Karpathy 2025 LLM Year in Review (karpathy.bearblog.dev)

2025 marked a transformative year for large language models (LLMs), with the introduction of significant paradigm shifts that reshaped the AI landscape. A standout development was Reinforcement Learning from Verifiable Rewards (RLVR), which added a new phase to traditional model training by optimizing LLMs against automatically verifiable rewards in diverse environments. This method enables LLMs to adopt complex reasoning strategies that mimic human-like problem-solving, enhancing their capabilities significantly. As RLVR required extensive computational resources, it shifted focus away from pretraining, allowing LLMs to achieve notable improvements in performance and reasoning depth. Additionally, the emergence of new LLM applications and interfaces, exemplified by tools like Cursor and Claude Code, showcased the evolution of user interaction with AI. Cursor has initiated a new layer of LLM applications designed for specific tasks, allowing for complex orchestration of LLM calls, while Claude Code exemplifies an on-device AI agent that integrates tool use and reasoning directly within a user's environment. The year also saw the rise of "vibe coding," where users could create software through natural language, democratizing programming beyond traditional professionals. These advancements illustrate LLMs' potential not only as intelligence entities but also as integral components of personal computing, ushering in a new era of human-computer interaction and creativity.