# 推文总结 2026-04-29 ## 总览 - 账号范围:18 个 - 活跃账号:15 个 - 总推文数:148 条 - 主要主题:中文 AI 圈 / AI 编码 / Agent / 模型与研究 ## 重点账号动态 - @bcherny: 当日未抓到新推文 - @karpathy: 当日未抓到新推文 - @trq212: 4 条;重点:@patrickc @link this is incredible [链接](https://x.com/trq212/status/2049602400873509243) - @gdb: 4 条;重点:a great codex tutorial: [链接](https://x.com/gdb/status/2049315404271469054) - @dotey: 23 条;重点:转译:深度拆解 Hermes Agent 的记忆系统:它如何修正 OpenClaw 的误区 如果你读过我之前关于 ChatGPT、Claude 以及 Clawdbot 记忆系统的文章,你就会知道我一直在钻研同一个问题:这些 AI 智能体(AI Agent)到底是怎么记事的? Hermes Agent [链接](https://x.com/dotey/status/2049534755729707205) - @oran_ge: 14 条;重点:为什么很多人要黑罗福莉呢? 能力超强,言语谦逊 未来一年,星光璀璨,不可限量 https://t.co/IdNMR9nKdG [链接](https://x.com/oran_ge/status/2049473260438557097) - @AnthropicAI: 3 条;重点:New on the Science Blog: We gave Claude 99 problems analyzing real biological data and compared its performance against an expert panel. On 23 problem [链接](https://x.com/AnthropicAI/status/2049624600741560340) - @dongxi_nlp: 3 条;重点:Interesting work! Alignment interventions may not erase misalignment; they may relocate it behind a contextual gate. That raises a deeper question: wh [链接](https://x.com/dongxi_nlp/status/2049581536815493277) - @jiangydev: 当日未抓到新推文 - @lifesinger: 5 条;重点:AI 产品的商业模式,正在快速趋同: 1、会员套餐,是收取壳的钱。类似买车,用户可选择是小轿车,还是 SUV,不同的车,价钱不一样。只是 AI 产品这里,是按月或按年收取车的钱。更像是租车或租房,定期收取租金。 2、然后是 Costco 的货架电商。提供各种大模型卖给你,或者把模型能力,结合 Har [链接](https://x.com/lifesinger/status/2049416668154397143) - @gregisenberg: 12 条;重点:The last 18 months changed more than the previous 10 years. The world is moving at an unprecedented speed I stopped making plans longer than 90 days i [链接](https://x.com/gregisenberg/status/2049293619601788988) - @garrytan: 24 条;重点:Saikat Chakrabarti made his fortune in tech. Now he wants to ban the data centers that power AI in San Francisco, the most important growth industry i [链接](https://x.com/garrytan/status/2049284944737009744) - @signulll: 14 条;重点:msft, goog, meta, & amazon are on track to spend ~$700b on ai infrastructure in 2026. this kinda spending usually happens via govts or wars whereas th [链接](https://x.com/signulll/status/2049614400860921877) - @thedankoe: 2 条;重点:Most people quit because they forget that you have to be bad at something before you can be good at it. It's so obvious. You suck. Of course you're no [链接](https://x.com/thedankoe/status/2049521752062185828) - @lidangzzz: 18 条;重点:diffusion model的商业革命奇点到了, 几百万条AI创作视频里,终于出现了第一条有机会冲击电影动画奖项的现象级作品, 虽然只有2分钟,但足够震撼,足够催泪,足够成功,全网播放量应该在2亿左右。 跨过这个临界点,接下来就是整个行业要严肃对待diffusion model带来的社会影响了。 [链接](https://x.com/lidangzzz/status/2049376312364552224) - @HiTw93: 18 条;重点:这篇文章里,我认为最有价值的部分其实是这一段: 非技术同学想用好 AI Coding,不能只会描述需求,最好补一点技术和产品的基本逻辑。 技术上至少知道 React、Vue、Next.js 在解决什么,终端、Git、VS Code、Chrome DevTools 怎么用,函数、变量、状态、拆文件是什 [链接](https://x.com/HiTw93/status/2049436090885357819) - @Khazix0918: 2 条;重点:https://t.co/RVSloneLvz [链接](https://x.com/Khazix0918/status/2049345300939391245) - @nash_su: 2 条;重点:llm_wiki 一天增加了1k star,总 star 来到了4.7k 🌟庆祝 同时 v0.4.4 更新发布: 新功能 - Wiki 页面元信息可视化:YAML frontmatter 不再裸露成代码,type 用色块徽章、tags 用 chip、sources/related 用卡片展示,可点击 [链接](https://x.com/nash_su/status/2049402407323660612) ## 重点推文 - @thedankoe: Most people quit because they forget that you have to be bad at something before you can be good at it. It's so obvious. You suck. Of course you're not going to win in 2 weeks. But (赞 10442 / 转 1549 / 回 504 / 看 412022) [链接](https://x.com/thedankoe/status/2049521752062185828) - @gdb: a great codex tutorial: (赞 3136 / 转 280 / 回 65 / 看 432830) [链接](https://x.com/gdb/status/2049315404271469054) - @gdb: terminal has been my primary interface to my computer for almost two decades. now it’s the Codex app. (赞 2432 / 转 83 / 回 227 / 看 370743) [链接](https://x.com/gdb/status/2049477747882254775) - @AnthropicAI: New on the Science Blog: We gave Claude 99 problems analyzing real biological data and compared its performance against an expert panel. On 23 problems, the experts were stumped. O (赞 1933 / 转 179 / 回 144 / 看 200354) [链接](https://x.com/AnthropicAI/status/2049624600741560340) - @AnthropicAI: In new Anthropic Fellows research, we discuss “introspection adapters": a tool that allows language models to self-report behaviors they've learned during training—including potent (赞 1238 / 转 119 / 回 125 / 看 169545) [链接](https://x.com/AnthropicAI/status/2049576143653929153) - @signulll: msft, goog, meta, & amazon are on track to spend ~$700b on ai infrastructure in 2026. this kinda spending usually happens via govts or wars whereas this time, it’s four companies r (赞 1335 / 转 76 / 回 82 / 看 80042) [链接](https://x.com/signulll/status/2049614400860921877) - @dotey: 转译:深度拆解 Hermes Agent 的记忆系统:它如何修正 OpenClaw 的误区 如果你读过我之前关于 ChatGPT、Claude 以及 Clawdbot 记忆系统的文章,你就会知道我一直在钻研同一个问题:这些 AI 智能体(AI Agent)到底是怎么记事的? Hermes Agent 对我来说格外有趣,因为这次我不需要只靠观察它的行为来搞“逆 (赞 938 / 转 209 / 回 26 / 看 108322) [链接](https://x.com/dotey/status/2049534755729707205) - @lidangzzz: diffusion model的商业革命奇点到了, 几百万条AI创作视频里,终于出现了第一条有机会冲击电影动画奖项的现象级作品, 虽然只有2分钟,但足够震撼,足够催泪,足够成功,全网播放量应该在2亿左右。 跨过这个临界点,接下来就是整个行业要严肃对待diffusion model带来的社会影响了。 https://t.co/IgyJ3Ba8Ci (赞 1098 / 转 100 / 回 47 / 看 159762) [链接](https://x.com/lidangzzz/status/2049376312364552224) - @gdb: build your own agents with codex app-server (赞 936 / 转 68 / 回 30 / 看 85738) [链接](https://x.com/gdb/status/2049609076351381580) - @HiTw93: 这篇文章里,我认为最有价值的部分其实是这一段: 非技术同学想用好 AI Coding,不能只会描述需求,最好补一点技术和产品的基本逻辑。 技术上至少知道 React、Vue、Next.js 在解决什么,终端、Git、VS Code、Chrome DevTools 怎么用,函数、变量、状态、拆文件是什么意思。 产品上推荐三本入门书:《启示录》看产品判断,《Li (赞 730 / 转 159 / 回 26 / 看 99950) [链接](https://x.com/HiTw93/status/2049436090885357819) - @signulll: there are 37 posters on my office wall that i have meticulously written & printed. this is the one i read almost every day: “great ideas are the output of the work, not the inp (赞 893 / 转 51 / 回 38 / 看 25279) [链接](https://x.com/signulll/status/2049548233832042781) - @dotey: OpenAI 上周发布 GPT-5.5 后,紧接着放出了一份官方提示词指南。这份指南传递的核心信息只有一个:别再写长提示词了。 GPT-5.5 距离上一代 GPT-5.4 只隔了六周,API 定价每百万输入 Token 5 美元、输出 30 美元,上下文窗口 100 万 Token,目前已面向 Plus、Pro、Business 和 Enterprise 用 (赞 594 / 转 121 / 回 31 / 看 66627) [链接](https://x.com/dotey/status/2049624930887614648) ## 标签分布 - china: 85 条 - ai: 81 条 - product: 65 条 - agent: 63 条 - ai-coding: 57 条 - startup: 38 条 - tools: 34 条 - yc: 24 条 - engineering: 23 条 - openai: 4 条 - gpt: 4 条 - model: 3 条 ## 一句话判断 - 2026-04-29 的信息流重点集中在 中文 AI 圈、AI 编码 / Agent,优先看上面的高互动链接即可。 ## 文件 - 原始抓取:`/Users/bytedance/Downloads/twitter_output/每日推文总结_raw_2026-04-29.json` - 信息源:`/Users/bytedance/Documents/info-bu/信息源/信息源.md`