Why this CEO thinks video games make better training data than the internet
When it comes to achieving artificial general intelligence (AGI), large language models just don’t have what it takes. Models like ChatGPT and Claude are great at text, but they’re less skilled at und
When it comes to achieving artificial general intelligence (AGI), large language models just don’t have what it takes. Models like ChatGPT and Claude
Read Full Story at TechCrunch →Why This Matters
The debate over training data quality is reshaping the race toward AGI, forcing a reckoning with the limitations of web-scraped content. If structured, interactive environments prove more effective than uncurated internet data, it could redefine how AI learns—and who controls that process. The implications stretch beyond technology, touching on education, ethics, and the very definition of intelligence in machines.
Background Context
For years, large language models have relied on massive, unfiltered datasets scraped from the web, despite their noise, bias, and lack of grounding in real-world tasks. Meanwhile, video games—especially those with physics engines and complex environments—have quietly emerged as a testing ground for embodied AI, offering controlled yet dynamic scenarios that mirror human learning. The disconnect between these approaches highlights a deeper tension: Should AI mimic human-like reasoning through exposure to raw information, or should it be shaped by tasks that demand interaction and adaptation?
What Happens Next
The shift toward game-based training could accelerate the development of AI systems that excel at spatial reasoning, collaboration, and even ethical decision-making—areas where text-only models falter. However, it also risks entrenching the dominance of a few corporations that control proprietary gaming environments, raising questions about transparency and accessibility. Policymakers and researchers will likely grapple with whether to regulate or incentivize this new frontier of AI training.
Bigger Picture
This moment reflects a broader pivot in AI research, from passive data ingestion to active, task-driven learning—a return to the roots of reinforcement learning but with far greater stakes. It also underscores the growing divide between AI "science" and AI "engineering," where theoretical breakthroughs must now prove their utility in simulated or real-world systems. The outcome may determine whether AGI remains a lab experiment or becomes a tool shaped by human-designed challenges.
