Authors: Packy McCormick (Not Boring) and Pim De Witte (CEO/co-founder, General Intuition) Published: 2026-03-19 on Not Boring Format: Substack essay (~30,000 words; co-written framing)
Core Thesis
World Models are a fundamentally new class of foundation model — interactive predictive systems that learn to simulate spatial-temporal environments in response to actions. They are not LLMs (which model language) and not video models (which predict statistically plausible next frames). They are the first model class that "computes the uncomputable" — handling stochastic, action-conditioned dynamics at fixed compute cost, regardless of scene complexity.
Formally:
- LLM: predicts $P(x_{t+1} | x_t)$ — next token given prior tokens
- Video model: predicts $P(x_{t+1} | x_t)$ over pixels — next frame given prior frames
- World Model: predicts $P(s_{t+1} | s_t, a_t)$ — next state given prior state and action
The action $a_t$ "is the magic." A video model is "a dream where you simply stood and watched." A World Model is "a lucid dream in which you were able to shape the story inside the mind-generated dreamscape."
The "Compute the Uncomputable" claim
Traditional simulation engines scale at $O(N)$ or $O(N^2)$ in the number of agents/objects/interactions. Imagine simulating a Manchester United match: thousands of fans, each with stochastic behavior, requires explicit per-entity rules. The cost grows with scene complexity.
In a World Model, the entire stadium is a fixed-cost forward pass through the neural network. The stochastic, messy, human reality is "already baked into the learned weights and absorbed from the millions of hours of video the model was trained on. The model doesn't calculate what a crowd should do. It has seen what crowds actually do and uses this information to make probable predictions."
This makes World Models the proposed solution to the long-standing robotics bottleneck: machines must respond to real-world situations in fixed time, but traditional simulation costs grow unbounded with complexity.
Why LLMs and Code Are Insufficient (the Castalia Argument)
McCormick and De Witte invoke Hesse's The Glass Bead Game: LLMs are "our Castalians" — exquisite manipulators of symbols, capable of drawing connections across the entirety of human textual knowledge, but operating "entirely in the realm of representation." They can describe clapping but not clap. They can talk about gravity but cannot know gravity the way a toddler does.
"Language is an incredibly lossy compression of reality."
Code is more precise but rule-bound and unable to handle the stochastic messiness of reality. The essay rejects the "scale-of-language-and-code-equals-AGI" thesis: "A code-based simulation is a poor version of a dream. It is rule-bound and unable to handle the stochastic messiness of reality. To know the world, you must interact with it."
Why Actions Are the Ultimate Form of Compression
A core insight: actions encode the output of all prior cognition. When you decide to step left to avoid a puddle, your brain processes the visual scene, predicts the immediate future, evaluates options, and selects one. An outside observer can't see inside your head, but they don't need to. They see the output: step left.
"If you could perfectly reconstruct someone's stream of observations and actions, you would have a nearly complete record of their interaction with reality. The World Model learns exactly this mapping."
Actions act as a form of compression that "unrolls future states" until the next action updates the picture. This explains why World Models can train Agents efficiently: the model learns from observation-action-observation triples, and the agent learns to predict actions given observations.
A (Brief) History — Four Waves
Wave 0 (1990-1991): Pre-deep-learning vision
- Schmidhuber, *Making the World Differentiable* (1990) — RNN that learns world dynamics + agent that trains inside the model.
- Sutton, Dyna (1991) — unifies learning, planning, and reacting in one architecture.
Wave 1 (2018-2019): "Can this even work?"
- Ha & Schmidhuber, World Models (2018) — V (vision) + M (memory) + C (controller). Train agent inside dream, transfer back to reality. It worked.
- SimPLe / Atari 100k benchmark (2019) — model-based RL learns 26 Atari games on sample efficiency.
Wave 2 (2020-2022): "Can it match human performance?"
- DreamerV2 (Hafner, Google DeepMind, 2020) — RSSM with discrete latents; first World Model agent at human-level on 55-game Atari benchmark, on a single GPU.
- MuZero (DeepMind, 2020, Nature) — predicts in abstract latent space (no observable dream), beats Atari + Go + chess + shogi, generalizing AlphaGoZero/AlphaZero progression. Establishes the generative vs. latent World Model split.
- JEPA (LeCun, 2022) — Joint Embedding Predictive Architecture; deliberately discards pixel-space prediction in favor of predicting meaning. Founding ideology of LeCun's AMI Labs.
- IRIS (Micheli, Alonso, 2022) — autoregressive transformer over discrete visual tokens; brings LLM-style scaling laws into World Modeling. Future co-founders of General Intuition.
Wave 3 (referenced but not yet detailed in this excerpt): scaling, Genie / Genie 2 / Cosmos / V-JEPA / Sora-style generative models, MolmoBot from Ai2 (sim-to-real zero-shot transfer).
Wave 4 (2026, the essay's present moment): action-conditioned interactive World Models for embodied agents — General Intuition's stated bet.
Players and Approaches
- General Intuition (De Witte, Rollins) — $133.7M Seed; action-conditioned World Models trained on gaming clips.
- World Labs (Fei-Fei Li) — raised $1B+; recently launched Marble (3D worlds from text/images/video/3D layouts).
- AMI Labs (LeCun) — raised $1.03B; JEPA-descended philosophy.
- Google DeepMind — Genie, V-JEPA, robotic policy work.
- NVIDIA Cosmos — World Foundation Models as compute infrastructure.
- Ai2 MolmoBot (2025) — open World Model suite for robotics; demonstrated sim-to-real zero-shot transfer for manipulation.
Quotes Worth Tracking
- Jim Fan (NVIDIA Director of Robotics): "Unfortunately, the most hyped use case of World Models right now is AI video slop (and coming up, game slop). I bet with full confidence that 2026 will mark the first year that Large World Models lay real foundations for robotics, and for multimodal AI more broadly."
- Alexandre LeBrun (CEO, AMI Labs): "My prediction is that 'World Models' will be the next buzzword. In six months, every company will call itself a World Model to raise funding."
The Embedded Definitional Argument
"LLMs learn the structure of language. World Models learn the structure of causality."
This distinction is the essay's most distinctive contribution.
Relationships
- supports: World Models (the cross-cutting concept page anchored by this essay and the four shorter overviews)
- supports: AI Robotics — articulates the embodied-AI bottleneck argument
- supports: General Intuition — De Witte is co-author; essay also serves as company manifesto
- related: Li Fei-Fei — World Labs framing
- related: Yann LeCun — JEPA / AMI Labs framing
- related: Jurgen Schmidhuber (candidate) — historical foundation
- related: Richard Sutton (candidate) — Dyna / Bitter Lesson author
- related: Scaling Laws (if exists) — IRIS demonstrates LLM-style scaling applies to World Models
- contradicts: strongest scale-of-LLMs-equals-AGI position — Argues LLMs are necessary but insufficient for AGI
- related: World Models Huckins Mit Tech Review, World Models Bechard Sciam, What Are World Models Nvidia (if/when these are folded as supporting summaries)
Confidence and Caveats
- Foundational essay, not peer-reviewed research. Treat as McCormick + De Witte's framing — specifically, General Intuition's stated approach and competitive thesis is woven through the essay.
- Self-interest disclosure baked in: De Witte runs General Intuition. The essay credits the GI team, says "Pim and the GI team have a perspective on the best way to build World Models," and acknowledges the future is undetermined. McCormick says GI has been "incredibly generous with their time and human intelligence."
- Some specific historical claims should be cross-checked against original papers (Ha-Schmidhuber 2018, MuZero 2020, JEPA 2022, IRIS 2022) before citing as foundational technical claims.
- The "World Models compute the uncomputable" framing is rhetorical — the actual claim is that World Models trade simulation cost for training-time compute that learns probable behavior, but the marketing claim is the framing the essay will be cited for.
Tracked claims
- Action-conditioned prediction as the defining feature of World Models vs. video models — high confidence (well-established formal distinction).
- General Intuition's action-conditioned-from-gaming-clips approach is one of three main schools (alongside generative-pixel like World Labs/DeepMind and latent-JEPA like LeCun/AMI) — confidence high (essay surveys all three).
- Ai2 MolmoBot achieved sim-to-real zero-shot transfer for manipulation in 2025 — confidence high (cited via Allen AI Twitter and Dhruv Shah).
- Funding figures (World Labs $1B+, AMI $1.03B, GI $133.7M Seed) — high confidence (publicly reported).
- The empirical claim that World Models will outperform LLMs as an AGI path — Krishnan-style position; confidence low; presented as essay's thesis, not established.