[ad_1]
Researchers Half they presented Toolformer: it is an AI language model that can self-learn the use of external tools, such as search engines, calendars, calculators and so on, without compromising the modeling capabilities of the AI language basis.
This made possible thanks to the ability to use APIs: Meta researchers, during the training of Toolformer, provided it with a small set of human-written examples demonstrating how each API should be used, subsequently allowing the AI to write a language modeling dataset with potential API calls. Through the “self-supervised” training technique, the AI was able to learn without the explicit need for a human being to “take it by the hand” in its learning path.
Thanks to this process, the AI has learned to interpret each text-based API call as if it were any other form of text: in this way it can insert API calls when necessary, pass them the appropriate arguments and, during the “responses ” to a human being, to choose the most suitable tool for the context and how to use it. Tools that can be, well… anything: translators, search engines, calculators, calendars and everything that can be controlled via API.
Toolformer is notably based on a template GPT-Jpre-trained with 6.7 billion parameters: by way of comparison, the GPT-3 model is based on 175 billion parameters, but from tests conducted by researchers Toolformer manages to express better performance precisely by virtue of the possibility of using external tools.
The novelty of the Meta model lies in the ability to learn to use the tools in a generic waywithout having to go through specialized training. This is quite an interesting approach, since while there already exist AI models capable of searching the web (the recent Bing Chat, for example), they have so far been based on large amounts of human input or specific settings to perform specific tasks.
The approach followed with Toolformer may lead to a future in which AI models using external applications will be able to act as versatile personal assistants and transform the way humans interact and interface with machines. In any case, it is a road that will have to be traveled with the necessary caution, given that greater potential and capacity inevitably bring with it greater risks.
.
[ad_2]
Source link
