To try to solve the problems of artificial intelligence errors or “hallucinations”, Meta proposes to entrust basic tasks to external tools.
As they evolve, new artificial intelligences such as ChatGPT or Google Bard are able to handle increasingly complex tasks. However, they show increasing difficulty with some of the most basic tasks, such as doing arithmetic or finding factual information.
Meta is currently trying to solve this problem by simply teaching an AI to use tools. Researchers from Meta AI Research have published an article introducing “Toolformer”, an AI based on the GPT-J language model, a model close to GPT-2 but developed by EleutherAI, a community of researchers . They taught him to use APIs (Application Programming Interfaces), tools that allow software to talk to each other.
AI more efficient than GPT-3
Researchers trained Toolformer by showing him some examples of using the APIs of various software, including a calculator, search engine, question-answer system, translator, and calendar. Artificial intelligence was able to learn on its own to integrate the use of these tools when making requests, deciding which API to call, when, what parameters to use, and how to integrate the results into its response.
AI Toolformer has been found to perform better than GPT-3 on many tasks, despite a large difference in complexity. GPT-3 has 175 billion parameters compared to only 6.7 billion for GPT-J. However, the researchers pointed out some limitations, since the artificial intelligence is not able to chain the software and use the result to make a request to another.