker;receiving a profile of a fifth target;producing a fifth explanation at least in part based on the trained LLM, the profile of the fifth seeker, and the profile of the fifth target; andcausing a display of the profile of the fifth target and the fifth explanation on a computing device of the fifth seeker.

8. An apparatus, comprising:Attorney Docket8463-010233a memory that stores a first explanation for a profile of a first seeker on an online dating service and that stores a profile of a first target on the on-line dating service; anda processor configured to train a large language model (LLM), at least in part based on the profile of the first seeker, the profile of the first target, and the first explanation, to produce a trained LLM, whereinthe trained LLM produces a second explanation, at least in part based on a profile of a second seeker on the on-line dating service and a profile of a second target on the on-line dating service.

9. The apparatus of claim 8, wherein the processor is further configured to train the trained LLM, at least in part based on the profile of the second seeker, the profile of the second target, and the second explanation, if a determination determines the second explanation is not satisfactory.

10. The apparatus of claim 8 or claim 9, whereinthe trained LLM produces a third explanation, at least in part based on a profile of a third seeker and a profile of a third target,the processor receives a revised explanation, at least in part based on the profile of the third seeker and the profile of the third target, andthe trained LLM is trained, at least in part based on the profile of the third seeker, the profile of the third target, the third explanation, and the revised explanation, to produce an intermediate LLM.

11. The apparatus of claim 10, whereinthe intermediate LLM produces a fourth explanation, at least in part based on a profile of a fourth seeker and a profile of a fourth target, andthe intermediate LLM is trained, at least in part based on the profile of the fourth seeker, the profile of the fourth target, the fourth explanation, and a revised explanation between the profile of the fourth seeker and the profile of the fourth target, if aAttorney Docket8463-010234determination determines the intermediate LLM is not satisfactory, at least in part based on the fourth explanation.

12. The apparatus of any of claims 8-11, whereinthe processor receives a profile of a training seeker and a profile of a training target,a prompting LLM produces a key connection point, at least in part based on the profile of the training seeker and the profile of the training target, andthe prompting LLM is trained, at least in part based on the profile of the training seeker, the profile of the training target, and the key connection point, to provide the LLM.

13. The apparatus of claim 12, whereinthe prompting LLM produces a prompting explanation, at least in part based on the profile of the training seeker and the profile of the training target, andthe prompting LLM is trained, at least in part based on the prompting explanation, to provide the LLM.

14. The apparatus of any of claims 8-13, whereinthe processor receives a profile of a fifth seeker and a profile of a fifth target, a fifth explanation is produced, at least in part based on the trained LLM, the profile of the fifth seeker, and the profile of the fifth target, andthe processor causes a display of the profile of the fifth target and the fifth explanation to the fifth seeker.

15. A computer-readable medium encoded with instructions that, when executed by a processor of a computing device, cause the computing device to perform operations comprising:receiving a first explanation for a profile of a first seeker of an on-line dating service and a profile of a first target of the on-line dating service;Attorney Docket8463-010235training the LLM, at least in part based on the profile of the first seeker, the profile of the first target, and the first explanation, to produce a trained LLM; and receiving a second explanation by the trained LLM, at least in part based on a profile of a second seeker of the on-line dating service and a profile of a second target of the on-line dating service.

16. The medium of claim 15, the operations further comprising:training the trained LLM, at least in part based on the profile of the second seeker, the profile of the second target, and the second explanation, if a determination determines the second explanation is not satisfactory.

17. The medium of claim 15 or claim 16, the operations further comprising: receiving a third explanation from the trained LLM, at least in part based on a profile of a third seeker of the on-line dating service and a profile of a third target of the on-line dating service;receiving a revised explanation, at least in part based on the profile of the third seeker and the profile of the third target; andtraining the trained LLM, at least in part based on the profile of the third seeker, the profile of the third target, the third explanation, and the revised explanation, to produce an intermediate LLM.

18. The medium of claim 17, the operations further comprising:receiving a fourth explanation by the intermediate LLM, at least in part based on a profile of a fourth seeker and a profile of a fourth target; andtraining the intermediate LLM, at least in part based on the profile of the fourth seeker, the profile of the fourth target, the fourth explanation, and a revised explanation between the profile of the fourth seeker and the profile of the fourth target, if a determination determines the intermediate LLM is not satisfactory, at least in part based on the fourth explanation.Attorney Docket8463-01023619- The medium of any of claims 15-18, the operations further comprising: receiving a profile of a training seeker and a profile of a training target; producing a key connection point by a prompting LLM, at least in part based on the profile of the training seeker and the profile of the training target;producing a prompting explanation by the prompting LLM, at least in part based on the profile of the training seeker and the profile of the training target; and training the prompting LLM, at least in part based on the profile of the training seeker, the profile of the training target, the key connection point, and the prompting explanation, to provide the LLM.

20. The medium of any of claims 15-20, the operations further comprising: receiving a profile of a fifth seeker;receiving a profile of a fifth target;producing a fifth explanation at least in part based on the trained LLM, the profile of the fifth seeker, and the profile of the fifth target; andcausing a display of the profile of the fifth target and the fifth explanation on a computing device of the fifth seeker.

Images