Yury Malkov

PhD in Laser physics, now doing CS.
Author of HNSW (approximate nearest neighbor search).

[Liam Fedus @LiamFedus]

GPT-4o is our new state-of-the-art frontier model. We’ve been testing a version on the LMSys arena as im-also-a-good-gpt2-chatbot 🙂. Here’s how it’s been doing.

[Matthias Niessner @MattNiessner]

Excited to share the #ICCV 2023 paper compilation of our TUM Visual Computing Lab & collaborators: https://t.co/sbtgMWPIRZ

Thanks so much everyone for all the hard work. It's going to be an awesome conference - really looking forward to Paris next week!

[Artem Sevastopolsky @ASevastopolsky]

Our paper has been recently accepted to ICCV '23, and it is my first accepted 1st-author paper to such a top-tier conference. Thank you @MattNiessner!

It was a fruitful collaboration with @MalkovYury, @nikitadurasov, and Luisa Verdoliva -- big thanks for their help and insights.

[Yury Malkov @MalkovYury]

@TheXeophon @cto_junior Not sure I understand. You can add, remove or update points in hnswlib without rebuilding the graph

[Yury Malkov @MalkovYury]

@pavan_ky It looks like it is mostly popularity-based for me.

[Yury Malkov @MalkovYury]

@jobergum Got it! I think this is not universal. At least there are works like https://t.co/zsrQKaKbc4 that say transformations might hurt

[Yury Malkov @MalkovYury]

@EmreSevinc @srchvrs And if the gas and oil is cheap, you buy it no matter where it came from?

[Ying Xiao @YingXiao]

Here's some actual insight into the Twitter algorithm: there is no NLP embedding inside the main ranking algorithm. https://t.co/E7feowtjPn i.e., Twitter ranks tweets without even really knowing what's inside them. Please stop the whole "replies boost 20x more than fav" thing.

[Yury Malkov @MalkovYury]

@ChrisKortge @OrwellsShaddow The later logic is explained in the HNSW paper. NSW struggled from log^2(N) which turned out to be due to degree hub scaling (that had a separate paper https://t.co/n9WrXwaxbn). And given NSW and skiplist are basically the same thing but skiplist had capped degree led to HNSW.

[Yury Malkov @MalkovYury]

@ChrisKortge @OrwellsShaddow NSW had an issue in 1D with load balancing (first nodes had the most load), so we look for the published distributed alternatives, which was skipgraph based on skiplist (also mentioned in the HNSW paper btw). That is how I learned about it.

[Yury Malkov @MalkovYury]

@ChrisKortge @OrwellsShaddow Also from the same source "The Hierarchical NSW idea is also very similar to a well-known 1D probabilistic skip list structure [27] and can be described using its terms." which gives away the inspiration. Sorry, it was not your work.

[Yury Malkov @MalkovYury]

@ChrisKortge @OrwellsShaddow Yes, I never saw your patent (until you referenced it now). I have no idea what is "mystery shell" who contacted you and I was never involved in buying/evaluated patents in MeraLabs.

[Yury Malkov @MalkovYury]

@alth0u Come join Josh :)

[Yury Malkov @MalkovYury]

@alth0u true story

[Ying Xiao @YingXiao]

Claiming 95% is too much (substantial engineering and modeling contributions from @alykhantejani, Yoan Russac, and others), but there's no denying @alth0u's critical contributions to Twitter Home Timeline Recommendation.

[Yury Malkov @MalkovYury]

@ChrisKortge @OrwellsShaddow If we were to extend authorship claims in such a manner, your work could be credited to William Pugh, who invented the skiplist in 1989.

[Yury Malkov @MalkovYury]

@ChrisKortge @OrwellsShaddow Even ignoring the fact I didn’t know your work and it had no opensource implementation, HNSW has a distinct fast parallel incremental construction (compared to O(N^2)) with an exact transition skiplist in 1D with major both practical and theoretical implications.