<?xml version="1.0" encoding="UTF-8"?><rss version="2.0"><channel><title>AGI Dawn</title><description>把智能分发给每一个人：研究深读、精选与前沿动态。</description><link>https://agidawn.cc/</link><language>zh-CN</language><item><title>官网改版：研究板块与 Claude Code 指南上线</title><link>https://agidawn.cc/news/2026-10-17-site-update/</link><guid isPermaLink="true">https://agidawn.cc/news/2026-10-17-site-update/</guid><description>新增研究板块（论文深读、正典清单、前沿动态、概念卡）、10 章 Claude Code 指南和团队与企业页；接入文档一并改版。</description><pubDate>Sat, 17 Oct 2026 00:00:00 GMT</pubDate><category>flash</category><category>agidawn</category></item><item><title>注意力就是全部：Transformer 如何改写了序列建模</title><link>https://agidawn.cc/research/attention-is-all-you-need/</link><guid isPermaLink="true">https://agidawn.cc/research/attention-is-all-you-need/</guid><description>它把「按顺序读」换成「同时看全句、按相关性取信息」，让训练可以完全并行；此后所有大模型都长在这副骨架上。</description><pubDate>Fri, 16 Oct 2026 00:00:00 GMT</pubDate><category>deep-read</category><category>基础架构与预训练</category></item><item><title>思维链提示：让模型「先想再答」为什么有效</title><link>https://agidawn.cc/research/chain-of-thought-prompting/</link><guid isPermaLink="true">https://agidawn.cc/research/chain-of-thought-prompting/</guid><description>在示例里写出中间推理步骤，大模型的算术与常识推理成绩就跃升；它确立了「让模型先想再答」的范式，也是推理模型时代的概念起点。</description><pubDate>Fri, 16 Oct 2026 00:00:00 GMT</pubDate><category>pick</category><category>推理与思维链</category></item><item><title>InstructGPT：用人类反馈把模型调教成助手</title><link>https://agidawn.cc/research/instructgpt-rlhf/</link><guid isPermaLink="true">https://agidawn.cc/research/instructgpt-rlhf/</guid><description>监督微调、奖励模型、强化学习三步，让 13 亿参数的模型在人类偏好上胜过 1750 亿的 GPT-3；这是 ChatGPT 的直接技术蓝本。</description><pubDate>Fri, 16 Oct 2026 00:00:00 GMT</pubDate><category>pick</category><category>对齐与后训练</category></item><item><title>ReAct：推理与行动交替，Agent 的最小原型</title><link>https://agidawn.cc/research/react-reasoning-and-acting/</link><guid isPermaLink="true">https://agidawn.cc/research/react-reasoning-and-acting/</guid><description>让模型在「想一步、做一步、看结果」之间循环，把思维链和工具调用合为一个回路；今天几乎所有 Agent 框架的循环都从这里来。</description><pubDate>Fri, 16 Oct 2026 00:00:00 GMT</pubDate><category>pick</category><category>Agent 与工具使用</category></item><item><title>规模律：为什么「把模型做大」从信仰变成了工程</title><link>https://agidawn.cc/research/scaling-laws-for-neural-language-models/</link><guid isPermaLink="true">https://agidawn.cc/research/scaling-laws-for-neural-language-models/</guid><description>语言模型的损失随参数量、数据量与算力按平滑的幂律下降，三者之间有可预测的最优配比；它让扩大规模成为可以提前算账的决策。</description><pubDate>Fri, 16 Oct 2026 00:00:00 GMT</pubDate><category>pick</category><category>规模律与数据</category></item><item><title>小模型价格降了一个数量级；缓存读减半；前沿模型开始做数学</title><link>https://agidawn.cc/news/issues/2026-w42/</link><guid isPermaLink="true">https://agidawn.cc/news/issues/2026-w42/</guid><description>Haiku 5.5 把小模型定价拉到 $0.10 / $0.50，Sonnet 5.5 缓存读减半；Anthropic 使用政策 11 月生效；OpenAI 公开模型产出的数学结果与 Lean 形式化证明。</description><pubDate>Fri, 16 Oct 2026 00:00:00 GMT</pubDate><category>issue</category><category>both</category></item></channel></rss>