Search

DE-102024133081-A1 - Method, device and computer program for processing a task using artificial intelligence

DE102024133081A1DE 102024133081 A1DE102024133081 A1DE 102024133081A1DE-102024133081-A1

Abstract

The present disclosure relates to a method, a device, and a computer program for processing a task using artificial intelligence (AI) (300). The method (10) for processing a task with the AI (300) comprises classifying (11) the task into a subject area and assessing (12) the suitability of one or more existing supplementary modules (34) for processing the task by the AI (300) using the one or more existing supplementary modules (34). The suitability reflects a minimum level of quality of a result of processing the task by the AI (300) using the one or more supplementary modules (34). The procedure (10) further comprises generating (13) one or more additional supplementary modules (36) by the AI (300) if the one or more existing supplementary modules (34) are not suitable, whereby the minimum quality standard of a result of the processing of the task by the AI (300) using the one or more additional supplementary modules (36) is met. The procedure further comprises processing (14) the task with the AI (300) based on the subject area using the one or more additional supplementary modules (36) in order to achieve a result for the task.

Inventors

  • Wolfgang Bau

Assignees

  • DEUTSCHE TELEKOM AG

Dates

Publication Date
20260513
Application Date
20241112

Claims (20)

  1. A method (10) for processing a task with an artificial intelligence, AI (300), comprising: classifying (11) the task into a subject area; assessing (12) the suitability of one or more existing supplementary modules (34) for processing the task by the AI (300) using the one or more existing supplementary modules (34), wherein the suitability reflects a minimum level of quality of a result of processing the task by the AI (300) using the one or more supplementary modules (34); generating (13) one or more additional supplementary modules (36) by the AI (300) if the one or more existing supplementary modules (34) are not suitable, wherein the minimum level of quality of a result of processing the task by the AI (300) using the one or more additional supplementary modules (36) is met; and Processing (14) the task with the AI (300) based on the subject area using one or more additional modules (36) to achieve a result for the task.
  2. The procedure (10) according to Claim 1 , furthermore, comprehensively address the task using the AI (300) based on the subject area, utilizing one or more existing supplementary modules (34) to achieve a result for the task, if the one or more existing supplementary modules (34) are suitable.
  3. The procedure (10) according to one of the Claims 1 or 2 , wherein the assessment (12) includes an estimation of the internal accuracy of the one or more existing supplementary modules (34) in the processing of the task and wherein the minimum level of quality of the result is a minimum level of internal accuracy.
  4. The procedure (10) according to one of the Claims 1 until 3 , wherein generating (13) includes creating executable software code for the one or more add-on modules (34).
  5. The procedure (10) according to Claim 4 , where the software code is adapted for an interpreter.
  6. The procedure (10) according to one of the Claims 4 or 5 , where the software code has an Application Programming Interface (API).
  7. The procedure (10) according to one of the Claims 4 until 6 , where the software code is executable on a virtual machine.
  8. The procedure (10) according to one of the Claims 1 until 7 , wherein the one or more add-on modules (34) comprise one or more plug-ins for AI.
  9. The procedure (10) according to one of the Claims 1 until 8 , wherein generating (13) includes creating one or more test data sets for the one or more additional add-on modules (36).
  10. The procedure (10) according to one of the Claims 1 until 9 , wherein at least one generated supplementary module (36) provides an intermediate result when the artificial intelligence processes the task.
  11. The procedure (10) according to one of the Claims 1 until 10 , whereby generation (13) also takes place if a corresponding specification is provided in the task description.
  12. The procedure (10) according to one of the Claims 1 until 11 , wherein the generation (13) also includes the generation of one or more interfaces to the one or more additional add-on modules (36).
  13. The procedure (10) according to Claim 12 , where one or more interfaces can be activated and deactivated by the AI.
  14. The procedure (10) according to one of the Claims 1 until 13 , furthermore comprehensive storage of one or more additional add-on modules (36) for further use.
  15. The procedure (10) according to one of the Claims 1 until 14 , furthermore, comprehensively generating a description of one or more additional supplementary modules for use in other tasks or by other AIs.
  16. The procedure (10) according to one of the Claims 1 until 15 , where generating (13) involves creating a questionnaire and interviewing selected experts.
  17. The procedure (10) according to one of the Claims 1 until 16 , furthermore comprehensively maintaining the minimum level of quality for a user of the AI.
  18. The procedure (10) according to one of the Claims 1 until 17 , furthermore, comprehensive selection of a data model based on the subject area.
  19. A computer program comprising program code for carrying out one of the methods (10) according to one of the preceding claims, when the program code is executed on a computer, a processor or a programmable hardware component.
  20. A device (20) for processing a task with an AI (300), with one or more interfaces (22) configured for communication with one or more users and/or communication systems; and one or more signal processing components (24) configured to perform one of the methods (10) according to one of the Claims 1 until 19 to execute.

Description

Technical field The present disclosure relates to a method, a device and a computer program for processing a task with an artificial intelligence (AI), in particular but not exclusively, a concept for the automated generation of supplementary modules for an AI to improve a quality measure of an answer to a specific task. background Various concepts from conventional engineering utilize AI. Generative AI receives a query text (prompt) and processes the query primarily based on the AI's underlying model and its pre-trained datasets. Since by default only the data in the request language is processed, the AI's internal response is always immediately available in the request language without the need for language translation, as it is assumed by default that the user wants to receive the response in the same language as the request. The quality of the answer (the internal accuracy of the AI's answer, not the external accuracy) depends not only on the AI's parameter model but also on the quantity and quality of the available training data in the query language, the exactness of the question in the current prompt, the context (unless explicitly prohibited) of the previous prompts (previous queries/instructions) and their associated answers, which of course applies equally to each of the query languages to be used and to each generative AI. Early generative AIs (e.g., up to the ChatGPT chatbot version 2) were already surprisingly good at generating text and understanding and processing small integers. Further development of generative AIs (gAI) then encompassed operations with larger integers, followed by floating-point numbers, and, with the use of pre-built plugins (including mathematical and business mathematics plugins), the correct processing of complex mathematical problems that the gAI must solve to provide the correct/best answer to a prompt. In parallel, gAIs were improved in areas such as reducing hallucinations, increasing the size or specialization of training data, increasing multilingual support, and enabling multimodal input (initially including speech in addition to text and numbers) for prompts. Later, images and videos were added as prompt input and as responses from the gAI. Starting with GPT4plus, users can create their own personalized AI via a user interface by selecting from a wide range of desired properties. Two of these properties relate, for example, to the role of the AI and the role of the user (e.g., as a root prompt: "From now on, behave as a math teacher talking to a 14-year-old child."). A chatbot can also consider the history of a conversation chain (prompt -> response -> prompt...) and use not only its training data but also data from the prompt and from links defined within the prompt. Currently, an AI can also generate programming elements in its known programming languages and create test datasets and target solutions. The latter is already used by many programmers, resulting in estimated work savings of 50% to 70% on average. Summary One existing problem is that there isn't always a best or suitable plug-in available for every prompt task that could be integrated into the gAI (Global AI). Users often don't even realize that their integrated plug-ins can't solve their prompts with the quality they expect. If the user is dissatisfied with the prompt response, they might suspect the problem lies with the gAI's core functionality rather than the plug-ins. Examples of implementation are based on the core idea that a specific task can be better solved with AI if the AI itself generates the most suitable add-on modules/plug-ins for the prompt currently being processed. This can be achieved, for example, by the AI generating the most suitable program, test data sets, and target solutions, and ultimately creating executable software as a plug-in that it can use. The following examples therefore establish a method for processing a task using artificial intelligence (AI). The method includes classifying the task within a subject area. It then assesses the suitability of one or more existing supplementary modules for the AI to process the task. Suitability reflects a minimum level of quality in the AI's task processing results using these modules. The method further includes generating one or more additional supplementary modules if the existing modules are unsuitable, ensuring that the minimum quality level of the AI's task processing results using these additional modules is met. Finally, the AI processes the task based on the subject area, utilizing these additional modules to achieve a solution. In this way, solutions to tasks can be improved through adaptively generated supplementary modules. The process can further include processing the task with AI based on the subject area, utilizing one or more existing add-on modules to achieve a result for the task, provided the one or more existing add-on modules are suitable. This ensures that a required minimum level of quality for the pro