Meta, announced the release of new AI models from its research division, including a groundbreaking "Self-Taught Evaluator." This model could potentially minimize human involvement in AI development.
The Self-Taught Evaluator was first introduced in an August paper, which explained its reliance on the "chain of thought" technique. This method, also used by OpenAI’s latest o1 models, breaks down complex problems into smaller steps, resulting in more accurate responses in areas such as science, coding, and math. Meta's researchers trained the evaluator using entirely AI-generated data, eliminating the need for human input at this stage.
This innovation provides a glimpse into the future of autonomous AI agents capable of learning from their own mistakes. According to Meta researchers, these models could replace the current, often costly process of Reinforcement Learning from Human Feedback (RLHF), which relies on human experts to label data and verify responses.
Other tech giants, including Google and Anthropic, are exploring similar concepts through Reinforcement Learning from AI Feedback (RLAIF). However, unlike Meta, these companies rarely release their models for public use.
Alongside the Self-Taught Evaluator, Meta also launched an update to its image-identification Segment Anything model, a tool for speeding up LLM response times, and new datasets to aid the discovery of inorganic materials.