Pascale Fung

@SenSanders The existential risk is sitting right before your eyes in a house made by former slaves. The existential risk is all those who support that. And those who voted for misogyny and racism, not to mention stupidity. There is the very present risk.

The existential risk of artificial intelligence.

Met the goat @ylecun Ty @amilabs for the lovely event~ so many lovely amazing people <3

I will be in Rio. I am looking to talk to former and current colleagues and students interested in working on vison language word models and human-agent interactions in the physical world.

Heading to ICLR in Rio 🇧🇷? We’re hosting our first networking mixer on April 24. Meet AMI’s technical team and cofounders, and learn more about what we’re building. Food, drinks, and great conversation included. Register at luma.com/np3x51zh

Be still my trekkie heart…

I am hiring researchers and builders for our #Paris team to build advanced machine intelligence that is fundamentally human-centered. amilabs.xyz

Meta’s former chief AI scientist has long argued that human-level AI will come from mastering the physical world, not language. His new startup, AMI, plans to prove it. wired.com/story/yann-lec…

AI pioneer Yann LeCun's new startup has raised $1 billion in initial funding bloomberg.com/news/articles/…

I am happy to share that I have joined forces with @ylecun and fellow founders as Co-Founder and Chief Research & Innovation Officer at AMI - Advanced Machine Intelligence. I will lead research initiatives that push AI to be genuinely human-centered - AI that perceives, learns, reasons and acts like we do and in our best interest. I am thankful for the trust placed in us and deeply aware of the responsibility we share to making the world a better place through our work everyday. Join us!

People tend to either overestimate or underestimate the power of LLMs. No, LLMs do not know the world first hand but are second hand learners from linguistic, symbolic and visual descriptions of human knowledge of the world. No, LLMs are not just “parrots” trying to string words together, NGrams are. LLMs can make correlations from the words they trained to predict. When you know these two fundamental characteristics of LLMs you would neither be surprised by hallucinations nor overly impressed or reliant on LLMs for everything. #LLM #llmhallucinations #GenerativeAI #LanguageModel

See our paper on arXiv (arxiv.org/abs/2601.10592) -- Action100M: A Large-scale Video Action Dataset by Delong Chen (@Delong0_0), Tejaswi Kasarla (@tkasarla_), Yejin Bang (@yejin_bang), Mustafa Shukor (@MustafaShukor1), Willy Chung (@willyhcchung), Jade Lei Yu, Allen Bolourchi (@AllenBolourchi), Théo Moutakanni (@TheoMoutakanni), and Pascale Fung (@pascalefung). Dataset can be acceesed from this repo: github.com/facebookresear…

We release Action100M, the hero behind VL-JEPA. It is a large dataset with O(100 million) dense action annotations on HowTo100M procedural videos. We hope it serves as a robust data foundation to advance physical world modeling research.

@dimadamen Will add. Thanks.

Curious that the demo you show here uses the video and annotations from the EPIC-KITCHENS dataset, but EK is not cited in your paper. In addition to acknowledging the work of the authors, this way others will miss referring to the results of this benchmark in comparative eval.

Introducing VL-JEPA: Vision-Language Joint Embedding Predictive Architecture for streaming, live action recognition, retrieval, VQA, and classification tasks with better performance and higher efficiency than large VLMs. • VL-JEPA is the first non-generative model that can perform general-domain vision-language tasks in real-time, built on a joint embedding predictive architecture. • We demonstrate in controlled experiments that VL-JEPA, trained with latent space embedding prediction, outperforms VLMs that rely on data space token prediction. • We show that VL-JEPA delivers significant efficiency gains over VLMs for online video streaming applications, thanks to its non-autoregressive design and native support for selective decoding. • We highlight that our VL-JEPA model, with an unified model architecture, can effectively handle a wide range of classification, retrieval, and VQA tasks at the same time. by @Delong0_0 @MustafaShukor1 @TheoMoutakanni @willyhcchung Jade Lei Yu Tejaswi Kasarla @AllenBolourchi @ylecun @pascalefung arxiv.org/abs/2512.10942

My father is gone suddenly on Oct 21st 2025. A painter, director, educator and a pioneer in Chinese animation, he was also a devoted father and husband to the very end, he is missed but will be remembered forever.

Thanks @_akhaliq for sharing! More about our VLWM: - Non-pixel-generative world model that reasons in abstract semantic space - Learned from 20k hours of unlabeled egocentric / web procedural videos with 5.7M action steps - System-2 planning with reasoning by cost-guided plan search Congrats to the whole team! @TheoMoutakanni @willyhcchung @yejin_bang , @ZiweiJi184538 @AllenBolourchi @pascalefung

New Vision Language World Model for planning and reasoning in the physical world

Our research on embodied AI agents that can perceive, learn, act and interact in the virtual and physical worlds. #metaAI #AIAgent #embodied #worldmodel #superintelligemce arxiv.org/abs/2506.22355