Meta Unveils New AI Models

By Katie Paul

NEW YORK – Facebook (NASDAQ:META) owner Meta announced on Friday the release of new AI models from its research division, including a Self-Taught Evaluator. This model aims to reduce human involvement in AI development.

The introduction follows a tool detailed in an August paper, where Meta described how the Self-Taught Evaluator uses the "chain of thought" technique similar to OpenAI’s recent models. This method breaks down complex problems into smaller, logical steps, enhancing accuracy in challenging subjects like science, coding, and math.

Meta’s researchers trained the evaluator entirely with AI-generated data, eliminating human input during training. This capability points towards creating autonomous AI agents capable of learning from their mistakes, according to Meta researchers who spoke to Reuters.

Many in the AI community envision these agents as intelligent digital assistants that perform various tasks independently.

Self-improving models could eliminate the costly and inefficient Reinforcement Learning from Human Feedback process, which relies on human annotators with specialized skills to label data accurately and verify complex queries.

Jason Weston, one of the researchers, stated, "We hope, as AI becomes more and more super-human, that it will get better and better at checking its work, so that it will actually be better than the average human."

He added, "The idea of being self-taught and able to self-evaluate is crucial to reaching a super-human level of AI."

Other companies, including Google (NASDAQ:GOOGL) and Anthropic, have researched Reinforcement Learning from AI Feedback (RLAIF), but unlike Meta, they typically do not release their models for public use.

In addition to the Self-Taught Evaluator, Meta released an update to the Segment Anything image-identification model, a tool for speeding up LLM response generation, and datasets for discovering new inorganic materials.