Inside the RSI Split: Anthropic Safety Lead Says No Scientific Plan, OpenAI Chief Scientist Expects Progress
The research community’s RSI debate went public this week, and the gap between capability expectations and safety readiness is now documented in direct quotes from researchers at both frontier labs.
Anna Wang, an AGI safety researcher at Anthropic, stated: “There is not yet a viable scientific plan to solve risks from recursively self-improving AI. Please look up!”
Jakub Pachocki, OpenAI’s chief scientist, wrote in “An Alien Mind” — his essay published September 6 — that he has “a strong expectation that this speed of progress could be sustained into recursive self-improvement.” The same essay framed RSI as a dangerous threshold requiring extreme caution, called for voluntary slowdowns, and advocated for shared safety bars and international coordination.
Both quotes are real. Their coexistence is the story.
What Is Actually Being Debated
Pachocki’s expectation is about trajectory: the rate at which AI capabilities are advancing suggests RSI is a foreseeable continuation rather than a distant hypothetical. His caution is about what happens when it arrives. Wang’s statement is about the state of the science: the safety tooling to evaluate, contain, and correct a recursively self-improving system does not yet exist at the level the problem requires.
Both Anthropic and OpenAI have formal frameworks that address RSI. Anthropic’s Responsible Scaling Policy (RSP) tracks self-improvement as a capability threshold. OpenAI’s Preparedness Framework covers AI self-improvement under its evaluation criteria. Wang’s point is not that these frameworks are absent but that the scientific solutions within them remain incomplete.
The Podcast That Made It Visible
The exchange gained attention through a Dwarkesh Patel podcast released September 11, featuring researchers John Schulman, Beren Millidge, and Charlie O’Neill discussing RSI timelines and risk. CNBC reported that OpenAI and Anthropic researchers are “ramping up calls for slowdown amid AI progress,” citing multiple internal voices asking for more time to resolve alignment questions before RSI-capable systems are deployed.
Schulman, who co-founded OpenAI and is now Chief Scientist at Thinking Machines Lab, was joined by Beren Millidge (CTO of Zyphra) and Charlie O’Neill (Head of Model Training at Baseten). The guest list signals that the RSI safety debate extends well beyond the two frontier labs driving the capability.
The Uncomfortable Math
The capability timeline and the safety timeline are not converging at the same rate. A companion OpenAI post, “Research acceleration: The view inside OpenAI,” documented that AI systems are now handling research tasks that previously required a few days of skilled research work. If that rate continues, RSI-capable systems arrive before the safety community has a tested, scalable plan for managing them.
That is not a controversial claim. It is Wang’s claim. It is also, implicitly, Pachocki’s: he called for caution precisely because he expects the capability to arrive.
OpenAI has not announced changes to its deployment timeline or internal RSI research schedule following the public debate. Anthropic has not publicly modified its RSP thresholds or timelines.