Epoch AI: post-training enhancements yield gains equal to 5-20x compute at under 1% the cost
CFGeek · x · 2026-11-15
An Epoch AI paper (with Open Philanthropy, UC Berkeley, ORCG) introduces "Compute-Equivalent Gain" (CEG): some post-training enhancements can improve LLM performance as much as 5-20x more training compute, at less than 1% of the cost.
The paper taxonomy of compute-free enhancement:
- Tool use: teaching models to use external tools (WebGPT, Toolformer)
- Prompting: steering reasoning via input (chain-of-thought)
- Scaffolding: programs structuring model reasoning (AutoGPT, LATS)
- Solution choice: generate candidates then select/verify (AlphaCode)
- Data: generating higher-quality fine-tuning data (Minerva, self-generated data fine-tuning)
Key takeaway: with frontier training compute affordable to only a few players, post-training enhancements remain a highly cost-effective path to capability gains beyond the Bitter Lesson.
More from Infra
- Upgrading from GTX 1070: Is RTX 3060 12GB Enough for Local Image Gen and LLMs? — Guilty_Elk4385 · 2026-11-15
- Qwen3 at 2.7 bits/weight with Leech lattice quantization beats IQ2_S on quality, loses on speed — Fuzzy-Layer9967 · 2026-11-15
- What 'Compute Multipliers' Mean: The Under-discussed Algorithmic Efficiency Metric in AI — CFGeek · 2026-11-15
- CoreWeave says GPU pricing jumped 30-50% QoQ, above original contract rates — Beth_Kindig · 2026-11-15
- Americans: no data centers in our backyard! SpaceX: no problem — draecomino · 2026-11-15
- The Starship math: 25,000 flights to deploy 250GW of space compute — JOBhakdi · 2026-11-15