Andrej Karpathy, a leading expert, reviewed the development of large-scale models in 2025. Key points: 1. In 2025, Verifiable Reward Reinforcement Learning (RLVR) will become the new mainstay of LLM training; 2. Cursors represent a completely new application layer. The rocket-like growth of curves proves the existence of a substantial and opportunity-rich "application layer" on top of basic large-scale models; 3. Claude Code twitter.com/qinbafrank/status/...