⚛️量子位•Stalecollected in 77m
Fei-Fei Li defines the future of World Models
💡Understand the new paradigm for world models that could define the next generation of embodied AI.
⚡ 30-Second TL;DR
What Changed
Integration of rendering, simulation, and planning
Why It Matters
This conceptual shift provides a roadmap for researchers to build more cohesive AI systems that can better understand and interact with the physical world.
What To Do Next
Review Fei-Fei Li's latest research papers on world models to align your architecture with the convergence of simulation and planning.
Who should care:Researchers & Academics
Key Points
- •Integration of rendering, simulation, and planning
- •Redefining the boundaries of world model capabilities
- •Advancing the foundational understanding of embodied AI
🧠 Deep Insight
Web-grounded analysis with 19 cited sources.
🔑 Enhanced Key Takeaways
- •Fei-Fei Li's World Labs, co-founded in 2024, has introduced 'Marble,' a multimodal frontier model capable of generating persistent, editable 3D worlds from various prompts, outputting both visual Gaussian splats and collision meshes for physics engines.
- •The new framework emphasizes 'spatial intelligence' as the next frontier beyond language, focusing on AI systems that understand geometry, persistence, occlusion, gravity, and object continuity in 3D space, which is crucial for embodied AI and robotics.
- •Li's vision for World Models aims to address the scarcity of spatial data by leveraging synthetic data generation and curated datasets, akin to how ImageNet catalyzed computer vision, to train models for complex real-world interactions.
- •The convergence of rendering, simulation, and planning is seen as a critical step towards Artificial General Intelligence (AGI), enabling AI to 'imagine' outcomes, plan safely, and interact reliably with the physical world, moving beyond the limitations of language-only models.
- •World Labs launched its 'World API' on January 21, 2026, providing developers and robotics firms access to Large World Models (LWMs) that understand fundamental physical laws and geometric structures, marking a pivotal shift towards interactive, navigable 3D environments.
📊 Competitor Analysis▸ Show
Competitor Analysis: World Models Landscape
| Feature/Aspect | Fei-Fei Li's World Labs (Marble) | Google DeepMind (Genie Series) | Yann LeCun's AMI Labs (JEPA, LeWorldModel) | NVIDIA (Cosmos Platform) |
|---|---|---|---|---|
| Core Offering | Generates editable, persistent, physics-aware 3D worlds from multimodal prompts (text, image, video, sketch). | Generates photorealistic, real-time interactive 3D worlds from text/image prompts. | Focuses on Joint Embedding Predictive Architectures (JEPA) for learning abstract representations and predictive models. | Platform of open-weight World Foundation Models (WFMs) trained on massive robotics and driving data. |
| Key Technology/Approach | Uses 3D Gaussian Splatting and proprietary techniques for 'discrete spatial state' and geometric consistency. | Autoregressive latent diffusion models for interactive 3D generation. | JEPA variants (e.g., V-JEPA 2, LeWorldModel) learn by predicting in latent space, avoiding hand-crafted heuristics. | Pretrained WFMs on real-world and synthetic data, focusing on reasoning and prediction based on physics laws. |
| Primary Goal | Build 'spatial intelligence' for embodied AI, robotics, AR/VR, and design/simulation tools. | General-purpose simulation and interactive environment generation. | Develop intelligence through predictive models of the world rather than pure pattern matching, supporting zero-shot robot control. | Streamline and scale training for autonomous machines in virtual environments. |
| Recent Milestones (2025-2026) | Launched Marble platform (Nov 2025), World API (Jan 2026), closed $1B funding round (Feb 2026). | Genie 3 released (Aug 2025), adopted by Waymo (Feb 2026). | LeWorldModel introduced (Mar 2026), AMI Labs founded (2026) with significant funding. | Launched Cosmos platform (2026). |
| Applications | Robotics, AR/VR, autonomous vehicles, creative design, gaming. | Gaming, autonomous driving simulation (Waymo World Model). | Robotics, video understanding, physical reasoning. | Robotics, autonomous driving, scientific discovery. |
| Pricing/Availability | Commercial offering via World API for developers and robotics firms. | Research preview/internal use, adopted by Waymo. | Research-focused, foundational architectures. | Platform for developers, likely enterprise-focused. |
| Benchmarks | Aims for persistence and geometric consistency in generated worlds. | Genie series set benchmarks for general-purpose simulation. | V-JEPA 2 achieved state-of-the-art in video understanding and physical reasoning. | Focus on robust, physics-based reasoning and prediction. |
🛠️ Technical Deep Dive
- Marble Model Architecture: World Labs' flagship model, Marble, is a multimodal frontier model. It takes diverse inputs including text, images, videos, or spatial sketches.
- Output and Representation: Marble generates explorable 3D environments. Its output includes Gaussian splats for visual rendering and collision meshes that a physics engine can operate on.
- Spatial Consistency: A key feature is the creation of a 'discrete spatial state,' ensuring objects maintain persistence and geometric consistency even when a virtual camera moves away and returns, unlike traditional video generation models.
- Underlying Technologies: The Marble model leverages a combination of 3D Gaussian Splatting and proprietary techniques to achieve its generative capabilities.
- Data Generation for Training: To overcome spatial data scarcity, World Labs is tackling the challenge through a hybrid approach of synthetic data generation and careful curation, drawing parallels to ImageNet's impact on computer vision. The BEHAVIOR Vision Suite is mentioned as a tool for generating customizable synthetic data for embodied AI tasks.
🔮 Future ImplicationsAI analysis grounded in cited sources
AI systems will achieve more robust and reliable interaction with the physical world.
By integrating rendering, simulation, and planning, World Models will enable AI to understand and predict physical dynamics, leading to more capable robots and autonomous systems.
The development of Artificial General Intelligence (AGI) will accelerate significantly.
World Models are considered essential for AGI, as they provide AI with an internal representation of reality, bridging the gap between language understanding and physical intelligence.
Creative industries and virtual environments will be revolutionized by easily generated, physics-aware 3D worlds.
Tools like Marble will empower architects, designers, and game developers to create dynamic, immersive, and physically coherent virtual spaces with reduced manual effort.
⏳ Timeline
1999
Received B.A. in physics from Princeton University.
2005
Received Ph.D. in electrical engineering from California Institute of Technology (Caltech).
2009
Joined Stanford University as an assistant professor.
2013
Served as Director of Stanford's AI Lab (SAIL) until 2018.
2017-01
Joined Google as Vice President and Chief Scientist of AI/ML at Google Cloud until September 2018.
2024
Co-founded and became CEO of World Labs, an AI company focusing on spatial intelligence and generative AI.
2025-11
World Labs launched its first commercial offering, 'Marble,' a platform for conjuring exportable 3D environments.
2026-01-21
World Labs launched its 'World API,' providing access to Large World Models (LWMs) for developers and robotics firms.
2026-02-18
Fei-Fei Li's World Labs closed a $1 billion funding round.
📎 Sources (19)
Factual claims are grounded in the sources below. Forward-looking analysis is AI-generated interpretation.
📰
Weekly AI Recap
Read this week's curated digest of top AI events →
👉Related Updates
AI-curated news aggregator. All content rights belong to original publishers.
Original source: 量子位 ↗