So how did we get to these amazing videos for Meta Movie Gen? One of the things I’m proudest of is that we released a very detailed technical report (ai.meta.com/research/movie……) Lets dive into a technical summary of what we did & learnt 🧵 1/n x.com/AIatMeta/statu…

This here is the important part: Unified architecture for text and media generation.

François Fleuret It looks like the name "entropy" in information theory was optimized for impact on the field 😉

Dearest LLM script kiddies who are slashdotting arXiv.org - if you're being blocked for being rude, the solution is not to try the same stupid thing again an hour later. You're spoiling the Commons. Please don't make us get irritated at your dumbfoolery.

Really excited to share what I've been working on with my colleagues at Physical Intelligence! We've developed a prototype robotic foundation model that can fold laundry, assemble a box, bus a table, and many other things. We've written a paper and blog post about it. 🧵👇

I'm giving a try to the other blue thing place, though not planning to leave this one as it has its own specific fun. Currently that place does have a vibe of early 2010s Google Plus as I start my bubble with mostly researchers.

If you are a MSc student looking for research internship in AI/CV in Paris, you should definitely check valeo.ai. This place takes a very good care of all of their interns. ❤️

Generative decisions in diffusion models are made at special critical time points. Missing these points with a fast sampler results in loss of diversity. It's like missing an exit in the highway! Paper: openreview.net/forum?id=lxGFG…

I always dreamed of a model that simultaneously 1. optimizes NLL of raw pixel data, 2. generates competitive high-res. natural images, 3. is practical. But it seemed too good to be true. Until today! Our new JetFormer model (arxiv.org/abs/2411.19722) ticks on all of these. 🧵

Better VQ-VAEs with this one weird rotation trick! I love papers like this: a simple change to an already powerful technique, that significantly improves results without introducing complexity or hyperparameters. arxiv.org/abs/2410.06424 (h/t lucidrains)

Rare sincere tweet: December can be tough in academia. As a student I thought everyone had it together. As an advisor you see that is very much not true. Generally, at least as a starting place, a really recommend finding someone who you can go on a long walk with to talk it

Our paper "Function-Space Learning Rates" is on arXiv! We give an efficient way to estimate the magnitude of changes to NN outputs caused by a particular weight update. We analyse optimiser dynamics in function space, and enable hyperparameter transfer with our scheme FLeRM! 🧵👇

Excited to share that I’ve been working on scaling up diffusion language models at Inception. A new generation of LLMs with unprecedented capabilities is coming!

No US Russia or Ukraine policy was ever going to be made or changed today. That is the only conclusion if a buffoon like Vance is permitted to yap and derail a meeting between the US president and a world leader. The UN vote with Russia was the real signal; this was noise.

In some research fields (including AI), it can be common for people to have similar ideas. You'll even talk research w/ people working on ideas similar to yours. You'll often get scooped unless you're working in some extremely niche field. But the mark of a good

We release a large scale study to answer the following: - Is late fusion inherently better than early fusion for multimodal models? - How do native multimodal models scale compared to LLMs. - How sparsity (MoEs) can play a detrimental role in handling heterogeneous modalities? 🧵

It's VLA day with open-source model releases today from both H & Hugging Face LeRobot 🦾🦾🦾 VLA is short for Vision, Language, Action models. These are the models that allow modern robots to see, hear, understand & take action thanks to AI. It's GPT but for

Vision transformers have high-norm outliers that hurt performance and distort attention. While prior work removed them by retraining with “register” tokens, we find the mechanism behind outliers and make registers at ✨test-time✨—giving clean features and better performance! 🧵