Imagine AI that understands the world like we do – that can predict, react, and even learn from its mistakes in a 3D environment. That's the promise of "world models," and they're poised to revolutionize artificial intelligence as we know it. But what exactly are they, and why are they suddenly so hot?
This story begins with Pim de Witte, the founder of Medal, a popular video game clipping platform. Around the middle of last year, de Witte started exploring potential collaborations with leading AI labs, offering them access to Medal's vast trove of gaming data for training their AI agents. He quickly discovered something extraordinary: his data was far more valuable than he initially thought. "We received multiple acquisition offers very quickly," he revealed, hinting at a feeding frenzy among AI giants. While he didn't disclose specific names, reports suggest that OpenAI was among those interested, allegedly offering a staggering $500 million.
Initially, de Witte considered these offers. "Initially, we were quite interested in them," he admitted, but that “was mostly a result of us not understanding what we were sitting on.” However, after delving into research, particularly a Google DeepMind paper demonstrating the potential of gaming data for teaching AI navigation in 3D environments, he realized the true potential of Medal's data. With roughly 2 billion video uploads annually from tens of thousands of video games, Medal possessed a unique foundation for extending AI's capabilities into the real world. And this is the part most people miss: it's not just about processing data; it's about understanding the context within that data.
This epiphany led to a bold move: the creation of General Intuition, a new AI lab spun out from Medal. This venture has secured a massive $133.7 million seed round, spearheaded by Vinod Khosla, the founder of Khosla Ventures and an early backer of OpenAI. Other investors include General Catalyst and The Raine Group. Moritz Baier-Lentz, from Lightspeed, is also joining the startup part-time as a founding team member.
Khosla sees General Intuition as potentially as transformative for AI agents as OpenAI was for large language models. He believes that the company's unique dataset and team position it for significant impact. "It’s a pretty big bet," he stated, emphasizing the high-stakes nature of the investment.
So, what are world models, exactly? In simple terms, they're AI systems trained to possess spatial understanding akin to humans. Imagine a robot that can anticipate the consequences of an action, like preventing a glass from falling off a table. More broadly, AI researchers are increasingly viewing world models as crucial for training agents that can reliably interact with and generate 3D spaces. Think of it as giving AI a sense of intuition about the physical world.
Google DeepMind CEO Demis Hassabis has been a prominent advocate for world models, emphasizing their importance in achieving Artificial General Intelligence (AGI). Google recently showcased Genie 3, an impressive model that generates interactive video game-like environments on the fly. Other players are also emerging, including Fei-Fei Li's World Labs, which recently unveiled its own real-time interactive video generation model.
General Intuition's vision is to create AI that can control devices using keyboard, mouse, or game controller-like inputs. De Witte envisions initial applications in search and rescue drones, with potential extending to humanoid robots and self-driving cars. Just as LLMs were initially trained on internet text data, de Witte believes that gaming environments will unlock AI’s ability to reliably predict the proper action to take in the physical world. “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.” He's essentially saying that games provide a controlled environment where AI can learn from its mistakes and refine its understanding of cause and effect.
But here's where it gets controversial... The path to developing effective world models is far from settled. As Khosla himself acknowledged, the optimal data and technical approaches remain uncertain. General Intuition faces competition from well-funded giants like Google, and the field is rapidly evolving. "Somebody will win big in this market," Khosla predicted, suggesting that multiple hundred-billion-dollar or even trillion-dollar companies could emerge.
De Witte anticipates that gaming companies will become attractive acquisition targets for AI labs as the focus on world models intensifies. His decision to launch General Intuition stems from his realization that Medal's data gives him a unique advantage, positioning him as more than just a data provider. However, he cautions that other gaming companies may find it difficult to resist lucrative licensing deals and acquisition offers from larger AI players.
“You are at an information disadvantage,” he warned, advising gaming companies to be aware of their valuable data assets and their potential impact on the future of AI. “The better these models get, the less data they’re likely going to need.” Essentially, he's suggesting that gaming companies need to act now to leverage their data before it becomes less valuable.
This raises a critical question: Will gaming companies become the new oil fields for AI development, or will they retain control of their valuable data and shape the future of AI themselves? What's your take on this? Are world models the key to unlocking true AI, or is this another overhyped trend? And who do you think will ultimately dominate the world model landscape – established AI giants or innovative startups like General Intuition? Share your thoughts in the comments below!
