Although current AI large models have shown astonishing performance in logical reasoning and task processing, a long-neglected technical flaw may be becoming a major obstacle on the path to artificial general intelligence (AGI). Recently, Jerry Tworek, a core researcher who participated in the development of OpenAI's inference models o1 and o3, stated in an interview that current AI models are fundamentally unable to learn from their mistakes.

Tworek pointed out that when existing AI models fail to solve a problem, developers often feel helpless because there is currently no effective mechanism for the model to update its internal knowledge base or beliefs based on failure experiences. He described the current AI training process as "extremely fragile," and once faced with new challenges outside of the training patterns, the model is likely to fall into a "reasoning collapse" situation. This contrasts sharply with the robust learning method of humans, which has self-stabilization and self-repair capabilities.

It was precisely to overcome this challenge that Tworek recently left OpenAI and is now focusing on finding a technical approach that allows AI to autonomously solve difficulties and get out of困境. He emphasized that if a model cannot evolve on its own when facing setbacks, it cannot be considered true AGI. In his view, the essence of intelligence should be like life itself—always finding a way out. However, current AI is clearly far from reaching this level.

Key Points:

  • 🧠 Core Defect: Former OpenAI researcher Jerry Tworek points out that current AI models lack a mechanism to learn from failures and update internal knowledge, unlike humans who can correct errors and improve.

  • 🚧 AGI Barrier: The "fragility" of models when dealing with unknown patterns and their tendency to "get stuck" are seen as major technical barriers to achieving artificial general intelligence (AGI).

  • 🛠️ Expert Movement: To solve this fundamental issue, Tworek, the behind-the-scenes contributor to the o1 series models, has left and is now working on developing a new generation of AI architecture that can autonomously escape difficult situations.