OpenAI's Former Research Chief Argues Current AI Models Hit Critical Limits in Basic Tasks

Former OpenAI research leader Jerry Tworek contends that the most significant constraint holding back artificial intelligence is not what today's models can accomplish at their peak, but rather their consistent failures when confronted with seemingly elementary challenges. Speaking at The Information's AI Agenda Live SF conference on September 23, 2026, in San Francisco, Tworek, now CEO of freshly launched startup Core Automation, articulated a fundamental flaw underlying the latest generation of reasoning systems: models cannot effectively learn from their mistakes.