Introducing MilliVid, our new method for long-context video generation! MilliVid creates videos that are consistent over long time spans, without using retrieval heuristics or 3D maps! (1/n) davidcharatan.com/millivid/#

Yet, at the same time, my key criticism of this report is that it focuses only on the short-term. Europe is in the position it is in today b/c it has never seriously built a research -> R&D -> product pipeline. The US has numerous mechanisms for seeding research topics that eventually emerge as products *decades* later: Consider the DARPA Grand self-driving challenge, which, more than *20 years ago*, planted the seed of autonomous driving research. A decade later, the US had an ecosystem of talent and experience to rely on when this technology started become ready for prime-time, and today, we have Waymo. Further, the US is based on *concentration* of talent and capital, something that is very much antithetical to Europe, but also to my home country Germany, which - due to an ancient political tradition - chooses to distribute resources across the whole country, where the US fosters hubs such as MIT & Harvard or the Bay Area to ensure a critical mass of R&D that is critical to actually get ambitious projects off the ground. I believe that the German way is better for society *if there is enough resources to go around*, unfortunately a situation that is very much not true in Europe anymore. In my opinion, for Europe to have a shot would require to invest into a single university and surrounding industry and startup ecosystem that concentrates talent and resources in a desirable location. There will be another revolution *after* this current "AI" wave has blown over, in 10-15 years from now. But if Europe doesn't change course, we will miss that one, too.

ArXiv paper now available here: arxiv.org/abs/2606.09056

A really cool idea! The question of how we can train sequence models such that they remember things that are T timesteps in the past without backpropping through T timesteps remains one of the core problems in ML, and this looks like an inspiring approach!

My students @RyuHyunwoooo and @evnkimm are presenting their paper “Scaling View Synthesis Transformers” today at 11:45 am at poster session 5. They are also brilliant and can chat about lots of things in the broader embodied intelligence landscape. Come by!!

I am incredibly grateful to be awarded the PAMI Young Researcher award. CVPR this year was amazing fun; I am very excited to be part of this community and feel honored for this vote of support in my students' and my work :) These are exciting times and I can't wait for next year!

Thanks a lot, Dima, I am so grateful to be in a research community with folks like you around!

If you’re @CVPR: come by our tutorial tomorrow, June 4th 8:30-5:00, on Analytic Understanding of Diffusion Models. We’ll be covering how and why diffusion models generalize, learning about state-of-the-art analytical theories for their behavior, and covering key open questions.