Who is betting on what in AI, and against whom.
If you think AI means chatbots, this page is for you. There are six serious bets on how to build intelligence, each with Turing Award winners and billions behind it. One paragraph each, in plain words, then where the money went. The fault lines page shows the four questions that separate them.
Take a transformer, train it to predict the next token on most of the written internet, then teach it manners with human feedback and reasoning with reinforcement learning on problems that have checkable answers. Give it tools. Repeat with more compute. That recipe produced every product a layperson has used: ChatGPT, Claude, Gemini, Grok, DeepSeek, Kimi. The camp's claim is that the recipe keeps working. Grounding, planning and memory are expected to arrive as side effects of scale plus post-training, not as separate research programs. The money reflects that confidence. Anthropic raised $65B at a $965B valuation in May 2026 and filed for an IPO on June 1; OpenAI filed a week later after a $122B round at $852B. Two Chinese labs, Zhipu and MiniMax, listed in Hong Kong in January. The open-weight labs in China keep the gap to the closed frontier under a year and have taken roughly half of third-party API traffic. The fight inside the camp is about where the next gains come from: bigger pretraining runs, or more reinforcement learning after pretraining. Every other camp in this atlas is a bet on where this recipe runs out.
The bet in one line: Internet-scale text pretraining, then reinforcement learning on outcomes, then tools. Grounding, planning and memory are expected to emerge from scale.
A language model predicts the next word. A world model predicts the next moment: what the scene will look like, where the objects will be, what happens if you push. Train that on video and sensor streams and you get something an agent can plan inside, by imagining outcomes before acting. Two sub-camps disagree about what to predict. LeCun's JEPA line predicts abstract latent features and argues that generating pixels wastes capacity on irrelevant detail. The video camp (Genie, Cosmos, Luma, Runway, Decart) generates frames directly and points out that its models exist, ship, and produce training data for robots. The money arrived before the results. AMI Labs raised a $1.03B seed with no product; World Labs raised $1B; Wayve, Waabi, Luma, Runway, Decart, Odyssey and General Intuition each raised hundreds of millions. Genie 3 can generate a playable world in real time, but no latent world model has yet planned a multi-step physical task from raw video better than a vision-language-action model does.
The bet in one line: Predict the world, not text. Learn the dynamics of the environment from video and sensors so an agent can imagine outcomes and plan.
Ideas are cheap now; experiments are the bottleneck. This camp wants a learned simulator good enough to replace most of the lab bench, and a loop around it that says which direction to improve a design. The signal is different from text: simulation data, plus the governing equations themselves, which act as a reward you can check. The representation is continuous fields in space and time, handled by neural operators that work at any resolution. The proof case already exists: learned weather models beat the best physics-based forecasts from 2023 on. The camp has two strands, and the lab pages below are sorted by them. Simulator-first labs (Prometheus, PhysicsX, Accelerated Understanding, Physical Superintelligence) bet on scale in the model: a learned simulator good enough to design inside. Lab-in-the-loop companies (Periodic, Lila, Radical, CuspAI, and the drug designers Isomorphic, Chai and Latent) bet on the loop itself: a model proposes, an automated lab tests, the result trains the next model. Bezos's Prometheus, at $41B, is the largest bet on the first strand; drug design is the most mature part of the second.
The bet in one line: Ideas are cheap now; experiments are the bottleneck. Replace the lab with a universal learned simulator that also tells you which direction to improve.
The other camps train once, then freeze. This one says that is the mistake. Intelligence is an agent learning continually from a stream of experience and reward, building its own features and sub-goals as it goes, with no line between training and deployment. Sutton's Bitter Lesson and the Alberta Plan are the founding documents; the 2025 essay 'Welcome to the Era of Experience' by Silver and Sutton is the manifesto. The engineering problem is that current deep networks lose plasticity and forget: a network trained on one thing for long enough stops being able to learn the next thing. Until 2026 the camp was deliberately tiny. Then David Silver left DeepMind, raised $1.1B for Ineffable Intelligence, and Sutton left Keen to found Oak Lab with a target of a trillion-parameter agent learning in real time on 20 watts. Prime Intellect sells the tooling; Adaption Labs sells models that learn on the fly. Public results are still at toy scale.
The bet in one line: Intelligence is online reinforcement learning from a reward stream. Static datasets and the train-then-deploy split are a detour, and current deep learning cannot learn continually.
The robotics camp mostly agrees with the LLM camp about architecture: a vision-language model with an action head, trained on teleoperated demonstrations, human video and simulation, then fine-tuned on the robot. What it needs is data, and the whole field is organized around getting it: fleets, teleoperation farms, simulators, and cheap hardware. The proof case is a robot doing a task in a home it has never seen. Physical Intelligence's pi0.5 showed generalization of that kind in 2025; Figure reports robots on a BMW line; Agility's Digit has the most warehouse hours. Humanoid demos are still teleop-heavy, and Tesla admitted in January 2026 that no Optimus was doing useful work. Money follows the platform thesis. Skild and Physical Intelligence are priced like AI infrastructure, not robot makers. China dominates installations, and Unitree's August 2026 listing in Shanghai gave the camp its first large public pure play. Levine's wing is philosophically closer to the experience camp: robots should learn from their own deployment.
The bet in one line: The architecture is roughly settled; data is the bottleneck. Scale vision-language-action models on teleoperation, human video and simulation until robots generalize.
Every other camp trains one giant network by gradient descent. This one says that is the wrong primitive. Intelligence is efficient skill acquisition on problems you have not seen, and that needs search over programs, explicit symbols, probabilistic models, a different computer, or a different kind of network. ARC-AGI is the camp's scoreboard. It is built to resist memorization, and it gets harder faster than models solve it: the best verified ARC-AGI-2 result in 2025 was 54%, achieved by wrapping a frontier model in a search loop, at $30 a task. ARC-AGI-3, launched in 2026, is interactive; the first milestone winner scored about 1%. The camp's best ideas keep leaking into camp 1 as test-time compute and tool use, which both validates and neutralises it. Its hardware wing (Extropic, Normal, Cortical Labs) is the only part with no overlap: thermodynamic chips and living neurons are not something a transformer can absorb.
The bet in one line: Gradient descent over one giant network is the wrong primitive. Intelligence is efficient skill acquisition on novel problems, which needs search, symbols, probabilistic models, or a different kind of computer.
Primary raises by camp from the rounds this atlas records. Listings, mergers and reported talks are left out. Camp 1 dwarfs the rest, which is the point of the other five.
Details, every round and the other side of the trade are on the Capital page.
Next: the four fault lines, then people, then the glossary when a word stops you.