World Models: The Next AI Revolution
▼ Summary
– Pim de Witte’s video game clipping platform Medal received multiple acquisition offers from AI labs after they showed strong interest in its data for training AI agents.
– Medal is spinning out a new AI lab called General Intuition, which raised a $133.7 million seed round primarily from Vinod Khosla, who believes it could be as impactful as OpenAI.
– General Intuition aims to develop foundational world models using Medal’s gaming data to enable AI to understand and interact with 3D environments, with applications in drones, robots, and self-driving cars.
– World models are a key research area in AI for achieving spatial understanding, advocated by leaders like Demis Hassabis, and involve training AI to predict and act in physical scenarios.
– The development of world models is competitive and risky, with uncertainty over the best technical approach and data value, but it has potential to create extremely valuable companies in the AI industry.
The emerging field of world models represents a pivotal frontier in artificial intelligence, focusing on teaching AI systems to understand and interact with three-dimensional spaces much like humans do. This technology promises to enable robots and autonomous agents to predict physical outcomes and take preventative actions, such as catching a falling object before it hits the ground. Pim de Witte, founder of the video game clipping platform Medal, discovered the immense value of his company’s data when several leading AI labs expressed strong acquisition interest last year.
Initially surprised by the rapid and substantial offers, including a reported $500 million bid from OpenAI, de Witte soon recognized that Medal’s repository of roughly two billion annual video uploads from tens of thousands of games offered a unique training ground for advanced AI. Rather than selling, he chose to launch a new venture named General Intuition, which has secured a $133.7 million seed round. Primary backing comes from Vinod Khosla of Khosla Ventures, an early OpenAI investor, with additional support from General Catalyst, Raine Group, and Lightspeed’s Moritz Baier-Lentz, who joins as a part-time founding member.
Khosla describes this investment as his firm’s largest seed commitment since funding OpenAI in 2018, emphasizing the potential for General Intuition to redefine AI agents. He stated, “It’s a pretty big bet. They have a unique dataset and a unique team.” World models aim to equip AI with spatial-temporal reasoning, allowing it to navigate and manipulate dynamic environments reliably. While still a specialized area, it has attracted influential proponents such as Google DeepMind CEO Demis Hassabis, who considers it essential for achieving artificial general intelligence (AGI).
Google recently showcased Genie 3, a model capable of generating interactive, game-like worlds from scratch. Other startups, including Fei-Fei Li’s World Labs, are also developing real-time interactive video models. For General Intuition, the initial goal involves controlling devices compatible with keyboard, mouse, or game controller inputs. Early applications may include search and rescue drones, with future possibilities spanning humanoid robots and autonomous vehicles.
De Witte argues that video games provide an ideal training environment because they offer clear, verifiable scenarios for distinguishing effective from ineffective actions. “Games are basically the only verifiable domain for spatial-temporal reasoning,” he explained. “You can separate a good action from a bad action, which is why it’s so valuable.” Still, the path to successful world models involves significant uncertainty. Technical approaches vary widely, and even Khosla acknowledges that the most valuable data types remain unclear.
Despite competing against well-resourced giants like Google, Khosla believes the rewards could be historic, predicting that “multiple hundred-billion-dollar and potentially even trillion-dollar companies will be built” in this sector. De Witte anticipates growing acquisition interest in gaming companies from AI labs but cautions that data providers may face diminishing leverage as models improve and require less training data. He advised, “You are at an information disadvantage. The better these models get, the less data they’re likely going to need.”
(Source: The Verge)
