Gaming’s Unlikely Gift: Can Video Games Train AI to Move in More Than Just Code?

The pursuit of artificial general intelligence (AGI) has been a decades-long endeavor, with many approaches yielding incremental progress but no silver bullet. Some researchers and entrepreneurs believe that video games may hold the key to AGI. This notion might seem far-fetched, but it’s gaining traction.

General Intuition, a startup backed by Jeff Bezos, has invested heavily in this idea, securing $320 million in funding from investors like Coatue and Eric Schmidt. The company’s CEO, Pim de Witte, recently discussed the potential of gaming-trained models on the TechCrunch podcast Equity. According to de Witte, traditional large language models struggle with understanding physical motion – a fundamental aspect of intelligence that generalizes.

The Gap Between Code and Reality

The gap in question is more than just a technical quibble; it’s a matter of cognitive scope. Language models excel at processing text but struggle to grasp spatial relationships and temporal dynamics. In contrast, video games offer experiential data that can teach AI systems how objects interact with their environment.

Games encompass diverse scenarios, from physics simulations to complex strategy games. This heterogeneity allows AI models to learn from multiple contexts and adapt more readily to real-world situations. Video games provide a training ground for AGI that’s far more extensive than static, text-based datasets used by traditional language models.

The Birth of General Intuition

General Intuition emerged from Medal TV, a gaming platform focused on streaming and community building. The startup’s founders recognized an opportunity to combine their expertise in game development and data analysis with the pursuit of AGI. By harnessing the collective knowledge embedded within games, they aim to create AI systems that can generalize across domains – not just process vast amounts of text.

Ethical Considerations: From Games to Defense

As researchers at General Intuition push the boundaries of gaming-trained models, questions arise about their potential applications. The company’s investors, including those from defense-oriented institutions like Google DeepMind and MIT, raise concerns about the ethics of developing AI that could be used for military purposes.

While some may argue that AGI is an inherently neutral pursuit, its development carries significant risks. As AI systems become increasingly sophisticated, their capabilities will inevitably attract attention from governments and industries with agendas beyond research or entertainment. General Intuition’s focus on gaming data takes on a different significance: it offers a potential means of creating AGI that can navigate the complexities of physical environments without being co-opted for nefarious purposes.

A New Era in AI Research?

General Intuition’s bet on gaming data marks an intriguing shift in the AGI landscape. By emphasizing experiential learning and adaptability, this approach has the potential to bypass some of the traditional hurdles faced by language models. As researchers continue to explore the possibilities of video game-trained AI, they may be forced to confront fundamental questions about what intelligence truly means – and how it can be cultivated.

If General Intuition’s gamble pays off, we might witness a transformative convergence of gaming, AI research, and real-world problem-solving. The notion that the next breakthrough in AGI will arise from the unlikely realm of video games is captivating, to say the least.