Upon stepping into the R&D area of General Intuition’s New York office, the company’s 31-year-old co-founder and CEO Pim de Witte pointed out a monitor positioned on a standing desk. It seemed as though someone was engaged in playing a game similar to Fortnite. However, it wasn’t a human.

“Our agent has been gaming for 100 consecutive hours,” Kent Rollins, the chief product officer, stated with a smile.

Before I could become captivated by the sight of an AI maneuvering through the game’s digital landscape, I heard the mechanical footsteps of a substantial quadrupedal robot approaching.

“The same intelligence that powers the agent playing the game is also driving the robot,” de Witte informed me.

Josh Duplantis, a data analyst equipped with a laptop that was streaming a live feed from the robot’s sole camera, interjected to clarify that the robot’s primary operation mode was “exploration.”

Utilizing that camera, its singular vision, the enormous insect-like robot approached me, orbited around me, and ventured further into the office. It occasionally bumped into chair legs or collided with an errant wastebasket, reminiscent of a toddler still discovering how to navigate their surroundings. Duplantis mentioned that merely eight minutes of physical robotics data were required to optimize an AI model for the quadruped. Additionally, that data had been collected outdoors, rather than within the office where the robot was autonomously exploring.

An agentic model capable of generalizing from gaming to simulation to embodiment is the central mission of General Intuition. The competence of that model in discerning its position in the environment has attracted the interest of significant investors.

On Thursday, General Intuition announced it had secured $320 million at a $2.3 billion valuation, validating prior reports from TechCrunch. This funding round brings the startup’s total disclosed capital to $454 million, following the $134 million raised at its inception last October.

The startup originated from de Witte’s previous venture, Medal, which permits gamers to upload and share video game clips. The hundreds of millions of hours of gameplay uploaded formed the foundational dataset used to educate General Intuition’s model in spatial-temporal reasoning — or comprehending movement through both space and time.

However, the pivotal element wasn’t the gameplay footage itself; it was the action labels integrated within those videos: detailed records of the exact buttons a player pressed and at what times. According to de Witte, most competitors strive to deduce actions solely from video data, which he contends is inadequate.

“We see this as merely a transition to the next phase of future pre-training,” de Witte remarked. “We possess a singular model that can react to Fortnite visuals on the screen and take action, in addition to real-world dynamics in a manner that an LLM could never achieve.”

At one point, de Witte connected me to a laptop running General Intuition’s world model, a simulation built frame-by-frame rather than rendered using a conventional game engine. As is usual for me when experimenting with world models, I walked directly into a series of walls. In other demonstrations I’ve encountered, the agents often pass through barriers, but this particular model did not. From the millions of hours of gameplay, it had somehow discerned that walls are solid, ladders serve a climbing purpose, and that shadows lengthen as the sun progresses.

For General Intuition, this world model is not the product; it serves as the training space (internally known as “the gym”). The company ultimately seeks to market the agentic model itself, and de Witte maintains that the action data embedded in gaming aids the model in distinguishing the “self” from the “environment” in a way that enriches its understanding of causality.

Although General Intuition’s technology appears impressive in demonstrations, it is not the sole entity attempting to solve this challenge. Furthermore, achieving such a model that performs reliably in the physical world at scale has yet to be accomplished. Most methods of this nature demand vast quantities of real-world data, which are typically gathered gradually and at high costs. General Intuition’s premise is that gameplay provides a scalable shortcut.

Investors are supportive of this premise as well. The latest funding round was spearheaded by Khosla Ventures, with backing from General Catalyst, Jeff Bezos, Eric Schmidt, Nico Rosberg, as well as researchers from Google DeepMind and MIT.

The majority of the funds from this round will be allocated to enhancing computational capabilities. General Intuition has a partnership with CoreWeave and aims to concentrate on pre-training the next iteration of its model. A portion has been set aside to broaden access to its API by the end of summer.

Vinod Khosla, whose firm led the funding round, expressed that he was attracted to de Witte’s vision and the company’s proprietary data advantages.

“If one examines LLMs, the onset of reasoning was a significant breakthrough,” Khosla shared with me during a phone conversation. “In the realm of world models, I believe the major breakthrough is the emergence of intuition within AI, akin to human intuition. The human action and reaction data present in games is crucial for this intuitiveness to develop.”

The vision is a generational company

General Intuition is not the only enterprise to recognize that Medal’s human action data is crucial in constructing dynamic world models and general agents. Brianna Martin, the startup’s chief of staff, mentioned that the company was partially founded after Medal declined an acquisition proposal from a prominent laboratory. Additionally, there have been other offers following that.

De Witte and his co-founders, Eloi Alonso, Adam Jelley, and Vincent Micheli, are not inclined towards being acquired, nor are the startup’s investors seeking an exit at this moment. The volume and quality of proprietary data General Intuition possesses through Medal is among the reasons Khosla believes the startup represents a generational investment rather than an M&A prospect; that it could evolve into the foundational platform for generalized agents and world models in both simulation and real-world applications.

“At this stage, it would merely constitute a data acquisition, which lacks excitement,” Khosla stated.

Part of that belief also entails faith in de Witte’s principles.

The entrepreneur dedicated three years working in the humanitarian field, including time with Doctors Without Borders. Consequently, he has established a clear boundary regarding the application of General Intuition’s technology: No agents will be utilized in any harmful way towards humans.

“We don’t intend to be a part of an escalatory system,” de Witte mentioned. “Hypothetically, if I were to announce, ‘We’re developing lethal autonomy,’ what do you suppose would be the reaction from other nations?”

This restriction on military applications comes at a time when Silicon Valley is increasingly enthusiastic about warfare, although de Witte states he is open to his models being employed for search and rescue operations.

De Witte hails from the Netherlands, and a significant portion of his team is European, which influences the identity of the company. He noted that he recruited Martin partly due to her choice to publicly resign from Palantir because of its collaboration with the United States Immigration and Customs Enforcement.

“I can’t comprehend the motivations behind Silicon Valley’s actions,” he remarked. “There’s a reason I’m not situated there.”

De Witte’s ethical stance doesn’t solely restrict the actions his models will avoid. As a gamer who earned $1.5 million by creating and managing a private RuneScape server during his adolescence, de Witte also considers the fate of individuals who might be marginalized by the advancements AI models can achieve.

Recently, General Intuition introduced a platform named Nerve, a job marketplace that allows gamers to earn income using their existing setups. Those who register start with data labeling tasks and can eventually progress toward robot teleoperation and other responsibilities. De Witte pointed out that Medal’s user community comprises the very generation most vulnerable to AI-induced job displacement, and he aims for them to have a stake in what the future holds.

A data flywheel

De Witte aspires for General Intuition to function as an ecosystem facilitator, akin to Anthropic or OpenAI — a model provider that empowers others to build upon its technology. As of now, the startup has several clients in gaming, simulation, and robotics.

“We’re not planning to establish a self-driving car company,” de Witte stated. “Our goal is to simplify the process for the next entity to create a self-driving car company significantly.”

The business asserts that once it makes its API accessible to a wider array of clients, it will be able to test its capabilities across various use cases — such as trialing a robot in a digital twin of a factory floor, powering a lifelike bot within a gaming studio, or deploying a quadruped to navigate perilous environments.

Though the quadruped represents the first tangible embodiment that General Intuition has tested in the real world, it has also experimented with drones and other devices, including assessing the model in driving simulations.

“It functions on any platform controlled via a game controller or a mouse and keyboard,” de Witte remarked.

The ambition to create a data flywheel is one of the objectives.

“We’ll select clients where we can diversify the embodiments that this generalized foundation model serves as the core for,” de Witte stated. “We’re going to prioritize choosing clients based on their ability to provide real-world data that will be intriguing and beneficial for advancing research. Moreover, we want to work with agile internal teams where we can function as genuine embedded partners and learn from one another.”

Khosla stated that General Intuition’s proprietary data has facilitated its progress up to this point, and its capability to continue gathering unique data will be crucial. Especially since, despite remarkable demonstrations, whether the transfer from simulation to the real world can sustain itself at scale remains a question that has not yet been fully addressed.

Correction: The headline previously misrepresented the amount raised by General Intuition in this round. The mistake has been amended.