Inception Labs CEO Stefano Ermon bets on diffusion LLMs over autoregressive decoding
No Priors · youtube · 2026-09-20
On a new No Priors episode, Inception Labs CEO Stefano Ermon argues diffusion LLMs are the winning path.
- Today's LLMs are autoregressive: token 10 can't exist until token 9 is generated, making decoding inherently serial.
- Diffusion LLMs generate tokens in parallel, removing that bottleneck.
- The conversation covers the reasoning behind this technology bet and its prospects.
More from Models
- New model ships Day-0 in Diffusers with novel KV-caching and disentangled prefill+decode — RisingSayak · 2026-09-20
- Robot Arm Showdown: Jev Completes Cube Task in 27s vs GPT-6 Astra's 1m11s — multiply_matrix · 2026-09-20
- Leak: Anthropic stealth-testing Opus 5.5 as claude-wafer-eap, release tipped for Tuesday — airesearch12 · 2026-09-20
- HN: Chat-based LLMs replicate the mechanisms of a psychic's con — petrusenko_max · 2026-09-20
- 6 clones of closed-source Jev appear in 2 days, each with a different architecture — IgorCarron · 2026-09-20
- Open-weights Hemmingway-1: a 27B creative writing model claimed to rival frontier models — Lukinator6446 · 2026-09-20