This is an excerpt from Source by Alex HeathA newsletter about AI and the tech industry, syndicated to The Verge subscribers only once a week.
Around the middle of last year, Pim de Witte began reaching out to some major AI labs to see if they were interested in using the data. medalTheir popular video game clipping platform to train their agents.
Within a few weeks, it became clear that Medel’s data was more valuable to the laboratories than they had anticipated. “We received several acquisition offers very quickly,” he told me. (He declined to be named, but it it has been reported OpenAI offered $500 million.) “Initially, we were quite interested in them,” he said of the offers, but “it was mostly a result of us not understanding what we were sitting on.”
He read Google DeepMind research paper This shows that gaming data can be used to teach AI how to navigate in 3D environments. However, the interest of AI labs led them to realize that their data from Medal, which receives approximately 2 billion video uploads from thousands of video games per year, could be used to develop a unique foundational model for extending AI to the real world.
“It’s a huge bet.”
Today, Pim de Witte announced that Medall is building a new AI lab called General Intuition, which has raised a seed round of $133.7 million. The funding for this round primarily comes from Vinod Khosla, founder of Khosla Ventures and one of the first investors in OpenAI. Other investors include General Catalyst and Rhine Group. Moritz Baer-Lentz, who oversees Lightspeed’s gaming investments, is also joining the startup part-time as a member of the founding team.
Khosla believes that General Intuition could be as influential on the field of AI agents as OpenAI was on how people use large language models. This is his company’s largest seed check since backing OpenAI in 2018. “It’s a huge bet,” he told me. “They have a unique dataset and a unique team.”
Unless you’re immersed in the world of AI, you probably haven’t heard much about world models yet. It is a branch of research that trains AI to have human-like spatial understanding. The idea is that a robot could, for example, predict when a glass of water will fall off the table and catch it before it falls. More practically, AI researchers are increasingly looking to world models as a way to train agents that can reliably generate and interact with 3D space.
Among major AI leaders, Google DeepMind CEO Demis Hassabis has been the most vocal proponent of world models and their importance in achieving AGI. Google recently showed off Genie 3, a model that generates a video game-like environment right from the start as you navigate. A handful of startups are also working on a similar model, including Fei-Fei Li’s World Labs, which Released its own demo this week A model that generates interactive videos in real time.
According to De Witte, for general intuition, the goal is to control any type of device that can be mapped to a keyboard and mouse or have an input scheme such as a game controller. He hopes the startup’s first models will be used by search and rescue drones, but he sees potential for applications in other areas, including humanoid robots and self-driving cars.
Just as LLM was initially trained on Internet text data, De Witte believes that gaming environments will unlock the potential of AI to reliably predict appropriate actions that would occur in the physical world. “Games are basically the only verifiable domain for spatial-temporal reasoning,” he explained. “You can separate a good deed from a bad deed, that’s why it’s so valuable.”
Still, it’s a risky bet. There is heated debate in the AI industry over the right technical path to develop world models, and as Khosla also told me, it is not clear which data will ultimately prove most valuable. Members of De Witte’s early research team has published Notable Research in the field, but the startup is still competing with better-funded giants like Google. “Someone will win big in this market,” Khosla said. He told me he believes this is an area where “several hundred billion dollars and potentially even trillions of dollars of companies will be created.”
De Witte predicts that as interest in world models grows, gaming companies will become prime acquisition targets for AI laboratories. His decision to start General Intuition was driven by the realization that, thanks to Medall’s data, he was uniquely positioned to become more than a data supplier. However, he cautioned me that others may find it challenging to resist license scrutiny and acquisition offers from large AI labs.
When I asked if he had any advice for the gaming industry he said, “You’re ill-informed.” “The better these models, the less data they need.”
