AI startup General Intuition CEO reveals: Why is "game data" the ultimate key to AGI?

Is the next breakthrough for large language models (LLMs) hidden in video games? According to the latest podcast interview released by TechCrunch today (8th), Pim de Witte, CEO of AI startup General Intuition, which is backed heavily by Jeff Bezos and valued at $2.3 billion, revealed that 'gaming data' with precise action labels is the ultimate solution for training AI world models. They use hundreds of millions of hours of game footage to train AI, not only allowing agents to play Fortnite for 100 hours straight, but also to transfer the same brain to real-world four-legged robots.
(Previous: Musk's SpaceXAI officially launches its strongest model 'Grok 4.5'! Joining forces with Cursor to aggressively target AI coding agents, fully integrating with Office software)
(Background supplement: Fed June FOMC meeting minutes: Removes dovish bias! AI frenzy and Middle East conflict fuel inflation, not ruling out another rate hike)

Table of Contents

Toggle

• Why gaming data crushes regular internet videos?
• From virtual worlds to driving physical robot dogs
• Top institutions place heavy bets and corporate vision

As large language models (LLMs) like ChatGPT and Claude gradually peak in text processing capabilities, the tech world is turning its attention to the next piece of the artificial general intelligence (AGI) puzzle — enabling AI to truly understand the physical world.

On July 8, 2026, TechCrunch senior reporter Rebecca Bellan hosted an episode of the Equity podcast, interviewing Pim de Witte, CEO of General Intuition, a startup that has recently made a name for itself in the AI circle. This company, spun off from game clip platform Medal TV, is proving to the world from a disruptive perspective that data generated from 'playing video games' will be the key fuel for the next generation of embodied AI.

Why gaming data crushes regular internet videos?

Pim de Witte pointed out in the interview that current LLMs, while proficient in human language, severely lack the ability of 'spatial-temporal reasoning,' which is the biggest gap for generalized intelligence in the physical world. To fill this gap, General Intuition chose the hundreds of millions of hours, even tens of billions of clips, of high-quality game footage accumulated on the Medal platform as training material.

Compared to traditional internet videos like YouTube, gaming data has incomparable advantages. Internet videos often have chaotic perspectives, inconsistent quality, and require massive computing power for human pose estimation. In contrast, gaming data not only provides consistent physics simulation and first-person perspectives, but most importantly, these videos have 'precise action labels' embedded at their core. The AI can know exactly which buttons the player pressed and what decisions they made at the moment the screen changes. Additionally, game clips often capture the player's 'peak moments' (extreme edge cases, both excellent and terrible), which is highly valuable for training models.

From virtual worlds to driving physical robot dogs

General Intuition is using this massive data to train powerful 'World Models.' These models do not rely on traditional game engines to operate, but can generate simulated environments frame by frame, allowing the AI to understand physical causal relationships such as 'walls are impenetrable,' 'ladders are usable,' and 'light and shadow change over time.'

More astonishingly, this intelligence cultivated in virtual space can be seamlessly transferred to robots in the physical world. Pim de Witte gave an example: the AI agents they trained can not only play Fortnite continuously for 100 hours, but the team can even implant the same 'brain' into a four-legged robot dog. By feeding it only about 8 minutes of real street data for fine-tuning, the robot dog can move smoothly in real-world environments. Their future business model will focus on providing APIs, allowing other companies to apply this technology to areas such as autonomous vehicles and search-and-rescue robots.

Top institutions place heavy bets and corporate vision

This revolutionary technical route has attracted the attention of top global capital. General Intuition includes the following astonishing funding and corporate background:

| Dimension | | --- | Key Data & Info | | --- | --- | | Latest Funding & Valuation | Raised $320 million at a $2.3 billion valuation (led by Khosla Ventures), with total funding reaching $454 million. | | Star Investor Lineup | Amazon founder Jeff Bezos, Coatue, former Google CEO Eric Schmidt, MIT and Google DeepMind researchers. | | Social Impact & Economy | Launched the Nerve platform, allowing Medal players to earn income by labeling data or remotely operating robots, offsetting AI's impact on employment. | | Tech Ethics Red Line | Strictly prohibits applying the model to lethal autonomous weapons, only supporting fields that benefit humanity such as search and rescue. |

Vinod Khosla, founder of Khosla Ventures, commented that labeled data containing 'human actions and reactions' in games is key to giving world models an 'intuition-like capability,' a breakthrough as important as the early demonstration of reasoning ability in LLMs. As General Intuition adheres to European values of not engaging in military applications (the CEO previously worked for Doctors Without Borders), this unicorn is trying to prove to the world that the data legacy left by the entertainment industry may be the most valuable stepping stone for humanity into the AGI era.