Meta releases AI model that can 'self-learn' and 'self-evolve'

(CLO) Meta, Facebook's parent company, announced Friday that it is launching a series of new AI models, including a "Self-Learning Evaluator" that can reduce human intervention in the AI ​​development process.

The announcement comes after Meta introduced the tool in a research paper in August, describing how it uses a “thought chain” technique similar to OpenAI’s new models to make accurate judgments about AI model responses. The technique breaks down complex problems into simpler logical steps, helping to improve accuracy in areas like science, programming, and mathematics.

Meta AI icon. Photo: Reuters

Meta researchers used entirely AI-generated data to train this rating model, completely eliminating human intervention at that stage.

The ability to use AI to evaluate AI itself shows the potential for developing autonomous AI agents that can learn from their own mistakes, according to two Meta researchers.

Many experts in the AI ​​field envision these intelligent digital agents as digital assistants capable of performing a variety of tasks without human intervention.

Self-improving models could eliminate the need for the ‘Reinforcement Learning from Human Feedback’ process, which requires highly skilled experts to label data and verify the accuracy of complex mathematical and written answers. This process is currently very expensive and inefficient.

“We hope that as AI becomes more and more superior to humans, it will get better at checking its own work, even surpassing human proficiency,” said Jason Weston, one of the project’s researchers.

“The ability to learn and self-evaluate is key to developing AI to superhuman levels,” he added.

In addition to Meta, other companies like Google and Anthropic have also published research on the concept of RLAIF, or “Reinforcement Learning from Feedback AI.” However, unlike Meta, these companies rarely release their models for public use.

Cao Phong (according to Reuters)

Source: https://www.congluan.vn/meta-phat-hanh-mo-hinh-ai-co-the-tu-hoc-va-tu-phat-trien-post317675.html