Why
This matters because it reframes two things jay's stack already leans on: what "planning" means for an agent, and where safety constraints should live. LeCun's position is that autoregressive token generation — pushing a fixed stack of layers once per token — is not reasoning, because humans think in an internal representation space and only externalize the result as language (14:19, 14:44, 14:58). The alternative he proposes, energy-based reasoning, treats acting as an optimization problem: propose a sequence of actions, let a world model predict the resulting state, and pick the sequence that minimizes an energy (cost) function subject to the goal and a set of guardrails (13:48, 15:19). The safety argument follows directly from that: an LLM can be jailbroken because its safety behavior is instructions competing with other instructions in the same token stream, but a guardrail wired into the energy function structurally rules out violating outputs during the optimization itself, rather than hoping the model declines (17:34, 18:23). That is a stronger claim than "add more RLHF" — it is a claim about where the constraint should sit architecturally.
How it works
Text scaling hits a data-volume ceiling
LLMs train on roughly 20–30 trillion tokens, on the order of 10^14 bytes — an amount of text a human could not finish reading in about 400,000 years (10:47). A 4-year-old receives a comparable order of magnitude, ~10^14 bytes, through vision and touch alone by that age (11:34). LeCun's point is not that text is worthless but that it cannot be the whole substrate for physical common sense — video and interaction data have to be part of the diet (02:45, 11:50).
Planning is optimization, not decoding
Instead of decoding tokens, an agent proposes candidate action sequences; a learned world model predicts the state each sequence leads to; an energy function scores each predicted state against the goal and the guardrails; the agent picks the action sequence with minimum energy (13:48, 15:19). Hierarchical planning follows from the same picture: getting from New York to Paris isn't planned as a sequence of 10ms muscle movements, it's planned as high-level sub-goals ("get to the airport," "board the flight") that get decomposed step by step — and LeCun flags this multi-level decomposition as robotics' core unsolved problem (19:07, 20:52).
Why generative video models are not world models
Video has effectively infinite plausible continuations (22:54): after a camera pan, trying to predict every pixel of the next frame forces the model to either collapse onto one arbitrary continuation or blur into the average of all of them (23:36, 24:41). LeCun's summary line is blunt: "don't call video generation a world model" (40:45, 41:00).
JEPA: predict in representation space, not pixel space
The Joint-Embedding Predictive Architecture sidesteps the pixel-reconstruction trap by encoding observations into an abstract representation space and predicting the next representation, not the next frame (25:39). The encoder is trained to discard whatever is unpredictable or irrelevant — noise, texture — and keep whatever carries physical causality (25:57, 40:19).
Stopping representation collapse
A predictor working purely in representation space has a degenerate shortcut available: map every input to the same constant vector and the prediction error goes to zero. LeCun's group blocks this with information-maximization terms and, more specifically, SIGReg (Sketched Isotropic Gaussian Regularization), which pushes the distribution of projected representations toward an isotropic Gaussian, combined with EMA-based teacher-student distillation as already used in I-JEPA, V-JEPA and DINO (28:48, 29:16, 42:54, 48:17).
Where it lands in Jayverse
- Theory: energy-based models, representation collapse and information maximization are new entries. They give the game-theory and optimization notes a concrete non-LLM example of "the objective shapes the behavior," which is the same lens worth applying to auction and matching mechanisms in Verex.
- Rabbit and the Auditor: guardrails belong in the objective, not the prompt. LeCun's jailbreak argument is the same shape as the case for
agentic-systems-need-ontologiesand the 7715 mandate enforcers — a session-key mandate that checks a structural constraint at execution time is closer to "guardrail inside the energy function" than a prompt-level instruction ever is. - Number: V-JEPA's "surprise" signal is an anomaly detector. A prediction error that spikes when reality violates the model's learned expectation (a ball vanishing mid-air) is the same primitive as a market-anomaly detector flagging a price path that violates a learned baseline — worth a note in Number's indicator research.
- Dark Horse: this is the theory layer under the robotics cluster. Microduck, GEN-1.5, Finn and Gemini Robotics 2 are the applied bets on physical AI; Finn's sub-goal images in particular are a JEPA-flavored move — predict the next useful state, not the next pixel.
- Eng: a two-sentence answer to "what is JEPA" is good interview material — predict in representation space instead of pixel space, and let the encoder learn to discard what isn't predictable.
Verified and unverified
Verified on 2026-09-19: Yann LeCun has published the JEPA line of work — I-JEPA (2023), V-JEPA (2024), V-JEPA 2 (2025) — and the position paper "A Path Towards Autonomous Machine Intelligence" (2022), which lays out the energy-based, hierarchical-planning, guardrail-in-the-objective architecture described in this talk; energy-based models are a framing he has used for years; Moravec's paradox and Piaget's account of intelligence as adaptation to novel situations are established references, not talk-specific claims; DINO, DINOv2 and DINOv3 are real Meta self-supervised vision models; LeCun left Meta at the end of 2025 to start a world-model-focused venture. Taken from the summary and not independently checked: the exact 10^14-byte figures for LLM training data and for a 4-year-old's sensory intake, the "~400,000 years to read" figure, the details of SIGReg as described, the claim that V-JEPA-based depth estimation beats DINOv3, every timestamp, and the specific name "AMI Labs" for the new lab — stated here only as the talk's own claim.
Sources: YouTube — Yann LeCun, "World Models: Enabling the next AI revolution" (ETH Zurich) · LeCun, "A Path Towards Autonomous Machine Intelligence" (2022) · I-JEPA, V-JEPA, V-JEPA 2 papers (Meta AI) · related items: Tech #63 (a model is weights plus an objective), Tech #62 (agentic engineering writes the boundaries), Tech #100 (Microduck), gen-1-5-one-shot-physical-prompting, finn-robotics-state-of-the-art-pi-0-7, gemini-robotics-2-whole-body, agentic-systems-need-ontologies.
Key expressions
| Expression | 뜻 · 쓰이는 자리 |
|---|---|
| world model | 월드 모델(행동의 결과 상태를 예측하는 내부 모델) · 이 항목 전체의 핵심 개념. "World Models: Enabling the next AI revolution" |
| JEPA | Joint-Embedding Predictive Architecture(결합 임베딩 예측 아키텍처, 표현 공간에서 다음 상태를 예측하는 구조) · 르쿤의 핵심 제안. "predicting the next representation, not the next frame" |
| I-JEPA / V-JEPA | Image-JEPA / Video-JEPA(이미지·비디오에 적용한 JEPA 모델 계열) · 실제 발표된 Meta AI 논문 이름. "I-JEPA, V-JEPA and DINO" |
| EBM | Energy-Based Model(에너지 기반 모델, 에너지 함수를 최소화해 답을 고르는 프레임워크) · 르쿤의 오랜 프레이밍. "energy-based reasoning" |
| energy function | 에너지 함수(후보 답을 채점해 최소화 대상으로 삼는 비용 함수) · 계획을 최적화 문제로 바꾸는 장치. "minimizes an energy (cost) function" |
| guardrail | 가드레일(위반해서는 안 되는 안전 제약) · 프롬프트가 아니라 목적함수 안에 두어야 한다는 논지의 핵심어. "guardrails belong inside the objective, not in the prompt" |
| autoregressive | 자기회귀적(이전 출력을 입력 삼아 한 토큰씩 생성하는 방식) · LLM 디코딩 방식을 가리키는 표준 용어. "autoregressive token generation" |
| representation space | 표현 공간(원시 입력이 아니라 학습된 추상 벡터가 사는 공간) · 토큰 공간과 대비되는 개념. "think in an internal representation space" |
| grounded / grounding | 접지된 / 접지(언어가 아니라 실제 감각·물리 데이터에 뿌리내림) · 텍스트만으로는 부족하다는 논지의 핵심어. "physically-grounded adaptability" |
| Moravec's paradox | 모라벡의 역설(사람에게 쉬운 일이 기계엔 어렵고, 그 반대도 성립한다는 관찰) · 이 강연의 출발점. "machines are already better... while lacking the commonsense" |
| commonsense | 상식(명시적으로 가르치지 않아도 아는 물리적·일상적 지식) · AGI 논의의 핵심 결핍 대상. "physical common sense" |
| hierarchical planning | 계층적 계획(하위 목표를 먼저 세우고 단계별로 세분화하는 계획 방식) · 로보틱스 미해결 과제로 언급됨. "high-level sub-goals... decomposed step by step" |
| representation collapse | 표현 붕괴(인코더가 모든 입력을 같은 상수로 매핑해 오차를 0으로 만드는 실패 모드) · JEPA 학습의 핵심 위험. "map every input to the same constant vector" |
| information maximization | 정보 최대화(표현이 입력의 정보를 최대한 보존하도록 강제하는 원리) · 표현 붕괴를 막는 방법. "information-maximization terms" |
| SIGReg | Sketched Isotropic Gaussian Regularization(투영된 표현 분포를 등방성 가우시안에 가깝게 만드는 정규화 기법) · 르쿤 팀의 붕괴 방지 기법 이름. "SIGReg... pushes the distribution... toward an isotropic Gaussian" |
| isotropic Gaussian | 등방성 가우시안(모든 방향으로 분산이 같은 정규분포) · SIGReg가 목표로 삼는 분포 형태. "an isotropic Gaussian" |
| EMA distillation | 지수이동평균 증류(느리게 갱신되는 교사 모델에서 학생 모델로 지식을 옮기는 자기지도 학습 기법) · I-JEPA·V-JEPA·DINO 공통 기법. "EMA-based teacher-student distillation" |
| DINO / DINOv3 | 메타의 자기지도 비전 모델 계열(라벨 없이 이미지 표현을 학습) · V-JEPA와 비교 대상. "beats DINOv3" |
| AGI | Artificial General Intelligence(범용 인공지능, 특정 과제에 국한되지 않는 인간 수준 지능) · 텍스트만으로는 도달 불가하다는 주장의 목표 지점. "human-level physical intelligence" |
| RL | Reinforcement Learning(강화학습, 시행착오와 보상으로 학습하는 방식) · 표본 비효율성 때문에 우선순위를 낮추라고 제언됨. "sample-inefficient RL" |
| sample-inefficient | 표본 비효율적(원하는 성능에 도달하는 데 매우 많은 시행이 필요함) · RL의 한계를 지적하는 표현. "sample-inefficient RL" |