Meta Launches an AI Model to Review Other AI Models

Meta has launched a groundbreaking AI model that can review other AI models without human involvement. Could this be the path toward self-improving AI systems that revolutionize both technology and research? Is this step a risk for the technology?

Share
Meta Launches an AI Model to Review Other AI Models

Meta has recently released a new set of AI models from its research division, including a "Self-Taught Evaluator" that could reduce the need for human involvement in AI development. The model is based on a technique called "chain of thought," which breaks complex problems down into smaller logical steps, improving the accuracy of AI's answers in fields like science, coding, and math.

AI Reviews AI

The big news with Meta's new AI model is that it's trained entirely on AI-generated data, meaning no human input is needed during the training phase. This is a potentially major step toward autonomous AI agents that can learn from their own mistakes. Such a development could replace current methods like "Reinforcement Learning from Human Feedback" (RLHF), a process that requires expensive, expert-dependent human feedback.

Jason Weston, one of the researchers behind the project, expressed hope that AI will in the future become "superhuman," better than humans at checking its own work.

Self-Improving AI Could Save Resources

Self-improving AI models could reduce the costs and inefficiencies associated with using human annotators to verify answers and train AI models. With self-learning, self-evaluating AI, the development of more advanced digital assistants capable of performing complex tasks without human involvement could accelerate.

Even though other companies like Google and Anthropic are also exploring similar techniques like "Reinforcement Learning from AI Feedback" (RLAIF), it's notable that Meta is openly sharing its models for public use, which could give researchers and developers around the world access to these tools.

Other AI Tools From Meta

In addition to the self-learning evaluator, Meta also released an update to its image identification tool, the Segment Anything model, as well as a technique that speeds up generation time for large language models (LLMs). Meta also introduced new datasets that can be used to discover new inorganic materials, which could contribute to research in materials science.