Summarize

Simplifying... Inshort

Meta has developed a new AI, the Self-Taught Evaluator, that can self-assess and monitor other AI models, potentially paving the way for autonomous AI agents that learn from their own errors.
This could replace the current, costly method of Reinforcement Learning from Human Feedback.
In addition, Meta has enhanced other AI tools, including an image-identification model and a tool that quickens language model response times.

Was a long read? Making it simpler...

Next Article

It was trained using only AI-generated data

Meta's new AI can self-evaluate, monitor other models

By Dwaipayan Roy

Oct 19, 2024

04:21 pm

What's the story

Meta has announced the development of a new artificial intelligence (AI) model, the "Self-Taught Evaluator." The innovative model, introduced in an August paper by the company's research division, aims to cut down on human involvement in AI development. The Self-Taught Evaluator uses a "chain of thought" approach similar to OpenAI's o1 models. The method simplifies complex issues into smaller logical steps to improve response accuracy on challenging problems in fields like science, coding and math.

Training process

AI model trained without human input

The Self-Taught Evaluator was trained using only AI-generated data, eliminating the need for human input at this stage. This capability to use AI to assess other AIs, reliably opens a possible path toward building autonomous AI agents that can learn from their own mistakes. The researchers behind the project believe such self-improving models could replace the current Reinforcement Learning from Human Feedback (RLHF) method, which is expensive and inefficient due to its reliance on human annotators with specialized expertise.

AI evolution

Meta's AI model aims for super-human level of intelligence

Jason Weston, one of the researchers on the project, was optimistic about the future of AI. "We hope, as AI becomes more and more super-human, that it will get better and better at checking its work," he said. Weston further noted that being self-taught and able to self-evaluate is key to reaching a super-human level of AI. Meta's development comes as other tech giants like Google and Anthropic, continue to research RLHF or Reinforcement Learning from AI Feedback.

AI advancements

Meta also updates other AI tools

Along with the Self-Taught Evaluator, Meta has also updated other AI tools. These include an improvement to the firm's image-identification Segment Anything model and a tool that speeds up Language Model (LM) response generation times. The tech giant has also released datasets to help discover new inorganic materials, further showcasing its commitment to advancing AI technology.