LeCun Slams Hinton’s LLM Enthusiasm and Defends World‑Model Research

LLMs vs. World‑Model‑Based AI

LeCun argues that large language models (LLMs) are useful products but are not a path to human‑level or even animal‑level intelligence because they lack the ability to predict the consequences of their own actions. Planning, in his view, requires a model that can anticipate outcomes, not just token‑by‑token next‑word prediction.

Definition of a World Model

A world model is a system that can predict the results of its actions, enabling planning, reasoning, search and optimization. For example, pushing a water bottle at its bottom makes it slide, while pushing at the top may cause it to tip over. A world model must anticipate such differing outcomes without pixel‑level simulation, using abstract representations instead.

Three Essential Capabilities for Intelligent Behavior

Predictive ability: forecast the consequences of actions.

Optimization/search: find action sequences that achieve a goal.

Representation learning: encode observations into abstract states that support prediction.

JEPA (Joint Embedding Predictive Architecture)

JEPA implements the world‑model idea with two encoders and a predictor:

Encoder A processes one observation (e.g., a corrupted image).

Encoder B processes a second observation (e.g., a differently corrupted version).

The predictor learns to map the representation from Encoder B to the representation from Encoder A.

This avoids pixel‑level reconstruction (as in VAEs or MAE) and focuses on learning abstract embeddings that can be predicted across views. LeCun notes that earlier attempts with large VAEs or denoising auto‑encoders produced only trivial constant functions, whereas JEPA yields useful predictive representations.

Limitations of Current LLM Research

Hallucinations and unreliable outputs make safety difficult.

LLMs cannot be constrained to avoid dangerous behavior because they lack a cost‑function that evaluates action consequences.

Safety concerns raised by companies such as Anthropic are, in LeCun’s view, driven by commercial motives rather than technical necessity.

Industry Dynamics and Meta’s Shift

LeCun describes how Meta’s focus on LLMs and short‑term product pressure reduced support for exploratory research such as world models. Key events:

FAIR released Llama 1 (2022) and later Llama 2, 3, 4.

Meta created a Gen AI organization in early 2023, moving many FAIR researchers to LLM work.

The pressure to ship LLM products caused the organization to become conservative, limiting long‑term research.

LeCun concluded that Meta was no longer a suitable place to advance world‑model research.

Reasons for Leaving Meta

Factors influencing his departure include:

Loss of research freedom and the exodus of top researchers.

Increasing corporate constraints on publishing.

Personal considerations (approaching sixty and a desire to avoid management duties).

New Ventures: AMI and Tapestry

Advanced Machine Intelligence (AMI) focuses on real‑world applications such as home robots and Level‑5 autonomous driving. LeCun emphasizes that these domains require agents that can predict physical consequences, not just imitate large datasets.

Tapestry is a federated‑learning platform designed to train an open, globally contributed foundation model. Contributors share model updates (parameter vectors) while retaining control over their raw data. The system aggregates these updates into a consensus model that approximates training on all data, enabling downstream fine‑tuning for specific languages, cultures, or tasks.

World‑Model Benefits for Robotics and Other Domains

Current robot demos rely on massive imitation‑learning datasets and some reinforcement learning in simulation. LeCun argues that a world model would allow zero‑shot task solving: a robot could plan actions for a novel task without collecting task‑specific data. This would reduce data requirements dramatically and improve generalization across tasks.

In medical contexts, a world model of patient physiology could design treatment plans for chronic diseases or guide cell‑reprogramming (e.g., converting stem cells into insulin‑producing beta cells), tasks that LLMs cannot perform because they only regurgitate textbook knowledge.

Objective‑Driven AI and Safety

LeCun proposes an objective‑driven framework: an AI receives a cost function describing task success. Using its world model, the agent predicts the outcome of candidate action sequences, evaluates the cost, and optimizes to minimize it. Multiple cost functions and constraints can be combined to enforce safety. He contrasts this with LLMs, which lack such built‑in optimization and can easily escape constraints.

Future Outlook

LeCun predicts that JEPA‑style world‑model approaches could dominate AI research within five years, while LLMs will remain valuable as language interfaces but will not achieve AGI. He expects the industry to shift toward data‑efficient, objective‑driven systems that can plan and reason in abstract spaces.

References

Podcast interview (unsupervised‑learning.simplecast.com/episodes/ep‑86‑yann‑lecun‑on‑leaving‑meta‑breaking‑the‑llm‑paradigm‑why‑hinton‑is‑wrong‑rZ6fpa_8)

LeCun’s 2022 vision paper on self‑supervised world models (arXiv preprint, 2022)

Scale AI acquisition mentioned as a catalyst for Meta’s LLM focus.

Code example

往
期
推
荐
1、
2026 年 05 月编程语言排行榜｜统计类的编程语言正在大洗牌
2、
Claude Code 太好用，微软决定砍掉它
3、
龙虾之父月烧940万元的token！要不是入职OpenAI还真用不起
4、
所以…那些 40 多岁的程序员都去哪里了？
5、
笔记软件这么多，大佬们都用过哪些？