Why I still haven’t bought into true RSI

In an Interconnects essay published September 19, 2026, Nathan Lambert laid out why he has not adopted the view that true recursive self-improvement (RSI) is imminent. He wrote that a few organizations, specifically frontier labs OpenAI and Anthropic, are using thousands of concurrent agents to improve processes and output, and that this environment amplifies AI concern. He said the step from that anxiety, and incidents such as OpenAI-HuggingFace, to extinction risks feels very religious, and that increased discussion of extinction risk seems very misplaced.

Lambert described his alternative scenario, which he calls "lossy self-improvement." Its stated components: automatable research is too narrow to achieve a massive net acceleration in progress given scaling laws' exponential costs; diminishing returns of more AI agents in parallel are real; and resource bottlenecks and politics are a major factor in building strong LLMs, with AI able to do much less to accelerate that. He holds high uncertainty, saying foundational, imagination-based breakthroughs would make him update his RSI timelines.

He cited Dwarkesh podcasts with Noam Brown and with John Schulman, Beren Millidge and Charlie O'Neill. On when AI reaches certain abilities, relative to the interview date: on a drop-in remote worker for broad white-collar work over a month, O'Neill said about one year with programmatic access to workplace tools, or about two years through a browser; Millidge said about three years for full generality, with 80-90% coverage sooner, citing online learning and the long tail of tasks; Schulman said about one year for an "okay" version. On a 10x productivity uplift for AI researchers, O'Neill said five to ten years, Millidge called Schulman's roughly two-year estimate plausible but gave no independent timeline, and Schulman said about two years. On AI surpassing top human experts across all computer-based work, including multiyear projects, O'Neill said five to ten years, Millidge about five years for areas labs focus on and potentially longer for every domain, and Schulman three to four years.

Lambert argued RSI is more helpful for efficiency than peak intelligence, and quoted Schulman on post-training being hard to fully automate because someone must decide how the model should behave, and on it being easy to screw up in ways benchmarks do not show. He cited the Claude Fable 5.1 & Mythos 5.1 System Card: "We believe that internal usage of recent AI models has been a key factor in maintaining the current rate of progress, but we do not yet see clear signs of dramatic acceleration beyond that rate." He concluded lossy self-improvement remains his baseline.