Diffusion LLM & Why the Future of AI Won't Be Autoregressive - Stefano Ermon (Stanford /Inception)

Stefano Ermon of Stanford and Inception discusses why diffusion-based language models could challenge the autoregressive approach behind today's large language models. The episode explains how diffusion models generate by iteratively refining noise, how discrete diffusion works on text with masking as only one possible noise process, and how score-matching mathematics carries over from continuous images. A central theme is inference: producing several tokens in parallel can make diffusion models much faster, which he argues shifts the cost-quality trade-off and is especially useful in reinforcement-learning post-training. The conversation ends with why transformers work so well, the state of AI coding tools and research advice for PhD students.