Genie, Cosmos and V-JEPA: three routes

English

If every lab uses "world model" in the same way, the term becomes useless. These three projects point to three different meanings. The useful question is not "which one is the real world model?" It is "what does this system predict, what can control it, and who can use the result?"

Route comparison

Genie: generated worlds you can act inside

The Genie route is closest to the intuitive picture: create an environment, let an agent or user move through it, and keep the world responsive over time. DeepMind's public materials describe Genie 3 as a general-purpose world model that generates dynamic worlds from text and makes them navigable in real time, with official claims around 24 fps, 720p and consistency over several minutes.

That is a meaningful shift from passive video. The boundary is still external validation: a public demo of interactive generation is not the same thing as proof that agents can train inside the world reliably for long tasks.

Cosmos: infrastructure for physical AI

Cosmos is better understood as a platform layer. It combines world foundation models, data processing, tokenization, guardrails and simulation workflows for robotics and autonomous vehicles. The question is not whether it behaves like a consumer app. The question is whether developers can plug it into real training, validation and synthetic-data pipelines.

Cosmos 3 pushes this even more clearly into infrastructure: NVIDIA describes omnimodal world models for physical AI, with language, video, action and audio tokens as part of the technical framing. That makes Cosmos less visually simple to explain, but more directly tied to real physical-AI workflows.

V-JEPA: predictive representations

V-JEPA is less visual as a product story because it does not center on generating pretty pixels. It predicts in representation space, aiming to learn useful structure for understanding, planning and robot control. That makes it easier to underestimate and harder to evaluate through demos.

Meta's V-JEPA 2 release frames the model as a world model for physical reasoning and reports robot-control transfer. The important distinction is that this route is not trying to be an interactive 3D world product. Its value would show up inside systems that need better predictive representations.

What this comparison proves

• The same world-model label now covers multiple technical bets, not one clean category.
• Action conditioning is the central dividing line between passive prediction and usable simulation.
• Product usefulness depends on downstream integration, not only on demo quality.

What it does not prove

• It does not prove that any one route has solved long-horizon physical simulation.
• It does not make official claims equivalent to external benchmark validation.
• It does not mean all world-model work should be judged by video realism.

Common misconceptions

• Genie is not just another video model. The differentiator is real-time navigation inside generated worlds.
• Cosmos is not a consumer demo. Its value is in physical-AI tooling, data and developer workflows.
• V-JEPA is not weak because it is less visual. Representation learning may matter most where pretty pixels are not the goal.

Next reading

中文

如果所有实验室都用同一种方式说“世界模型”，这个词就会失去解释力。Genie、Cosmos 和 V-JEPA 之所以适合放在一起，是因为它们把三种不同含义暴露出来了。真正有用的问题不是“谁才是正宗世界模型”， 而是：这个系统预测什么、由什么控制、预测结果能被谁使用。

路线对比

Genie：可以行动其中的生成世界

Genie 路线最接近大众直觉：生成一个环境，让智能体或用户在里面移动，并让世界随时间保持响应。 DeepMind 公开资料把 Genie 3 称为通用世界模型，强调它能由文本生成动态世界，并支持实时导航； 官方说法包括 24 fps、720p，以及数分钟级一致性。

这已经明显不同于被动视频。边界在于外部验证：可交互生成演示，不等于已经证明智能体可以长期、稳定地在其中训练。

Cosmos：物理 AI 的基础设施

Cosmos 更适合被理解为平台层。它把世界基础模型、数据处理、tokenizer、guardrails 和仿真工作流 组合起来，服务机器人和自动驾驶。问题不是它像不像一个消费级应用，而是开发者能否把它接入真实的 训练、验证和合成数据流程。

Cosmos 3 进一步把这个方向推向基础设施：NVIDIA 把它描述为面向物理 AI 的全模态世界模型， 技术叙事中包含语言、视频、动作、音频等 token。这让 Cosmos 不如演示视频直观，但更贴近物理 AI 的真实工程链路。

V-JEPA：预测表征

V-JEPA 的产品叙事没那么直观，因为它不是以生成漂亮像素为中心。它在表征空间里预测，目标是学习 对理解、规划和机器人控制有用的结构。这让它更容易被低估，也更难通过演示视频来评价。

Meta 的 V-JEPA 2 发布把它放在物理推理和机器人控制迁移的世界模型路线里。关键区别是： 它不是要做一个可观看、可游玩的 3D 世界产品，而是要成为系统内部更好的预测表征。

这个对比能证明什么

• “世界模型”已经覆盖多条技术路线，不是单一产品类别。
• 是否受动作控制，是区分被动预测和可用模拟的重要分界线。
• 产品价值取决于下游集成，而不只是演示效果。

这个对比不能证明什么

• 不能证明任何一路线已经解决长期物理模拟。
• 不能把官方能力声明直接等同于外部基准验证。
• 不能说明所有世界模型都应该用视频真实感来评价。

常见误解

• Genie 不只是另一个视频模型。它的差异点是能在生成世界中实时导航。
• Cosmos 不是消费级演示。它的价值在物理 AI 工具链、数据和开发者工作流。
• V-JEPA 不直观不等于弱。在不追求漂亮像素的系统里，表征学习可能更关键。

下一步阅读

Evidence table / 证据表