Training compute cost per task may be mere cents, argues 1a3orn — if models generalize from small runs
1a3orn · x · 2026-11-22
1a3orn argues the compute cost per task during training is likely in the cents — possibly fractions of a cent, or under a dollar — with a key caveat: this assumes models learn facts about large runs from much smaller runs. He compares it to math models, where training cost per task and the cost of using them to prove something like RH are separate quantities.
More from AGI Musings
- Ex-NSA's Perlroth: AI is multiplying cyber risk, exploits now beat patches by 7 days — chrisrohlf · 2026-11-22
- AI Is Transforming Society Because It Works as Advertised — iskander · 2026-11-22
- Szegedy predicts proofs without formal verification won't count within a year — aran_nayebi · 2026-11-22
- Quant firms 'print money' with reasoning models — and early access looks like cheating — scaling01 · 2026-11-22
- Takeoff math: separate training compute from the compute to actually explore — 1a3orn · 2026-11-22
- Two compute costs that decide AI takeoff: training the improver and paying for its exploration — 1a3orn · 2026-11-22