Twitxy Saved

Alberto Bietti

@albertobietti

Pretraining @MistralAI. Prev @FlatironInst CCM (on leave), @MetaAI, @NYUDataScience, @Inria, @Quora.
New York alberto.bietti.me Joined December 2007
1.7K Following 2.2K Followers 1.9K Posts

Latest posts 5

Alberto Bietti @albertobietti
Don't forget to submit your work on scaling laws to this year's HiLD workshop at ICML! Deadline extended to May 11.
Elliot Paquette Verified account @poseypaquet May 7
The High-Dimensional Learning Dynamics Workshop @ ICML 2026 馃嚢馃嚪, with a special focus on scaling laws, is coming up, July 10!
Submission Deadline: May 11 AoE (extended).

@Locchiu @albertobietti @JustinLin610 @inbarser @BachFrancis @ShamKakade6 @andrewgwils @blake__bordelon @_brloureiro @Qiuzihanhan
View quoted post on X
0 replies 1 reposts 15 likes 3.9K views View on X
Alberto Bietti @albertobietti
Why does Muon work so well in language modeling? Our perspective based on associative memories: early in training, Muon can "pack" heavy-tailed knowledge into a matrix much more efficiently than GD.
A real tour de force lead by @junokim_ai and @EshaanNichani !
Juno KIM @junokim_ai Mar 30
Excited to share our new paper on sharp capacity scaling of the Muon optimizer! Joint work with @EshaanNichani Denny Wu @albertobietti @jasondeanlee:
arxiv.org/abs/2603.26554 (1/7)
View quoted post on X
0 replies 5 reposts 74 likes 7.6K views View on X
Alberto Bietti @albertobietti
Our new paper on understanding hyperparameter transfer: when does it work, when is it useful, and a conjecture for why fast transfer happens in practice, backed by extensive experiments.

Check out the 馃У below by the great @nikhilghosh101 for more details!
Nikhil Ghosh @nikhilghosh101 Jan 8
Sharing our recent work on understanding the mechanisms underlying the empirical success of hyperparameter transfer using 渭P! (1/11)

with Denny Wu and @albertobietti
View quoted post on X