Six major paradigm shifts defined LLM progress in 2025: RLVR emerged as a new training stage that enables reasoning capabilities through verifiable rewards; LLMs developed jagged intelligence profiles distinct from human cognition; Cursor revealed a new layer of vertical-specific LLM applications; Claude Code demonstrated agents running locally on user computers; vibe coding enabled programming through natural language; and multimodal models like Gemini Nano hint at visual LLM interfaces beyond text chat.

9m read timeFrom karpathy.bearblog.dev
Post cover image
Table of contents
1. Reinforcement Learning from Verifiable Rewards (RLVR)2. Ghosts vs. Animals / Jagged Intelligence3. Cursor / new layer of LLM apps4. Claude Code / AI that lives on your computer5. Vibe coding6. Nano banana / LLM GUI

Sort: