From Attention to Agents
Five papers that trace how today’s assistants and coding agents came to be: a parallel architecture, a law that made scale predictable, a training recipe that made models follow instructions, a prompting trick that made them think first, and the minimal loop that joined thinking to acting. Read in order; each explains why the next one happened.
- Deep read · 22 min
注意力就是全部:Transformer 如何改写了序列建模
Architecture: attention replaces recurrence; training becomes parallel.
- Pick · 2 min
规模律:为什么「把模型做大」从信仰变成了工程
Law: loss falls as a power law of scale; scaling becomes a calculation.
- Pick · 2 min
InstructGPT:用人类反馈把模型调教成助手
Training: human feedback turns a text continuer into an assistant.
- Pick · 2 min
思维链提示:让模型「先想再答」为什么有效
Reasoning: write out the steps and large models reason.
- Pick · 2 min
ReAct:推理与行动交替,Agent 的最小原型
Action: alternate thinking and acting, the minimal agent loop.
Picks in this series are expanded into deep reads over time; the order stays the same. All research →