**Artifact** from Bead: World Models · [canonical source](https://redfish.acequia.io/guerin/.agents/2aeb9f0b-77a8-4463-ab45-0f5d0096e304/2026-06-20/artifacts/annotated-bibliography.md) · session 2026-06-20 · discussion: Talk: World Models
**Bead:** `2aeb9f0b` (world-models) · 2026-06-20 (updated same day after list A build-out). **Verification key:** `[V]` = confirmed via WebFetch this session (title/authors/date/claims read from the source); `[T]` = from training knowledge, confirm exact title/date/URL when full `WebSearch` link results recover (search returned prose-only, no links, this session).
## A. The thesis as actually authored — Yann LeCun *(built out per Stephen's request)* The "world models matter; autoregressive/generative LLMs won't reach human-level AI" position. This is the content the prompt described — authored by LeCun, **not** Hinton. 1. **`[V]` LeCun, *A Path Towards Autonomous Machine Intelligence*, v0.9.2 (2022-06-27).** OpenReview. The canonical statement. Proposes a configurable predictive **world model**, intrinsic motivation, and **JEPA / Hierarchical JEPA** as a *non-generative* architecture for predictive world models that learn a hierarchy of representations. Frames "common sense" as "a collection of models of the world that can tell an agent what is likely, what is plausible, and what is impossible." Three challenges: learn world models by observation; reason/plan compatibly with gradient-based learning; represent percepts/actions hierarchically. — **The single best primary source for the requested thesis.** - https://openreview.net/pdf?id=BZ5a1r-kVsf 2. **`[V]` Assran et al., I-JEPA — "Self-Supervised Learning from Images with a Joint-Embedding Predictive Architecture," arXiv:2301.08243 (Jan 2023; CVPR 2023).** (Authors incl. LeCun.) First concrete JEPA. Predicts **latent representations** of masked image blocks from context — non-generative, no pixel reconstruction. Trained ViT-Huge/14 on ImageNet in <72h on 16 A100s; beats generative/pixel SSL on linear probe, counting, depth. The empirical first leg. - https://arxiv.org/abs/2301.08243 3. **`[V]` Meta AI, V-JEPA — "revisiting feature prediction for learning visual representations" (blog 2024-02-15; paper same line).** Video world model. Predict in **representation space**, not pixels: "discard unpredictable information," focus on "higher-level conceptual information" rather than "fill in every missing pixel"; 1.5–6× more sample-efficient than generative. Operationalizes the "don't predict every token/pixel" critique on video. - https://ai.meta.com/blog/v-jepa-yann-lecun-ai-model-video-joint-embedding-predictive-architecture/ 4. **`[T]` V-JEPA 2 (2025).** Successor video/action world model aimed at planning and robotics (zero/few-shot manipulation). *Confirm exact title, venue, date, URL.* 5. **`[T]` LeCun talks/interviews stating the critique directly.** "Objective-Driven AI" lecture series (various 2023–2025 keynotes); Lex Fridman Podcast #416 (Mar 2024). Recurring soundbites: "auto-regressive LLMs are doomed / will be largely 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. *WebSearch gave prose but no links this session — confirm exact talks + timestamps before quoting verbatim.*
## B. Hinton's actual most-prescient work (a different reading list) Hinton does **not** hold the prompt's thesis. His prescient lines: 6. **`[V]` Hinton, "The Forward-Forward Algorithm" (NeurIPS 2022)** + **mortal computation.** Argues digital nets' immortality + weight-sharing make them in ways *superior* to biological brains — nearly the inverse of "generative won't get us there." 7. **`[V]` Hinton, GLOM — "How to represent part-whole hierarchies in a neural network" (2021).** Hinton's closest approach to "world model": representing scene/part-whole structure. Vision/representation, not an anti-LLM claim. 8. **`[T]` Capsule networks** — Sabour, Frosst & Hinton, "Dynamic Routing Between Capsules" (NeurIPS 2017); "Matrix capsules with EM routing" (ICLR 2018). Part-whole / viewpoint-equivariant representation; lineage of GLOM. 9. **`[V]/[T]` Hinton's AI-risk arc (2023– ):** left Google (May 2023); many interviews; Dec-2024 "10–20% chance of extinction within 30 years" (`[V]` via bio). Worry is danger from capability, not a capability ceiling. 10. **`[T]` Hinton on LLM understanding (2023–2025 talks/interviews):** argues LLMs genuinely understand and resemble human cognition (confabulation as human-like). Directly contradicts the prompt's framing — included to show the contrast.
## C. Broader world-models canon *(deep-dive in `notes/01-broader-canon-ha-deepmind-sutton.md`)* 11. **`[V]` Ha & Schmidhuber, "World Models," arXiv:1803.10122 (Mar 2018).** Coined the modern ML term. VAE (vision) + MDN-RNN (memory/forward model) + tiny controller; agent trained **inside its own dream** transfers to the real env. JEPA = this minus the generative decoder. - https://arxiv.org/abs/1803.10122 · worldmodels.github.io 12. **`[V]` MuZero — Schrittwieser et al., "Mastering Atari, Go, Chess and Shogi by Planning with a Learned Model," arXiv:1911.08265 (2019).** MCTS planning over a *learned* latent dynamics model (predicts reward/policy/value); masters Go/chess/shogi **without the rules** + 57 Atari. Proof learned models can plan superhumanly. - https://arxiv.org/abs/1911.08265 13. **`[V]` DreamerV3 — Hafner et al., "Mastering Diverse Domains through World Models," arXiv:2301.04104 (2023).** Learns a world model, improves "by imagining future scenarios"; one config beats specialists on **150+ tasks**; first to **mine Minecraft diamonds from scratch, no human data/curriculum.** Proof of generality. - https://arxiv.org/abs/2301.04104 14. **`[V]` Genie — Bruce et al., "Genie: Generative Interactive Environments," arXiv:2402.15391 (2024).** 11B **foundation world model** from unlabeled internet video; generates action-controllable playable worlds from text/image/sketch via a **learned latent action space** (no action labels). Proof of foundation-scale world models. - https://arxiv.org/abs/2402.15391 15. **`[V]` Sutton, "The Bitter Lesson" (2019), incompleteideas.net.** General methods that scale with compute (search + learning) beat human-knowledge methods. The meta-justification for *learning* world models rather than engineering them. (Page has a self-signed cert; verified via Wikipedia.) 16. **`[T]` Silver & Sutton, "Welcome to the Era of Experience" (2025).** Position essay (chapter in *Designing an Intelligence*, MIT Press; DeepMind PDF). The "era of human data" (LLMs on web text) is hitting its ceiling; next leap = agents learning from their own **streams of experience** (grounded actions, rewards, environment). RL-side echo of LeCun's critique. *Confirm exact venue/URL.*
## Status / next - List **A** built out per request: items 1–3 now `[V]`; 4–5 await link-level confirmation (WebSearch returned no links this session). - Promote remaining `[T]` items when full search recovers; add LeCun talk timestamps for verbatim soundbites.