Artem Frenk

Representation Learning, Diffusion, & Optimization | MS @ NYU Courant

[Artem Frenk @afrenkai]

And somehow only 1 person made the low-hanging 322 joke. Is this happens because we are from SSI?????

[elie @eliebakouch]

What did Ilya see?

[Artem Frenk @afrenkai]

@yuzu_4ever Start only reading journal papers in ML. TMLR & JMLR are (usually) pleasant to read and very informative due to the quality standards upheld. Still not perfect but remember that a preprint has NOT gone through peer review (altho tbf some of the best work never went to a conf!)

[Artem Frenk @afrenkai]

@Keleesssss Yeah super true. I have started using Mathlib + Lean4 for a lot of longer proofs but both Fable + Sol have pretty terrible looking logic to me in my domain. Writing by hand for me is prob faster (bulk of time is reading docs) since LLMs do a lot of trivial/obvious proofs.

[Artem Frenk @afrenkai]

@benglickenhaus @BenDenny 🥲 add the Lando Norris Monster pls

[Artem Frenk @afrenkai]

@henrylhtsang obsidian/notion/whichever foss thing I'm coping on. Integrate comet/wandb into it

[Artem Frenk @afrenkai]

@max_spero_ @xeophon There is a happy medium (for now) imo. Example: AI plots look terrible but I don't want to read all of the docs for strange edge cases. So if I can write the first ~30%, have a model write the next ~40-60% and then add the final touches myself this is perfect for me. Case by case

[Artem Frenk @afrenkai]

@drisspg needed this like 2wk ago 😭. Nice work

[Artem Frenk @afrenkai]

@pangram Goat

[Artem Frenk @afrenkai]

@gaunernst @henrylhtsang Even better is when it recompiles from scratch everytime so you have to give it a dummy location in the pyproject.toml for different cu versions for the cache. who knows which one triggers recompile...

[Artem Frenk @afrenkai]

@eigenron Finding this out the hard way... Why did no one tell me I would have to look at triple integrals I thought we left that shit behind man. Bring back SDEs please

[Artem Frenk @afrenkai]

@eigenron Wait a couple days...

[Sander Dieleman @sedielem]

New blog post: continuous diffusion for language is back!

This research direction receded into the background for a while, but as of this year, it is once again a hot topic. I wrote down a historical perspective and some thoughts on the recent revival.

https://t.co/wNafED6aDB

[Artem Frenk @afrenkai]

@barrowjoseph I have gotten dFlash to work a bit but had to write it myself. Very very rough...

[vik @vikhyatk]

i don’t understand why we need more datacenters. just write faster kernels it’s not that hard

[Artem Frenk @afrenkai]

@cloneofsimo @JulienBlanchon Yes but more conditional infilling like mask git. Like mdlm/bd3lms text diffusion is down to ddpm ~Cat(z), only thing that changes is whether you are in simplex or token space w.r.t pi for the parameterization for the ones in continuous diffusion (~N(z ))

[Artem Frenk @afrenkai]

@purshow04 I would say, initially, that a lot of this is valid but the argument of copying images to DDP ranks in specific might not apply. I am assuming you refer to NaViT or newer implementations, but variable resolution/sequence handling has been possible since FA2 with cu_seqlens.

[Artem Frenk @afrenkai]

@benglickenhaus Didn't pass banger threshold 🥲

[Artem Frenk @afrenkai]

@papayathreesome @ben_golub Я думаию что это меняетсо в других странах. Другие комнани пробайют купить месты па Гоогле. Е-коммерсе всегда также роботал)

[Artem Frenk @afrenkai]

@dingchilling @pangram ty bro