Igor Babuschkin explains his departure from xAI to found River AI, driven by a desire to distribute AI's benefits and address power concentration. River AI focuses on three bets: a reinforcement learning API, personalized AI agents, and local hardware for running frontier models. He argues that personal AI, which learns from individual user signals, will be the next major application, moving beyond coding and math into everyday life. On the enterprise side, he predicts a shift toward companies training their own models as open-source improves, leveraging proprietary data for competitive advantage. He also discusses the challenges facing proprietary labs, including GPU and data constraints, and expresses skepticism about RL generalizing across domains. Igor highlights the dominance of Chinese open-weight models, calling for a top US open model, and reflects on his time at xAI, including the rapid construction of the Colossus data center and the strategic acquisition of Cursor. He identifies the lack of training methods for long-horizon, non-verifiable tasks as a key bottleneck, and emphasizes the need for alignment research to keep humans relevant as AI advances.