Summary
In this No Priors episode, host sits down with Stefano Ermon, Stanford professor and co-founder/CEO of Inception, one of the researchers credited with developing diffusion models. Ermon traces his research path from early generative models of images in 2014, through GANs and score-based generative models, to founding Inception, a company betting that diffusion-based language models can outcompete today’s autoregressive LLMs on speed.
Ermon explains the core technical argument: diffusion models generate text by refining from noise rather than token-by-token, which he believes can deliver major inference speed and efficiency gains without sacrificing quality — a growing priority as AI companies hit compute and power constraints. He discusses which use cases already demand this speed, citing voice-agent customers like Open Call, where latency across the ASR-to-LLM-to-text-to-speech pipeline directly affects user experience.
The conversation covers why efficiency, once a secondary concern behind pure capability scaling, is becoming central to frontier labs’ compute strategy, and why Ermon believes the field hasn’t yet found the best way to build these systems. It’s a substantive look at diffusion as a genuine architectural alternative to autoregressive LLMs, from one of the researchers who helped invent it, aimed at listeners interested in model architecture and inference economics.
📺 Source: No Priors: AI, Machine Learning, Tech, & Startups · Published September 18, 2026
🏷️ Format: Interview
