NVIDIA GTC 2025 - Building LLM-Powered Applications
Eugene Yan
Eugene Yan and Chip Huyen spoke on a GTC panel about the messy reality of building LLM apps. It's a rare practitioner voice cutting through NVIDIA's keynote hype machine.
Based on reporting by Eugene Yan — read the original for the full story.
Summary, retelling and take written by AI under human oversight; images are AI-generated illustrations. How we work · Report an error
NVIDIA's GTC is usually a parade of new GPUs, new frameworks, and enough marketing gloss to make anyone forget how hard it actually is to ship an LLM feature that works. So it's notable that this year's lineup included a panel titled "Insights and Lessons Learned From Building LLM-Powered Applications," featuring Eugene Yan alongside Chip Huyen, two people who've spent years in the trenches rather than on stage selling silicon.
The panel itself isn't described in granular detail in Yan's writeup, and that's telling in its own way. What's on record is simpler: a session happened, it was recorded, and for now it lives behind a login wall on NVIDIA's attendee portal. Eventually it'll surface on NVIDIA On-Demand for anyone outside the conference bubble to watch. That staggered release is standard practice for big vendor conferences, but it also means the most useful, unvarnished advice from people actually building these systems takes weeks longer to reach the engineers who need it most.
Yan's decision to publish a formal citation for the talk, complete with a BibTeX entry, says something about how he treats his own work. Blog posts about building AI products don't usually come with academic-style attribution instructions, but Yan runs eugeneyan.com like a body of research rather than a stream of hot takes, and that consistency has built him a following of nearly 12,000 subscribers who track his writing on machine learning, recommendation systems, and LLM engineering.
The format matters here more than the content we can currently see. A panel of practitioners swapping war stories about production LLM systems is a very different animal from a keynote demo. It's slower, messier, and full of caveats, which is exactly what makes it useful to the people actually wiring these models into real products.
My take — AI-written commentary, not fact-checked reporting
I'll take a scrappy practitioner panel over another keynote demo any day. NVIDIA's GTC is built to sell hardware roadmaps, so when actual engineers get five minutes to say what breaks in production, that's the part worth chasing down once the recording escapes the login wall.
Read more about this at: Eugene Yan