Meta releases AI model that can check work of other models

meta

Facebook owner Meta on Friday released a series of new AI models, including one that can evaluate on its own with less human involvement.

The new model uses a “thought chain” technique to break down complex problems into smaller steps, Reuters knows. That increases accuracy in science, programming and math.

“Self taught model”

This model is fully trained with AI-generated data, without any human involvement. Thus, Meta provides a preview of autonomous AI agents that can learn from their mistakes. Those agents may soon be intelligent assistants who can autonomously perform a wide range of tasks without manual input.

read also

AI agents: iPhone moment or old wine in new bags?

That would mean that the cumbersome and expensive process called Reinforcement Learning from Human Feedback would be unnecessary. “We hope AI will get better and better at controlling its work so that it eventually becomes better than the average human,” said one of the researchers.

Meta also released other AI tools Friday, including an update to the Segment Anything model for image recognition, a tool that speeds up the response time of large language models, and datasets that can help discover new inorganic materials.