Source虎嗅•Stalecollected in 23m
Five Schools Assaulting LLMs with World Models
$2B+ funded world model schools by LeCun/Li redefine AI beyond LLMs
30-Second TL;DR
What Changed
AMI raises $1.03B seed funding, Europe's AI record, for JEPA-based world models.
Why It Matters
These heavily funded efforts signal a paradigm shift from text-based LLMs to embodied AI, potentially accelerating robotics and simulation applications. Practitioners gain new tools for physical reasoning beyond pattern matching.
What To Do Next
Test Marble on World Labs site to generate editable 3D scenes from sketches.
Who should care:Researchers & Academics
Key Points
- •AMI raises $1.03B seed funding, Europe's AI record, for JEPA-based world models.
- •World Labs launches Marble for editable 3D worlds from text/images, raises $1B.
- •V-JEPA 2 achieves 65-80% robot success in novel environments with just 62 hours data.
- •Genie 3 generates persistent, interactive 3D environments like Venice canals from text.
Deep Insight
AI-generated analysis for this event — not the original article.
Enhanced Key Takeaways
- •The shift toward world models represents a fundamental architectural pivot from next-token prediction to latent space predictive modeling, specifically designed to mitigate the 'hallucination' and 'lack of common sense' inherent in autoregressive LLMs.
- •AMI's JEPA architecture utilizes a non-generative approach, focusing on predicting missing information in representation space rather than pixel space, which significantly reduces computational overhead compared to diffusion-based video generation models.
- •The integration of Marble and Genie 3 into robotics pipelines suggests a move toward 'sim-to-real' transfer learning, where synthetic 3D environments are used to pre-train agents before deployment in physical, unstructured environments.
Competitor Analysis
Primary Focus
- AMI (JEPA)
- Abstract Reasoning
- World Labs (Marble)
- 3D World Reconstruction
- Google DeepMind (Genie)
- Interactive Simulation
- OpenAI (Sora/Video)
- Generative Video
Architecture
- AMI (JEPA)
- Latent Predictive
- World Labs (Marble)
- Neural Radiance Fields
- Google DeepMind (Genie)
- Latent Action Model
- OpenAI (Sora/Video)
- Diffusion Transformer
Benchmark
- AMI (JEPA)
- Robot Success Rate
- World Labs (Marble)
- 3D Fidelity/Editability
- Google DeepMind (Genie)
- Interaction Latency
- OpenAI (Sora/Video)
- Visual Coherence
| Feature | AMI (JEPA) | World Labs (Marble) | Google DeepMind (Genie) | OpenAI (Sora/Video) |
|---|---|---|---|---|
| Primary Focus | Abstract Reasoning | 3D World Reconstruction | Interactive Simulation | Generative Video |
| Architecture | Latent Predictive | Neural Radiance Fields | Latent Action Model | Diffusion Transformer |
| Benchmark | Robot Success Rate | 3D Fidelity/Editability | Interaction Latency | Visual Coherence |
Technical Deep Dive
- V-JEPA 2 Architecture: Employs a hierarchical encoder-decoder structure where the encoder maps input video patches into a latent space, and the predictor operates solely within this latent space to forecast future states.
- Marble Reconstruction: Utilizes a hybrid approach combining sparse point cloud generation with neural surface reconstruction, allowing for real-time editing of geometry and lighting parameters.
- Genie 3 Latent Action Space: Implements a discrete latent action space that maps user inputs to environment transitions, enabling the model to maintain temporal consistency across long-horizon interactions.
Future ImplicationsAI analysis grounded in cited sources
World models will replace LLMs as the primary foundation for autonomous robotics by 2027.
Current LLMs lack the spatial and causal reasoning required for physical interaction, which world models explicitly solve through latent space simulation.
The cost of training foundation models will shift from data volume to compute-intensive simulation cycles.
As models move toward world simulation, the bottleneck becomes the generation of high-fidelity, interactive synthetic training data rather than scraping static internet text.
Timeline
2023-06
Yann LeCun publishes 'A Path Towards Autonomous Machine Intelligence' (AMI) whitepaper.
2024-02
Google DeepMind introduces Genie, a foundation world model capable of generating interactive 2D worlds.
2024-09
Fei-Fei Li officially announces the founding of World Labs to focus on spatial intelligence.
2025-11
AMI releases V-JEPA 2, demonstrating significant improvements in physical reasoning benchmarks.
2026-02
World Labs unveils Marble, enabling text-to-3D environment generation.
- 2023-06Yann LeCun publishes 'A Path Towards Autonomous Machine Intelligence' (AMI) whitepaper.
- 2024-02Google DeepMind introduces Genie, a foundation world model capable of generating interactive 2D worlds.
- 2024-09Fei-Fei Li officially announces the founding of World Labs to focus on spatial intelligence.
- 2025-11AMI releases V-JEPA 2, demonstrating significant improvements in physical reasoning benchmarks.
- 2026-02World Labs unveils Marble, enabling text-to-3D environment generation.
Weekly AI Recap
Read this week's curated digest of top AI events →
AI-curated news aggregator. All content rights belong to original publishers.
Original source: 虎嗅 ↗
This is a summary, not the original. Read the source, or get the weekly briefing.
The weekly digest
One email a week. Unsubscribe anytime.