Discussing Compute Bottlenecks at ICML
avt_im · x · 2026-07-20
In a conversation at ICML, the author heard a student of Jimmy Ba say that his biggest bottleneck in doing LLM reinforcement learning research was not having enough compute to run meaningful experiments. If even he faces this, it's much harder for other researchers to push forward.\n\nA follow-up point added that without Chinese open-source toolchains helping to squeeze every drop of performance out of existing GPUs, he likely wouldn't have been able to complete this work or acquire the skills needed for today's cutting-edge experiments.
Related event: Academic Compute Bottlenecks Partly Eased by Chinese Open Source Tools(3 posts)→
More from Infra
- AWS on agent prompt caching: explicit + implicit patterns to slash token costs — AWS ML Blog · 2026-10-27
- Thieves spoofing GPS steal $100M+ in data center cargo since June — Ars Technica AI · 2026-10-27
- Lam Research: weekly token usage up ~250x in 18 months — Beth_Kindig · 2026-10-27
- AWS shows how to build enterprise agentic retrieval on Bedrock with CDK — AWS ML Blog · 2026-10-27
- Encrypted GPT-2 speedrun challenges FHE inference: best is 3.031 s/token, goal sub-1s — tensorqt · 2026-10-27
- Open-source Latch compiles PyTorch into training megakernels, 2.75x faster on H100 — sguada · 2026-10-27