🔥Telegram İfşa

This is really impressive work, and congrats to the team, but as an aside...does anyone else find it weird that "world model" evolved from meaning "the minimal model needed to plan in an environment" to "action-conditional video model"?

Congratulations, Blake Bordelon ☕️🧪👨‍💻! Working with you has been a privilege! I’ll miss having you at Harvard, but I’m excited to see the groundbreaking discoveries your new group will make. Wishing you every success!

Please apply, Hanze is an amazing person to work with

🏙️ Want to spend your summer in London? Applications are open for our 10-week internship in quantitative research 2026. Work with leading researchers on projects in modelling, deep learning and optimisation. Apply here: eu1.hubs.ly/H0mvF8v0 #Internship #GResearch

Excited to announce I'll be starting this September 2025 as a Lecturer (Assistant Professor) at the University of Bath! I will continue my research on deep learning foundations, and am open to ideas for collaborations. (Pictured: Bath. Not pictured: University of Bath)

Super exciting that my former student Edward Milsom is starting his lab at the University of Bath. Ed has led super-fundamental work on representation learning and learning dynamics. And he’s great fun to work with, so I definitely recommend collaborating!

Mr Bloomberg please buy us more GPU thx 🙏

I'm looking for an informal PhD supervisor in LLMs/post-training — any recommendations? My supervisor is leaving academia & the rest of the dep't doesn't work on LLMs, so I'm hoping to find someone external to collaborate with More info 👇, RTs appreciated! 🙏

Possibly one of the best PhD programs in the world if you are passionate about machine learning and AI:

🚨How to build a Self-Improving Coding Agent A free 18-page PDF:

(1/4) Cerebras Hot off the presses 🔥📄arxiv.org/abs/2509.25087 If you're spending $1B to train an LLM, you need to know it’s on track—every step of the way. With optimal AdamW τ + fixed TPP, loss curves collapse to a universal path → an early-warning signal for training.

Beautiful work on pretraining science using scaling collapse to precisely predict, debug, and tune LLM training from small-scale and partial runs. So much insights on going beyond μP!

Thrilled to announce Mistral AI's new formal math team following our $2B funding round. AI for formal math researchers wanted. We offer: - Work on creating the state-of-the-art prover, autoformalizer, and automatic proof agent all in one model - World class team (with

*Scale-invariant attention* by Ben Anson Laurence Aitchison Attention is "scale-invariant" if it acts similarly at different (length) scales. This paper proposes a simple modification to make it scale-invariant & improve length generalization. arxiv.org/abs/2505.17083

i just want to say, i fucking called it re: sqrt(log(pos))

Recent work from Nvidia uses the information gain of reasoning chain as a reward function during pre-training (left most). Our recent work considers a similiar quantity but uses it as a signal for early exiting the reasoning chain arxiv.org/abs/2509.26522

Wooo so Laurence Aitchison and I proposed micro adam for batch size invariance and linear batchsize-learning_rate scaling opt-ml.org/papers/2024/pa… a while ago, and this paper from Google provides an efficient implementation and more in-depth analysis and extensions!

I'm not sure how to put it, but even if you scale your model with different compute, fit a function to these points, and present it on a log-log scale, it's not necessarily a scaling law! Thank you for this matter