General Intuition is spending $134 million to teach agents to think spatially using video game clips

Medal, a platform for uploading and sharing video game clips, has established a pioneering recent artificial intelligence research lab that uses its game videos to train and build basic models and artificial intelligence agents that may understand how objects and entities move through space and time – a concept often called spatio-temporal inference.

A startup called General Intuition is betting that Medal’s dataset – which consists of two billion videos a yr from 10 million monthly energetic users across tens of hundreds of games – will outperform alternatives like Twitch or YouTube for training agents.

“When you play video games, you essentially transfer your perception, usually through a first-person camera view, to different environments,” Pim de Witte, CEO of Medal and General Intuition, told TechCrunch. He noticed that players who upload clips tend to post very negative or positive examples that function really useful edge cases for training. “You get this selection bias towards exactly the kind of data that you actually want to use in your training work.”

This moat of knowledge has reportedly attracted the attention of OpenAI, which attempted to acquire Medal for $500 million late last yr with Information. (Neither OpenAI nor General Intuition commented on the report.)

It also led to General Intuition raising a whopping $133.7 million in seed funding, led by Khosla Ventures and General Catalyst with participation from Raine.

The startup intends to use the funding to develop a team of researchers and engineers focused on training a general agent that may interact with the world around it, with initial applications in gaming and search and rescue drones in mind.

De Witte says the founding team has already made progress: the General Intuition model can understand environments it has not been trained in and accurately predict actions in them. It is only able to do this through visual data; agents only see what a human player would see and move in space following the controller’s signals. The company says this approach will be naturally transferred to physical systems corresponding to robots, drones and autonomous vehicles, which are often manipulated by humans using video game controllers.

General Intuition’s next milestone is twofold: generating recent simulated worlds for training other agents and autonomously navigating completely unknown physical environments.

This technical approach shapes the way the company plans to commercialize its technology and distinguishes it from competitors building global models.

Although General Intuition also builds models of the world on which they train their agents, such models are not the product. Unlike other global modelers corresponding to DeepMind and World Labs who sell their global Genie models and Marblefor training agents and content creation respectively, General Intuition focuses on other use cases to avoid copyright issues.

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

Instead, the startup’s gaming applications revolve around creating bots and NPCs that may outperform traditional “deterministic bots” or pre-programmed characters that produce the same result every time.

“[The bots] can scale to any level of difficulty,” Moritz Baier-Lentz, founding father of General Intuition and partner at Lightspeed Ventures, told TechCrunch. “It’s not necessary to create a bot that can beat everyone, but if you can incrementally scale and replenish liquidity for each player situation so that their win rate is always around 50%, that will maximize their engagement and retention.”

De Witte also has a background in humanitarian work, which indicates the startup is focused on powering search and rescue drones, which sometimes need to navigate unfamiliar environments and extract information without GPS.

Ultimately, de Witte and Baier-Lentz see General Intuition’s core functionality—spatial-temporal reasoning—as a key component of the race toward artificial general intelligence (AGI). While mainstream AI labs focus on building increasingly powerful large language models, General Intuition believes that true AGI requires something that an LLM fundamentally lacks.

“As humans, we create text to describe what is happening in our world, but in doing so we lose a lot of information,” de Witte said. “You lose your general intuition about spatio-temporal reasoning.”