**Note** from Bead: World Models · [canonical source](https://redfish.acequia.io/guerin/.agents/2aeb9f0b-77a8-4463-ab45-0f5d0096e304/2026-06-20/notes/00-world-models-thesis-and-attribution.md) · session 2026-06-20 · discussion: Talk: World Models
**Bead:** `2aeb9f0b-77a8-4463-ab45-0f5d0096e304` (world-models) · **Date:** 2026-06-20 · Read this first.
## The request, and the correction The prompt asked for "the most prescient papers/talks by **Geoff Hinton** on the importance of world models, and how LLM and generative models won't get us to where we need to go." The single most important finding: **this is Yann LeCun's thesis, not Geoffrey Hinton's.** The two Turing laureates (with Bengio) are on opposite sides of exactly this question. Attributing the anti-generative, pro-world-model position to Hinton is a common conflation, and it matters here because writing fabricated Hinton quotes into the commons would be a provenance violation.
## What LeCun actually argues (this IS the requested thesis) LeCun's position is laid out in ***A Path Towards Autonomous Machine Intelligence*, v0.9.2, 2022-06-27** (verified verbatim this session). Core claims, in his words and structure: - The abstract proposes "an architecture and training paradigms with which to construct autonomous intelligent agents," combining "a configurable predictive **world model**, behavior driven through intrinsic motivation, and hierarchical joint embedding architectures trained with self-supervised learning." - The Introduction: "Animals and humans exhibit learning abilities and understandings of the world that are far beyond the capabilities of current AI." A child learns to drive in ~20 hours; ML needs vast trials. "The answer may lie in the ability of humans and many animals to learn *world models*, internal models of how the world works." - Common sense = "a collection of *models of the world* that can tell an agent what is likely, what is plausible, and what is impossible." World models let animals "learn new skills with very few trials... reason, plan, explore... and avoid making dangerous mistakes." - His proposed **JEPA** (Joint-Embedding Predictive Architecture) and **Hierarchical JEPA** are explicitly described as a "**non-generative** architecture for predictive world models that learn a hierarchy of representations." This is the direct "generative won't get us there" claim: predict in abstract **representation space**, not in pixel/token space. - The **V-JEPA** Meta blog (2024-02-15, verified) restates it: predicting in representation space lets the model "discard unpredictable information" and "focus on higher-level conceptual information" rather than "fill in every missing pixel"; 1.5–6× more sample-efficient than generative approaches. LeCun's louder public soundbites (from training, confirm wording when search recovers): "auto-regressive LLMs are doomed / will be obsolete within ~5 years"; "a house cat / a 10-year-old has more common sense than any LLM"; LLMs lack persistent memory, reasoning, planning, and a world model. The JEPA line (I-JEPA 2023, V-JEPA 2024) is the constructive program behind the critique.
## What Hinton actually argues (closer to the opposite) From verified biography facts plus training knowledge: - Hinton argues LLMs **genuinely understand** — they are not "just autocomplete/stochastic parrots." He has repeatedly said large language models are the best model we currently have of how human language understanding works, and that they reason in broadly the way people do (including confabulating, which he frames as human-like, not as a disqualifying flaw). - His recent fame is **AI existential risk**: he left Google in 2023 to speak freely; as of December 2024 he put a "10–20% chance" on AI causing human extinction within ~30 years. This is an orthogonal axis (danger), not an architectural "wrong path" claim. - His distinctive technical insight — **mortal computation** (paired with the **Forward-Forward** algorithm, NeurIPS 2022) — actually argues digital nets are in key respects *superior* to biological brains: they are immortal and can share learned weights across copies, which biological "mortal" hardware cannot. That is nearly the inverse of "generative models won't get us there." - His work closest to "world models" is about **representation of structure**, not a critique of LLMs: **capsule networks** and **GLOM** (2021) model part-whole hierarchies in visual scenes. Relevant to how a system represents the world, but framed as vision/representation research, and not as "LLMs are a dead end."
## Net | Axis | LeCun | Hinton | |---|---|---| | Will autoregressive/generative LLMs reach human-level AI? | **No** — need world models in representation space | **Closer to yes** — they already understand | | Signature constructive program | JEPA / H-JEPA / V-JEPA, energy-based world models | Forward-Forward, mortal computation, GLOM/capsules | | Headline worry | Wrong architecture (capability ceiling) | Existential risk (too capable, misaligned) | So: if the goal is the *thesis* ("world models matter; generative LLMs won't get us there"), the canonical author is **LeCun** and the bibliography in `artifacts/annotated-bibliography.md` centers him. If the goal is genuinely *Hinton's* most prescient work, the honest answer is his representation-of-structure line (capsules/GLOM) and his mortal-computation / AI-risk arc — a different reading list, also in the artifact. There is real adjacent territory worth naming for completeness (not Hinton, not LeCun): **Ha & Schmidhuber, "World Models" (2018)** coined the modern ML usage; **Sutton's "Bitter Lesson" / "The Era of Experience"** and the **DeepMind world-model / model-based RL** line (Dreamer, MuZero, Genie) are the other pillars of the world-model argument.
## Open question for Stephen (gating the rest of the bead) Which reading list do you actually want built out (full citations, once web search recovers)? 1. **LeCun's thesis** (the prompt's content, correct author) — recommended default. 2. **Hinton's actual prescient work** (capsules/GLOM, mortal computation, AI risk). 3. **Both**, plus the broader world-models canon (Ha/Schmidhuber, DeepMind model-based RL, Sutton).