A structured, evolving journey through AI research.
Follow the most important papers, understand the ideas that shaped the field, then try them yourself — in an engaging way.
408 papers · 1763–2025 · every idea touchable
408 papers, 1763–2025, one connected sky.
The stars led you to this paper
DeepSeek-R1: Incentivizing Reasoning Capability in LLMs via Reinforcement Learning · 2025
Imagine a student who, instead of copying a teacher's worked solutions, just gets told "right" or "wrong" after each attempt. Over thousands of tries, the student discovers her own problem-solving tricks — pausing to double-check, trying a different approach when stuck, even talking herself through hard steps. DeepSeek-R1 does exactly this: it learns to reason not by imitating humans, but by trial-and-error with reinforcement learning.
Read the full paper →Every paper: the idea in simple terms, its practical application, and its relationship to previous and subsequent research.