Meta, the parent company of Facebook, has just announced an exciting new development in the world of artificial intelligence: a model that can evaluate the performance of other AI models. This innovative “Self-Taught Evaluator,” released on Friday, could significantly reduce the need for human oversight in AI development.
This groundbreaking model builds on research introduced in an August paper, where Meta highlighted its reliance on a “chain of thought” approach—similar to techniques used in OpenAI’s latest models. This method breaks down complex challenges into smaller, logical steps, enhancing the model’s ability to tackle difficult problems in areas like science, coding, and mathematics.
Remarkably, Meta’s researchers trained the evaluator using entirely AI-generated data, eliminating human input from the training process. This shift opens up exciting possibilities for creating autonomous AI agents that can learn from their own errors, as explained by two of the researchers involved in the project.
The prospect of self-improving AI models is intriguing. These digital assistants could handle a wide range of tasks without needing constant human intervention. According to researcher Jason Weston, “We hope that as AI becomes more super-human, it will get better at checking its work, surpassing the average human’s capabilities.” He emphasized that self-evaluation is key to achieving a truly advanced level of AI.
While other tech giants like Google and Anthropic have explored similar concepts—termed Reinforcement Learning from AI Feedback (RLAIF)—Meta distinguishes itself by making its models publicly accessible, rather than keeping them behind closed doors.
In addition to the Self-Taught Evaluator, Meta also rolled out several other AI tools on Friday, including updates to its image-identification model, Segment Anything, and enhancements that speed up response times for large language models. With these developments, Meta is not just pushing the envelope on AI capabilities; it’s also paving the way for a future where AI can take a more autonomous role in its own evolution.
