By ByteBot
Yann LeCun just walked away from Meta after 12 years to bet €500 million on AI that understands physics instead of language. His startup AMI Labs is raising funds at a €3 billion valuation—before shipping a single product. The timing matters. Fei-Fei Li launched a commercial world model in November. General Intuition raised $134 million in October. Google DeepMind shipped Genie 3 in August. Within six months, the industry’s elite pivoted from large language models to “world models,” systems that simulate physical environments instead of predicting text. LeCun’s take? LLMs will be obsolete “within three to five years.” World models aren’t hype anymore. They’re the race.
World Models: Physics Over Prediction
World models are neural networks that learn how the physical world works by watching video, digesting simulation data, and absorbing spatial inputs. They don’t predict the next word. They predict what happens next in an environment when you take an action. The architecture splits into three parts: a vision encoder compresses video frames, a recurrent neural network maintains memory and predicts future states, and a controller picks actions. The result? Systems that achieve 50 to 100 times better sample efficiency than traditional reinforcement learning.
The difference from LLMs is foundational. Language models excel at statistical text patterns. World models simulate physics. For text generation and chatbots, LLMs shine. For robotics, autonomous vehicles, and spatial reasoning, they fail. Hard.
The 2026 Lineup
LeCun’s AMI Labs is targeting a €3 billion valuation for AI systems that “understand physics, maintain persistent memory, and plan complex actions.” He’s already named Alexandre LeBrun, founder of health-tech startup Nabla, as CEO. Headquarters will launch in Paris early this year. Why did he leave? Meta pivoted toward LLM-based models under Alexandr Wang, Scale AI’s founder who became Meta’s new chief AI officer. LeCun wasn’t interested in building better chatbots. He’s betting the next decade belongs to systems that simulate the world, not summarize it.
Fei-Fei Li’s World Labs launched Marble in November 2025, the first commercial world model product. Feed it a text prompt, photo, or video, and it generates downloadable 3D environments compatible with Unreal Engine and Unity. Pricing starts at free (four generations) and scales to $95 per month for 75 generations with commercial rights. Game developers and filmmakers are the targets, but the infrastructure matters more. Marble exports Gaussian splats and polygonal meshes, making spatial AI accessible to anyone with a Unity license.
General Intuition raised $134 million in seed funding—one of the largest seed rounds in AI history. Led by Khosla Ventures and General Catalyst, the company is training spatial-temporal reasoning models using two billion gaming videos per year from Medal, a platform for uploading gameplay clips. Applications start with gaming but extend to search-and-rescue drones, robotic arms, and autonomous vehicles. The model already generalizes to unseen environments using pure visual input.
Google DeepMind’s Genie 3 generates interactive 3D environments from text prompts at 24 frames per second in 720p resolution for multiple minutes. Unlike hard-coded physics engines, Genie 3 uses self-supervised learning to teach itself how physics works. It remembers what it previously generated, maintaining physical consistency over time. DeepMind positions it as a “crucial stepping stone toward AGI.” Access is limited to a research preview, but the signal is clear. Google is all-in on world models.
LLMs Hit a Wall
Language models lack spatial reasoning. Research shows they produce “spatially invalid paths” because they lack direct real-world experience. Text-only input and output is insufficient for robots navigating 3D spaces. When tested on navigation tasks, models fail to establish stable spatial correspondences. Worse, they fail differently each time. Run the same prompt 100 times, and you get 100 unique failure modes. LLMs hallucinate in text. In physical systems, hallucinations become crashes, collisions, and catastrophic failures.
The 2026 shift is toward pragmatism. Real-world applications. Measurable outcomes. Sustainable costs. LLMs are excellent for language tasks, but physical AI—robotics, drones, autonomous vehicles—requires physics understanding, not pattern matching. World models provide that.
What Developers Need to Know
World models introduce new architectures: variational autoencoders, recurrent state-space models, and controllers instead of transformers. The skill set shifts toward video data processing, physics simulation, and spatial reasoning. The APIs are already here. World Labs’ Marble offers a free tier for experimentation. NVIDIA’s API catalog provides access to world foundation models. The Model Context Protocol, donated to the Linux Foundation’s Agentic AI Foundation in December, has 10,000 published servers connecting AI agents to external tools.
Use cases are emerging fast. Robotics teams need spatial planning. Autonomous vehicle engineers need scene understanding. Game developers want procedural 3D world generation. All of them are adopting world models in 2026, not later. LeCun predicts world models will dominate AI architectures within three to five years. Whether that timeline holds doesn’t matter. The commercial products are here. The funding is locked. The race is on.
The Paradigm Shift Is Here
Four major players. Six months. Hundreds of millions raised. AI’s next era isn’t about better chatbots. It’s about systems that understand the physical world. World models are already commercial (Marble), massively funded ($134 million seeds), and positioned as AGI pathways (Genie 3). For developers building physical AI—robotics, autonomous systems, spatial applications—ignoring world models in 2026 means falling behind. The paradigm shift isn’t coming. It’s here.
