{"slug": "the-ai-world-has-a-new-obsession-and-it-isn-t-chatbots", "title": "The AI world has a new obsession and it isn't chatbots", "summary": "Investors are pouring billions into world models, a new AI category focused on spatial intelligence, as Stanford professor Fei-Fei Li's World Labs raised $230 million in 2024 and a larger 2026 round from Nvidia, AMD, and others, while Yann LeCun's AMI Labs secured a $1.03 billion seed round in March 2026. These systems generate 3D worlds from text or images, aiming to power robotics, design software, and simulations, signaling a shift away from chatbots toward AI that understands physical space.", "body_md": "*World models are getting the kind of money that tells you investors are looking past chatbots. The bet is simple: the next valuable AI system won't only answer you, it will understand the room it's standing in.*\n\nFei-Fei Li has spent her career teaching machines to see. The Stanford professor helped build ImageNet, the dataset that pushed modern computer vision into the mainstream, and in 2024 she co-founded World Labs to work on what she calls spatial intelligence. By 2026, that idea had stopped sounding like a research slogan and started looking like a funding category.\n\nWorld Labs raised $230 million in 2024, then added a much larger 2026 financing round that brought in investors including Nvidia, AMD, Fidelity, Emerson Collective and Autodesk. Autodesk said in February 2026 that it invested $200 million in the company and would work with World Labs on research and development. You don't write that check for a chatbot wrapper. You write it because design software, robotics, games and simulation all have the same problem: text alone doesn't tell a machine how objects occupy space, collide, move, break or stay put.\n\nMarble, World Labs' first commercial product, shows the difference. As The Verge reported when it launched in November 2025, Marble lets users generate downloadable 3D worlds from text, image or video prompts, with integrations aimed at tools such as Unity and Unreal Engine. The product also has paid plans, including a $95-a-month Max tier for heavier world-building and commercial use. That pricing detail is worth keeping. It makes the story less airy. This isn't only a professor's manifesto about the future of intelligence. It's a product people can try, export from and build with.\n\nLi's own framing is cleaner than most of the market chatter around this category. In a June 2026 Substack essay, she divided world models into renderers, simulators and planners. Renderers create what you can see. Simulators model what happens next. Planners choose a sequence of actions toward a goal. The useful systems will blur those lines, because a robot, a warehouse twin or a game engine doesn't get to separate appearance from consequence. If a shelf falls, the model has to know more than what a shelf looks like.\n\n## LeCun is making the harder argument\n\nYann LeCun is taking the sharper position. The former Meta chief AI scientist left Meta in late 2025 and launched Advanced Machine Intelligence Labs, or AMI Labs, around the view that large language models won't be enough for general-purpose AI. He has said for years that machines need internal models of the world, not only systems trained to predict the next token. Frankly, he's right to force the distinction. A model that writes a paragraph about gravity hasn't proved it understands gravity.\n\nAMI Labs announced a $1.03 billion seed round in March 2026, with Business Insider reporting that Alex LeBrun joined as CEO and that Cathay Innovation, Greycroft, Hiro Capital, HV Capital and Bezos Expeditions co-led the financing. Le Monde reported the round at 890 million euros, valuing the Paris-based company at about 3 billion euros, with backers also including Nvidia, Samsung and Toyota. That is a remarkable amount of money for a company whose core claim is still scientific, not commercial.\n\nIt also tells you where the pressure is building. Chatbots have become familiar very quickly. Coding assistants, search assistants and office copilots are now part of the software stack, and the biggest labs are spending heavily to keep small gains coming. World models offer investors a different story: AI that can help a car train for rare road events, let an architect walk through a generated building, or give a robot a safer way to learn before it touches the real floor.\n\nGoogle DeepMind has already pushed in that direction with Genie, its line of models for interactive environments. Academic work is moving too. A May 2026 paper on WorldAct described a method for turning static generated 3D worlds into editable, interaction-ready scenes with object-level meshes and collision-aware manipulation. Another April 2026 paper, HY-World 2.0, described a framework that generates navigable 3D Gaussian splatting scenes from text, images or video. Some of this will stay in research. Some of it will become products faster than people expect.\n\n## The name is new, the problem isn't\n\nThe skeptic's case is still fair. Video generation, physics simulation, digital twins and model-predictive control all existed before investors started saying world models every ten minutes. A new label doesn't magically solve contact physics, long-horizon planning or data quality. Anyone who has watched a generated video melt a hand into a coffee cup knows the gap between convincing output and usable understanding.\n\nBut don't dismiss the category as a rebrand. The money is arriving because several old threads are finally being pulled together: generative 3D, self-supervised video learning, cheaper inference, simulation data and robotics demand. General Intuition, a New York startup focused on world models trained from game data, announced a $133.7 million seed round. Overworld is working on real-time diffusion world models for games. Causal Labs is applying physics-modeling ideas to weather and climate. These companies aren't all chasing the same customer, which is exactly the point.\n\nIf you're building or buying AI software, this is the shift to watch. The easy version of AI was asking a model to describe the world. The harder version is asking it to predict what happens when something moves inside it. That second task is where factories, autonomous systems, games, architecture and robotics live.\n\n**Also read:** [Qualcomm bets its future on a 250-core data center chip and a $3.9 billion software acquisition](https://startupfortune.com/qualcomm-bets-its-future-on-a-250-core-data-center-chip-and-a-39-billion-software-acquisition/), [Micron's $41 billion quarter confirms the AI memory supercycle is nowhere near done](https://startupfortune.com/microns-41-billion-quarter-confirms-the-ai-memory-supercycle-is-nowhere-near-done/), [Anthropic's distillation problem reveals that export controls alone cannot hold the line in the US-China AI race](https://startupfortune.com/anthropics-distillation-problem-reveals-that-export-controls-alone-cannot-hold-the-line-in-the-us-china-ai-race/)", "url": "https://wpnews.pro/news/the-ai-world-has-a-new-obsession-and-it-isn-t-chatbots", "canonical_source": "https://startupfortune.com/the-ai-world-has-a-new-obsession-and-it-isnt-chatbots/", "published_at": "2026-06-24 22:56:34+00:00", "updated_at": "2026-06-24 23:23:13.426686+00:00", "lang": "en", "topics": ["artificial-intelligence", "computer-vision", "ai-startups", "ai-research", "ai-products"], "entities": ["World Labs", "Fei-Fei Li", "Nvidia", "AMD", "Autodesk", "Yann LeCun", "AMI Labs", "Google DeepMind"], "alternates": {"html": "https://wpnews.pro/news/the-ai-world-has-a-new-obsession-and-it-isn-t-chatbots", "markdown": "https://wpnews.pro/news/the-ai-world-has-a-new-obsession-and-it-isn-t-chatbots.md", "text": "https://wpnews.pro/news/the-ai-world-has-a-new-obsession-and-it-isn-t-chatbots.txt", "jsonld": "https://wpnews.pro/news/the-ai-world-has-a-new-obsession-and-it-isn-t-chatbots.jsonld"}}