SignalGraphI build SignalGraph around one rule: a sentence rests on reviewed evidence, or it is labelled as my own reading. Reading a year of world model news, I kept seeing the opposite problem. The best models produce futures that look right, and nothing in the frame says whether they are.
World Labs offers a useful way to sort them, by what they produce. Renderers make pixels for people. Simulators make state that software can use. Planners make the next action. The three are moving at very different speeds.
Renderers are shipping. DeepMind's Genie 3 builds a world you can walk through and keeps it steady for a few minutes, and Google now offers it to subscribers as Project Genie. OpenAI went the other way, closing Sora so its team could work on world simulation for robots.
Simulators are drawing the money. World Labs' Marble exports worlds with rough collider meshes for physics, NVIDIA opened Cosmos 3, and Wayve's GAIA-4 now puts its driving model inside the loop. In September, AMD agreed to buy World Labs for about $8.2 billion in stock.
Planners are still mostly demos. Meta's V-JEPA 2 planned simple pick and place moves after less than 62 hours of robot video. Dreamer 4 mined Minecraft diamonds from recorded play alone. 1X's humanoid imagines a task as video before it moves, and by 1X's own account still leans on teleoperators. It seems each is real progress, and each is narrow.
The most useful research this year measured the gap. WorldBench found every model it tested lacked the physical consistency real interaction needs. Hansen and Wang argued the failures are predictable: they cluster where training data is thin. If that holds, the bottleneck may be coverage more than architecture.
What I would watch next is not another beautiful clip. It is a closed loop evaluation someone else can rerun: a model, a policy, a protocol and a number. Until those appear, I think the honest score is open. The models that matter will be the ones that show their work.