As better chatbots get harder to build, AI turns to simulated worlds
AI researchers are turning to simulated 3D worlds to train smarter chatbots, as progress from text-only data slows. These virtual environments teach AI physical reasoning and common sense. Major labs like Google DeepMind and Meta are investing heavily in this approach.
Background
- The article discusses how AI companies are increasingly using simulated 3D environments (virtual worlds) to train next-generation AI models, since progress from simply scaling up text data and compute ("bigger models, more data") is showing diminishing returns.
- This shift is driven by a shortage of high-quality, diverse training text on the public internet. Simulated worlds (like Minecraft, or custom-built 3D spaces) can generate limitless, varied training data for tasks involving planning, reasoning, physical intuition, and multi-step problem-solving — capabilities that pure language models struggle with.
- Key players in this space include DeepMind (Google), OpenAI, Meta, and startups like Sam Altman-backed 1X (robotics); their goal is to build "world models" — AI systems that understand the causal and physical structure of environments, not just statistical patterns in text.
- The approach is seen as a possible path toward more general artificial intelligence (AGI), but critics question whether skills learned in simulation will transfer reliably to the messy, unpredictable real world.