JP7641032B2 - Generative AI management system, generative AI management method, and generative AI management program - Google Patents

Description

The present invention relates to a generation AI management system, a generation AI management method, and a generation AI management program, and in particular to a generation AI management system, a generation AI management method, and a generation AI management program that can change a generation AI to another generation AI during use.

Currently, generative artificial intelligence (generative AI) is making revolutionary advances, processing huge amounts of data and enabling advanced pattern recognition and prediction, and it is expected to be used in a variety of fields, including business, science, medicine, and automobiles. For example, Patent Document 1 discloses technology related to generative AI, and claims that the technology disclosed in this document can use a language model to generate information composed of text appropriate to the situation.

However, when using generative AI, which is used as a general-purpose technology, new generative AI is constantly being released or updated by various organizations and companies, and it is currently unclear which generative AI should be used.

The generation AI takes past conversational experience into account when generating new answers, but when changing from the generation AI currently in use to another generation AI midway, it can be difficult to have the new generation AI inherit the conversational experience that had been exchanged between the generation AI up to that point.

JP 2023-043129 A

The present invention aims to provide a generation AI management system, a generation AI management method, and a generation AI management program that, when changing a generation AI currently in use to another generation AI midway, makes it easy to have the changed generation AI inherit the conversational experience exchanged between the generation AI currently in use.

That is, the generation AI management system according to the first aspect is a generation AI management system that connects to a server of multiple generation AIs and provides a user terminal with a function for utilizing the multiple generation AIs, and is characterized by comprising: a recording control unit that records the multiple generation AIs; an input information acquisition unit that acquires information to be provided to one of the multiple generation AIs as input information via the user terminal; an answer information acquisition unit that acquires an answer of one generation AI corresponding to the input information from the server as answer information; a dialogue data storage unit that associates the input information with the answer information corresponding to the input information and stores it as dialogue data; and a learning unit that, when changing one generation AI to another generation AI via the user terminal, uses dialogue data stored in the dialogue data storage unit, which is obtained by communicating with one generation AI, to train the other generation AI.

In a second aspect, in the generation AI management system according to the first aspect, the generation AI management system further includes an input information sending unit that sends input information to a server, and a vector data conversion unit that converts text data included in the input information acquired by the input information acquisition unit into vector data, and the input information sending unit may send the input information including the vector data to the server after it has been converted from text data into vector data.

In a third aspect, in the generative AI management system according to the second aspect, the vector data conversion unit may convert text data included in the answer information acquired by the answer information acquisition unit into vector data, and the dialogue data storage unit may associate input information including the vector data with answer information that corresponds to the input information and includes the vector data, and store the input information as dialogue data.

In a fourth aspect, in the generation AI management system according to the second aspect, the generation AI management system further includes a learning data acquisition unit that acquires learning data used for training the generation AI from a user terminal, a learning data transmission unit that transmits the learning data acquired by the learning data acquisition unit to a server, and a learning data storage unit that stores the learning data transmitted to the server by the learning data transmission unit, and when changing one generation AI to another generation AI via the user terminal, the learning unit may train the other generation AI using the learning data stored in the learning data storage unit and used for training the one generation AI.

In a fifth aspect, in the generation AI management system according to the fourth aspect, the generation AI management system may further include a vector data conversion unit that converts text data included in the learning data acquired by the learning data acquisition unit into vector data, and the learning data storage unit may store the learning data including the vector data after it has been converted from text data into vector data.

In a sixth aspect, in the generation AI management system according to the fourth aspect, there are multiple generation AIs, and the generation AI management system is connectable to the APIs of the multiple generation AIs, and may further include a display control unit that assigns an avatar to each of the APIs of the generation AI and displays a list of the avatars assigned to each API on the display unit of the user terminal, thereby displaying a list of the APIs on the display unit.

In a seventh aspect, in the generation AI management system according to the sixth aspect, the display control unit may display a list of the learning data stored in the learning data storage unit and used to learn the generation AI on the display unit of the user terminal, thereby displaying the learning history of the generation AI on the display unit.

In an eighth aspect, in the generation AI management system according to the first aspect, when constructing a clone of a machine learning model configured by the generation AI, the learning unit may use dialogue data stored in the dialogue data storage unit, obtained by communicating with the generation AI that configures the machine learning model that is the source of the clone, for learning the generation AI that configures the clone.

In a ninth aspect, in the generation AI management system according to the fourth aspect, when constructing a clone of a machine learning model configured by the generation AI, the learning unit may use, for learning the generation AI that constitutes the clone, learning data stored in the learning data storage unit, which learning data was used for learning the generation AI that constitutes the machine learning model that is the source of the clone.

In a tenth aspect, in the generation AI management system according to the sixth aspect, there are multiple generation AIs, the generation AI management system is connectable to APIs of the multiple generation AIs, the input information sending unit sends the same input information to multiple servers, the answer information acquisition unit acquires answer information corresponding to the same input information from the multiple servers, and the display control unit may display a list of the multiple answer information acquired by the answer information acquisition unit on the display unit of the user terminal.

In an eleventh aspect, in the generation AI management system according to the second aspect, there are multiple generation AIs, and the generation AI management system is connectable to APIs of the multiple generation AIs, and the input information sending unit sends input information for obtaining an answer regarding the evaluation of the answer information acquired by the answer information acquisition unit from one server to another server different from the one server from which the answer information was acquired, and the answer information acquisition unit may acquire an answer from the other server regarding the evaluation of the answer from the one server as second opinion information.

In a twelfth aspect, in the generation AI management system according to the sixth aspect, the generation AI management system further includes a token number acquisition unit that acquires from the server the number of tokens generated by a tokenizer mounted on the server based on input information as token number information, and a fee calculation unit that calculates the usage fee for the generation AI calculated based on the token number information acquired by the token number acquisition unit as fee information, and the display control unit may display the fee information calculated by the fee calculation unit on the display unit of the user terminal.

In a thirteenth aspect, in the generation AI management system according to the sixth aspect, there are a plurality of generation AIs, and the generation AI management system is connectable to the APIs of the plurality of generation AIs, and further includes a token number acquisition unit that acquires the number of tokens generated by the tokenizer mounted on the plurality of servers that acquired the input information as token number information from each of the plurality of servers, a fee calculation unit that calculates the usage fee of the generation AI calculated based on the token number information acquired by the token number acquisition unit as fee information for each of the plurality of generation AIs, and a minimum fee extraction unit that extracts the generation AI associated with the cheapest fee information based on the fee information for each of the plurality of generation AIs calculated by the fee calculation unit, and the display control unit may display the generation AI extracted by the minimum fee extraction unit on the display unit of the user terminal.

The generation AI management method according to the fourteenth aspect is a generation AI management method used in a generation AI management system that connects to a server of multiple generation AIs and provides a user terminal with a function for using the multiple generation AIs, and the generation AI management system is characterized by executing a recording control step of recording the multiple generation AIs, an input information acquisition step of acquiring information to be given to one of the multiple generation AIs as input information via the user terminal, an answer information acquisition step of acquiring an answer of one generation AI corresponding to the input information from the server as answer information, a dialogue data storage step of associating the input information with the answer information corresponding to the input information and storing them as dialogue data, and a learning step of learning the other generation AI using the dialogue data stored in the dialogue data storage step and obtained by communicating with one generation AI via the user terminal when changing one generation AI to another generation AI.

The generation AI management program according to the fifteenth aspect is a generation AI management program used in a generation AI management system that connects to a server of multiple generation AIs and provides a user terminal with a function for utilizing multiple generation AIs, and is characterized by realizing the following in the generation AI management system: a recording control function that records multiple generation AIs; an input information acquisition function that acquires information to be given to one of the multiple generation AIs as input information via the user terminal; an answer information acquisition function that acquires an answer from one generation AI corresponding to the input information from the server as answer information; a dialogue data storage function that associates the input information with the answer information corresponding to the input information and stores it as dialogue data; and a learning function that, when changing one generation AI to another generation AI via the user terminal, uses the dialogue data stored in the dialogue data storage function, which is obtained by communicating with one generation AI, to have the other generation AI learn.

The generation AI management system according to the present invention is a generation AI management system that connects to a server of multiple generation AIs and provides a user terminal with a function for using multiple generation AIs, and includes a recording control unit that records multiple generation AIs, an input information acquisition unit that acquires information to be given to one of the multiple generation AIs as input information via the user terminal, an answer information acquisition unit that acquires the answer of one generation AI corresponding to the input information from the server as answer information, a dialogue data storage unit that associates the input information with the answer information corresponding to the input information and stores it as dialogue data, and a learning unit that, when changing one generation AI to another generation AI via the user terminal, uses the dialogue data stored in the dialogue data storage unit and obtained by communicating with one generation AI to train the other generation AI, so that when changing the generation AI in use to another generation AI midway, it becomes easy to have the changed generation AI inherit the dialogue experience exchanged with the generation AI in use.

FIG. 1 is a conceptual diagram showing the usage environment of a generation AI management system according to this embodiment. FIG. 2 is a block diagram showing an example of the hardware configuration of the generation AI management system and the generation AI server according to this embodiment. Figure 3 is a block diagram showing an example of the software configuration of the generation AI server in this embodiment. FIG. 4 is a block diagram showing an example of the software configuration of the generation AI management system according to this embodiment. FIG. 5 is a diagram showing an example of a display screen of a display unit of a user terminal according to this embodiment. FIG. 6 is a diagram showing an example of a display screen of the display unit of the user terminal according to this embodiment. FIG. 7 is a flowchart of the generation AI management program according to this embodiment.

A generation AI management system 10 according to an embodiment of the present invention will be described with reference to FIGS. 1 to 6. FIG.
First, the usage environment of the generation AI management system 10 according to the present embodiment will be described with reference to Fig. 1. Fig. 1 is a conceptual diagram showing the usage environment of the generation AI management system 10 according to the present embodiment.

The generative AI management system 10 according to this embodiment is connected to a server (generative AI server) 30 of multiple generative AIs (generative artificial intelligence) 33, and provides a user terminal 20 with the functionality to utilize the multiple generative AIs 33 (see FIG. 3).

Generative AI 33 is a machine learning technique that automatically generates something from sample data provided by prompts, etc., and a machine learning model trained by this machine learning technique. The something is, for example, text, images, music, etc. The main difference between generative AI and conventional AI is that conventional AI extracts features of sample data by providing a large amount of sample data and makes inferences, while generative AI generates something new from a small amount of sample data.
The generation AI 33 is, for example, ChatGPT (registered trademark) 3.5, ChatGPT 4.0, and bard (registered trademark), which are built based on a large-scale language model (LLM). The large-scale language model is a machine learning model of natural language processing (NLP) trained using a large amount of text data. The large-scale language model can be adapted to natural language processing tasks such as text classification, sentiment analysis, information extraction, text summarization, text generation, and question answering by fine-tuning after pre-learning using a large amount of text data.

The use of the functions of the generation AI 33 by external applications is performed in accordance with the API (Application Programming Interface) of the generation AI 33. The API is a mechanism for applications to exchange information with each other, a protocol for connecting applications with each other, and an agreement on the specifications of the connection part. The generation AI server 30 (30a, 30b, 30c) is a server equipped with the generation AI 33 and provides the API of the generation AI 33, and implements and operates the API so that others who connect to the API can use the functions of the generation AI 33. The generation AI management system 10 uses the functions of the generation AI 33 by connecting to the generation AI server 30 and calling the API provided by the generation AI server 30.

The generating AI management system 10 is connected to the generating AI server 30 via a network 11 including the Internet. In FIG. 1, for convenience, the generating AI management system 10 is shown as being connected to three generating AI servers 30 (a first generating AI server 30a, a second generating AI server 30b, and a third generating AI server 30c), but the number of servers is not limited to three, and the system may be connected to two or four or more generating AI servers 30.

The generation AI management system 10 provides the functions of the generation AI 33 to user terminals 20 (20a, 20b) connected via a network 11. In FIG. 1, for convenience, the generation AI management system 10 shows an example of connection to two user terminals 20 (a first user terminal 20a and a second user terminal 20b), but this is not limited to two, and one or three or more user terminals 20 may be connected to the generation AI management system 10.

The user terminal 20 is, for example, an information processing terminal such as a personal computer (hereinafter referred to as a PC), a notebook PC, a tablet PC, or a smartphone. A web browser runs on the user terminal 20, and the user terminal 20 can connect to the generation AI management system 10 through the web browser.

The generation AI management system 10 functions as a web server, and provides a display screen (web page) 61 configured using HTML (Hyper Text Markup Language) or the like to the display unit (monitor) 60 of the user terminal 20 connected via the network 11 (see Figures 5 and 6). The generation AI management system 10 is configured, for example, using a general-purpose computer such as a workstation or PC. Furthermore, the generation AI management system 10 is not limited to being configured in an individual, independent form, but may be configured as a module that is part of another system, or may be configured using cloud computing.

Next, referring to FIG. 2, an example of the hardware configuration of the generation AI management system 10 and the generation AI server 30 will be described. FIG. 2 is a block diagram showing an example of the hardware configuration of the generation AI management system 10 and the generation AI server 30 according to this embodiment.

The generation AI management system 10 and the generation AI server 30 are equipped with a communication interface 10a, a ROM (Read Only Memory) 10b, a RAM (Random Access Memory) 10c, a storage unit 10d, a CPU (Central Processing Unit) 10e, and an input/output interface 10f. Furthermore, a dedicated computing element (processor) specialized for the generation AI may be implemented in place of or together with the CPU 10e.

The communication interface 10a has a function for transmitting and receiving data of the generation AI management system 10 and the generation AI server 30 between other devices via a network 11 including the Internet. The data of the generation AI management system 10 and the generation AI server 30 includes learning data, input information, response information, token number information, etc., which will be described later.

The memory unit 10d can be used as a storage device for the generation AI management system 10 and the generation AI server 30, and is composed of, for example, a hard disk drive, a solid state drive, and a flash memory, and the memory unit 10d can also be composed of cloud storage. The memory unit 10d may also be composed of a database. By configuring the memory unit 10d as a database, it is possible to efficiently manage huge amounts of data such as learning data, input information, and answer information, and it is possible to quickly refer to the required information with excellent searchability and accessibility.

In addition, the memory unit 10d of the generation AI management system 10 stores the generation AI management program described below, the OS (Operating System) required for the generation AI management system 10 to operate, various applications, and various data used by the applications.

The memory unit 10d of the generation AI server 30 stores the OS, various applications, and various data used by the applications required for the generation AI server 30 to operate.
The OS is a type of software built into the generation AI management system 10 and the generation AI server 30, and has the function of managing the basic control of the generation AI management system 10 and the generation AI server 30.

The generation AI management system 10 stores the generation AI management program described below in the ROM 10b or the memory unit 10d, and imports the generation AI management program into a main memory such as the RAM 10c. The CPU 10e accesses the main memory into which the generation AI management program has been imported, and executes the generation AI management program.

The generation AI server 30 stores the OS in the ROM 10b or the memory unit 10d, and imports the OS into a main memory such as the RAM 10c. The CPU 10e accesses the main memory into which the OS has been imported and executes the OS.

The input/output interface 10f transmits and receives data to and from external devices of the generation AI management system 10 and the generation AI server 30. The external devices are the input device 10g and the output device 10h, which input and output data to and from the generation AI management system 10 and the generation AI server 30. The input device 10g is, for example, a keyboard and a mouse, and the output device 10h is, for example, a monitor, a printer, and a speaker.

Next, an example of the software configuration of the generation AI server 30 will be described with reference to Fig. 3. Fig. 3 is a block diagram showing an example of the software configuration of the generation AI server 30.
By executing the OS, the generation AI server 30 has functional units such as an input information receiving unit 31, a tokenizer 32, a generation AI 33, and an answer information sending unit 34 on the CPU 10e.

The input information receiving unit 31 receives input information such as a prompt for the generation AI 33 from the generation AI management system 10.
The input information to the generation AI 33 is information given to the generation AI 33 to generate something. For example, when the generation AI 33 is made to generate a sentence, an image, music, etc., text, keywords, images, and voices are given as input information. Text is used when generating sentences, dialogues, program codes, etc. Text is analyzed by natural language processing (NLP) and understood by the generation AI 33. Keywords are used when generating images, music, etc. Keywords indicate the content and features that you want the generation AI 33 to generate. For example, when the keyword "cat" is input, the generation AI 33 generates an image of a cat. The image is used when processing the image or generating an image different from the image. The image is analyzed by computer vision and understood by the generation AI 33. The voice is used when processing the voice or generating a voice different from the voice. The voice is analyzed by speech processing and understood by the generation AI 33.

The tokenizer 32 provides a function of dividing the text data included in the input information accepted by the input information accepting unit 31 into tokens and assigning IDs to the tokens. A token is the smallest unit for dividing text data, and includes words and phrases. The tokenizer 32 uses, for example, Sentence Piece, BPE (Byte Pair Encoding), BBPE (Byte-level BPE), Word Piece, Unigram Language Model, ByT5, etc.

The generation AI 33 takes in the input information as an ID converted by the tokenizer 32, and generates a response to the input information by performing various processes such as natural language processing, image processing, and voice processing.

The answer information sending unit 34 sends the answer (completion) to the input information generated by the generation AI 33 to the generation AI management system 10 as answer information.

Next, referring to FIG. 4, an example of the software configuration of the generation AI management system 10 will be described. FIG. 4 is a block diagram showing an example of the software configuration of the generation AI management system 10 according to this embodiment. By executing the generation AI management program described below, the generation AI management system 10 has functional units such as a recording control unit 40, a learning data acquisition unit 41, a learning data transmission unit 42, a learning data storage unit 43, an input information acquisition unit 44, a vector data conversion unit 45, an input information transmission unit 46, a response information acquisition unit 47, a dialogue data storage unit 48, a learning unit 49, a token number acquisition unit 50, a fee calculation unit 51, a display control unit 52, and a minimum fee extraction unit 53 in the CPU 10e.

The recording control unit 40 records a plurality of generated AIs.
The recording control unit 40 records, for each of the multiple generated AIs, a uniform resource identifier (URI) for connecting to the generated AI's server, information for identifying the generated AI such as an API key, and attribute information of the generated AI such as the creation date, last use date, and the organization or company providing the API in the memory unit 10d of the generated AI management system 10.
The learning data acquisition unit 41 acquires learning data used for training the generation AI 33 from the user terminal 20.
The learning data mainly includes data for reference when the generation AI 33 generates an answer, and a data set for training the machine learning model constructed by the generation AI 33. The former data includes text data, images, videos, audio, music, program code, etc. The latter data set includes learning data that is a set of input data and correct answer data (label) used in supervised learning, and data that is only input data used in unsupervised learning. Supervised learning is a learning method in which an answer generated by a machine learning model configured by the generation AI 33 is adjusted so as to approach correct answer data. Unsupervised learning is a learning method in which, for example, a large amount of text data is given as learning data to enable understanding of the grammar and context of the text.
The learning data includes dialogue data of other generation AIs. That is, the generation AI 33 may learn by reading input information such as prompts for other generation AIs 33 and answer information that responds to the input information.

The learning data transmission unit 42 transmits the learning data acquired by the learning data acquisition unit 41 to the generation AI server 30.
The learning data sent to the generation AI server 30 is used for learning the generation AI 33.
The generation AI 33 uses a generative adversarial network (GAN), GPT-3 (registered trademark), GPT-4, etc. to extract features and patterns from the given training data, digitize them, and use them as parameters of the internal neural network.

The learning data memory unit 43 stores the learning data transmitted to the generation AI server 30 by the learning data transmission unit 42.
The learning data is stored in the memory unit 10d of the generation AI management system 10 by the learning data memory unit 43.
The generation AI management system 10 also has a vector data conversion unit 45.
The vector data conversion unit converts the text data included in the learning data acquired by the learning data acquisition unit 41 into vector data.
The learning data storage unit 43 may store learning data including vector data after the text data is converted into vector data.
By converting the learning data of the generation AI 33 into vector data, when the learning data is used to learn a new generation AI 33 to be used by the learning unit 49 described below, the learned generation AI 33 can achieve effects such as expanding the range of application of natural language processing and improving accuracy. Vector data in natural language processing is data in which words and sentences in natural language are converted into text data, and the words and sentences are further quantified to make them easier for a computer to process. By converting words and sentences into vector data, it is possible to visualize the similarity in meaning of words and make their features apparent.

The input information acquisition unit 44 acquires, via the user terminal 20, information to be provided to one of the multiple generation AIs 33 as input information.
The user sends input information to the generation AI management system 10 from a web page (display screen 61) of the generation AI management system 10 displayed on the display unit 60 of the user terminal 20. The input information is a prompt for causing the generation AI 33 to generate something, and includes instructions and inputs to the generation AI 33. For example, when causing the generation AI 33 to generate a poem, a story, an image, etc., the content, format, and style (form), etc. are specified by the prompt. The input information may be created in the file format of a specific application, such as pdf (registered trademark), MS-Word (registered information), and MS-Excel (registered trademark).

The vector data conversion unit 45 converts the text data included in the input information acquired by the input information acquisition unit 44 into vector data.
There is a limit to the total number of tokens used in one exchange with the generation AI 33, i.e., one prompt and the content (completion) of the answer to the prompt. For example, GPT-3 has a limit of 4097 tokens. The vector data conversion unit 45 converts text data included in the prompt, which is input information, into vector data, thereby almost eliminating the above-mentioned limit on the total number of tokens. Methods for converting text data into vector data include, for example, bag-of-words, TF-IDF, and word embedding, and the vector data conversion unit 45 is configured using a machine learning model or application (algorithm) that realizes these methods.
The conversion of text data contained in the input information into vector data is performed as pre-processing of a processing task executed by the input information transmitting unit 46 described below.

The input information sending unit 46 sends the input information to the generation AI server 30.
The input information sending unit 46 may be configured to send the input information including the vector data to the generation AI server 30 after the input information has been converted from text data to vector data.

The answer information acquisition unit 47 acquires an answer of one generation AI 33 corresponding to the input information from the generation AI server 30 as answer information.
The generation AI 33 predicts the next appropriate vector by inputting the input vector data into a machine learning model (language model). The generation AI 33 converts the predicted vector back into tokens and generates answer text by connecting the tokens.

The dialogue data storage unit 48 associates the input information with the response information corresponding to the input information and stores the information as dialogue data.
The dialogue data memory unit 48 associates the input information transmitted from the input information transmitting unit 46 with the answer information corresponding to the input information and acquired by the answer information acquisition unit 47, and stores them as dialogue data in the memory unit 10d of the generation AI management system 10.
The vector data conversion unit 45 converts the text data included in the answer information acquired by the answer information acquisition unit 47 into vector data, and the dialogue data storage unit 48 may associate input information including vector data with answer information that corresponds to the input information and includes vector data and store them as dialogue data.
By storing the dialogue data as vector data, when the dialogue data is used to train a new generation AI 33 to be used by the learning unit 49 described below, the trained generation AI 33 can achieve effects such as expanding the range of application of natural language processing and improving accuracy.
The dialogue data is data that associates input information to the generation AI with answer information that the generation AI generates as an answer in response to the input information.
For example, when text data is input to the generation AI as input information, the generation AI generates answer information corresponding to the input information (text data) using text data, voice data, image data, etc. The combination of this input information and this answer information becomes dialogue data. Note that the input information to the generation AI may be voice data and image data in addition to text data, or may be a combination of text data, voice data, and image data. The generation AI may generate answer information using text data, voice data, and image data, or a combination of these, as an answer corresponding to the input information.

When changing one generation AI 33 to another generation AI 33 via the user terminal 20, the learning unit 49 learns the other generation AI 33 using the learning data stored in the learning data memory unit 43 and used to learn the one generation AI 33.
For example, when changing the generated AI 33 of the first generating AI server 30a used by the user terminal 20 to the generated AI 33 of the second generating AI server 30b, the generated AI 33 of the newly used second generating AI server 30b is trained using the learning data stored in the learning data memory unit 43, which is the learning data used to train the generated AI 33 of the first generating AI server 30a and which is stored in the memory unit 10d of the generating AI management system 10.
By using the learning data used to train one generation AI 33 to train another generation AI 33, the learning (learning history) experienced by the generation AI (one generation AI 33) that has been used until now can be passed on to the newly used generation AI (other generation AI 33).

For example, suppose that the generated AI 33 of the second generation AI server 30b is GPT-4, and the generated AI 33 of the first generation AI server 30a is GPT-3. By training GPT-3 and GPT-4 with the same learning model, the things generated by both are based on the same learning data, so they will have similar content, format, and style. In addition, both are generated by reflecting the biases and errors contained in the learning model. Furthermore, both are not generated based on information and knowledge that is not in the learning data.

On the other hand, GPT-4 outperforms GPT-3, so its products are more natural, more accurate, more diverse and creative, and produce higher quality results with less prompting input information than GPT-3 products.

When changing one generation AI 33 to another generation AI 33 via the user terminal 20, the learning unit 49 may use the dialogue data stored in the dialogue data storage unit 48, which is obtained by communicating with the one generation AI 33, to train the other generation AI 33.

The generation AI 33 generates a new answer taking into account the history (experience) of past dialogues. When changing the generation AI 33 used by the user terminal 20, the new generation AI 33 can learn using dialogue data obtained through communication with the generation AI 33, so that the new generation AI 33 can inherit the history (experience) of past dialogues.
By learning the newly used generation AI (other generation AI 33) using the dialogue data obtained from communication with the previously used generation AI (one generation AI 33), i.e., input information such as the query content (prompt) for the generation AI 33 and the answer information corresponding to the input information, the newly used generation AI (other generation AI 33) can inherit the content of communication (input information and corresponding answer information) between the newly used generation AI (one generation AI 33) and the previously used generation AI, the newly used generation AI (other generation AI 33) can generate new answer information by taking into account the past dialogue history of the previously used generation AI (one generation AI 33).

As described above, when changing the generation AI server 30 (generation AI 33) used by the user terminal 29, the learning data used to train the original generation AI 33 and the dialogue data obtained by communicating with the original generation AI 33 are stored in the memory unit 10d of the generation AI management system 10, so that this learning data and dialogue data can be easily used to train the newly used generation AI 33.

When generating a clone of the machine learning model constructed by the generation AI 33, the learning unit 49 uses the learning data stored in the learning data storage unit 43 to train the generation AI 33 that constitutes the clone, and the learning data used to train the generation AI 33 that constitutes the machine learning model that is the source of the clone.

When generating a clone of a machine learning model constructed by the generation AI 33, the learning unit 49 may use dialogue data stored in the dialogue data storage unit 48 to learn the generation AI 33 that constitutes the clone, and the dialogue data obtained by communicating with the generation AI 33 that constitutes the machine learning model that is the basis of the clone.
A clone of a machine learning model is a machine learning model that is composed of the same type of large-scale language model as the original machine learning model and is trained using the same training data as the original machine learning model, and is a copy (duplicate) of the original machine learning model.

As described above, when generating a clone of a machine learning model constructed by the generation AI 33, the learning data used to train the original generation AI 33 and the dialogue data obtained by communicating with the original generation AI 33 are stored in the memory unit 10d of the generation AI management system 10, so that this learning data and dialogue data can be easily used to train the generation AI 33 that constitutes the clone.

There are multiple generation AIs 33, and the generation AI management system 10 can be connected to the APIs of the multiple generation AIs 33.
The display control unit 52 assigns an avatar to each API of the generated AI 33, and displays a list of the avatars assigned to each API on the display unit 60 of the user terminal 20, thereby displaying a list of APIs on the display unit 60 (see Figures 5 and 6).
The avatar in this embodiment embodies the subject of question answering, which is one of the functions of the generation AI 33. For example, an avatar represented by an image such as a photo or illustration of a robot, animal, anime character, mark, fruit, etc. is displayed to the user as if it is answering a question. The avatar may be displayed as an icon or as an animation with movement. By assigning an avatar to the subject of question answering, which is a function of the generation AI 33, the user of the generation AI 33 can distinguish between multiple generation AIs 33, and further, by embodying the avatar, the user can have feelings such as attachment, trust, and favor toward the avatar, and can use the generation AI 33 with enjoyment.
The display control unit 52 displays a list of learning data used to learn the generated AI 33 on the display unit 60 of the user terminal 20, thereby displaying the learning history of the generated AI 33 on the display unit 60 (see Figure 6).

An example of the display screen 61 of the display unit 60 of the user terminal 20 according to this embodiment will be described with reference to Figures 5 and 6. Figures 5 and 6 are diagrams showing an example of the display screen 61 of the display unit 60 of the user terminal 20 according to this embodiment.

As shown in Figures 5 and 6, a menu bar 62 is displayed on the left side of the display screen 61, and a main display section 63 is displayed on the right side of the display screen 61. New registration, login, main menu, etc. are displayed on the menu bar 62, and when selected by the user's selection operation, a page corresponding to new registration, login, main menu, etc. is displayed on the main display section 63.

A sub-menu bar 64 is displayed on the left edge of the main display section 63. The sub-menu bar 64 is used to select the display contents of the sub-display section 65. The sub-menu bar 64 displays buttons for avatars, fees, and settings, and when selected by the user's selection operation, pages corresponding to avatars, fees, and settings are displayed on the sub-display section 65.

In the sub-display unit 65 shown in Figures 5 and 6, the avatar button is selected in the sub-menu bar 64, and an avatar search input window 66, a tag list 67, and an avatar list 68 are displayed in the sub-display unit 65. The avatar button in the sub-menu bar 64 is surrounded by a thick border, indicating that it has been selected by the user. The avatar list 68 represents a list of APIs of the generation AI server 30 (generation AI 33) registered in the generation AI management system 10.

Information about an avatar selected by the user from the avatar list 68 in the sub-display unit 65 is displayed in the main display unit 63. The content to be displayed in the main display unit 63 is selected from the headings 69, and the dialogue history is selected in FIG. 5. "Avatar 001" is selected from the avatar list 68, and a dialogue list 70 between "Avatar 001" and the user is displayed in the main display unit 63. The button for "Avatar 001" in the avatar list 68 in the sub-display unit 65 is surrounded by a thick border, indicating that it has been selected by the user. An avatar creation button 74 is displayed at the bottom of the sub-display unit 65.

In the dialogue list 70, an avatar icon 72 is displayed to the left of the answer generated by avatar 001, and a user icon 73 is displayed to the left of input information such as the user's inquiry, making it possible to identify who wrote what. A message input window 71 is displayed at the bottom of the main display section 63.

In FIG. 6, an API is selected from the headings 69, and a list 75 of APIs 76 and an API creation button 77 are displayed in the main display section 63. The API 76 displays the name of the API, the creation date, the last use date, and the API key.

Based on the input information, the token number acquisition unit 50 acquires the number of tokens generated by the tokenizer 32 mounted on the generation AI server 30 as token number information from the generation AI server 30.
The fee calculation unit 51 calculates the usage fee for the generation AI 33 (generation AI server 30) based on the token number information acquired by the token number acquisition unit 50 as fee information.
The display control unit 52 displays the fee information calculated by the fee calculation unit 51 on the display unit 60 of the user terminal 20 .

According to the above-mentioned token number acquisition unit 50 and fee calculation unit 51, if the number of tokens contained in the text string included in the input information can be determined from the token number information, it can be determined whether the text string included in the input information is too long to be processed, and the fee information can be determined how much it will cost to use the generation AI server 30.

There are multiple generation AIs 33, and the generation AI management system 10 is capable of connecting to the APIs of multiple generation AIs 33, the input information sending unit 46 sends the same input information to multiple servers 30, the answer information acquisition unit 47 acquires answer information for the same input information from the multiple servers 30, and the display control unit 52 may display a list of the multiple answer information acquired by the answer information acquisition unit 47 on the display unit 60 of the user terminal 20.
In this case, the user can obtain answers to the same input information from multiple generation AIs 33, and can compare the multiple answers to obtain a better answer.

There are multiple generation AIs 33, and the generation AI management system 10 is capable of connecting to the APIs of multiple generation AIs 33. The input information sending unit 46 sends input information for obtaining an answer regarding the evaluation of the answer information acquired by the answer information acquisition unit 47 from the generation AI server 30 (generation AI 33) to another generation AI server 30, which is a generation AI server 30 other than the generation AI server 30 that acquired the answer information, and the answer information acquisition unit 47 can acquire the answer of the other generation AI server 30 (generation AI 33) regarding the evaluation of the answer of the generation AI server 30 (generation AI 33) as second opinion information.
In this case, the user can have another generation AI 33 evaluate the answer obtained from the generation AI 33 to obtain second opinion information that serves as a second opinion, which is useful when the answer from the generation AI 33 is incorrect and a better answer is desired.

There are multiple generation AIs 33, and the generation AI management system 10 can connect to the APIs of the multiple generation AIs 33. The token number acquisition unit 50 acquires the number of tokens generated by the tokenizers 32 installed in the multiple servers 30 that acquired the input information as token number information from each of the multiple servers 30. Next, the fee calculation unit 51 calculates the usage fee for the generation AI 33 calculated based on the token number information acquired by the token number acquisition unit 50 as fee information for each of the multiple generation AIs 33.

The minimum fee extraction unit 53 extracts the generated AI 33 related to the cheapest fee information based on the fee information for each of the multiple generated AIs 33 calculated by the fee calculation unit 51.
The display control unit 52 displays the generated AI33 extracted by the minimum fee extraction unit 53 on the display unit 60 of the user terminal 20. The user can select the generated AI33 with the lowest fee for obtaining an answer from among the multiple generated AI33.

Referring to FIG. 7, a generation AI management program according to one embodiment of the present invention will be described together with a generation AI management method. FIG. 7 is a flowchart of the generation AI management program according to this embodiment. The generation AI management method is executed by the CPU 10e of the generation AI management system 10 based on the generation AI management program.

The generation AI management program enables the CPU 10e of the generation AI management system 10 to realize a recording control function, a learning data acquisition function, a learning data transmission function, a learning data storage function, an input information acquisition function, a vector data conversion function, an input information transmission function, an answer information acquisition function, a dialogue data storage function, a learning function, a token number acquisition function, a fee calculation function, a display control function, and a minimum fee extraction function. These functions are executed in the order shown in FIG. 7, but the order can also be changed as appropriate. Note that each function overlaps with the explanation of the various functional parts of the generation AI management system 10 described above, so detailed explanations thereof will be omitted.

The recording control function records a plurality of generated AIs (S40: recording control step).
The learning data acquisition function acquires learning data to be used for training the generation AI 33 from the user terminal 20 (S41: learning data acquisition step).

The learning data transmission function transmits the learning data acquired by the learning data acquisition function to the generation AI server 30 (S42: learning data transmission step).

The learning data storage function stores the learning data transmitted to the generation AI server 30 by the learning data transmission function (S43: learning data storage step).

The input information acquisition function acquires information to be provided to one of the multiple generation AIs 33 as input information via the user terminal 20 (S44: input information acquisition step).

The vector data conversion function converts the text data contained in the input information acquired by the input information acquisition function into vector data (S45: vector data conversion step).

The input information transmission function transmits the input information to the generation AI server 30 (S46: input information transmission step).

The answer information acquisition function acquires an answer from one generation AI 33 corresponding to the input information as answer information from the generation AI server 30. (S47: Answer information acquisition step).

The dialogue data storage function associates the input information with the response information corresponding to the input information and stores the information as dialogue data (S48: dialogue data storage step).

When changing the generation AI server 30 (generation AI33) used by the user terminal 20, the learning function uses the learning data stored in the learning data storage function and used to learn the generation AI33 to learn the newly used generation AI33 (S49: learning step).

The token number acquisition function acquires the number of tokens generated by the tokenizer 32 installed in the generation AI server 30 as token number information from the generation AI server 30 based on the input information (S50: token number acquisition step).

The fee calculation function calculates the usage fee for the generation AI 33 (generation AI server 30) as fee information, calculated based on the token number information acquired by the token number acquisition function (S51: fee calculation step).

The display control function displays the fee information calculated by the fee calculation function on the display unit 60 of the user terminal 20 (S52: display control step).

The minimum price extraction function extracts the generated AI33 associated with the cheapest price information based on the price information for each of the multiple generated AI33 calculated by the price calculation function (S53: minimum price extraction step).

The present invention is not limited to the generation AI management system 10, the generation AI management method, and the generation AI management program according to the above-described embodiments, and can be implemented by various other modified examples or application examples without departing from the gist of the present invention described in the claims.

10 Generation AI management system 10a Communication interface 10b ROM (Read Only Memory)
10c RAM (Random Access Memory)
10d Storage unit 10e CPU (Central Processing Unit)
10f Input/Output Interface 10g Input Device 10g Output Device 11 Network 20 User Terminal 20a First User Terminal 20b Second User Terminal 30 Generation AI Server 30a First Generation AI Server 30b Second Generation AI Server 30c Third Generation AI Server 31 Input Information Receiving Unit 32 Talker 33 Generation AI
34 Answer information transmission unit 40 Recording unit 41 Learning data acquisition unit 42 Learning data transmission unit 43 Learning data storage unit 44 Input information acquisition unit 45 Vector data conversion unit 46 Input information transmission unit 47 Answer information acquisition unit 48 Dialogue data storage unit 49 Learning unit 50 Token number acquisition unit 51 Fee calculation unit 52 Display control unit 53 Minimum fee extraction unit 60 Display unit 61 Display screen 62 Menu bar 63 Main display unit 64 Sub-menu bar 65 Sub-display unit 66 Avatar search input window 67 Tag list 68 Avatar list 69 Heading 70 Dialogue list 71 Message input window 72 Avatar icon 73 User icon 74 Avatar creation button 75 List 76 API
77 API creation button

Claims (15)

A generation AI management system that connects to a server of a plurality of generation AIs and provides a user terminal with a function for utilizing the plurality of generation AIs,
A recording control unit that records the plurality of generated AIs;
An input information acquisition unit that acquires information to be provided to one of the plurality of generation AIs as input information via the user terminal;
an answer information acquisition unit that acquires an answer of the one generation AI corresponding to the input information from the server as answer information;
a dialogue data storage unit that stores the input information and the answer information corresponding to the input information as dialogue data in association with each other;
A learning unit that, when changing the one generation AI to another generation AI via the user terminal, uses the dialogue data stored in the dialogue data storage unit and obtained by communicating with the one generation AI to learn the other generation AI;
A generative AI management system comprising: The generating AI management system includes:
an input information transmitting unit that transmits the input information to the server;
a vector data conversion unit that converts text data included in the input information acquired by the input information acquisition unit into vector data;
Further comprising:
The generation AI management system according to claim 1, characterized in that the input information sending unit sends the input information including the vector data to the server after the input information is converted from the text data to the vector data. the vector data conversion unit converts text data included in the answer information acquired by the answer information acquisition unit into vector data;
The generation AI management system according to claim 2, characterized in that the dialogue data storage unit associates the input information including the vector data with the answer information corresponding to the input information and including the vector data, and stores them as dialogue data. The generating AI management system includes:
A learning data acquisition unit that acquires learning data used for learning the generation AI from the user terminal;
a learning data transmission unit that transmits the learning data acquired by the learning data acquisition unit to the server;
a learning data storage unit that stores the learning data transmitted to the server by the learning data transmission unit;
Further comprising:
The learning unit is
A generative AI management system as described in claim 2, characterized in that when one generative AI is changed to another generative AI via the user terminal, the other generative AI is trained using the learning data stored in the learning data storage unit, which is the learning data used to train the one generative AI. The generating AI management system includes:
a vector data conversion unit that converts text data included in the learning data acquired by the learning data acquisition unit into vector data;
5. The generation AI management system according to claim 4, wherein the learning data storage unit stores the learning data including the vector data after the text data is converted into the vector data. The generated AI is a plurality of
The generating AI management system includes:
A plurality of APIs of the generation AI can be connected to the generation AI;
The generation AI management system of claim 4, further comprising a display control unit that assigns an avatar to each of the APIs of the generation AI and displays a list of the avatars assigned to each API on the display unit of the user terminal, thereby displaying a list of the APIs on the display unit. The generative AI management system according to claim 6, characterized in that the display control unit displays a list of the learning data stored in the learning data storage unit and used to train the generative AI on the display unit of the user terminal, thereby displaying the learning history of the generative AI on the display unit. The learning unit is
When constructing a clone of the machine learning model configured by the generating AI,
The generative AI management system described in claim 1, characterized in that the generative AI that constitutes the clone is trained using the dialogue data stored in the dialogue data storage unit, the dialogue data being obtained by communicating with the generative AI that constitutes the machine learning model that is the source of the clone. The learning unit is
When constructing a clone of the machine learning model configured by the generating AI,
The generative AI management system described in claim 4, characterized in that the learning data stored in the learning data memory unit is used to learn the generative AI that constitutes the clone, and the learning data used to learn the generative AI that constitutes the machine learning model that is the source of the clone is used. The generated AI is a plurality of
The generation AI management system is connectable to APIs of the plurality of generation AIs;
the input information transmitting unit transmits the same input information to a plurality of the servers;
the answer information acquisition unit acquires the answer information corresponding to the same input information from a plurality of the servers;
The generation AI management system according to claim 6, characterized in that the display control unit displays a list of the multiple pieces of answer information acquired by the answer information acquisition unit on a display unit of the user terminal. The generated AI is a plurality of
The generation AI management system is connectable to APIs of the plurality of generation AIs;
the input information sending unit sends input information for obtaining an answer regarding an evaluation of the answer information acquired by the answer information acquiring unit from one server to another server different from the one server from which the answer information was acquired;
The generation AI management system according to claim 2, characterized in that the answer information acquisition unit acquires an answer from the other server regarding an evaluation of the answer from the one server as second opinion information. The generating AI management system includes:
a token number acquisition unit that acquires, from the server, the number of tokens generated by a tokenizer installed in the server based on the input information as token number information;
A fee calculation unit that calculates a usage fee for the generated AI based on the token number information acquired by the token number acquisition unit as fee information,
The generation AI management system according to claim 6, characterized in that the display control unit displays the fee information calculated by the fee calculation unit on a display unit of the user terminal. The generated AI is a plurality of
The generating AI management system includes:
A plurality of APIs of the generation AI can be connected to the generation AI;
a token number acquiring unit that acquires, as token number information from each of the plurality of servers, the number of tokens generated by a tokenizer installed in the plurality of servers that acquired the input information;
a fee calculation unit that calculates a usage fee for the generated AI calculated based on the token number information acquired by the token number acquisition unit as fee information for each of the generated AIs;
A minimum fee extraction unit extracts the generated AI associated with the cheapest fee information based on the fee information for each of the plurality of generated AIs calculated by the fee calculation unit,
The generated AI management system according to claim 6, characterized in that a display control unit displays the generated AI extracted by the minimum price extraction unit on a display unit of the user terminal. A generation AI management method used in a generation AI management system that connects to a server of a plurality of generation AIs and provides a user terminal with a function for utilizing the plurality of generation AIs,
The generating AI management system includes:
a recording control step of recording the plurality of generated AIs;
An input information acquisition step of acquiring information to be provided to one of the plurality of generation AIs as input information via the user terminal;
An answer information acquisition step of acquiring an answer of the one generation AI corresponding to the input information from the server as answer information;
a dialogue data storage step of storing the input information and the answer information corresponding to the input information as dialogue data in association with each other;
A learning step of learning the other generation AI using the dialogue data stored in the dialogue data storage step, which is obtained by communicating with the one generation AI, when changing the one generation AI to another generation AI via the user terminal;
A generative AI management method comprising: A generation AI management program used in a generation AI management system that connects to a server of a plurality of generation AIs and provides a user terminal with a function for utilizing the plurality of generation AIs,
The generating AI management system includes:
A recording control function for recording the plurality of generated AIs;
An input information acquisition function that acquires information to be provided to one of the plurality of generation AIs as input information via the user terminal;
An answer information acquisition function that acquires an answer of the one generation AI corresponding to the input information from the server as answer information;
a dialogue data storage function for storing the input information and the answer information corresponding to the input information as dialogue data in association with each other;
A learning function that, when changing the one generation AI to another generation AI via the user terminal, uses the dialogue data stored in the dialogue data storage function, which is obtained by communicating with the one generation AI, to make the other generation AI learn;
A generating AI management program characterized by realizing the above.

Images