Matthew Muckley
@mattmucklm
Followers
965
Following
775
Media
15
Statuses
230
Research Engineer, Meta Fundamental AI Research (FAIR). ML for compression, computer vision, medicine.
New York, NY
Joined December 2018
Very excited to share V-JEPA 2! I've been working on the encoder pretraining pipeline and data curation for this model the last few months, and am excited for it to finally be out!.
2
0
8
RT @mervenoyann: Meta just released V-JEPA 2: new open-source image/video world models ⏯️🤗. > based on ViT, different sizes (L/G/H) and res….
0
82
0
Very strong results on SSv2 and action anticipation, plus zero-shot robotics planning! And we also attached an LLM to the vision encoder and got strong numbers on PerceptionTest!. Check out the blog post (with link to paper and GitHub) above!.
0
0
2
RT @brandondamos: If you prompt an LLM and stop in the middle of a token, what happens? ❌ The generated response doesn't correctly complete….
0
7
0
Tokenization is a limitation of modern LLMs. How to make it better? Find a way to convert your LLM's probabilities to byte-level probabilities! . Details in the thread below! ⬇️. Work done by our great intern @buutphan.
🤔Tokenization (1) makes your LLMs produce odd text when prompts are cut off mid-token? (2) gives you problems when ensembling different LLMs?. 💡We solve both by converting tokenized LLMs into equivalent byte-level LLMs! No training required!. 📎Paper:
1
0
10
One example here by the responsible AI team!.
Our responsible AI team is hiring 3 research scientist interns this cycle (2 in Montreal, one in NYC). We're seeking enrolled PhD students who are excited to spend their summer figuring out how to ensure vision and/or language models work for everyone!.
0
0
0
I should clarify: I am not personally hosting an intern, but I am happy to vouch for all my colleagues that will be hosting interns!.
2
0
2
Research scientist positions at FAIR are open for the coming year! Come join us!.
2
14
158
To use them, please follow the instructions in this README and use either "msillm_quality_vlo1" or "msillm_quality_vlo2" (the README will be updated).
1
0
0
We've just open-sourced a couple of very low bitrate versions of MS-ILLM that were used in our PerCo paper on generative compression. The MS-ILLM models aren't super strong at these rates (they were mostly used to highlight how great PerCo is :) ), but we've gotten some requests.
2
0
3
RT @karen_ullrich: Even with preference alignment, LLMs can be enticed into harmful behavior via adversarial prompts 😈. 🚨 Breaking: our t….
0
20
0
Appearing at ICML this week!.
Preprint and code of my internship @Meta on neural-augmented residual quantization is now online: ⚡️. 🚨We heavily improve SOTA compression and search performance by conditioning the codebook in each residual quantization step on selected codewords so far.
0
0
4
RT @buutphan: Why do LLMs fail simple completion tasks, but not on a harder task? . Learn about tokenization bias in LLMs and how to fix it….
0
11
0
RT @Piovrasca: Are sota image generative models effective world models?. Consistency-diversity-realism Pareto fronts show they're not (yet)….
0
42
0
RT @hall__melissa: Interested in Responsible AI + GenAI? Stop by our poster "DIG In" @iclr_conf on Tuesday @ 10:45a. We introduce indicator….
0
4
0
The result is sharp, realistic images at bitrates as low as 0.0032 bits per pixel!
1
1
2