Frontier lab compute shifts hard: pre-training falls from 67% to 7%, post-training/RL now 55%
akbirthko · x · 2026-11-10
A chart of Anthropic + OpenAI combined compute by workload (1Q24–4Q26E) shows pre-training's share collapsing from 67% to a projected 7%, while post-training/RL (mostly bandwidth-bound) surges from 5% to 55% and inference holds at 29–39%. Commenting on it, danielrock argues most RSI talk is really the effect of this bandwidth-bound shift plus automating schlep — a whack-a-mole loop of talking to the big world and exponentially generating more wants, shipping x.1, x.2, x.5… until the loop must escape the bowels. The data shows frontier labs' compute emphasis has decisively moved from pre-training to post-training and inference.
More from AGI Musings
- Negotiation and contracting are AI's most underrated use case — and bullish for BTC — curious_vii · 2026-11-10
- A concise summary of the core argument in 'If Anyone Builds It, Everyone Dies' — DavidSKrueger · 2026-11-10
- Prediction: a major lab will solve robotic intelligence by end of 2027 — alexmacgregor__ · 2026-11-10
- "We haven't seen the bitter end of the bitter lesson yet" — andreisavu · 2026-11-10
- AI safety circlejerk: any model behavior gets read as a paperclipper omen — sebkrier · 2026-11-10
- e/acc founder Beff Jezos: Anthropic's power concentration risks regulatory capture — beffjezos · 2026-11-10