Beyond Autoregressive: Stefano Ermon on Where Diffusion LLMs Win

Описание: Why are diffusion-based LLMs emerging as the next major shift in AI inference — and how can they outperform traditional autoregressive models on speed and cost?

In this fireside conversation, Stefano Ermon, co-founder of Inception Labs and former Stanford professor, joins the moderator to discuss how his pioneering work on diffusion models evolved from image and video generation into high-speed diffusion-based language models.

They explore why classic GPT-style autoregressive models are reaching efficiency limits, how diffusion LLMs can generate tokens in parallel and reach 1000+ tokens per second, and why the next frontier of AI may be scaling inference — not just training. Stefano also shares his journey from academia to founding Inception Labs, why frontier AI companies require venture capital, compute, and data at scale, and how faster, cheaper inference could reshape the landscape for startups, frontier labs, and AI agents.

Topics include the evolution of generative models, diffusion vs. autoregressive architectures, parallel token generation, the “bitter lesson” of parallelism, building an R&D-heavy AI startup, inference cost and latency, context length and scalability, and Inception Labs’ role in a future where LLMs increasingly reason, use tools, and may eventually self-improve.

Key Links:
Inception Labs: https://www.inceptionlabs.ai
Stefano Ermon:   / ermon  
Ash Kumra:   / ashkumra  

Timecodes
01:37 – Stefano’s journey: working on generative models before they became mainstream
02:15 – Inventing diffusion models at Stanford: images, video, and music generation
03:05 – 2024 breakthrough: diffusion-based LLMs matching GPT-style quality at 10× speed
04:05 – How diffusion LLMs differ from autoregressive models: parallel token generation
04:50 – Founding Inception Labs: scaling models commercially and reaching 1000+ tokens/sec
05:24 – Where Stanford research ends and Inception Labs’ work begins
05:41 – New math and algorithms: extending diffusion from images to text and code
06:20 – Inception Labs as an R&D engine: training methods, serving stack, and proprietary IP
06:51 – Industry shift: from training-focused scaling to inference at scale
07:35 – The “bitter lesson” of parallelism: why more parallel solutions tend to win
08:21 – How diffusion changes the efficiency profile of inference workloads
09:16 – Moving from research lab to company: why raise capital and buy data/compute
10:00 – Scale and resources: competing with big labs and hyperscalers
10:38 – From research idea to product: where Inception Labs stands today
11:01 – Looking toward 2027: agents, reasoning, complex tasks, and potential self-improvement
12:04 – Audience Q&A: context length and scalability
12:09 – Answer: transformer under the hood, quadratic attention, and ~130k context window today
13:00 – Three key LLM dimensions: quality, speed, and cost — and their trade-offs
13:30 – How diffusion-based LLMs shift the Pareto frontier on quality, latency, and price
14:05 – Core value proposition of Inception Labs: faster and cheaper inference at competitive quality
14:09 – How to use Inception Labs: website, OpenAI-compatible API, and token-based pricing
15:09 – Closing remarks and thanks