Meta(META.US), Facebook's parent company, announced the release of a series of new artificial intelligence models, including one that can self-assess, which may reduce human involvement in the AI development process. The release follows Meta's paper in August that detailed how the tool uses the same "chain-of-thought" technology as OpenAI's o1 model to improve accuracy in answering challenging problems such as science, coding, and math. Meta researchers trained the assessment model on entirely AI-generated data, eliminating the need for human input.
Meta's move marks a significant step forward for the company in AI's ability to learn and assess itself. Self-assessment technology not only improves model accuracy but also may reduce the cost and inefficiencies of the reinforcement learning process that currently relies on human feedback. Meta researchers say using AI to assess its own capabilities is key to creating intelligent agents that can learn from their mistakes and improve autonomously. These agents are envisioned as digital assistants capable of performing a wide range of tasks without human intervention.
Meta researcher Jason Weston emphasized that as AI advances, it will become increasingly adept at self-checking. The ability to learn and assess oneself is crucial for AI. Besides Meta, companies like Google and Anthropic are also researching reinforcement learning concepts based on AI feedback, but they typically do not publicly release models.
In this release, Meta also introduced other AI tools, including an update to the image recognition model Segment Anything, which can speed up response generation time for large language models (LLM), and a dataset for discovering new inorganic materials. The tools' release further demonstrates Meta's efforts and commitment to advancing AI technology.