Devansh Tandon works on foundation models and large-scale recommendation systems at Meta. His research and talks explain how content can become a compact vocabulary that lets a language model select videos and respond to stated user preferences.[1][2][3]
Start with Teaching Gemini to Speak YouTube →
X Career
Tandon is a principal product manager at Meta. Previously at Google and YouTube, he worked on adapting Gemini for video recommendations. His conference biography describes study in computer science and economics at Yale and seven years at Google across advertising, Search, Discover, YouTube and Google DeepMind. The common product question is how to connect a large content catalog with the interests of an individual user.[1]
Contributions
Identifiers that carry structure
Tandon coauthored Better Generalization with Semantic IDs. Conventional random item identifiers give a model little information about the relationship between two videos. The paper learns compact, discrete identifiers from content embeddings, so related items can share structure. It then adapts pieces of those identifier sequences for a YouTube ranking model. The experiments report better generalization to new and infrequent items without sacrificing overall quality in that setting.[3]
A model that speaks the catalog’s language
In his YouTube talk, Tandon explains adapting a language model to understand both English and video identifiers. Training connects identifiers to descriptions and viewing sequences. To limit serving cost, some next-video candidates can be generated ahead of time and retrieved through inexpensive lookups. That architecture separates the cost of preparing recommendations from the cost of returning them to a user.[4]
Ideas
His later Meta talk argues that language-aware recommendation could let a user add an interest directly instead of communicating it only through clicks and watch time. A model combines that request with the existing interaction history, then selects content from a catalog. He distinguishes better offline model quality from engagement or revenue observed after a production launch. The broader consumer-AI opportunity is his thesis, not a settled forecast.[2]
Tools, writing & talks
Writing, research & talks
Start here
Teaching Gemini to Speak YouTubeai.engineerExplains adapting a language model to video tokens and serving recommendations at large scale.
Begin with the YouTube talk for the practical sequence from video content to tokens, training and low-cost serving. Basic familiarity with recommender systems helps, but it introduces the representation problem before the research paper’s details.
Speaker[4]A talk on semantic catalogs, language steering and the economics of selecting existing content.
Speaker[2]Original sourceResearch on content-derived identifiers that help ranking models handle new and infrequent items.
Coauthor[3]