General Intuition’s $2.3 billion bet that video games can train AI agents for the real world

As I entered General Intuition’s R&D department in its New York office, the company’s co-founder and CEO, Pim de Wit, 31, directed my attention to a monitor sitting on a standing desk. It looks like someone is playing something like Fortnite. It wasn’t a person.

“Our agent was playing for 100 hours straight,” said Kent Rollins, the company’s chief product officer, beaming.

Before I could immerse myself in the sight of the AI ​​navigating the game’s virtual environment, I heard the electronic footsteps of a large, four-legged robot approaching.

“The same brain that powers the agent playing the game is the one that powers the robot,” De Wit told me. Josh Duplantis, a data analyst carrying a laptop broadcasting a live feed from the robot’s only camera, explains that the robot’s default mode is “exploration.”

Relying on that camera, and its unique eye, the giant insect-like robot advanced toward me, circled me, and continued into the office. Sometimes he clipped chair legs or bumped into a trash can, like a young child who has not yet learned how his body relates to the world around him. Duplantis said it took just eight minutes of real-world robotics data to fine-tune the four-legged AI model. Moreover, that data was collected on the street, not inside the office where the robot was currently navigating itself.

An effective model that can generalize from gameplay to simulation to embodiment is the raison d’être of general intuition. This model’s ability to find its place in the world has received support from some big hitters.

General Intuition said Thursday it has raised $320 million at a $2.3 billion valuation, confirming Previously reported by TechCrunch. The round brings General Intuition’s total disclosed funding to $454 million, after $134 million round It was raised at its launch last October.

The startup spun out of De Witte’s other company, Medal, which allows gamers to upload and share video game clips. Hundreds of millions of hours of uploaded gameplay provided the raw data set to train General Intuition’s model of spatio-temporal reasoning — or understanding how to move through space and time.

But the main element wasn’t the gameplay footage; Embedded in those clips were action labels: records of exactly which buttons the player pressed and when. De Wit says most competitors try to infer actions from video alone, which he says is insufficient.

“We see this as just the next phase of future pre-training,” De Wit said. “We have a single model that can respond to Fortnite on-screen information and take action, but also to real-world dynamics in a way that MBAs never could.”

At one point, De Wit provided me with a laptop running General Intuition’s Universal Model, a simulation environment that is generated frame by frame rather than rendered by a traditional game engine. As I often do when Testing global modelsI walked straight into a series of walls. In other demos I’ve tried, sometimes the agents you control will pass right by, but that didn’t happen. And from millions of hours of play, I’ve somehow learned that walls are walls, stairs are for scale, and shadows lengthen as the sun moves.

For general intuition, that is Universal model Not the product; It is the training environment (internally referred to as the “gym”). The company eventually wants to sell the agent model itself, and de Wit says the action data embedded in gameplay helps the model distinguish “self” from “environment” in a way that gives it a richer understanding of causality.

Although General Intuition’s technology has been shown in demos, the company isn’t the only one trying to solve this problem. Moreover, getting such a model to hold up in the physical world, on a large scale, has not yet been done. Most methods of this type require huge amounts of real-world data that are collected slowly and at great expense. General Intuition’s bet is that gameplay is a scalable shortcut.

And its investors are okay with that bet, too. General Intuition’s latest round was led by Khosla Ventures, with participation from General Catalyst, Jeff Bezos, Eric Schmidt, Nico Rosberg, and researchers at Google DeepMind and MIT.

The vast majority of the round will go toward expanding computing power. General Intuition has struck a deal with CoreWeave, and plans to focus on pre-training the next version of the model. Segment is set to make its API more widely available by the end of the summer.

Vinod Khosla, whose company led the round, says he was attracted to De Witt’s vision and the company’s data position.

“If you look at the MBAs, when the logic came in, it was a quantum leap,” Khosla told me in a phone interview. “In global paradigms, I think the quantum leap is the emergence of intuition in artificial intelligence, an ability similar to human intuition. The human action data and reaction data in games is the key part of the emergence of intuition.”

The vision is a generational company

General Intuition isn’t the only company to note that Medal’s human action data is a key piece of the puzzle for building dynamic world models and public agents. The company came about in part after Medal rejected a takeover offer from a major lab, said Brianna Martin, the startup’s chief of staff. There have been other shows since then as well.

De Witte and his co-founders, Eloi Alonso, Adam Jelley and Vincent Micheli, do not want a takeover, and investors in the startup are not looking for a way out yet. The amount and quality of private data General Intuition has via Medal is one reason Khosla is convinced the startup is a generational bet, not a target for mergers and acquisitions; They can become the backbone of generalized agents and global models in simulation and the real world.

“At this point, it’s going to be a data acquisition process, which is kind of uninteresting,” Khosla said.

Part of this bet also includes trust in De Witt’s values.

The businessman spent three years working in the humanitarian field, including with Doctors Without Borders. As such, he drew a clear line for how General Intuition’s technology would be used: no agents would be used to harm humans.

“We don’t want to be an escalatory part of the system,” De Wit said. “Let’s say I came out and said: We’re exercising lethal independence.” What do you think will happen in other countries?

This limitation of military use cases comes as Silicon Valley grows increasingly optimistic about war, though De Wit says he’s happy to use his models for search and rescue missions.

De Witt is Dutch, and most of his team members are European, which shapes the company’s identity. He says he brought Martin in part because of her decision to do so He left Palantir publicly About her work with US Immigration and Customs Enforcement.

“I don’t know why Silicon Valley does what it does,” he said. “There’s a reason I’m not there.”

De Witt’s ethics does not simply limit what models will not do. As a gamer who made $1.5 million by creating and hosting a private RuneScape server as a teenager, De Witte also thinks about what happens to people left behind by what AI models can do.

General Intuition recently launched a platform called Nerve, a job marketplace that allows players to earn money using their existing setup. Those who sign up start by sorting data and can eventually move toward remote operation of the robot and other tasks. De Wit noted that Medal’s user base is precisely the generation most vulnerable to AI displacement, and he wants them to have a stake in what comes next.

Flywheel data

De Witte wants General Intuition to be an ecosystem enabler, like Anthropic or OpenAI — a model provider that enables others to build on its technology. Today, the startup has a few clients in the field of gaming, simulation and robotics.

“We’re not going to build a self-driving car company,” De Wit said. “We will make it 10 times easier for the next person to build a self-driving car company.”

The company says that once it gets its API into the hands of more customers, it will be able to test its power with a variety of use cases — such as testing the robot in a digital twin on a factory floor, operating a human-like robot inside a game studio, or sending a quad to navigate hazardous environments.

While the quad is the first physical incarnation General Intuition has tried in the real world, it has also experimented with drones and other devices, including testing the model in driving games.

“It works on anything you can control with a game controller or keyboard,” De Wit said.

The possibility of building a data flywheel is one goal.

“We will choose customers where we can diversify models for which this generalized foundation model serves as the backbone,” De Witte said. “So we will prioritize choosing clients based on whether they can provide real-world data that will be interesting and useful to move the needle in research. And if they have an agile internal team where we can be true embedded partners and learn from each other.”

General Intuition’s proprietary data is what has gotten it this far, and its ability to continue collecting data that no one else has will be essential, Khosla said. Especially since, despite the impressive demonstrations, whether the transfer from simulation to the real world can continue on a large scale is an open question that no one has yet fully answered.

Correction: The title previously misstated how much public intuition was raised in this round. The error has been fixed.