The new AI model developed by Facebook's parent company is based on the so-called “chain reasoning” technology, which OpenAI also uses in new developments. This method allows the system to make reliable judgments about model responses while breaking down complex problems into smaller logical steps. This technology has proven particularly effective in difficult fields such as science, programming, and mathematics.
Meta researchers exclusively used AI-generated data to train the evaluation model.
We hope that as AI becomes more powerful, it will control its work more and more, so that it will actually be better than the average human.
– said Jason Weston, one of the lead researchers on the project.
Self-development models have the potential to replace the costly and often ineffective procedures currently in place that rely on human feedback.
This can be particularly important in fields that require special expertise to accurately label data and validate answers to complex mathematical and written queries.
In addition to Meta, other tech giants, including Google and Anthropic, are researching the concept of RLAIF (Reinforcement Learning from AI Feedback). However, unlike Meta, these companies do not make their models publicly available.
As part of Friday's announcement, Meta also unveiled additional AI tools. This includes an update to the company's Segment Anything image recognition model, a tool that speeds up LLM response generation times, and datasets that can help discover new inorganic materials.
Source: Reuters
The cover image is an illustration. Cover image source: Portfolio
