AI at Meta's Recent LinkedIn Posts

A few days ago, Yann LeCun's new venture — AMI - Advanced Machine Intelligence — raised $1.03 billion. Not to build a better chatbot. Not to generate prettier images. But to build what he believes today's AI cannot become. His conviction: AI will never truly understand the world by predicting the next word. It needs to learn how reality works — the way a child learns that objects fall, that water flows, that pushing something makes it move. He co-founded AMI Labs to build this. His approach, called JEPA, works like this: build an internal model of the world and reason within it, rather than relying on pixel-level generation as the main learning objective. Eric Xing — President of MBZUAI (Mohamed bin Zayed University of Artificial Intelligence), Professor at Carnegie Mellon University — shares the conviction that AI needs world models. But he sees a fundamental limitation in reasoning purely in abstract space. His research suggests that if a model only reasons internally without checking its predictions against what actually happens, its representations risk drifting away from reality. His team built PAN, a world model that combines latent prediction with video reconstruction so that reasoning remains anchored in the observable world. One says: understand the world in your mind. The other says: understand it, but keep your eyes open. ↓ Next week they discuss this live. Position talks. Rebuttals. No script. As Eric Xing wrote: "Don't expect a shouting match or drama." What to expect instead: two scientists who clearly respect each other, confronting a question that may define the next era of AI. March 23–25 · UM6P · Ben Guerir, Morocco Ecole Marocaine des Sciences de l'ingénieur EMINES - School of Industrial Management / Mohammed VI Polytechnic University Chair of Industrial Processes École Polytechnique OCP Group #WorldModels #JEPA #AIResearch #AIForImpact

Merci Julien ! Enhance Lab is sharp, as we can all see 😄

We’re live now. Join the discussion 👇 : https://lnkd.in/gTx3_tsM

👩‍💻 AI Pioneers #IWD2026 À l'occasion de la Journée internationale des droits des femmes, nous mettons chaque jour en avant une femme qui nous inspire dans l'écosystème de l'IA en France Aujourd’hui, focus sur celle qui structure l’excellence de la recherche française : Isabelle Ryl. Directrice de l'institut PRAIRIE et figure historique de l'Inria, Isabelle est une architecte de l'écosystème IA. Ce qui nous inspire chez elle : - Son leadership stratégique : Elle pilote l’un des centres de recherche les plus prestigieux au monde, faisant de Paris un hub incontournable de l’IA mondiale. -Le lien Recherche-Industrie : Elle ne voit pas l'IA uniquement en laboratoire. Elle œuvre sans relâche pour que les découvertes scientifiques se transforment en innovations concrètes pour les entreprises et la société. -Sa vision de l'excellence : Professeure d'informatique, elle incarne la rigueur et la transmission. Elle prouve que pour diriger l'innovation, il faut d'abord en comprendre les fondements les plus complexes. Pour Isabelle, l’IA est une aventure collective et interdisciplinaire. Elle est la preuve que le leadership féminin est indispensable pour coordonner les talents qui dessinent notre futur technologique. Merci Isabelle de porter si haut les couleurs de l'IA ! #AIPioneers #WomenInAI #IWD2026

j'aime bien l'image.

Feel the AGI 😅 [I suspect the thing was fine-tuned with an example I used in some of my talks/podcasts/interviews in which I used a sheet of paper, not a rigid object like a pen] LLMs mostly memorize knowledge and retrieve answers, which is one reason they need to be so big, with so many parameters. They don't understand reality at a deep level. They don't really have a mental model of reality that allows them to predict outcomes. They do work well in domains where the language itself serves as a substrate for reasoning, such as coding and mathematics.

Intelligence is not a collection of skills nor an accumulation of declarative knowledge. Intelligence is the ability to accomplish new tasks with no prior training or with fast training. This points to the necessity of System 2, world models, and planning. https://lnkd.in/eBky2ayw

A regularizer to train JEPAs so that trajectories in representation space have minimum curvature. This helps gradient-based planning.

Excellente interview de notre inégalable CEO Alex LeBrun parlant d'AMI Labs sur BFM TV. https://lnkd.in/evhg-66i

We’re looking forward to an important discussion on the future of world models on March 25th at the Spring School - AI For Impact 2026, featuring MBZUAI President and Professor Eric Xing and Executive Chairman of AMI - Advanced Machine Intelligence Professor Yann LeCun. This session will explore two emerging paradigms shaping next-generation AI systems: generative latent prediction (GLP) architectures and joint embedding predictive architectures (JEPA). Both move beyond token-level autoregression, but diverge in how they represent and predict the world. At MBZUAI's Institute of Foundation Models, our work on GLP focuses on learning structured latent variables that explicitly model dynamics and uncertainty, enabling systems to generate, simulate, and reason over possible futures. This generative framing allows for planning, counterfactuals, and interactive adaptation. Rather than a binary choice, we see this as a critical moment to examine where these approaches diverge, where they may converge, and how they collectively advance the development of scalable, general-purpose world models. For those interested in the technical foundations behind GLP, you can read more here: https://lnkd.in/dpG8Dr_B If you're working on next-generation AI systems, this is a conversation you won’t want to miss. Register to attend here: https://lnkd.in/ezDshVKR

Je serai l'invité de Benjamin Duhamel dans la Matinale de France Inter à 7h50 mardi 10 mars. Je parlerai du lancement de notre nouvelle entreprise AMI Labs. https://lnkd.in/eTS-5cXc

The human brain saves energy by leaving most neural pathways inactive at any given moment. Researchers applied this biological concept of sparsity to artificial intelligence. CDS PhD student Yilun Kuang, CDS founding director Yann LeCun, former CDS Faculty Fellow Tim G. J. Rudner, and others demonstrated that computer vision models can ignore up to 90 percent of their learned data without losing accuracy. The team bypassed the need for negative samples by introducing the Rectified Generalized Gaussian distribution. This mathematical technique forces the vast majority of a vector’s entries to become zero, acting as an anti-collapse mechanism. “In natural images, there is a lot of redundancy you can exploit,” Kuang said. “You can readily learn a vector with 10 percent of its entries being non-zero, the other 90 percent being zero, and you still are able to use the features for a variety of tasks.” https://lnkd.in/e34q6EY7 #ArtificialIntelligence #ComputerVision #MachineLearning #DataScience #RepresentationLearning

My keynote at the AI Research Symposium held at the AI Impact Summit in India last month. https://lnkd.in/eN7ibzae

Excellent article de Guillaume Grallet dans Le Point sur AMI Labs et son équipe de choc.

Today we're introducing TRIBE v2, a foundation model trained to predict how the human brain responds to almost any sight or sound. Building on our Algonauts 2025 award-winning architecture, TRIBE v2 draws on 500+ hours of fMRI recordings from 700+ people to create a digital twin of neural activity. It enables zero-shot predictions for new subjects, languages, and tasks, consistently outperforming standard modeling approaches. We’re releasing the model, codebase, paper, and an interactive demo to help researchers advance neuroscience, apply brain insights to build better AI, and use computational simulation to speed up breakthroughs in neurological disease diagnosis and treatment. Try the demo and learn more here: https://go.meta.me/tribe2

starting in 20 minutes

The great world model debate: generative architectures vs JEPA. A live debate between Eric Xing and me at the AI for Impact Spring School (I'll be remote).

Unveiling our new startup Advanced Machine Intelligence (AMI Labs). We just completed our seed round: $1.03B / 890M€, one the largest seeds ever, probably the largest for a European company. We're hiring! [the background image is the Veil Nebula - a picture I took from my backyard, most appropriate for an unveiling] More details here: https://lnkd.in/e3cxdfvb

Boom! A clean recipe to train JEPA world models.

V-JEPA 2.1: a new JEPA model trained on video. Uses features from multiple levels in the hierarchy. Better performance than V-JEPA 2 on some dense vision tasks, such as semantic segmentation, object tracking, and short-term action anticipation. Better performance on action planning. https://lnkd.in/eF-7_sjY

Episode 3 of Tom Mitchell's podcast "Machine Learning: How did we get here" in which I discuss how we got there! https://lnkd.in/eR8cAbRX

Position paper with Emmanuel Dupoux and Jitendra MALIK on how to get machines to learn like humans and animals.

🎉 We’re proud to announce our co-leading of AMI Labs’ €890M seed round — Europe's largest AI seed round to date. For AI to move beyond screens and into the physical world, it needs to reason about how situations evolve — understanding the relationship between actions and outcomes rather than simply producing outputs. This is the direction behind world models, the next chapter of the AI revolution. The team brings together exceptional scientific vision and proven execution: Created by Turing Award-laureate Yann LeCun, and led by CEO Alex LeBrun, with whom we’ve formed a close partnership at CI-backed Nabla. For us, this investment furthers a long-standing commitment: backing European scientific excellence with global ambition 🇪🇺 🌎 Congrats to the whole team — excited to have a front-row seat to the new frontier of AI. More on AMI from the CI team soon, but for now scroll through some highlights and articles linked in comments 👇 CC: Denis Barrier Ming-Po Cai Jacky Abitbol Cathay Capital #VentureCapital #WorldModels #AI