Seven Mental · 心智七篇
← Knowledge Atlas · Source

World Models Race 2026

世界模型竞赛:AMI Labs、Genie 3、Marble、Cosmos
SOURCE · WORLD MODELS RACE · Introl · 2025-2026 surge · 2026-01-03

The World Models Race · 2026

A paradigm shift from text processing → physical simulation / video generation / embodied reasoning

LeCun’s central claim — LLMs learn statistical patterns, not an understanding of physical reality, from text, so they will never reach AGI. Concrete failure modes: factual hallucination, broken physical reasoning, causal confusion, temporal inconsistency.

Five contenders · leaderboard
01
AMI Labs
LeCun · €500M funding
I-JEPA architecture scale-up
Physical understanding · persistent memory · action planning
02
Genie 3
DeepMind
Autoregressive frame generation
24fps real-time 3D · self-learned physics
03
Marble
World Labs · first commercial
Text/image → 3D environment
Downloadable, editable, persistent 3D worlds
04
Cosmos
NVIDIA · 2M+ downloads
Physics-aware video generation
Synthetic training data for AV / robotics
05
Sora 2
OpenAI
Video generation + physical compliance
Agents under simulated physical constraints
Complements, not replaces, LLMsLLMs handle language reasoning + tool use · world models supply physical environment understanding — directions for the multimodal evolution of agentic systems
→ world-models · yann-lecun · ami-labs · i-jepa · meta-i-jepaintrol.com/blog

World Models Race 2026

摘要

综述 2025-2026 年世界模型领域的爆发:LeCun 离开 Meta 创办 AMI Labs(融资 5 亿欧元),DeepMind 发布 Genie 3(首个实时交互式世界模型),World Labs 推出 Marble(首个商用世界模型产品),NVIDIA Cosmos 下载量超 200 万次。核心论点:世界模型代表了从文本处理向物理模拟、视频生成、具身推理的计算范式转移。

LLM 的天花板论

文章引用 LeCun 的核心主张:LLM 从文本中学习统计模式而非物理现实的理解,因此永远不会达到通用智能。具体局限包括事实幻觉、物理推理失败、因果混淆、时序不一致。

关键参与者

项目方法关键能力
AMI Labs (LeCun)I-JEPA 架构扩展物理理解、持久记忆、行动规划
Genie 3 (DeepMind)自回归帧生成24fps 实时 3D 世界、自学物理
Marble (World Labs)文本/图像→3D 环境可下载可编辑的持久 3D 世界
Cosmos (NVIDIA)物理感知视频生成自动驾驶/机器人合成训练数据
Sora 2 (OpenAI)视频生成+物理合规模拟物理约束下的智能体行为

世界模型 vs LLM:互补而非替代

虽然文章倾向”世界模型可能超越 LLM”的叙事,但从 agent 工程角度看,两者更可能互补:LLM 处理语言推理和工具使用,世界模型提供物理环境理解。这对 agentic systems 的多模态演进方向有启示。

References

  • sources/introl-world-models-race-2026.md