World models: 42AI predicts people, Genie 3 builds worlds

On November 2, TechCrunch reported that Israeli startup 42AI came out of stealth with $62.5 million in funding and “a world model that can predict human behavior.” 2026 has been a year of large funding rounds for world model developers: World Labs raised $1 billion in February, and Yann LeCun’s AMI Labs raised $1.03 billion in March.

Yet the same term “world model” covers very different things. Google’s generates 3D worlds as video, while 42AI’s predicts what brokerage clients will buy and sell. I compared what is inside 42AI’s model with the models of other developers that deal with the physical world, along with the uses and limits of each, based on each company’s official announcements.

A world model predicts “the next state”#

In Google DeepMind’s description, world models are AI systems “that can use their understanding of the world to simulate aspects of it.” They can predict how an environment will evolve and how their own actions will affect it. Where large language models such as ChatGPT and Claude predict “the next word,” a world model predicts “the next state after an action is taken.”

The idea itself is not new. The 2018 paper “World Models” by David Ha and Jürgen Schmidhuber built an agent from three parts: one that compresses what it sees, one that predicts what comes next, and one that decides actions. The paper showed that an agent trained entirely inside a “dream” generated by the world model could carry that policy into the real environment. Trial and error inside a predicted world spares you repeated failures in reality.

Today’s world models fall roughly into four groups by what they predict.

What it predicts or generates	Main examples	Output
Continuation of video	Google DeepMind’s Genie 3, NVIDIA’s Cosmos 3	Video that responds to input, robot actions
3D spaces	World Labs’ Marble	3D worlds you can walk around in (exportable as files)
Continuation of the “meaning” of video	Meta’s V-JEPA 2, AMI Labs	Prediction in an abstract representation, not pixels
Human decisions	42AI, Simile	The probability of what a person will choose in a given situation

The first three deal with the physical world, mainly for robots and self-driving. 42AI applied the same term to individual psychology and behavior.

What 42AI’s world model predicts#

Ehud Ezra, 42AI’s co-founder and CTO, told TechCrunch that “in our world model, each person is a world.” What it predicts is the probability of which decision a person will make in a specific situation.

According to TechCrunch, the model tries to estimate not an extension of past behavior but the psychological tendencies behind it. For a brokerage client, its inputs are reactions to seemingly unrelated events such as identity verification, account opening, calls to support, money transfers, and stock trades. For conversations with support, it looks at the emotion in the speech, the voice, and even the number of calls. These observations become numbers representing personality traits, and feeding them into a Transformer-based world model yields the probability of each decision per situation. A separate language model writes the explanation of the reasoning behind a prediction.

Training data started with psychological tests and behavioral scenarios from a few thousand people, to which it added data on more than 2 million people from governments and biobanks. The company says client data now covers nearly 30 million people. In February 2026, Ezra and others published the paper “Decoding the Human Factor,” a model that predicts individuals’ strategic choices conditioned on personality profiles built from psychological tests.

On usage, the official site says it “predicts each client’s trades, down to the security, direction, and size, up to 24 hours in advance.” According to CEO Amit Gal-Or, 10 major brokerages use it to manage liquidity, margin, and hedging, and one brokerage generated about $30 million in value through dynamic hedging.

The 98% accuracy figure and who put up the money#

Gal-Or himself explains that the figure “98% accuracy in predicting market trades” is the share of cases where trade flow fell within the predicted range, not the share of individual trades predicted correctly. For predicting trade direction and size, he says the F1 score (a metric that accounts for both misses and false alarms) is 0.8 to 0.85. Both are the company’s claims, and no external validation has been published.

The $62.5 million did not come only from outside investors either. It came from the founders and Infinity Group, raised over three years. Infinity Group was founded by chairman and co-founder Amir Gal-Or (the CEO’s father).

Limits and privacy concerns#

CEO Amit Gal-Or also admits that accuracy has a ceiling, because the model cannot access everything a person experiences. According to CTO Ezra, it also struggles with situations such as someone changing a trade on advice from a friend or colleague. The company plans to expand beyond finance into gaming, fraud prevention, and cybersecurity, but each domain requires retraining on its rules and context.

On privacy, Amit Gal-Or says “the model doesn’t know who the person is.” TechCrunch noted, however, that the company still links client behavior to build a profile of each individual. Could brokerages anticipate client moves and use that against the clients? Ezra called this a “legitimate concern” and pointed to regulation, oversight, and an internal ethics committee. But he gave no specifics on how misuse would be prevented.

In predicting human behavior, Simile also raised $100 million in February. Its founders are the researchers behind “Smallville,” a study in which 25 AI agents lived in a virtual town. Simile builds digital twins based on real people and has them answer how they would react to new products or interface changes.

Who else is building world models#

The main moves by developers dealing with the physical world, as of November 2026, are as follows.

Developer (model)	What it can do	Availability
Google DeepMind (Genie 3, Genie 4)	Generates explorable worlds from text in real time	Limited release. US AI Ultra subscribers can try it in Project Genie
Waymo (Waymo World Model)	Based on Genie 3, generates driving scenes as camera and LiDAR data	Internal development use
NVIDIA (Cosmos 3)	Understands and generates text, images, video, audio, and robot actions	Open model
World Labs (Marble)	Builds 3D worlds from text, images, video, and more	Generally available. Agreed to join AMD
Meta (V-JEPA 2)	Predicts what comes next from video and runs robots in unfamiliar environments	Code and model released
AMI Labs (unreleased)	Researches models that understand reality, based on the JEPA approach	No product

Google DeepMind announced Genie 3 in August 2025, the first Genie able to generate worlds in real time in response to input. It generates at 24 frames per second in 720p. Its visual memory reaches back one minute, and it keeps the world consistent for a few minutes. In January 2026, an experimental web app, “Project Genie,” began rolling out to Google AI Ultra subscribers in the US aged 18 and over.

On October 27, 2026, Google announced its successor, Genie 4. It also generates ambient sound, music, and dialogue in the world at the same time, and accepts natural-language instructions to move the camera, change the weather, or add characters. It is currently open only to some trusted testers, and Google says it will come to Project Genie and the Gemini API soon.

NVIDIA announced Cosmos 3 at GTC Taipei on May 31, 2026. It combines a Transformer for reasoning with a Transformer for generation, understanding how things move and where they are before generating video or robot actions. By use case, there is Super, which favors accuracy, Nano, which responds quickly, and Edge, for edge devices, coming soon.

World Labs was co-founded by Fei-Fei Li. Marble became generally available in November 2025, and the worlds it creates can be exported as Gaussian splats (a format that draws 3D scenes as collections of points), meshes, or video. It raised $1 billion in February 2026, but on September 28 it announced a definitive agreement to join AMD. Li is expected to join AMD as Executive Vice President and Chief Scientist, and the deal is expected to close by the end of 2026, subject to regulatory approval and other conditions.

Meta’s V-JEPA 2 is a model built on JEPA, an architecture that predicts what comes next in an abstract representation rather than in raw video pixels, and it has 1.2 billion parameters. LeCun, who proposed JEPA, left Meta to co-found AMI Labs, which raised $1.03 billion in March. CEO Alexandre LeBrun told TechCrunch that because the company starts from fundamental research, commercialization could take years.

What world models might be able to do#

Creating hard-to-collect scenes for training#

Closest to practical use today is training for self-driving cars and robots. In February 2026, Waymo announced the Waymo World Model, based on Genie 3. It generates scenes that real driving almost never records, such as a tornado, a flooded residential street, a snow-covered Golden Gate Bridge, or an elephant in the road, as both camera and LiDAR data. According to Waymo, most simulators in the self-driving industry learn only from their own driving data, so they cannot create anything outside what they have experienced. Waymo explains that drawing on what Genie 3 learned from vast amounts of video lets it create situations the car has never seen. The generated content can be changed with language prompts, driving inputs, and road layouts.

NVIDIA also offers Cosmos 3 as a model for generating synthetic data to train robots and self-driving cars. NVIDIA says it shortens the time to train and evaluate physical AI “from months to days.”

Letting robots “think through the outcome before moving”#

World models do more than create data; they can also serve as a simulator before a robot decides what to do. Meta announced that, using V-JEPA 2, it had robots pick up and place objects in environments never seen during training. The robot predicts the outcome of each candidate action and chooses the one that gets closest to the goal. Cosmos 3 can also serve as a base for models that teach robots specific tasks.

3D production and games#

3D worlds made with World Labs’ Marble can be exported as files such as meshes. In Project Genie, you set up a world and a character from text or images, choose how to move, such as walking, flying, or driving, and explore inside it. In Google DeepMind’s view, world models are at the stage of starting to affect both AI research and generative media.

Estimating human decisions in advance#

When the target is human behavior, as with 42AI, uses expand to financial risk management, fraud detection, and predicting reactions to products. Simile CEO Joon Sung Park told Bloomberg that in a mock earnings call of a listed company, it predicted 8 of 10 analyst questions.

What they cannot do yet#

At announcement, Genie 3 sustained continuous interaction for only a few minutes, not hours. The range of actions an agent can take on its own is also limited. For Project Genie, Google itself writes that each generation is capped at 60 seconds, worlds may not match prompts or real-world physics, and controls may lag.

AMI Labs’ LeBrun joked that “in six months, every company will call itself a world model to raise money.” Eight months later, 42AI showed up under the same term, using psychological and trading data. When an article mentions a “world model,” it seems worth first checking what kind of “next state” the model predicts.