An illustration taken on September 28, 2023, shows the Meta AI logo. Reuters
Meta, the company behind Facebook, has announced the release of several new artificial intelligence (AI) models from its research division, including a significant innovation called the "Self-Taught Evaluator." This tool is designed to reduce the need for human involvement in the AI development process, a potential game-changer for the field. The development comes after Meta initially introduced the tool in a research paper back in August, explaining how it employs a "chain of thought" technique to improve accuracy in evaluating other AI models’ responses.
This technique involves breaking down complex tasks into smaller, logical steps, which makes it easier for AI to handle difficult subjects, including science, math, and coding. The approach also mirrors one used by OpenAI in their recent AI models, which similarly aim to improve AI judgment. What’s particularly striking about Meta’s new evaluator model is that it was trained entirely with AI-generated data, removing the need for human input in the training process. This marks a significant step toward creating more autonomous AI systems.
The ability for AI to evaluate itself reliably hints at a future where AI agents could learn from their own mistakes. This concept excites many in the AI field, as these self-improving agents could eventually become digital assistants capable of performing a wide range of tasks independently, without human oversight. Two Meta researchers involved in the project shared their thoughts, noting that the development could lead to cutting out a costly and time-consuming process known as Reinforcement Learning from Human Feedback (RLHF). This traditional method involves human experts who must carefully label data and verify that AI-generated answers are correct, which requires specialized knowledge, particularly in areas like math and writing.
Jason Weston, one of the leading researchers on the project, expressed optimism about AI's future capabilities, saying, "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." He emphasized that the concept of self-teaching and self-evaluation is critical in pushing AI towards achieving superhuman-level intelligence.
Other major tech companies, including Google and Anthropic, are also working on similar ideas, specifically focusing on a method called Reinforcement Learning from AI Feedback (RLAIF). However, unlike Meta, these companies tend not to make their models publicly available, which sets Meta apart in terms of transparency and accessibility.
In addition to the Self-Taught Evaluator, Meta released updates to other AI tools, including its image-recognition model called Segment Anything, which speeds up the response times of large language models (LLMs), and new datasets designed to assist in the discovery of new inorganic materials.
