Lossy self-improvement

In three linesNathan Lambert argues self-improvement in AI models is real but doesn't trigger fast takeoff. Improvement is gradual with diminishing returns, contradicting exponential growth scenarios.

Summary generated by Claude — human-verified