Nathan Lambert is still not convinced AI will bootstrap itself

Recursive self-improvement, the idea that an AI system will start upgrading itself and set off an intelligence explosion, has become the field's favorite reason to expect a sudden jump in capability. Nathan Lambert does not buy the fast version. His argument rests on where the bottlenecks actually sit. The research that AI can automate today is narrow, he writes, and current methods are good at solving problems we already know how to pose, not at generalizing to harder unknown ones. Running thousands of AI agents in parallel hits real diminishing returns, and the slow parts of building frontier models, from politics to physical infrastructure, are exactly the parts AI cannot speed up.

What he expects instead is what he calls lossy self-improvement: AI making itself cheaper to run and taking over routine work like monitoring experiments and writing software, which shortens iteration cycles on problems you can already check. That is useful, but it does not push the ceiling on raw intelligence upward the way the explosion story promises. He points to Anthropic's own Claude system card, which says the company does not yet see clear signs of dramatic acceleration beyond the current rate of progress. Post-training stays hard to automate because it needs human judgment about how a model behaves across messy situations.

Why it matters

If your planning assumes a near-term takeoff where AI redesigns AI, Lambert gives you concrete reasons to slow down and budget for steady, compounding gains instead. The near-term payoff is cheaper inference and faster engineering loops, not a machine that rewrites itself past human help.