Source: Dwarkesh Patel
The industry has shifted from scaling static datasets to training AI systems through active trial-and-error across millions of verifiable tasks. This treats environment interaction as essential for AGI rather than relying on bigger models trained on larger corpora. Resources are consolidating around reinforcement learning infrastructure and simulation environments. The next 18-24 months will show whether labs like OpenAI and DeepMind can execute this transition or whether gains plateau again. Whoever builds scalable on-the-job learning first likely controls the economic moat for the next wave of AI capability. This, not chat quality or multimodal features, is where the competition matters.