Diffusion LMs were 5-10x faster but never loved — timing, not speed, wins
eliza_luth · x · 2026-09-22
elizaluth offers a timing observation: agents took off while models were powerful but slow, whereas diffusion LMs — 5-10x faster than autoregressive ones — never caught the same love.
The takeaway: ordering matters. Speed alone isn't the hook; what matters is matching the current bottleneck of the field.
More from AGI Musings
- AI researcher: calling multiagent systems 'swarms' reveals our deep unease about AI — jachiam0 · 2026-09-22
- "Intelligence Starts Accelerating Intelligence": Watching the Curve Go Vertical — Dr_Singularity · 2026-09-22
- Researcher: environment design, not just agent training, is key to AI safety — yaringal · 2026-09-22
- AI safety talkers doubled in two weeks, but the sane experts halved — jankulveit · 2026-09-22
- World models need theory of mind: Persimmon pairs user modeling with world modeling — CShorten30 · 2026-09-22
- "I don't want to read what you didn't write": an engineer's case against AI ghostwriting — mooreds · 2026-09-22