Meta researchers have introduced a new artificial intelligence language model called Toolformer, which has the power to learn to work with some applications, including search engines, calculators and calendars.
New language models, including ChatGPT, which have recently received public attention, have created a huge revolution in the field of natural language processing, but due to the newness of this phenomenon, they still have minor problems in performing some tasks such as calculation and verification.
The key thing that separates Toolformer from others is that it can use APIs. During their experiments, Meta researchers provided Toolformer with a small set of human handwritten samples that can demonstrate how to use each API, thereby allowing interpretation of large language modeling datasets based on potential API commands.
The researchers designed this artificial intelligence in such a way that Toolformer managed to do its work in a “self-supervising” manner, meaning that it can learn without the need for human guidance. In addition, Toolformer can also decide by itself which tool to use to get the right context and how to use it.
Another feature of the API call is that it allows Toolformer to use external software such as search engines, calculators, and more. This tool can help those using large language models (LLMs) who are notoriously computationally inefficient and overcome this limitation.
Based on the researchers’ tests on this new and efficient tool, Toolformer has been found to have a much stronger and more accurate performance than the much larger GPT-3 model, which contains 175 billion parameters.
In the end, it should be noted that this is not the first time that researchers have tried to eliminate the limitations of artificial intelligence language models, but their efforts have failed. According to what has been said, the new Bing browser that supports the advanced version of ChatGPT and It can perform web searches on its own when needed.
