Nikita Zhiltsov

Ph.D. in CS / Staff Machine Learning Engineer @LinkedIn 🇺🇸 The opinions are my own. Husband, father of 2 boys

[Nikita Zhiltsov @nzhiltsov]

We shipped semantic search for LinkedIn Hiring Assistant over 1B+ member profiles: free-text qualifications, no relevance labels at launch, only InMail engagement as a proxy.

A few things surprised us. Thread 🧵

Full write-up: https://t.co/onDiQzHBGA

[Nikita Zhiltsov @nzhiltsov]

Other lessons in the post: contrastive post-training alignment beat model size for embedding FT; offline judge replay predicted online A/B direction; relevance supervision at retrieval + engagement at ranking are complementary.

https://t.co/onDiQzH3R2

[Nikita Zhiltsov @nzhiltsov]

We shipped semantic search for LinkedIn Hiring Assistant over 1B+ member profiles: free-text qualifications, no relevance labels at launch, only InMail engagement as a proxy.

A few things surprised us. Thread 🧵

Full write-up: https://t.co/onDiQzHBGA

[Nikita Zhiltsov @nzhiltsov]

Classic recruiter search trade-off: faceted search had coverage but many bad matches; boolean had precision but ~half of queries returned nothing.

EBR-powered semantic retrieval became the best single strategy — faceted-level liquidity with the best pre-ranking relevance.

[Elon Musk @elonmusk]

The SpaceX team is incredible!

[Nikita Zhiltsov @nzhiltsov]

@elonmusk Yet still disappointing voice recognition

[Nikita Zhiltsov @nzhiltsov]

While chatting with a friend recently, I finally managed to pinpoint the one real edge humans still have over AI: taking full responsibility for decision-making. Everything else feels like the moat has already shrunk to almost nothing

[Nikita Zhiltsov @nzhiltsov]

Built my first AI-generated video technology #CareerReel end-to-end at an internal LinkedIn hackathon. Treated it like Lean Startup: ship fast, validate with real viewers, iterate hard. All learnings in 3‑min recap video 👇 #GenAI #Higgsfield #VideoResume https://t.co/RQS0om2KcD

[Nikita Zhiltsov @nzhiltsov]

"High-Dimensional Probability" -- the Second Edition by Roman Vershynin is out! https://t.co/myNN4YIv32 I am a huge fan of the first ed.

[Nikita Zhiltsov @nzhiltsov]

@yaroslavvb Yay, back to SVMs!

[Nikita Zhiltsov @nzhiltsov]

My career reflection with #CareerReel Details: https://t.co/AY43NeFLY1

[Nikita Zhiltsov @nzhiltsov]

At #KDD2025 poster session

[Nikita Zhiltsov @nzhiltsov]

All set for @aclmeeting Vienna! #ACL2025 If you’re into industry-scale embedding-based retrieval, LLMs for search and recsys, LLMs-as-judge, or @LinkedIn’s semantic search and agentic experience for Recruiter Search, let’s connect

[Nikita Zhiltsov @nzhiltsov]

GPU retrieval can rival HNSW / IVFPQ on both latency and recall. @LinkedIn's LiNR (Feed OON recsys) and SJS and @Meta's MoL already deploy exhaustive k-NN at production scale on A100/H100 nodes. See my full post for the technical details ⬇️ https://t.co/LjE7lkZqYV #VectorSearch

[Nikita Zhiltsov @nzhiltsov]

@SamuliGloersen @ylecun @elonmusk Groundless statements from someone who does not have a clue what he is talking about. I am no longer interested in this discussion

[Nikita Zhiltsov @nzhiltsov]

@chrisgpt You’re forcing the open door: it’s widely used in industry, not necessarily for LLMs and particularly for cases of non-stationary distribution like ads ranking. For example, production Google ads https://t.co/SCvrnpS5dn

[Nikita Zhiltsov @nzhiltsov]

@kellerjordan0 Bold statement, but close to the truth. If one has to pick only one name to follow true innovations in this field: it’s Yuri Nesterov. For example, his cubic regularized Newton method from mid-2000s is still waiting for practical adaptation for deep learning use cases

[Nikita Zhiltsov @nzhiltsov]

@SamuliGloersen @ylecun @elonmusk I hope you understand how stupid and arrogant it sounds. I would not question the cognitive and critical thinking capabilities of a nation who launched the first sputnik and piloted space flight

[Nikita Zhiltsov @nzhiltsov]

My two cents to the famous discussion: LLMs imitate reasoning (it’s how they are trained after all), but it works for practical use cases, as we saw @LinkedIn while working with job qualifications, often qualitatively better and surely faster than human. Fake it until you make it