Former OpenAI researcher says current AI models can't learn from mistakes, calling it a barrier to AGI

Jerry Tworek, one of the minds behind OpenAI's reasoning models, sees a fundamental problem with current AI: it can't learn from mistakes. "If they fail, you get kind of hopeless pretty quickly," Tworek says in the Unsupervised Learning podcast. "There isn't a very good mechanism for a model to update its beliefs and its internal knowledge based on failure."

The researcher, who worked on OpenAI's reasoning models like o1 and o3, recently left OpenAI to tackle this problem. "Unless we get models that can work themselves through difficulties and get unstuck on solving a problem, I don't think I would call it AGI," he explains, describing AI training as a "fundamentally fragile process." Human learning, by contrast, is robust and self-stabilizing. "Intelligence always finds a way," Tworek says.

Other scientists have described this fragility in detail. Apple researchers recently showed that reasoning models can suffer a "reasoning collapse" when faced with problems outside of the patterns they learned in training.