💼 [Career Update] I feel fortunate to have spent the past 4 years at Meta, working with some of the brightest minds to build Meta’s multimodal foundation models — MovieGen, Emu, and more. It shaped me not only as a researcher but as a person. Now, it’s time for a new chapter. I

Meet our third @MicrosoftAI model: MAI-Image-1 #9 on LMArena, striking an impressive balance of generation speed and quality Excited to keep refining + climbing the leaderboard from here! We're just getting started. https://t.co/33BiNfIjPg

Vision tokenizers are stuck in 2020🤔while language models revolutionized AI🚀 Language: One tokenizer for everything Vision: Fragmented across modalities & tasks Introducing AToken: The first unified visual tokenizer for images, videos & 3D that does BOTH reconstruction AND

🐺 Introducing the Werewolf Benchmark, an AI test for social reasoning under pressure. Can models lead, bluff, and resist manipulation in live, adversarial play? 👉 We made 7 of the strongest LLMs, both open-source and closed-source, play 210 full games of Werewolf. Below is

A picture now is worth more than a thousand words in genAI; it can be turned into a full 3D world! And you can stroll in this garden endlessly long, it will still be there.

what a time we live in

What if you could not only watch a generated video, but explore it too? 🌐 Genie 3 is our groundbreaking world model that creates interactive, playable environments from a single text prompt. From photorealistic landscapes to fantasy realms, the possibilities are endless. 🧵

🚀Introducing Hierarchical Reasoning Model🧠🤖 Inspired by brain's hierarchical processing, HRM delivers unprecedented reasoning power on complex tasks like ARC-AGI and expert-level Sudoku using just 1k examples, no pretraining or CoT! Unlock next AI breakthrough with

Introducing AlphaEarth Foundations an AI model that integrates petabytes of satellite data into a single digital representation of Earth. It'll give scientists a nearly real-time view of the planet to incredible spatial precision, and help with critical issues like food security,

Scaling CLIP on English-only data is outdated now… 🌍We built CLIP data curation pipeline for 300+ languages 🇬🇧We train MetaCLIP 2 without compromising English-task performance (it actually improves! 🥳It’s time to drop the language filter! 📝 https://t.co/pQuwzH053M [1/5] 🧵

Excited to share what I worked on during my time at Meta. - We introduce a Triton-accelerated Transformer with *2-simplicial attention*—a tri-linear generalization of dot-product attention - We show how to adapt RoPE to tri-linear forms - We show 2-simplicial attention scales

Self-supervised Learning (SSL) vs Contrastive Language-Image (CLIP) models is a never-ending battle What about using both? This Google paper does exactly that, and results are really good on many different tasks TIPS is a model trained with a CLIP loss and 2 SSL losses [1/9]

Transition Matching (TM) -- a discrete-time, continuous-state generative paradigm that unifies and advances both diffusion/flow models and continuous AR generation. Claimed to be the first fully causal model to match or surpass the performance of flow-based methods on

We're taking a big step towards medical superintelligence. AI models have aced multiple choice medical exams – but real patients don’t come with ABC answer options. Now MAI-DxO can solve some of the world’s toughest open-ended cases with higher accuracy and lower costs.

Q-learning is not yet scalable https://t.co/hoYUdAAeGZ I wrote a blog post about my thoughts on scalable RL algorithms. To be clear, I'm still highly optimistic about off-policy RL and Q-learning! I just think we haven't found the right solution yet (the post discusses why).

We're excited to launch Scouts — always-on AI agents that monitor the web for anything you care about.

Seedance 1.0 has just claimed the #1 spot on the @ArtificialAnlys Video Arena Leaderboard! 🚀 Proudly topping the charts in BOTH “text-to-video” and “image-to-video”categories, Seedance 1.0 sets a new standard for AI video creation. Stay tuned as Seedance 1.0 will soon be

Is the new king of video generation emerging? ByteDance’s Seed team is showing serious potential with their latest model, possibly outshining DeepMind’s Veo3, which was just announced weeks ago. I have been impressed by the consistent stream of inspiring work from Seed.

new paper from our work at Meta! **GPT-style language models memorize 3.6 bits per param** we compute capacity by measuring total bits memorized, using some theory from Shannon (1953) shockingly, the memorization-datasize curves look like this: ___________ / / (🧵)

A recent clarity that I gained is viewing AI research as a “max-performance domain”, which means that you can be world-class by being very good at only one part of your job. As long as you can create seminal impact (e.g., train the best model, start a new paradigm, or create