The SlothAI argument: a lab whose researchers think 50x less would be far slower, so compute alone won't cap faster AI researchers.

The claim

Dwarkesh Patel's "SlothAI" footnote (13 Nov 2026, adapted from Greenblatt and Lifland) argues compute will not cap a research speed-up: a lab with today's compute but researchers working one hour a week would make algorithmic progress at least ~5x slower, so researcher thinking time matters even at fixed compute; and since "Human thinking speed is a biological accident", humans are unlikely to sit exactly where faster thinking stops helping. Dan Robinson (Paradigm), who had hoped compute would bottleneck RSI, called it a sledgehammer. Weak points: the 5x is an intuition, not a measurement; slowing researchers down shows labour matters, not that speeding them up keeps paying (the relation need not be symmetric); and experiments still cost compute, so faster thinkers may mostly queue for GPUs.