Distillation is necessary but not sufficient, and it drains rivals' capital

fleetwood___ · x · 2026-09-25

In a debate about whether model gains come "largely" from distillation, the author argues distillation is a necessary but not sufficient component of training strong models. Its strategic value: exhausting competitors of significant capital, since human-annotated data is the most expensive part (citing Mercor's revenue as evidence). The author concedes IQ-level gains aren't mostly from distillation, but says the strategy gives Chinese labs an advantage.

Related event: AI Researchers Debate Whether Distillation Drives China's Model Gains(5 posts)→

Original post →

More from Models

Models channel →