Meta AI Evaluator to Automate AI Development Checks

Meta has announced the release of new AI models from its research division, including the “Self-Taught Evaluator.” This tool employs the “chain of thought” technique utilized by OpenAI’s o1 models, enabling it to make reliable assessments of other models’ responses.

Introduced in an August paper, the Self-Taught Evaluator breaks down complex problems into smaller, logical steps. This method has enhanced accuracy in challenging areas such as science, coding, and mathematics.

Meta’s researchers have trained this evaluator model using entirely AI-generated data, significantly reducing human involvement in the early stages of AI development. As per two meta-researchers who spoke to Reuters, this development hints at the possibility of fully autonomous AI agents independently learning from their mistakes.

These self-improving models could eventually eliminate the need for the current method known as Reinforcement Learning from Human Feedback (RLHF). RLHF is costly and inefficient as it depends on human experts to label data and verify complex responses.

Jason Weston, a researcher at Meta, expressed optimism about the future of AI: “We hope that as AI becomes increasingly superhuman, it will excel at checking its work, potentially surpassing the average human in accuracy.”

He emphasized the significance of being self-taught and capable of self-evaluation as essential to achieving a super-human level of AI.

Unlike Meta, companies like Google and Anthropic, which have explored similar concepts through Reinforcement Learning from AI Feedback (RLAIF), typically do not make their models publicly available.

In addition to the Self-Taught Evaluator, Meta also released updates to its image-identification tool, Segment Anything. This tool accelerates response generation times for large language models and datasets to discover new inorganic materials.