World models: spatial intelligence as a base-model category

status: watch, opened 2026-09-13, last moved 2026-10-16. AI world-labs models

• The thread: the anti-token-stream thesis now has consumer-lab infrastructure — World Labs' Atlas (omni world model: camera-controlled generation, sparse-view 3D reconstruction, real-to-sim for robotics; launched Sep 1, early access, powers Marble) alongside Genie 3 (DeepMind), Cosmos (NVIDIA), V-JEPA (Meta), LeCun's AMI Labs, and the Graepel/Hafner ventures (deepmind-exodus thread). Cross-beat: the robotics real-to-sim application is where this touches the robots digest.
• 2026-09-13: opened from the Sep 13 ai digest — a Guardian-tier synthesis ("AI first learnt the words. Now it is reaching for the world," via MSN syndication) makes the LLM→world-models frame mainstream press 12 days after Atlas shipped. The launch itself never entered our feed — another pressroom-shaped gap.
• Watch for: Atlas early-access partner workloads (company-run benchmarks are self-reported), GA/pricing (undisclosed), whether LeCun's AMI Labs ships a competing product, Genie 3 productization, and whether any frontier lab adopts a world-model component for agentic planning (vs the token-stream stack).
• 2026-09-24 (ai-social dig): MotionJEPA released (Markus Karmann RT, project page): "JEPA-style world models are biased towards learning slow, simple features" — MotionJEPA built for a balanced static+dynamic representation. Independent academic work in the V-JEPA lineage, not a lab announcement; no benchmark claims to grade yet. Watch for adoption beyond the authors.
• 2026-09-13 (dig addendum): AMI Labs funding/affiliation — TNW characterizes LeCun as "leaving Meta's orbit to bet $1bn on world models"; RuntimeWire lists AMI Labs as LeCun's lab pursuing "controllable models designed to understand and plan in the physical world." Exact Meta relationship unverified — check affiliation before printing anything beyond "AMI Labs (LeCun)."
• 2026-09-28: AMD is acquiring World Labs (Fei-Fei Li) for $8.2B (GlobeNewswire primary + TechCrunch; 8-K filed) — "scaling our reach, and getting closer to the hardware," per the lab's own note. The spatial-intelligence category's leading independent lab joins a chipmaker — the world-model bet gets a hardware home (AMD's AI-compute push), and the base-model-category race loses its flagship startup: Genie-3 (DeepMind), V-JEPA lineage (AMI Labs), Marble/Atlas (World Labs→AMD). Watch: Fei-Fei Li's role post-close, whether Marble/Atlas access changes, NVIDIA's counter.
• 2026-09-29 (X layer, ai-social dig): AD-E2E-JEPA (arXiv 2609.34085; Haoran Zhu, joint work with LeCun and Choromanska, ~11:49 AM ET): a JEPA world model for end-to-end autonomous driving — "can a world model drive without training a driving policy?" The JEPA lineage's driving-native instantiation, with LeCun's own name on the author line, six days after MotionJEPA added a second academic entry to the same lineage.
• 2026-09-30: deal refinements (TechTimes re-report): $8.2B all-stock confirmed and Fei-Fei Li named AMD chief scientist — the role-watch from the Sep 28 entry partially answered on paper. The base-model-category race now runs through a chip company's science arm; watch Marble/Atlas access under AMD unchanged and NVIDIA's counter.
• 2026-10-03 (robots-brief cross-ref): Runway introduces Praxis-1, a "world action model for robotics" (Robot Report, Oct 2 ~2:32 PM ET) — the video-gen company joining the spatial/base-model category; ships open weights when released publicly (ledgered).
• 2026-10-06 (X layer): H-JEPA (Balestriero, LeCun RT; arXiv 2610.06805): "the first end-to-end learned hierarchical world model for long-horizon visual planning," atop SIGReg/LeWM — the academic JEPA layer thickens again after MotionJEPA and AD-E2E-JEPA; no benchmark claims graded.

• 2026-10-07→09 (X layer, ai-social dig): the JEPA stack runs a paper-cadence — six more entries in three days on LeCun's feed: an alphaXiv page for H-JEPA with LeCun named advisor; RoboJEPA (8B — "robot world models now have scaling laws," the lineage's first scaling claim); LeAVJEPA (sound+vision self-supervised, shared encoder); DSReg (provable recovery of individual latent variables — the theory answer to why latent prediction works); financial world models (Balestriero, NeurIPS — "financial data is RIPE for JEPAs"); LeWAM (Oct 9) — JEPAs that "jointly imagine what happens and what to do next": claimed 32.3× faster planning, 19 ms/plan, success 28.6%→89.7% on contact-rich manipulation. From critique-proof demos toward scaling-law and theory papers — the consolidation move the transformer era made in 2017–2020. Academic/vendor-adjacent (LeCun-amplified by construction); no independent replications in feed. ELLIS's world-models sessions and the Jordan Libération interview are the community layer. Watch: replications of the AntMaze/LeWAM numbers, wire-grade pickup of the scaling-law claim.

• 2026-10-10→11 (RT layer, via LeCun/Zholus): RoboJEPA — "the first scaling law for multi-embodiment world models"; JEPA predictors scaled to 8B, curves extrapolate, compute correlates with planning performance; all checkpoints released (https://x.com/ylecun/status/2108983952316694970). Development of the spatial-intelligence-as-base-category layer (the JEPA stack cadence, Oct 10). Single-thread claim family; watch the paper artifact + whether the scaling law reproduces.

• 2026-10-16 (~3:24 PM ET, LeCun RT of NYU Center for Data Science — https://x.com/ylecun/status/2111200521201537026): AdaJEPA — NYU CDS PhD student Ying Wang with LeCun and Mengye Ren — a world model that "keeps learning after training ends," built "to adapt during planning" (Medium write-up), the JEPA lineage's continual-learning turn where it meets the Oak Lab thesis. Academic release, not a lab announcement; no benchmark claims graded in the window.