General Intuition lands $134M seed to teach agents spatial reasoning using video game clips

Medal, a platform for uploading and sharing video game clips, has created a new frontier AI research lab that is using footage from its gaming videos to train and build foundational models and AI agents that understand how objects and entities move through space and time — a concept known as spatio-temporal reasoning.

Called General Intuition, the startup is betting that Medal’s dataset — which consists of 2 billion videos per year from 10 million monthly active users across thousands of games — outperforms alternatives like Twitch or YouTube for training agents.

“When you play video games, you essentially transfer your perception, usually through the first-person view of the camera, to different environments,” Pim de Wit, CEO of Medel and General Intuition, told TechCrunch. He notes that gamers who upload clips tend to post very negative or positive examples, which serves as a really useful edge case for training. “You get this selection biased towards the kind of data you actually want to use for the training task.”

It was this data mining that attracted the attention of OpenAI, which sought to acquire the medal for $500 million late last year. Information. (OpenAI or General Insights will not comment on the report.)

General Intuition has raised a whopping $133.7 million in seed funding led by Khosla Ventures and General Catalyst with Rain’s participation.

The startup intends to use the funding to grow its team of researchers and engineers to train a general agent that can interact with the world around it, with the goal of early applications in gaming and search-and-rescue drones.

De Wit says the founding team has already made progress: General Intuition’s model can understand environments it wasn’t trained in and accurately predict actions in them. It is able to do this purely through visual input; Agents see only what a human player would see, and they move through space following controller input. This approach, the company says, could naturally transfer to physical systems like robotic arms, drones and autonomous vehicles, which are often piloted by humans using video game controllers.

The next milestone for general intuition is twofold: creating new simulated worlds to train other agents and navigate autonomously in completely unfamiliar physical environments.

This technological approach is shaping how the company plans to commercialize its technology and differentiates it from competitors in the world’s model making.

Although General Intuition is creating world models on which to train its agents, such models are not products. Unlike other world model makers like DeepMind and World Labs, who are selling their world models the genie And MarbleFor training agents and content creation, respectively, common intuition is focusing on other use cases to avoid copyright issues.

“Our goal is not to create models that compete with game developers,” says de Wit.

Instead, the startup’s gaming applications are focused on creating bots and non-player characters that can outperform traditional “deterministic bots” or pre-programmed characters that produce the same output every time.

“[The bots] Any level of difficulty can be,” Moritz Baier-Lentz, a founding member of General Intuition and partner at Lightspeed Ventures, told TechCrunch. “It’s not mandatory to build a god bot that beats everyone, but if you can slowly scale and fill liquidity for any player’s situation so that their win rate will always be at most 5%.”

De Witte also has a background in humanitarian work, which informs the startup’s focus on powering search-and-rescue drones, which sometimes need to navigate unfamiliar environments and extract information without GPS.

Ultimately, de Witte and Baer-Lentz see the core function of general intuition — spatio-temporal reasoning — as a critical part of the race toward artificial general intelligence (AGI). While major AI labs focus on building more powerful large language models, General Intuition believes that true AGI requires something that LLMs fundamentally lack.

“As humans, we create text to describe what’s going on in our world, but in doing so you lose a lot of information,” says de Wit. “You lose the general intuition about spatio-temporal reasoning.”