David Fan

[1/9] What happens when you treat vision as a first-class citizen during multimodal pretraining? To find out, we studied the design space of training Transfusion-style models that input and output all modalities, from scratch. Here is what we learned about visual representations, data, world modeling, architecture, and scaling behavior! Paper: arxiv.org/abs/2603.03276 Website: beyond-llms.github.io @TongPetersb, @DavidJFan, @__JohnNguyen__, @ellisbrown, @GaoyueZhou, @JasonQSY, @boyangzheng, @webalorn, @han_junlin, @rob_fergus, @NailaMurray, @gh_marjan, @ml_perception, Nicolas Ballas, @_amirbar, Michael Rabbat, Jakob Verbeek, @LukeZettlemoyer, @koustuvsinha, @ylecun, @sainingxie

Our world modeling team at FAIR is hiring! Do reach out / spread the news!

@purshow04 Thanks for stopping by the poster 😁 Really enjoyed our chat

This was such a wonderful conversation with David @DavidJFan. I’ve gained so many insights and discovered that there are some overlaps in our ideas. I firmly believe that even if Vision faces temporary decline, it will always be the inevitable path leading into the real world!

@JinWeiyang18434 Really enjoyed our chat today; thanks for stopping by! WISE is an awesome benchmark 😁

Really enjoyed my conversation with @DavidJFan at ICML 2026 today! We found tons of overlapping thoughts on pre-training, VLMs, video models, and MoE/MoT design. It’s clear computer vision still has decades of exciting research ahead of us—building more grounded, real-world-aware evaluation tasks is absolutely essential for future advances. So many takeaways from this talk! If you also interested with vision and multimodal, check our work WISE (arxiv.org/abs/2503.07265). It will be showcased at Poster Session 8 #2203 on July 9. Come chat with us if you’re interested!

@ayaannmalik @amilabs @ylecun So glad to have you Ayaan!! Welcome to nyc

I just graduated from Stanford, and surprisingly, I feel happier, more energized, and more excited than ever. The SF bubble can be an echo chamber. In search of more original, creative, and grounded thinking, I’m moving to New York to join @amilabs alongside @ylecun, @sainingxie, and an incredible team researching on world modeling and physical intelligence. Excited to branch out and meet people from all walks of life in NYC. Please reach out :)

see you at ICML in Seoul and the AMI × SBVA mixer on thursday evening! looking forward to meeting everyone :) luma.com/rv0q6b5q

@__JohnNguyen__ and I are presenting the Beyond Language Modeling paper as an ICML spotlight in 30 minutes at the 10:30 AM poster session! I’ll also be at the AMI Mixer on Thursday and hanging around in Seoul until early next week :)

@gberta227 TimeSformer was really influential! I came across it right as I was starting to do research into video action recognition and it made me appreciate how flexible of an architecture the ViT is. You can bake in inductive bias or not, completely up to the use case.

ICML 2026 is coming up, which means it's been 5 years since I presented TimeSformer at ICML 2021 (arxiv.org/pdf/2102.05095). It turned out to be the most cited paper I've published, and it's still the only ICML paper I have. A bit of the story behind it. 🧵👇

@ZiqiPang Congrats Ziqi!!

Career Update: After defending my Ph.D., I am joining Amazon FAR, working on building models for embodied AI. Looking forward to the journey ahead!

@WeijiaShi2 @LukeZettlemoyer @nlpnoah @lucyluwang @uwnlp Congrats Weijia!!

Defended my PhD recently! Grateful to my advisors @LukeZettlemoyer and @nlpnoah, committee members Colin and @lucyluwang, and the many mentors, collaborators, and the @uwnlp community for all the support and friendship along the way ❤️ Slides: Break the language model monolith (drive.google.com/file/d/1na8ofU…)

(1/n) Excited to share our new paper from @AIatMeta FAIR: "EgoBabyVLM: Benchmarking Cross-Modal Learning from Naturalistic Egocentric Video Data" 👶Human infants learn language from sparse, noisy multimodal input. Today's VLMs can't. We built a benchmark + challenge to close that gap. w/ @rust_phillip, @angelvillar96, @juanmiguelpino, @mcxfrank, Emmanuel Dupoux and many other amazing collaborators. 🧵

@garridoq_ Welcome!! So happy to have you 😁

After 4.5 formative years at FAIR, I am thrilled to join AMI Labs as a Member of Technical Staff ! I'm looking forward to pushing the boundaries of JEPA and World Models, alongside an amazing team of familiar and new faces 🚀

@jocelynjshen @GoogleDeepMind Welcome to NYC!

Delayed post-grad update 👩🏻‍🎓but I’ve joined @GoogleDeepMind (NYC) as a Research Scientist in socio-affective systems😊! See ya in the big 🍎🌃

@ayaannmalik @maximsiv @MAliyanIshfaq Congrats!!

failed to be a dropout 💔 more soon 🗽

@_amirbar @imperialcollege @amilabs So glad to have you Amir!! Welcome back ;)

A triple life update: 1. I recently moved from NYC to London 2. I’m starting a new lab at @imperialcollege 3. I’ve joined my friends and mentors at @amilabs as a founding Member of Technical Staff It's time to build 🏗️🧑‍🍳

@raktimgg @amilabs Excited to have you!!

I’m happy to share that I’ve joined @amilabs as a Research Intern for the summer, working on world models! Excited to learn and contribute to the team! 🚀

@artemZholus Might be because people aren't using the same recipes for high-dim diffusion!

In the WM community the disagreement peaks right now and this is exciting! This paper (arxiv.org/abs/2605.23993) claims high-dim representations are bad for diffusion WMs and we should use VAEs instead. Another paper a week before (arxiv.org/abs/2605.06388) says the opposite.

@USCIS @StateDept The entire big tech industry is going to crumble after this 😂 good luck staying competitive in STEM

USCIS is applying long-standing law and prior court decisions to require certain aliens with temporary visas who decide they want to permanently reside in the U.S. to return to their home countries to apply for permanent visas through the @StateDept. We're returning to the original intent of the law to ensure aliens navigate our nation’s immigration system properly. Here’s what you should know: uscis.gov/newsroom/news-…

In Oct last year, Representation Autoencoders provided an elegant solution to unified tokenization for understanding and generation. Today we make them a bit more simple. a bit more general. Result: >10x faster convergence, better reconstruction, better generation. And yes we test them on T2I and world models :) Introducing RAEv2

Today we released the code for our CVPR 2026 paper, Flowception. Flowception bridges fully bidirectional sequence modeling and autoregressive generation by inserting frames via learned order, then denoising them with continuous flow. Website: flowception-meta.github.io Code: github.com/facebookresear…

@__JohnNguyen__ Congrats John!!