AI Engineer 2025 - Improving RecSys & Search with LLM techniques
Eugene Yan
Eugene Yan hosted the RecSys track at AI Engineer World's Fair 2025 and posted his slides plus the full session recording. It's a rare inside look at how teams are actually using LLMs to fix search and recommendations, not just talk about them.
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
Conference talks about recommendation systems don't usually get much attention outside the RecSys crowd, but Eugene Yan's recap of hosting the RecSys track at AI Engineer World's Fair 2025 in San Francisco is worth a look for anyone building search or personalization products right now. Yan, who writes one of the more technically grounded newsletters in the ML space, put together the opening slides for the track and has now published both those slides and the full YouTube recording of the sessions that followed.
The framing matters more than it might seem. RecSys and search have historically been solved with embeddings, collaborative filtering, and a pile of ranking heuristics tuned over years. LLMs are now getting pulled into that stack, not to replace the old machinery entirely, but to patch the parts that always felt brittle: cold-start problems, query understanding, re-ranking, and generating richer signals from sparse user behavior. Yan positioning an entire conference track around 'improving RecSys and search with LLM techniques' signals that this is no longer a fringe experiment. Enough teams are doing it in production that it warranted a dedicated room full of practitioners at one of the bigger AI engineering gatherings of the year.
What's notable is that Yan didn't just drop a summary and move on. He's asking people to formally cite the write-up if they use it, complete with a BibTeX entry, which is the kind of thing academics do and industry practitioners almost never bother with. It's a small detail, but it says something about how seriously this corner of applied ML is starting to take its own output. RecSys used to be the part of the stack nobody wrote papers about. That's shifting.
For anyone who couldn't make it to San Francisco, the real value here is the recording itself rather than the announcement. A full track's worth of talks on applying LLM techniques to search and recommendations, curated by someone who's spent years in the trenches of production ML at places like Amazon, is not something you get every week. The slides are a preview. The recording is where the actual substance lives.
My take — AI-written commentary, not fact-checked reporting
I'll take a practitioner's slide deck over another lab's benchmark press release any day of the week. RecSys is the least glamorous, most economically important part of the AI stack, and it's been quietly absorbing LLM techniques while everyone else argues about AGI timelines. The teams shipping real revenue impact right now are the ones bolting LLMs onto boring recommendation pipelines, not the ones chasing frontier model headlines.
Read more about this at: Eugene Yan