Finally, I found somebody who shares my soul: "Транспонирование создает опасную иллюзию того, что Kn естественно изоморфно n K. В действительности, этот изоморфизм совершенно не является естественным."

Again in my twitter feed I see this discussion about number of papers versus quality of papers and how all this leads to positions and salary. But what about Abel? Wasn't his research papers good enough?

Never gets old

A year ago, I left DeepMind to chase a crazy dream with Spellbrush: to bring anime characters to life. Last week, my colleague and I put our in-house AI models to the test. We made this short anime from scratch in 48 hours. 1/3

Cinema plot: a scientist came to conclusion that machine learning research will at the end lead to new materials for climate save, but on the way this search will accelerate climate change due to go gpu going brrr and therefore it would be too late. Shocked, he became a Luddite.

It is natural to think about each process as causal, ie perturbation that leads to something is in the past. So it feels that a lot of things are fix points and there is no start and finish.

If you think this is good, just wait a bit! Yesterday academy society bun papers from Russian universities. Tomorrow cut partnership with Israel institutions. And believe me if you think you country has 0 probability to do some shit you are wrong. Or may be just Trump can ban all

It feels awful to write a rejection review. :( I can see that paper is not ready in all ways, but still people worked on it, spend their life. Even if the work is bad the time was spend, some hope may be lives in hearts.

Workshop: "It's happening again!" Associative memory: youtube.com/watch?v=JPdZNC…

Okay

Best

Check out my first PhD paper!

I always surprised how Francis Bach does this last mile from theoretical to practical considerations. This was always a problem from my point of view in "bayesian" approach in deep learning, connecting this and that is good but so what?

Жив выходит

(1/n) Introducing KaVa (arxiv.org/abs/2510.02312) – the first latent reasoning framework leveraging compressed KV-Cache to guide the latent generation. We beat previous approaches, especially on a realistic, Natural Language GSM8K dataset.

(5/n) Shoutout to my wonderful co-authors: Anna Kuzina, Paul Whatmough, Babak Ehteshami Bejnordi, who I worked with on this project during my internship at Qualcomm. I will definitely miss Amsterdam…