General Intuition is an American artificial intelligence research company that trains AI agents and world models on gameplay video from Medal, a platform where players record and share clips of video games [1][2]. It calls itself "the frontier lab for acting in space and time" and says it is building "a new class of models that can perceive, predict, and act in virtual and physical environments" [3]. The company's central bet is on action labels: a Medal clip comes with a record of the in-game actions the player took, so a model can learn which inputs led to which outcomes instead of guessing the actions from pixels [2][4]. General Intuition was announced on 16 October 2025 with a $133.7 million seed round led by Khosla Ventures and General Catalyst, and it is based in New York [1][2].

General Intuition logo, from the official website generalintuition.com. Trademark of its owner.

Pim de Witte, a co-founder of Medal, is the chief executive. His co-founders are Eloi Alonso, Adam Jelley and Vincent Micheli, researchers whose earlier work includes the IRIS and DIAMOND world models [5][2][6][7]. In 2026 the company announced two more rounds: $320 million at a $2.3 billion post-money valuation, led by Khosla Ventures and announced in June, and $220 million at a reported $6.2 billion valuation, announced on 29 September [8][9]. The three announced rounds add up to $673.7 million; the company's website says it "raised over $650M in the past year" [3]. As of October 2026 its models were open to a small number of commercial partners through an API, with a waitlist for wider access, and its main public research release so far was MIRA, an open research world model of the game Rocket League built with Kyutai in collaboration with Epic Games [3][10][11][12].

General Intuition
Type	Private company; its website is operated by General Intuition US Inc., a Delaware corporation [13][14]
Founded	2025; announced 16 October 2025 [1][8]
Founders	Pim de Witte, Eloi Alonso, Adam Jelley, Vincent Micheli [5][2]
Chief executive	Pim de Witte [2]
Headquarters	New York City; GamesBeat also listed offices in Geneva, London and Paris [13][15]
Team size	25 people (GamesBeat, June 2026) [15]
Focus	Action models and world models trained on action-labeled gameplay video, for games, simulation and robotics [3][15]
Notable release	MIRA multiplayer world model (July 2026) [12]
Announced funding	$673.7 million across three announced rounds (sum of the reported round sizes), October 2025 to September 2026 [1][8][9]
Latest reported valuation	$6.2 billion (September 2026) [9][16]
Related company	Medal (gameplay-clip platform) [1][15]
Website	generalintuition.com

History

Medal and the idea for the company

Medal was started in 2018 [17][4]. In an essay he co-wrote with Packy McCormick, de Witte said that in 2018 he and former colleagues from his teenage RuneScape private-server business built a competitive game called Get Wrecked. To keep enough players online, they added a way to watch game clips; that clip feature went viral on the Rocket League subreddit and became Medal, and the game itself was never released [4].

According to Alex Heath, around the middle of 2024 de Witte began asking AI labs whether they would want to use Medal's data to train their agents. "We received multiple acquisition offers very quickly," de Witte told Heath, adding that the company was initially interested but that this "was mostly a result of us not understanding what we were sitting on" [18]. TechCrunch, citing The Information, reported that OpenAI tried to buy Medal for $500 million in late 2024; neither OpenAI nor General Intuition would comment on that report [1]. De Witte told Heath he had read a Google DeepMind paper showing that gaming data could teach AI to navigate 3D environments, and that the labs' interest made him see Medal's data as the basis for a foundation model of its own rather than something to license [18]. Brianna Martin, General Intuition's chief of staff, later told TechCrunch that the company was born in part after Medal turned down an acquisition offer from a major lab, and that other offers had followed [2].

The co-written essay says the company is named after a remark by Demis Hassabis about mimicking "the intuition of these gamers, who are, by the way, only amateur biologists" [4].

Launch and seed round (October 2025)

On 16 October 2025 Medal announced that it had spun out General Intuition with $133.7 million in seed funding [1][18]. TechCrunch reported that the round was led by Khosla Ventures and General Catalyst, with participation from Raine [1]. Heath wrote that the money came primarily from Vinod Khosla and that it was Khosla Ventures' largest seed check since the firm backed OpenAI. "It's a pretty big bet," Khosla told him. "They have a unique dataset and a unique team." [18] Moritz Baier-Lentz, a partner at Lightspeed Venture Partners who oversees its gaming investments, joined part-time as a founding team member [1][18].

At launch the company said it would spend the money on researchers and engineers to train a general agent, with gaming and search-and-rescue drones as the first applications [1]. De Witte said the model could already understand environments it had not been trained on and predict actions in them from visual input alone, and that the approach could carry over to robotic arms, drones and autonomous vehicles, which people often operate with game-style controllers [1]. "Games are basically the only verifiable domain for spatial-temporal reasoning," he told Heath. "You can separate a good action from a bad action, which is why it's so valuable." [18]

Series A (announced June 2026)

In December 2025 Heath reported that General Intuition was in late-stage talks to raise several hundred million dollars at a valuation above $2 billion [19]. On 18 June 2026 TechCrunch reported talks to raise about $300 million at just over $2 billion, with Jeff Bezos and Eric Schmidt among the backers [5]. A week later, on 25 June, the company said it had raised $320 million at a $2.3 billion valuation [2]. Axios described the round as a Series A at a $2.3 billion post-money valuation, led by Khosla Ventures and joined by General Catalyst, Hedosophia, Bezos Expeditions, Innovation Endeavors and Nico Rosberg [8]. TechCrunch's list of participants also named Eric Schmidt and researchers at Google DeepMind and MIT, and a later TechCrunch podcast said Coatue had joined the investor list [2][20]. GamesBeat, which interviewed de Witte, reported that the round had been completed in January 2026 and took six to eight weeks to close [15].

TechCrunch reported that most of the money would go to compute. The company had a deal with CoreWeave, planned to pre-train the next version of its model, and set aside part of the round to make its API more widely available by the end of the summer [2]. "It shouldn't have been possible to start a frontier lab in 2025," de Witte said in a statement quoted by Axios, crediting Khosla for backing the founding team [8].

September 2026 round

On 24 August 2026 TechCrunch reported that General Intuition was in talks to raise at a $6 billion pre-money valuation from new investors including Valor Equity Partners, Point72 Ventures and Seven Seven Six, with Khosla Ventures and General Catalyst also taking part. A source close to the deal described the round as oversubscribed, according to TechCrunch, and the company intended to spend the money on compute and hiring, with a focus on robotic embodiments [21]. On 29 September 2026 the company announced that it had raised $220 million at a $6.2 billion valuation from Valor Equity Partners, Atreides, Seven Seven Six, Point72, Khosla Ventures and General Catalyst [9][16]. The coverage did not say whether the $6.2 billion figure was pre-money or post-money. The company said it would grow its teams in New York and Europe and hire more researchers, and it opened a waitlist for its models alongside the announcement [9][16].

Funding

Round	Announced	Amount	Valuation	Investors named in coverage
Seed	16 October 2025	$133.7 million	Not disclosed	Khosla Ventures and General Catalyst (leads), Raine [1]
Series A	25 June 2026 (completed January 2026, per GamesBeat)	$320 million	$2.3 billion post-money [8]	Khosla Ventures (lead), General Catalyst, Hedosophia, Bezos Expeditions, Innovation Endeavors, Nico Rosberg; also Eric Schmidt, Coatue, and researchers at Google DeepMind and MIT [8][2][20][15]
Third round (series not named in coverage)	29 September 2026	$220 million	$6.2 billion, basis not stated [9]	Valor Equity Partners, Atreides, Seven Seven Six, Point72, Khosla Ventures, General Catalyst [9][16]

Relationship with Medal

Most coverage describes General Intuition as a spin-out of Medal [1][9]. GamesBeat, after interviewing de Witte in June 2026, reported a different arrangement: General Intuition "was not spun out of Medal; the two companies operate under a shared parent company" [15]. TechCrunch described de Witte as chief executive of both companies at the launch [1]. The two share a careers page, listed as "General Intuition & Medal", whose openings in October 2026 were based in New York City, with some roles also offered in Geneva, London and Paris [22]. Medal's blog has called General Intuition "our research lab" and described it as a public benefit corporation, while the General Intuition website names its operator as General Intuition US Inc., a Delaware corporation with an address on West 21st Street in Manhattan [23][13][14]. GamesBeat reported in June 2026 that General Intuition had a 25-person team and Medal had 65 employees [15].

Medal's reported scale has grown over the period covered by these sources:

Figure reported in	Medal usage	Source
October 2025	About 2 billion videos a year from 10 million monthly active users, across tens of thousands of games	TechCrunch [1]
December 2025	A 12-million-user clipping platform with 3.8 billion highlight clips	Latent Space episode notes [24]
March 2026	More than 15 million gamers a month; more than 2.5 billion clips expected that year	Medal [23]
June 2026	More than 17 million monthly active players	GamesBeat [15]
September 2026	On track for 3 billion uploaded videos a year	PocketGamer.biz [16]

Founders and team

Name	Role	Background
Pim de Witte	Co-founder and chief executive	Co-founded Medal in 2018 [5][17]; ran a RuneScape private server as a teenager; worked with Doctors Without Borders (MSF) and co-founded the MapSwipe disaster-mapping project [17][4]
Eloi Alonso	Co-founder and researcher	PhD in François Fleuret's group at the University of Geneva, working on reinforcement learning and world models; co-author of IRIS, Δ-IRIS and DIAMOND [25][6][7]
Adam Jelley	Co-founder and member of technical staff	PhD (2025) at the University of Edinburgh, supervised by Amos Storkey and by Sam Devlin of Microsoft Research Cambridge; joint first author of DIAMOND [26][7]
Vincent Micheli	Co-founder	PhD on world models for decision making, supervised by François Fleuret; first author of IRIS and Δ-IRIS [27][6][28]

In the Not Boring essay de Witte wrote that he learned to reverse-engineer and code to build a private RuneScape server, that the business was making about $1.5 million a year when he shut it down in 2015, and that he then joined MSF at 19 and stayed three years working on Ebola and humanitarian mapping, followed by time at Google Crisis Response [4]. TechCrunch reported that he is Dutch and that much of his team is European [2].

Other people named in coverage include Kent Rollins, the chief product officer, and Brianna Martin, the chief of staff [2]. The Not Boring essay says that Anthony Hu, who led the GAIA-1 driving world model research at Wayve, leads world modeling at General Intuition [4]. Hu is the second author of Wayve's GAIA-2 report and a joint first author of MIRA [29][4][11].

Technology and research

Action-labeled gameplay

Medal's desktop app can record what a player does alongside the video. In a July 2025 post introducing keyboard and controller overlays, Medal said the app translates physical inputs that are "provably game actions" into simple verbs such as jump or rudder-left, throws away the raw key data on the device, and stops recording when the player tabs out of the game [30]. De Witte argues that these labels, not the footage, are the key ingredient. He told TechCrunch that most competitors infer actions from video alone, which he considers insufficient, and that action data helps a model tell the "self" apart from the environment [2]. In the Not Boring essay he wrote that without action labels "video data is like shadows" and that "while inferred actions might look good on benchmarks, they struggle deeply on edge cases" [4].

He has also said that players tend to upload unusually good or bad moments, which he described as useful edge cases for training [1]. When GamesBeat asked whether that selection skewed the data, de Witte called it a shallow criticism and said the company had access to many kinds of data, including action data [15].

Action models and world models

The company works on two kinds of model, which its website describes as "action models that decide what actions to take" and "world models that predict the outcomes of actions" [3]. It treats the world model as a training environment, referred to internally as "the gym", rather than as the product; what it intends to sell is the agent [2]. "We don't build models that predict pixels or compete with game developers. We build models that predict actions," de Witte told GamesBeat [15]. In a December 2025 Latent Space interview he described fully vision-based agents that see only frames and output actions in real time, and said its first customers were replacing brittle behavior trees in games, engines and controller-based robots with a "frames in, actions out" API [24]. Khosla, in a statement for the June 2026 round, called this category "large action models" and described it as "the next big leap" [15].

At launch the company said its gaming applications would center on bots and non-player characters that improve on traditional scripted, "deterministic" bots. "[The bots] can scale to any level of difficulty," Baier-Lentz said, adding that bots which keep a player's win rate around 50 percent would maximize engagement and retention [1]. TechCrunch contrasted this with DeepMind and World Labs, which sell their world models Genie and Marble, and reported that General Intuition was focusing on other uses partly to avoid copyright issues. "Our goal is not to produce models that compete with game developers," de Witte said [1].

Robotics demonstrations

During a June 2026 visit to the New York office, TechCrunch watched a General Intuition agent play a game resembling Fortnite ("Our agent has been playing for 100 hours straight," said Kent Rollins, the chief product officer) and a quadruped robot walk around the office using a single camera. "The same brain powering the agent playing the game is powering the robot," de Witte said. A data analyst said that fine-tuning the model for the quadruped took eight minutes of real-world data, collected on the street rather than in the office [2]. The company has also tried the model on drones and in driving games. "It works on anything that you can control using a game controller or a keyboard mouse," de Witte said [2]. The reporter noted that the robot still clipped chair legs and bumped into a bin, and that getting such a model to hold up in the physical world at scale had not yet been done [2].

MIRA

On 6 July 2026 General Intuition and Kyutai, a Paris AI research lab, posted a technical report on MIRA (Multiplayer Interactive World Models with Representation Autoencoders), a project made in collaboration with Epic Games [12][11][31]. MIRA simulates Rocket League, Epic's car-football game, for four players in a 2v2 match. It is a 5-billion-parameter diffusion transformer paired with a 600-million-parameter video codec, and it generates 20 frames per second on a single NVIDIA B200 GPU, at 576p split across the four players' views [11][12]. The model has "no physics engine, no rendering engine, and no explicit 3D representation at all"; it learns the game from video and actions [11].

Aspect	Details from the MIRA report and blog
Training data	About 10,000 match-hours of 2v2 Rocket League, all played by the Nexto bot with no human gameplay; the game's physics state was logged for evaluation only [11][12]
Training method	Diffusion forcing in the latent space of a video codec, unrolled autoregressively at inference [11]
Codec	The largest gain came from a "video representation codec", which replaces the usual encoder with a pretrained self-supervised image representation model (the blog gives DINO as the example); the authors say this is when the model became stable over long rollouts [11]
Stability	Quality held steady to five minutes, the longest horizon measured, and rollouts were observed running for hours [12]
Evaluation	Fréchet Inception Distance on frames, probes that read the game's physical state from model activations, and an inverse-dynamics test of how well actions are followed [11]
Known limits	A context of roughly four seconds, so post-goal replays are invented; single-player variants forget off-screen cars; play far from the training distribution makes the image "melt", although the model often recovers [11]
Release	Training and inference code under Apache 2.0, a live demo, and the "Rocket Science" dataset (1,000 hours of matches at 720p with actions and physics state) under CC BY-NC-SA 4.0, distributed with Epic's permission [32][33][11]

The authors present MIRA as a research demonstration and "a stepping stone to physical AI", and they state that the technology is not being used to develop Rocket League [11]. The report calls it "the first multiplayer world model for highly dynamic environments governed by complex physical interactions" [12]. The claim is limited to highly dynamic, physics-driven environments, and other multiplayer world models came earlier: Odyssey's Agora-1, released in May 2026, let up to four players share one generated world, and Odyssey's announcement cited earlier multi-agent systems such as Multiverse and Solaris [34].

Earlier research by the team

The company's website lists four papers as prior research by its team [3].

Work	Venue	Team members among the authors	Contribution
IRIS, "Transformers are Sample-Efficient World Models"	ICLR 2023 (rated notable, top 5%)	Vincent Micheli, Eloi Alonso	An agent that learns inside a world model built from a discrete autoencoder and an autoregressive Transformer; a mean human-normalized score of 1.046 on the Atari 100k benchmark with the equivalent of two hours of gameplay [6]
Δ-IRIS, "Efficient World Models with Context-Aware Tokenization"	ICML 2024	Vincent Micheli, Eloi Alonso	Encodes stochastic changes between time steps; state of the art on the Crafter benchmark at several frame budgets and an order of magnitude faster to train than earlier attention-based approaches [28]
DIAMOND, "Diffusion for World Modeling: Visual Details Matter in Atari"	NeurIPS 2024 (spotlight)	Eloi Alonso, Adam Jelley, Vincent Micheli	A reinforcement learning agent trained inside a diffusion world model; a mean human-normalized score of 1.46 on Atari 100k, and a playable world model of Counter-Strike: Global Offensive [7]
GAIA-2	Wayve technical report, March 2025	Anthony Hu	A controllable multi-view generative world model for autonomous driving [29][4]

François Fleuret of the University of Geneva, in whose group Alonso did his PhD, is the last author of IRIS, Δ-IRIS and DIAMOND [25][6][28][7]. DIAMOND's other authors were Amos Storkey of the University of Edinburgh and Anssi Kanervisto and Tim Pearce of Microsoft Research, with Pearce and Fleuret credited as joint supervisors; Alonso and Jelley met during an internship at Microsoft Research Cambridge [7].

Products and commercial activity

General Intuition plans to sell access to its models rather than build end products. De Witte told TechCrunch he wants it to be a model provider that others build on, comparing it with Anthropic and OpenAI: "We're not gonna build a self-driving car company. We're gonna make it 10 times easier for the next person to build a self-driving car company." [2] In June 2026 it had a handful of customers in gaming, simulation and robotics, and de Witte said it would favor customers that could supply real-world data and diversify the embodiments the model serves [2]. As of October 2026 its website said a small number of partners in games, simulation and robotics had been onboarded to its commercial API, and its waitlist page said "Our first models are coming soon" [3][10]. SiliconANGLE reported in September 2026 that the commercial version was being tested with a limited number of customers in robotics, simulation and entertainment [9].

The company also runs Nerve, a data-collection and jobs platform for gamers. TechCrunch reported that users start with data labeling and can move on to robot teleoperation and other tasks, and that de Witte presented it as a way to give Medal's users, the generation he sees as most exposed to AI-driven job loss, a stake in the technology [2]. GamesBeat said Nerve adds to the company's pre-training data and lets people earn money "within games, tele-operated embodiments in simulation, and the real world" [15].

Data use and privacy

General Intuition's privacy notice says that when it processes Medal data for research and model development, that processing is governed by Medal's privacy policy and the arrangement between the two companies [14]. Medal said in March 2026 that the models use data from uploaded clips and controller overlays, including actions such as "player moved forward" and events such as "goal was scored", and that users can opt out at any time in Settings, "which will remove all of your past data" [23]. The opt-out is a setting called "AI Research and Development" [35]. Medal also said the research would power its own features, including smart trimming, captions, automatic clipping and montages [23]. In a section addressed to game developers it wrote, "We have no interest in building games, game engines, or asset generation systems," and said that General Intuition is focused on foundation models for spatial intelligence [23].

Stated policies

TechCrunch reported that de Witte has ruled out using the company's agents to harm people. "We don't want to be an escalatory part of the system," he said, adding that he is happy for the models to be used in search-and-rescue work [2]. TechCrunch reported that he hired Martin partly because she had publicly quit Palantir over its work with US Immigration and Customs Enforcement [2]. On gaming, de Witte told GamesBeat that gamers are "rightfully upset because AI has only taken from us", while arguing that this should not block new ideas [15].

Reception

Khosla has compared the company with OpenAI. Heath reported that Khosla thought General Intuition could matter as much for AI agents as OpenAI did for how people use large language models [18]. In June 2026 Khosla told TechCrunch he saw the startup as a long-term bet rather than an acquisition target: "At this point, it would be a data acquisition, which is sort of uninteresting." [2] In August 2026, when the company was in talks at a $6 billion pre-money valuation two months after its $2.3 billion round, TechCrunch described it as "one of the hottest startups tackling physical AI" [21].

Others have pointed to the risks. Heath called it a risky bet, noting that the right technical path for world models is disputed, that even Khosla said it was unclear which data would prove most valuable, and that the company competes with better-funded labs such as Google [18]. TechCrunch noted that whether skills learned in simulation transfer to the real world at scale is an open question [2]. In a February 2026 essay, Jeff Hawke, a co-founder of the world-model lab Odyssey, grouped General Intuition with behaviour models, which learn which action to take, rather than with world models in his sense, describing it as focused on agent behaviour in games "with the belief that this will generalise into broader environments" [36]. Debarghya Das's "The Ultimate Neolab List" includes General Intuition among the new AI labs it tracks [37].

See also

References

Rebecca Bellan, "General Intuition lands $134M seed to teach agents spatial reasoning using video game clips", TechCrunch, 16 October 2025.
Rebecca Bellan, "General Intuition's $2.3B bet that video games can train AI agents for the real world", TechCrunch, 25 June 2026.
General Intuition, "General Intuition: The frontier lab for acting in space and time" (company homepage). Accessed 27 October 2026.
Packy McCormick and Pim de Witte, "World Models: Computing the Uncomputable", Not Boring, 19 March 2026.
Rebecca Bellan, "General Intuition in talks to raise $300M at around $2B valuation", TechCrunch, 18 June 2026.
Vincent Micheli, Eloi Alonso and François Fleuret, "Transformers are Sample-Efficient World Models", arXiv:2209.00588, ICLR 2023.
Eloi Alonso, Adam Jelley, Vincent Micheli, Anssi Kanervisto, Amos Storkey, Tim Pearce and François Fleuret, "Diffusion for World Modeling: Visual Details Matter in Atari", arXiv:2405.12399, NeurIPS 2024.
Dan Primack, "General Intuition raises $320 million to develop AI from gaming", Axios, 26 June 2026.
Maria Deutscher, "World model startup General Intuition closes $220M investment", SiliconANGLE, 29 September 2026.
General Intuition, "Enter the waitlist". Accessed 27 October 2026.
General Intuition and Kyutai, "MIRA: Multiplayer Interactive World Models with Representation Autoencoders" (project blog post), July 2026.
Anthony Hu, Václav Volhejn, Adrien Ramanana Rahary, Chris Mulder, et al., "Multiplayer Interactive World Models with Representation Autoencoders", arXiv:2607.05352 (technical report), submitted 6 July 2026.
General Intuition, "Terms of Use". Accessed 27 October 2026.
General Intuition, "Privacy Notice". Accessed 27 October 2026.
Dean Takahashi, "General Intuition raises $320M at $2.3B valuation for AI frontier models based on gameplay | exclusive interview", GamesBeat, 25 June 2026.
Isa Muhammad, "General Intuition raises $220m at $6.2bn valuation", PocketGamer.biz, 30 September 2026.
Pim de Witte, personal website. Accessed 27 October 2026.
Alex Heath, "Why world models are the next big thing in AI", The Verge (excerpt of the Sources newsletter), 17 October 2025.
Alex Heath, "A new world model startup is quietly raising big money", Sources, 18 December 2025.
Rebecca Bellan and Theresa Loconsolo, "Your gaming data could be the secret to AGI, according to this Bezos-backed startup", TechCrunch Equity podcast, 8 July 2026.
Rebecca Bellan, "Valor, Point72 back General Intuition at $6B valuation as AI startup pushes into robotics", TechCrunch, 24 August 2026.
General Intuition & Medal, careers page (Ashby). Accessed 27 October 2026.
Medal, "Medal x General Intuition: How Clips And Gameplay Actions Create Spatial Intelligence", Medal blog, 19 March 2026.
Latent Space, "World Models & General Intuition: Khosla's largest bet since LLMs & OpenAI" (podcast episode with Pim de Witte), 6 December 2025.
Eloi Alonso, personal website. Accessed 27 October 2026.
Adam Jelley, personal website. Accessed 27 October 2026.
Vincent Micheli, personal website. Accessed 27 October 2026.
Vincent Micheli, Eloi Alonso and François Fleuret, "Efficient World Models with Context-Aware Tokenization", arXiv:2406.19320, ICML 2024.
Lloyd Russell, Anthony Hu, Lorenzo Bertoni, George Fedoseev, Jamie Shotton, Elahe Arani and Gianluca Corrado, "GAIA-2: A Controllable Multi-View Generative World Model for Autonomous Driving", arXiv:2503.20523, 26 March 2025.
Medal, "Privacy and security on Medal's new overlay features", Medal blog, 7 July 2025.
Kyutai, homepage. Accessed 27 October 2026.
mira-wm, "MIRA: Multiplayer Interactive World Models with Representation Autoencoders" (code repository, Apache 2.0), GitHub.
Kyutai, "Rocket Science" dataset (CC BY-NC-SA 4.0), Hugging Face.
Odyssey, "Agora-1: The Multi-Agent World Model", 18 May 2026.
Medal Support, "Opt-Out of AI Training". Accessed 27 October 2026.
Jeff Hawke, "On the Origin of Species (of World Models)", Odyssey, 27 February 2026.
Debarghya Das, "The Ultimate Neolab List". Accessed 27 October 2026.

Editing: General Intuition