Chart: how compute shifted from pretraining to post-training over two years
japie06 · reddit · 2026-10-11
A Reddit post shares a chart showing how AI training compute allocation has shifted from pretraining to post-training over the past two years, echoing the industry view that pretraining gains are flattening while post-training becomes the main compute battleground.
More from Infra
- Meme: inference bill too high, so the manager decides to self-host GPUs — zainhas · 2026-10-11
- Dev influencer jokes about building data centers in the Dolomites — KevinNaughtonJr · 2026-10-11
- a16z: Agents burn 5x more tokens than humans, up 14x in six months — GregCook2011 · 2026-10-11
- Codebase test: Telnyx-hosted GLM runs 9% cheaper, 5% faster than OpenAI — SucceededMind · 2026-10-11
- Hive Compute Grid MCP auctions solver jobs across io.net, Akash, Render with signed receipts — modelcontextprotocol · 2026-10-11
- Nvidia said to go head-on with frontier labs as labs build their own ASICs — pzakin · 2026-10-11