Cooper Saye joins OpenAI to lead recursive self-improvement safety
In brief
- Cooper Saye joins OpenAI as Member of Technical Staff focused on recursive self-improvement safety evaluations
- Role addresses loss-of-control risks as AI systems accelerate their own research and capability development
- OpenAI, Anthropic, Google DeepMind, and Meta investing heavily in RSI safety mechanisms
The safety frontier
OpenAI is building a team whose entire job is to figure out whether AI systems are starting to accelerate their own research capabilities, and if so, whether humans can still maintain control over the process. Loss of control is the specific risk category that these evaluations are designed to address.
Saye announced his move on X and highlighted his enthusiasm for contributing to efforts aimed at accelerating AI research through RSI evals. The shift signals OpenAI's conviction that recursive self-improvement isn't a distant theoretical concern—it's an immediate engineering problem.
Industry convergence on RSI safety
OpenAI, Anthropic, Google DeepMind, and Meta have all been increasingly vocal about the urgent need for safety mechanisms in response to RSI risks. OpenAI is actively recruiting for roles in Recursive Self-Improvement Safety, with job postings based in San Francisco that mirror the scope of Saye's appointment.
The convergence reflects a broader industry recognition: when an AI system can improve its own code, training methods, or architecture without human intervention, the traditional safety playbook breaks down. Humans can't review every iteration if the system is self-modifying faster than oversight can keep pace.
Why Saye matters
Saye's background at Jane Street—one of the most quantitatively rigorous trading firms in the world—suggests OpenAI is prioritizing mathematical rigor in RSI evaluations. His fintech work at Ramp, a corporate card and expense management platform, signals experience with systems that scale and handle edge cases. That combination of precision and real-world complexity is exactly what RSI safety requires.
