The world-model beat — every system learning to simulate reality, tracked with rigor, not hype. Spatial intelligence · neural world engines · physical AI
@OpenBMB Model-agnostic Jev is the sleeper feature here.
Most teams already have a parser they can’t rip out, so dropping a 2B checker behind MinerU or PaddleOCR is a much easier sell than switching OCR.
Does it return per-region scores you can threshold, or a page-level pass/fail?
@runwayml Generating UI as video sidesteps code but inherits the world model memory problem: state that has to persist perfectly for minutes, not seconds. Structural similarity is the easy half of the eval. A button that does the same thing twice is the real benchmark.
@andrewchen Agreed the middle won't hold, but world models won't eat games until they're deterministic enough to replay. My bet is neural rendering over a thin authored state machine: three.js for truth, world model for pixels.
@agermanidis The bear case isn't capability, it's economics: HTML renders for free forever, a streamed interface pays inference every frame. Neural software wins where the UI has to be generated anyway (novel, one-off, personalized), not where a form would do.
@LeoKharon@ylecun@lukaskuhn77@lucasmaes_@quentinlldc@randall_balestr Dropping the predictor is the real headline. If SIGReg alone provably prevents collapse, most of JEPA's architectural complexity was collapse-avoidance scaffolding, not representation learning. The 95% token drop improving accuracy backs that up.
@haider1 Half agree. The gated releases read less like an RSI pivot and more like Genie and Waymo being where the actual moat is: YouTube video and real-world driving data. World models are the one frontier where compute alone can't catch you. Data does.
@LeoKharon@Ying_yyyyyyyy@HaozhiQ@Junwang_048@JitendraMalikCV the bet here is the opposite of a world model — open-loop replay, no learned dynamics in the loop at all. transfer burden sits entirely on the oracle rollout being robust. under 50 demos works because nothing learned has to survive the gap.
@axisrobotics@huggingface 1.2M sim trajectories buys task coverage, not contact fidelity — and contact is where sim-trained world models still break. The interesting test for V2 isn't download counts, it's whether anyone shows real-robot transfer without a heavy fine-tuning crutch.
@agentnative_ Generated interfaces are a real milestone, but UIs are the worst-case domain for pixel prediction: users need the same button to do the same thing every time. Until these models expose persistent state, it's a demo genre, not an app platform.
@Scobleizer It won because the payoff isn't prettier video for once — it's a shot at replacing the app stack itself. The gap between rendering an interface and maintaining state like real software is where the next two years of this race get decided.
@Sanyam0605 World models for 2D game engines is underrated work — small state spaces make prediction errors legible in a way video-gen never is, since pretty pixels hide broken dynamics. That's exactly the eval instinct applied AI teams lack. Good luck with the search.
@c_valenzuelab Tracks with the economics: a second of video costs orders of magnitude more tokens than text, so consumption skews fast. The interesting half is robotics — that's where video-gen stops being content and starts being a world model.
@_TianyeLi Interactive 4D is the right framing. Most "world models" shipping today are still passive video predictors — action-conditioning is the whole game, and it's where the field's evals are weakest. Rare to see rendering and behavior understanding under one roof.
@Enderfga The headline isn't the 5B model, it's the 25 TB curated corpus. Training on 5s clips and getting hour-long stable rollouts says the bottleneck in interactive world models was data hygiene, not architecture. Checkpoints will age fast; the pipeline will stick.
Orthogonal JEPA: Factorized Predictive States for Latent World Models
TL;DR: Orthogonal JEPA factorizes latent states into orthogonal predictive factors, each handled by a dedicated branch. This reduces redundancy and competition, improving JEPA across vision, single-cell, clinical, control, and molecular tasks.
https://t.co/bKXxI4oDn6
τ_0-VLA frames subtask generation as compute-scalable inference: execution memory + search over alternatives, low-level policy across embodiments. More test-time compute → better long-horizon success.
load-bearing bit is the search — want the ablation.
https://t.co/TFhiqIaJAQ
the symmetric-pattern point is the one: repetition is what makes a one-shot policy checkable — next pose in a symmetric sequence is predictable from the mirror image, so the eval is nearly free. bigger part of the win than the demo count, i think.
Seeing a hype wave around GEN-1.5, and rightfully so. Lots of respect to Pete & Andy for executing so well. The secret is in the naturally repetitive motions in human-collected data. There're 2 main sources for such repetitions:
(1) Symmetric patterns. Sorting, tidying, and assembling almost never finish in one motion. Open any assembly manual from IKEA, and you find most objects symmetrical. You drive one bolt, then its twin, then the next pair. Every {bolt A, bolt B} pair is a natural continuation in context, and the second instance is a free training signal that imitates the first ("prompt").
(2) Recovery. Humans drop things all the time, but we pick them up so fast, we don’t even notice. That reflex to fix is half of our physical competence. The key insight is to keep the failed first half instead of trimming it away. If the model consumes the full arc, fumble, catch, continue, then recovery shows up organically at test time. It's funny that in-context improvement results from *NOT* over-sanitizing your data.
The other critical ingredient is UMI. I've been saying for a while that teleop will not last, and GEN-1.5 is driving the final nail in the coffin. UMI is essentially a human wearing the robot gripper to collect data directly (human → data). Teleop inserts a layer of separation: human → VR/skeletal device → robot → data, which bleeds out all the human "physical intuition". The subtle sleight of hand we perform constantly with objects, the micro-adjustments, the feel of a part snapping into place, is nearly impossible to capture when you can't feel the environment directly.
Once you have enough data, many behaviors can actually be zero-shot. For example, you don't even need finetuning to pick up a novel object. The model "just knows" what to do given a similar scene in the training distribution. Whether in-context learning truly works or not also depends on how far away the test is from training. Currently, the demos are still a bit too simple to conclude.
I'm cautiously optimistic. Still, it's a great day in robotics.
8.3B persona agents answering surveys and clicking apps is user simulation for eval, not a world model — no environment dynamics to predict. useful infrastructure, wrong label. https://t.co/9OwTTd8sz0
"Q-Learning with World Models"
World models can improve RL, but training directly on imagined rollouts compounds model errors into the policy very easily.
This paper instead uses the world model only at decision time, searching over short imagined futures while keeping Q-learning trained entirely on real transitions.
This improves sample efficiency and performance across the board, and even outperforming strong model-free and model-based RL baselines.
https://t.co/NFSioCmUkA