Genie 4 launched, and the feeds flooded. The praise sorts into three kinds: it understands physics; world models are the road to AGI; the simulator is the intelligence.
All three are excited; none answers the question that actually matters:
In a world model, is environment prediction and generation intelligence — or 3D computation plus physical-law constraints?
The answer separates two routes: if it is intelligence, then scaling compute and data leads to AGI; if it is physics learned by another means, the world model’s position must be redrawn. This piece takes three cuts at the question with the human-referential fitting theory. The conclusion first:
Wherever physics suffices, it is computation; the test of intelligence begins where physics runs out.
1. Physics Is a Free Lunch — and Free Goods Test Nothing
A technical fact too often ignored: physics is obtainable by both routes.
One route is to write it in. The game industry has built “world simulators” for forty years: physics engines (the Havok/PhysX line) compute collisions, fluids, cloth with deterministic integration and constraint solvers — precise, verifiable, bounded error. Nobody ever said PhysX was AGI.
The other route is to learn it. The Sora/Genie line learns physical regularities statistically from massive video: no equations written, regularity pressed into parameters by the density of observations.
One and the same low-entropy regularity — written in or learned out, converging paths. The physical world is the most lawlike part of the universe; compressing it is computation’s home turf. The corollary is sharp: wherever physics suffices, nothing about intelligence is being tested — precisely in those places, exact computation and learned regularity give equally good answers, so the capability discriminates no route. Awarding it the title “intelligence” adds no information, like awarding a calculator.
How, then, to tell learned physics from understood physics? By the signature of its failure. Genie 4’s own model page lists its limits: environmental detail and object persistence drift over time; real-world locations cannot be simulated precisely; continuous interaction is supported for minutes only. That is not how understood physics fails — drift is the signature of learned physics; consistency under intervention would be the signature of understanding, and that has not been demonstrated.
(One more detail worth savoring: Genie 4 wires Gemini reasoning models in to maintain coherence. Patching generative drift with linguistic reasoning is a sideways admission that the generative side itself drifts.)
2. The Test Begins Where Physics Runs Out
After physics runs out, what remains in the environment? Intentions, attempts, failures, corrections.
A hand reaches — grasping the cup or steadying it? What counts as the first miss? How is a slip repaired? These are the parts physics equations do not cover — and the parts of the environment that genuinely belong to practice: what happens in environments is mostly someone trying, erring, repairing.
So two families of world models part ways at the root — not in architecture, but in the object of fitting:
Video world models fit finished footage — the outcome-side projection of practice. A finished clip is an edited product: the filmmaker’s failures, retakes, and cut struggles do not exist in the material. The model has seen only “what success looks like,” never “how success grew out of failure.”
Interactive world models fit play traces and interaction records — the flow of practice. Microsoft’s Muse was trained on seven years of continuous gameplay; every player attempt, death, and retry is a training signal — failures and corrections are first-class citizens in this kind of data. The Genie line generates action-conditioned worlds. Roots in interaction grow capabilities in interaction.
From this, the theory makes one clean, testable prediction:
At equal scale, action-conditioned models trained on real interaction traces will systematically outperform video-only models on intervention and counterfactual queries.
An intervention query: change one condition — swap the ball for a paper ball, move the table half a meter — does the world change with it, or carry on as if nothing happened? A counterfactual query: the same scene, a different play — what then?
Reality has already begun to keep score: action-conditioned world models advertise controllability, while video-only generation has long struggled to be steered at all — at equal scale, whether you can steer it traces back to the object of fitting.
And the falsification channel, as always, stays open: if a video-only model ever matches interactive models on intervention queries, this prediction is voided and this branch of the theory is overturned. The weight of a claim comes from its being overturnable — a rule the theory set for itself.
3. Intelligence Enters Through Use
The last cut settles where “intelligence” can attach at all. The same Genie 4 has two modes of existence:
Closed into a loop: predictions consumed, checked, corrected — robots rehearse inside it before touching hardware, agents practice inside it before entering real tasks, humans reshape the world in natural language and the model follows. Errors have somewhere to go; the fitting lag converges. Here it is a load-bearing component of an intelligent system.
Rendering alone: generating high-resolution worlds whose errors nobody consumes and nobody corrects — drift, then, just drifts. Here it is a more gorgeous physics engine — no insult in that; tools do not need the honorific “intelligence” to be priceless.
The criterion is not render quality; it is whether the loop is closed. The direction worth noting: Genie 4 itself is moving toward the loop — Gemini reasoning wired in, event orchestration, multi-step task demos. Its real claim to fame is not “it understands physics” (physics is the free lunch) but that it puts interaction at the center of training and use — the object of fitting turning from the projection of works toward the flow of practice. That is the watershed, and it lands exactly on the axis the theory drew.
Closing: Replace the Checklist
World models deserve this year’s fire — but most people state the reason backwards. Next time you evaluate any world model, don’t just ask “how realistic.” Ask three things:
- Intervention consistency: change one condition — does the world change with it?
- Failure coverage: do the training data contain attempts and corrections, or only finished products?
- The loop: are errors consumed and corrected, or left to drift?
All three point the same way: where physics suffices, it is computation — no cause for awe; the test of intelligence begins where physics runs out, and what remains there is all intention, attempt, failure, correction — the interior of practice.
The answer to the world-model question lies not in the prediction layer, but in the object of fitting and the closing of the loop.
This is the theory’s second public application. The definition piece is AGI Fitting Theory: Fitting Practice, Not Works; the engineering protocols are in the companion Fitting the Practice: A Data Methodology for Frontier Model Training — in the readiness audit of eight 2026 flagships, the distribution line averaged 0.062: “whom to cover, and which experiences” is nearly blank industry-wide, and world models are this exam’s newest candidate.