Meta releases an AI model that can check the work of other AI models by Reuters

By Katie Paul

NEW YORK – Facebook (NASDAQ: ) owner Meta said Friday it was launching a batch of new AI models from its research division, including a “self-learning evaluator” that may offer a path to less involvement human in the AI ​​development process.

The release follows Meta’s introduction of the tool in an August article, which detailed how it relies on the same “chain of thought” technique used by OpenAI’s recently released o1 models to make it reliable judgments about model responses.

This technique involves breaking down complex problems into smaller logical steps and appears to improve the accuracy of answers to difficult problems in subjects such as science, coding and math.

Meta researchers used fully AI-generated data to train the rater model, also eliminating human input at this stage.

The ability to use AI to reliably evaluate AI offers insight into a possible path toward creating autonomous AI agents that can learn from their own mistakes, two of the Meta researchers told Reuters behind the project.

Many in the AI ​​field envision agents as digital assistants intelligent enough to perform a wide range of tasks without human intervention.

Machine-enhancing models could eliminate the need for an often expensive and inefficient process used today called Reinforcement Learning from Human Feedback, which requires input from human annotators who must have specialized expertise to label data accurately and verify that answers to complex math and writing queries. they are correct

“We hope that as the AI ​​becomes more and more superhuman, it will get better and better at checking its work, so that it will actually be better than the average human,” said Jason Weston, one of the researchers

“The idea of ​​being self-taught and being able to self-assess is basically crucial to the idea of ​​getting to this kind of superhuman level of AI,” he said.

Other companies, including Google (NASDAQ:) and Anthropic, have also published research on the concept of RLAIF, or Reinforcement Learning from AI Feedback. Unlike Meta, however, these companies tend not to release their models for public use.

Other AI tools released by Meta on Friday included an update to the company’s Segment Anything image identification model, a tool that speeds up LLM response generation times, and datasets that can be used to help to the discovery of new inorganic materials.