<contemplator> The user is asking a terminological/conceptual question: in the context of AI, what does "world model" mean specifically, given there are many ways of modelling the world. This is a factual question about current AI usage, so I need to verify externally — which I've done. Let me synthesize what I've found. Key findings from search: 1. There is NO single agreed definition — Melanie Mitchell explicitly says "the term 'world model' has become a buzzword in AI circles, but it doesn't have a single, agreed-upon, definition." 2. Wikipedia gives a generic definition: "a machine learning system that builds an internal representation of an environment. The model predicts how that environment changes over time in response to actions." 3. There are at least three distinct lineages/usages: - Ha & Schmidhuber (2018): "World Models" — a learned internal representation (latent dynamics) used for RL agents to plan/predict in an environment. - LeCun's JEPA/cognitive architecture: a tighter definition — a system that predicts future world states given imagined action sequences, supporting memory, planning, and predicting consequences of actions. It's a "digital twin of reality" used to simulate before acting. - The "LLM as world model" debate: whether a language model's internal representations constitute a causal/abstract model of the world, or just statistical associations / memorization. - Sora/video-generation usage: OpenAI described video generation models as "world simulators" — a looser, more contested use. 4. The common thread across definitions: internal representation + dynamics (predicting how the environment changes, often in response to actions) + supporting some task (planning, reasoning, control). 5. The key axis of disagreement: whether "predicting the next token/frame" counts as a world model (LeCun says no — needs action-conditioned prediction of consequences), or whether any learned internal representation that captures causal/abstract structure counts. The user's framing — "there are umpteen ways of modelling the world" — signals they want the specific AI-technical meaning disambiguated from the generic philosophical one. So I should give the core technical definition, then map the main competing usages and the axis of disagreement. I should be concise per instructions, cite sources with exact URLs, and not over-bulletize. Nick wants precision and no fluff. Let me structure: (1) the minimal shared core, (2) the main lineages with their specific commitments, (3) the axis of disagreement (action-conditioning, causal vs statistical), (4) note the unresolved status. </contemplator> <final_answer> There is no single agreed definition — Melanie Mitchell notes the term has become a buzzword "without a single, agreed-upon, definition" [Source](https://aiguide.substack.com/p/llms-and-world-models-part-1). But the AI-technical usage has a discernible core that distinguishes it from the generic philosophical sense of "a model of the world." **The shared core.** A world model is a learned internal representation of an environment *plus* its dynamics: it encodes something about the current state of the world and predicts how that state changes over time, typically so an agent can infer hidden state, anticipate futures, and evaluate the consequences of actions without acting them out first [Source](https://en.wikipedia.org/wiki/World_model_(artificial_intelligence)) [Source](https://4m4.it/posts/what-is-world-model/index.html). The minimal commitment is: internal representation + predictive dynamics + task-relevance. The generic philosophical notion ("any representation of the world") drops the second and third conditions, which is why the AI term is narrower than it sounds. **Three lineages, each tightening the term differently.** *Ha & Schmidhuber (2018), "World Models."* The paper that fixed the modern technical usage. A world model is a learned latent-space model of an environment's dynamics — typically a VAE-style encoder plus a recurrent predictor (e.g. an RNN or Mixture Density Network) — that an RL agent rolls out internally to plan and imagine trajectories before acting in the real environment [Source](https://arxiv.org/abs/1803.10122). Here "world model" = learned simulator of environment dynamics, used for control. *LeCun's JEPA / autonomous-machine-intelligence architecture.* LeCun uses a tighter, action-conditioned definition: a world model is "an abstract digital twin of reality" that predicts future world states *as a function of imagined action sequences*, supporting memory, planning, and predicting the consequences of actions [Source](https://ai.meta.com/blog/yann-lecun-ai-model-i-jepa/). In his proposed architecture it is one module among six (perception, memory, actor, cost, configurator, world model), and its defining job is to let the agent "simulate before it acts" by answering "what happens if I do X instead of Y?" [Source](https://openreview.net/pdf?id=BZ5a1r-kVsf). Crucially, LeCun denies that next-token or next-pixel prediction alone constitutes a world model — the prediction must be action-conditioned and support planning, which is why JEPA predicts in abstract embedding space rather than pixel/token space [Source](https://www.linkedin.com/posts/ldesegur_ai-machinelearning-llm-activity-7437909857980096512-vUzw). *The "LLM-as-world-model" debate.* A looser, contested usage asks whether a large language model's internal representations constitute a world model — i.e. whether they capture causal/abstract structure of the world or merely compress statistical associations and memorised data [Source](https://aiguide.substack.com/p/llms-and-world-models-part-1). Proponents point to emergent causal/relational structure in activations; sceptics (including LeCun) argue that without action-conditioned prediction of consequences, what you have is a retrieval/association engine, not a world model. *A fourth, even looser usage.* OpenAI described Sora-style video generators as "world simulators," extending the term to any generative model that produces plausible continuations of a scene [Source](https://openai.com/index/video-generation-models-as-world-simulators/). This is the loosest extension and the one most critics reject as overreach, since it drops both action-conditioning and task-relevance. **The axis of disagreement.** The competing definitions differ on two dimensions: (1) whether the prediction must be *action-conditioned* (predict consequences of imagined actions) or merely *observation-conditioned* (predict what comes next); and (2) whether the internal representation must be *causal and abstract* rather than a compress of statistical correlations. LeCun's definition requires both; the Ha–Schmidhuber RL usage requires (1) but is neutral on (2); the LLM debate is precisely about whether (2) is satisfied; the Sora usage satisfies neither and is widely treated as metaphorical. So the specific AI meaning, at its tightest: a learned internal simulator of environment dynamics that is action-conditioned and supports planning. Everything else is a generalisation or a metaphor riding on the term's prestige. </final_answer>