web analytics
Home » Technology » Facebook » Meta creates an AI language bot that utilizes third-party software

Meta creates an AI language bot that utilizes third-party software

ChatGPT is a language model. It was introduced in November 2022. Since then, it has marked the beginning of several changes in the natural language field. Despite the great potential and capacity of language models, the ability to perform basic tasks is unsatisfactory.

The last week, on Thursday, Meta researchers unveiled Toolformer. Toolformer is an AI language model. It has the capacity to learn how to use external tools. External tools like calculators, calendars, and search engines can be used by Toolformer. It can do so without compromising on its core language modeling capacities.

The main idea behind this is that Toolformer can utilize APIs. APIs are basically application programming interfaces. They are a set of protocols. These protocols permit the apps to communicate with one another in a streamlined and programmed approach. Now the question here is how Toolformer uses APIs.

During the training phase, the researchers provided the Toolformer with a small set of examples written by humans. The idea was to show how individual API is utilized. It was followed by the permission to annotate a large language modeling dataset with likely API calls. Toolformer performed the task in a self-supervised manner. Thus, it indicates that it can learn without the requirement of guidance from humans.

In this way, the model was trained to take each text-based API call as some form of text. During operation, calls can be inserted when required. In addition to this, the model can decide which tools to use for which context and how to utilize them on its own.

Based on the API-calling ability, the model can utilize external sources or software tools like calculators, search engines, factual references, and language translators. Where the large language models are not well-equipped for performing arithmetic operations. Toolformer presents a solution to this limitation. Similarly, the model can deal with adding a date to the calendar by using the API link to a calendar app.