How Close Is Self-Improving AI? OpenAI & Thinking Machines Researchers Explain
Sarah Wang, Mark Chen, John Schulman
OpenAI's Mark Chen and Thinking Machines' John Shulman characterize the current landscape of recursive self-improvement as a "centaur era" where AI acts as an accelerant for human-led research pipelines rather than a fully autonomous system. Despite significant progress in automated alignment fixes and a reported 2x speedup in research velocity at Anthropic, experts identify critical barriers such as a deficit in "research taste" and the inability of models to generate novel hypotheses without human intuition. Looking toward the next decade, the field anticipates a transition to full automation driven by a two-phase acceleration strategy, though researchers maintain that human judgment remains essential for validating open-ended discoveries and defining safety protocols.