Hu Xu (back on robotics)

@Hu_Hsu

FAIR, @AIatMeta, #MetaCLIP (scaling #CLIP data from scratch as the #imagenet moment for web) #llama 3, DINO, SAM3, PE; coined #posttraining in 2019 #AI #GenAI

Bellevue, WA

ai.meta.com/people/7761982…

Replies Reposts Media

Pinned
Hu Xu (back on robotics)
@Hu_Hsu
Aug 3, 2025
The cross-modal multilingual capability in Meta CLIP 2 is naturally from scaling. It is a reflection of the Bitter Lesson(@RichardSSutton) on building a scalable learning environment, and let optimization and scale do the heavy lifting. Inspired by Jensen Huang’s scaling law plot
Jason Weston and 8 others
Jason Weston
@jaseweston
Jul 30, 2025
🌿Introducing MetaCLIP 2 🌿 📝: arxiv.org/abs/2507.22062 code, model: github.com/facebookresear… After four years of advancements in English-centric CLIP development, MetaCLIP 2 is now taking the next step: scaling CLIP to worldwide data. The effort addresses long-standing
3
Hu Xu (back on robotics)
@Hu_Hsu
Aug 22
What a legendary run, colleague @ZeyuanAllenZhu ! Your leave reminds me a few leftover videos of phys. of LLMs series to watch. I am truly motivated by your ambition to support open research on fundamental questions—the kind of work frontier model companies often lack the time,
Zeyuan Allen-Zhu, Sc.D.
@ZeyuanAllenZhu
Aug 21
I've decided it's time to resign from FAIR. 🫡 I'm especially grateful for the compute that made much of my research possible: 400 H100/200 GPUs allocated to me by FAIR; over a thousand H100s borrowed from FAIR Europe's CodeGen team led by Gabriel (@syhw ); thousands more
1
Hu Xu (back on robotics)
@Hu_Hsu
Aug 22
True self-supervised learning hasn’t arrived yet. Today’s SSL still relies on humans to define the learning task, objectives, and training data. Language/class label (Y) is even embedded into the distribution of (X) in SSL. The current SSL is better understood as unimodal
Gabriele Berton
@gabriberton
Aug 22
Finally watched Lucas Beyer's lecture on "Vision in the Age of LLMs" What I found most interesting: 1. This plot is why he gave up on SSL. They trained SSL vs SSL + few labeled samples, the latter won by far. So his focus became how to get cheap labels and use them well [1/3]
Hu Xu (back on robotics)
@Hu_Hsu
Aug 22
100% agree. Filtering is a poor tool for quality control in training data. Our illustration shows the difference between filtering and curating a distribution: x.com/Hu_Hsu/status/… In MetaCLIP 1 and 2, we propose to focus on the training distribution consumed by the model and
Lucas Beyer (bl16)
@giffmana
Aug 22
Filters for 100% obvious garbage like "unknown.jpg" make sense, but many things that may look like obvious garbage actually aren't, for example "2012-03-21.jpg". For any "smarter" filtering, i will disagree except in two scenarios: (1) you simply can't train on everything or (2)
1
Hu Xu (back on robotics)
@Hu_Hsu
Aug 11
Open source is back.
AI at Meta
@AIatMeta
Aug 10
Introducing Muse Glimmer, an open-weight 30B-parameter model optimized for local, always-on agent workflows. Muse Glimmer delivers strong performance on key agentic use cases and benchmarks compared with leading models in its size category, and is designed to run entirely on
1