Saining Xie joining AMI Labs (LeCun's new world-model lab) — announcement

TL;DR

Saining Xie (NYU; DiT, ConvNeXt, Cambrian) announces he is joining AMI Labs, the new frontier-AI research lab co-founded with Yann LeCun. AMI’s stated mission is to build AI systems that understand the real world, have persistent memory, can reason and plan, and are controllable and safe — explicitly framed as a non-generative, world-model-first alternative to LLM-centric labs. The lab disclosed a $1.03B (~€890M) seed-equivalent round and is hiring across three continents. No technical artifacts have been released; this is a personnel-and-positioning datapoint, not a paper.

Key claims

• AMI Labs is “not a conventional lab” and “[doesn’t] intend to become one” — explicit positional framing against the dominant lab template [tweet body].
• AMI is targeting four capability axes simultaneously: (1) understanding the real world, (2) persistent memory, (3) reasoning and planning, (4) controllability and safety [amilabs.xyz mission statement].
• AMI’s stated technical thesis: generative architectures trained by self-supervised next-token prediction are “astonishingly successful” for language but do not work well on continuous, high-dimensional, noisy sensor data — therefore AMI is building world models that learn abstract representations and predict in representation space [amilabs.xyz].
• Action-conditioned world models are positioned as the substrate for agentic planning under safety guardrails — i.e., the JEPA-line bet from LeCun’s 2022 “A Path Towards Autonomous Machine Intelligence” position paper, now operationalized as a company [amilabs.xyz].
• Application targets named: industrial process control, automation, wearable devices, robotics, healthcare [amilabs.xyz].
• Funding: $1.03B (~€890M) raised from global investors [quoted @amilabs tweet of 2026-03-10].
• Distribution model: “open publications and open source” alongside industry partners, contrasted (implicitly) with the closed-flagship pattern of GDM / OpenAI / Anthropic [amilabs.xyz].

Method

Not applicable — no paper, no model, no benchmark. The artifact is a recruitment-and-positioning announcement consisting of (a) Xie’s quote-tweet, (b) AMI Labs’ pinned mission tweet from 2026-03-10, and (c) the amilabs.xyz landing page. The substantive content is the mission statement and the founding-team signal.

Results

Personnel signal worth filing for the wiki’s lab-landscape model:

• Yann LeCun (Meta FAIR Chief AI Scientist; JEPA / V-JEPA / VL-JEPA line; LeCun 2022 “A Path Towards Autonomous Machine Intelligence” position paper) co-founded the lab. LeCun’s departure from Meta to lead a JEPA-first startup had been rumored through Q1 2026; this is the public confirmation of the structural shape.
• Saining Xie (NYU CILVR; lead author on DiT — the diffusion-transformer backbone now used in virtually every frontier video / image generator filed on this wiki — plus ConvNeXt, MAE, Cambrian, REPA) is the most recent named hire.
• AMI is operating “across three continents from day one” — implying significant non-US presence, consistent with LeCun’s France connections and the €890M figure quoted in euros.
• The mission statement effectively reads as a charter for the latent-predictive world-model thesis (V-JEPA, VJEPA-2 as physics-plausibility reward in Inference-time Physics Alignment of Video Generative Models with Latent World Models, VL-JEPA on Action100M: A Large-scale Video Action Dataset) scaled into a full company.

Why it’s interesting

This is the highest-signal lab-formation datapoint for the world-foundation-models thesis the wiki has filed. AMI’s mission statement is essentially the position taken in The flavor of the bitter lesson for computer vision — that generative-rollout / latent-predictive world models, not LLMs scaled further, are the right substrate for embodied AI — converted into a $1B research organization. It also concentrates risk on a previously diffuse research bet: where the JEPA line had been a Meta-internal research program plus a few academic groups, it is now also the explicit charter of a well-capitalized frontier lab. Worth tracking against Project Genie: Experimenting with infinite, interactive worlds (Genie 3 / Project Genie), which represents the generative-rollout closed-flagship counter-thesis from GDM, and against On the Slow Death of Scaling (Hooker’s “slow death of scaling”), whose argument that the LLM scaling regime is plateauing is consistent with the framing AMI uses to justify a non-LLM bet. Saining Xie’s hire is also notable for the architecture lineage — DiT is the dominant generative-side backbone, while AMI is publicly betting against generative-side modeling for sensor data; the internal reconciliation of those two positions is something to watch.

See also

• World Foundation Models — AMI’s stated mission is a charter for the latent-predictive WFM thesis tracked by this concept
• The flavor of the bitter lesson for computer vision — Sitzmann’s “bitter lesson of CV” essay; closest filed articulation of AMI’s stated technical position
• Inference-time Physics Alignment of Video Generative Models with Latent World Models — VJEPA-2 used as physics-plausibility reward; concrete instance of the predictive-latent thesis AMI is doubling down on
• Action100M: A Large-scale Video Action Dataset — VL-JEPA on Meta’s Action100M; the most recent public scale-up of the JEPA-line representation learning
• Project Genie: Experimenting with infinite, interactive worlds — closed-flagship generative-rollout WFM; the counter-thesis to AMI’s predictive-latent framing
• On the Slow Death of Scaling — argument that pure-LLM scaling is plateauing; macro-rationale consistent with AMI’s non-LLM bet
• The Adolescence of Technology — comparable “frontier-lab-as-civilizational-bet” framing from the LLM side, useful as a contrast piece