Diffusion makes inference look like training—the shape GPUs were built for
victor_explore · x · 2026-09-21
Citing Inception CEO and Stanford professor Stefano Ermon: Transformers replaced RNNs and fixed training, but inference never got that fix—a model can't write token 10 until the first nine exist, so GPUs mostly move weights instead of doing math. Diffusion is that fix for inference: it makes inference look like training, the shape the hardware was built for. The author argues inference speed is an architecture choice, not a bigger-cluster problem.
More from Infra
- Four Intel B70s vs a 128GB MacBook for Local LLMs: A Buyer's Dilemma — Rokett · 2026-09-21
- Jeff Dean says RL plus new EDA tooling could compress chip design from 2 years to 3 months — ycombinator · 2026-09-21
- fal tests Serverless cluster mode for on-demand 1K H200 RDMA clusters — isidentical · 2026-09-21
- Meta data center floods Louisiana school district with $50,000 teacher bonus checks — victor_explore · 2026-09-21
- Agent Substrate open-sourced: agent runtime with 10x container density, sub-500ms resume — rakyll · 2026-09-21
- Huawei's Guo Ping: data centers next need 'book nerds' to route tokens, not coders — teortaxesTex · 2026-09-21