Meta, the parent company of Facebook, recently introduced a new AI model called the “Self-Taught Evaluator,” capable of evaluating the work of other AI models. This represents a significant advancement in the development of autonomous AI systems, minimizing the need for human involvement in the process of verifying and improving AI models.

The new model utilizes a method known as “chain of thought,” breaking down complex problems into smaller, logical steps. This approach has been shown to enhance accuracy, especially in areas such as science, coding, and mathematics. Instead of relying on human evaluators to verify AI-generated results, the Self-Taught Evaluator is trained entirely on AI-generated data, potentially leading to AI systems that learn from their own mistakes.

According to Meta’s researchers, these self-improving AI systems could one day act as digital assistants, capable of performing complex tasks without human oversight. This development could also eliminate the need for Reinforcement Learning from Human Feedback (RLHF), which often requires expensive and specialized human involvement.

Meta distinguishes itself by making its models publicly available, a contrast to competitors like Google and Anthropic, which tend to keep their systems more closed.