Why diffusion models could win AI inference — Inception's Stefano Ermon

Key points

Key takeaways from a No Priors interview with Inception Labs co-founder & CEO Stefano Ermon (September 2026):

Transcript unavailable — summary based on title and description only.

The bet. Ermon, a Stanford professor and diffusion pioneer, argues diffusion architectures will win AI inference as generative AI hits hardware and latency bottlenecks.

Beyond images. Inception is applying diffusion beyond images and video into discrete text and code generation, challenging the autoregressive LLM paradigm.

Scaling edge. Parallel token generation in diffusion models offers superior inference scaling and hardware utilisation on standard GPUs, per Ermon.

Mercury models. He shares details on Inception's Mercury models and real-world voice agent applications already running on them.

Systems work. The episode covers the software stack required to serve diffusion-based models at scale, a gap between research promise and production reality.

Efficiency era. Ermon argues the next era of AI competition will be defined by efficiency, not raw model size, and discusses academia's continued role at the frontier.

Read more: No Priors — Why Diffusion Will Win AI Inference