timur kharisov

DL optimisation researcher at day, amateur artist at night

[timur kharisov @normisdeath]

bright minds and AI are capable of destroying obsolete pharma corporations. i wish i had a chance to participate in this revolution.

[Isaak @isaakfreeman]

We built the lab that's able to go from AI-led drug design to data in 24h.

GPT-8 won't be bottlenecked by intelligence. It needs a biological compute layer.

This is Capable. We're turning AI capabilities into human capabilities--starting with short-sleeper peptides.

[Bastian Hagedorn @HagedornBastian]

Nsight Python 1.0 is here 🎉
GPU performance analysis in a few lines of Python: sweep kernel configurations, collect Nsight Compute metrics, compare implementations, derive custom metrics, and generate DataFrames, CSVs, and publication-ready plots.
https://t.co/9V006bqXeV

[timur kharisov @normisdeath]

western europe can't understand that global warming happened years ago and laws had to be changed

[Aram Hăvărneanu @aramh]

Opposition to AC is mostly a western Europe thing, led by Germany and France. As for what permission is required, it depends a lot on where you are. In my case, I need the permission of:

1. all my neighbours, 2. the local city government (which requires a doctor's notice for why I "need" an AC). 3. quantified proof (i.e. measurements) that I tried other ways of dealing with the problem that didn't work, 4. some dude that comes to measure himself to certify my measurements.

Depending on whether the building is new or old, you have to get the approval of different parties for the second point. Requests in old buildings will get denied because you are not allowed to modify the facade. In new buildings (under the purview of a different authority) will get denied because "the building was already designed to be livable without AC", and you have the onus to prove that in fact it's not livable, which is basically impossible.

Plus of course, if you're renting, the landlord will refuse.

But it's all academic, the primary problem is that the neighbours will not agree.

Even if you have your own detached house it gets more and more difficult to obtain permission for installing an AC. The regulations used to deal mostly with noise levels, now there is some environmental regulation in place that tries to stop you as well.

I say "(western) Europe" because it's very much a cultural things that is common across many countries. The specific laws and regulations differ from place to place, but exist for the same reason. They don't want you to have an AC.

[timur kharisov @normisdeath]

@Creative_Math_ LLM-pretraining is unstable as well (DeepseekV4 report); final val loss criteria may be uninformative, as applying cooldown is only helpful for fine-tuning on the chosen dataset, i.e. post-training capabilities are better without cooldown

[timur kharisov @normisdeath]

@octonion there is one literally called Oil (Нефть) targeted at female customers

[timur kharisov @normisdeath]

@HarmonicMath wdym by software verification? it seems practically impossible due to unknown hardware specs/arbitrary inputs and etc.

[Michal Valko ✈️ ICML 2026 until July 20th 🔥 @misovalko]

Congrats to our student Giorgio Racca on his first paper, accepted at @icmlconf 2026. A learning-theoretic view of model collapse: when LLMs train on their own outputs, replay is benign for uniform generation but provably breaks the weaker notions. @AmartyaSanyal https://t.co/KGvKHV0EHM

[timur kharisov @normisdeath]

effective warmup and high stable lr is all you need! (decay is an optional fine-tuning on pre-training data)

[Jeremy Cohen @deepcohen]

The recent Microsoft AI report noted that too much learning rate decay during pretraining hurts post-RL performance. This is actually just the latest of several papers this year pointing out that small learning rates can be harmful in LLM pretraining. (Thread)

[timur kharisov @normisdeath]

hot summer rainy day + full moon = insomnia 🌝

[timur kharisov @normisdeath]

Collaborating with non-git LaTeX users just got a whole lot easier! https://t.co/iN5KYTP8jO

[Overleaf @overleaf]

Collaborating with non-LaTeX users just got a whole lot easier! 🤝 You can now import .docx (Word) files directly into Overleaf (we'll automatically convert them to LaTeX for you!), and export your work back to Word when you're done.

[timur kharisov @normisdeath]

@yule_gan for non-convex non-smooth landscapes (such as post-training) 0-order optimisation is more robust

[timur kharisov @normisdeath]

@tonysilveti @gowerrobert @GuillaumeG_ @CrichaelMawshaw @aaron_defazio nonsmooth convex rates could predict only lr cooldown phase of pre-training.

[timur kharisov @normisdeath]

at least thorough reviews are rewarded. will attend icml virtually this year.

[timur kharisov @normisdeath]

1st time reviewing A* AI conference
1/6 paper to review have an AI generated reference in the first paragraph of the introduction
other 3/6 papers are also likely AI generated with chatgpt-like proofs and weird experiments
such a strange time to start an academic career in AI...
#icml2026

[timur kharisov @normisdeath]

@YouJiacheng where is this figure from? do they compare final validation loss with constant lr schedule with fixed number of steps?

[timur kharisov @normisdeath]

@VictorTaelin try codex-5.3-spark

[Gautam Kamath @thegautamkamath]

Accepted papers for #COLT2026:
https://t.co/H6YROvhSBe

[timur kharisov @normisdeath]

@YouJiacheng Zurich canton has, check out Zug taxes

[timur kharisov @normisdeath]

@gowerrobert @CV_novel_plume @MLCommons algoperf v2 when?

[timur kharisov @normisdeath]

@__paleologo never understood the point of reading engineering books... it doesn't contain proofs and reading code on paper is kinda strange

[timur kharisov @normisdeath]

@ShiqianMa congrats!