Meta Unveils AI Model to Assess Other AI Models
Meta introduces "Self-Taught Evaluator" AI model to assess other AI models.The model utilises a "chain of thought" technique for improved accuracy in complex problem-solving.Meta's researchers trained the evaluator model using AI-generated data, reducing human involvement.
Among these releases is the "Self-Taught Evaluator," a tool that could potentially reduce human involvement in the AI development process.
The Self-Taught Evaluator, introduced in an August paper by Meta, utilises a "chain of thought" technique similar to OpenAI's latest models. This approach involves breaking down complex problems into smaller logical steps, enhancing the accuracy of responses in challenging subjects like science, coding, and math.
Meta's researchers trained the evaluator model solely on AI-generated data, eliminating the need for human input during the training phase. This advancement paves the way for autonomous AI agents capable of learning from their own mistakes, according to two Meta researchers.
The prospect of using AI to evaluate AI reliably offers a glimpse into a future where digital assistants can perform a wide range of tasks independently, without human intervention. Such self-improving models could potentially replace the costly and inefficient process of Reinforcement Learning from Human Feedback, which relies on human annotators with specialised expertise.
Jason Weston, a researcher at Meta, expressed optimism about the evolution of AI, stating, "We hope, as AI becomes more and more super-human, that it will get better and better at checking its work, so that it will actually be better than the average human." Weston emphasised the importance of self-teaching and self-evaluation in achieving a super-human level of AI.
While other companies like Google and Anthropic have explored the concept of Reinforcement Learning from AI Feedback (RLAIF), Meta stands out by making its models publicly available. The company's recent AI tool releases also include an updated image-identification Segment Anything model, a tool to accelerate LLM response generation times, and datasets to aid in the discovery of new inorganic materials.
Meta introduces "Self-Taught Evaluator" AI model to assess other AI models
The model utilises a "chain of thought" technique for improved accuracy in complex problem-solving
Meta's researchers trained the evaluator model using AI-generated data, reducing human involvement
Source: REUTERS