# 推文总结 2026-05-05 ## 总览 - 账号范围:18 个 - 活跃账号:14 个 - 总推文数:147 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: 3 条;重点:See you all soon! We've got some fun announcements ahead. I'll also be doing a workshop on "how we Claude Code" with some workflows I'm excited to sha [链接](https://x.com/trq212/status/2051808772423819507) - @gdb: 2 条;重点:Major ChatGPT upgrade rolling out now, in the form of GPT-5.5 Instant: [链接](https://x.com/gdb/status/2051759920467312940) - @dotey: 10 条;重点:https://t.co/cFZS7VsNIv [链接](https://x.com/dotey/status/2051688341519609897) - @oran_ge: 18 条;重点:俞浩一针见血地指出了美团的问题 其实问题这个词是不严谨的,是 pattern ,中性词,不好不坏 中国的企业家做梦的时候,一般会得到其他人的嗤之以鼻的反应 安分守己的思维,做好自己的事情 这是在资源稀缺的时代生存本能带来的思维惯性 在资源丰饶的时代反而成了束缚 https://t.co/Md5g3a [链接](https://x.com/oran_ge/status/2051584248268886476) - @AnthropicAI: 7 条;重点:As AI takes on work humans can't fully check, a capable model could deliberately hold back—and we'd never know. New Anthropic Fellows research finds t [链接](https://x.com/AnthropicAI/status/2051718308702081047) - @dongxi_nlp: 1 条;重点:ProgramBench 在这个 benchmark 中,一众 frontier LLM 悉数零分! 预感到下一波前沿模型得分 1% 会越来越多,让 LLM 的故事和潜力继续。 [链接](https://x.com/dongxi_nlp/status/2051785581751890219) - @jiangydev: 当日未抓到新推文 - @lifesinger: 1 条;重点:五一期间在红果看了一部修真漫剧。发现真好看啊。同时细思很有道理: 1、这部漫剧每集 2-3 分钟,画面很一般,第四部甚至画风都变了。但不影响剧情和想往下看完的急切。剧情其实也一般,闭着眼猜接下来剧情,也能猜中七八成。然而人脑就想着看下一集,想着看大结局。于是很自然一集一集往下看。 2、偶尔会有 7s [链接](https://x.com/lifesinger/status/2051661857459761270) - @gregisenberg: 10 条;重点:Coinbase is now testing 1 person teams + AI agents and announced laying off 700 employees. Other companies doing this (layoffs + AI): 1. Shopify: No n [链接](https://x.com/gregisenberg/status/2051683558486392890) - @garrytan: 9 条;重点:What makes gbrain genuinely different? It's not a memory layer OR a code tool OR a search engine. It's all three unified under one graph with one quer [链接](https://x.com/garrytan/status/2051525161380364315) - @signulll: 20 条;重点:great marketing should feel more like a drum beat than a complex melody. like deep house, it should be one beat, repeated, until the message stops sou [链接](https://x.com/signulll/status/2051481457986842635) - @thedankoe: 当日未抓到新推文 - @lidangzzz: 19 条;重点:如果让我设计一款中国大学: 1. 只有一个计算机专业; 2. 微积分、线性、概率、复变、大学物理(上下)、基础编程语言、基础数据结构,全部下放到高中,AP修完入读; 3. 大一开学一顶门,直接上算法、操作系统、体系架构、编译器四门课,大二大三大四全部专业课,删掉所有政治课; 4. machine l [链接](https://x.com/lidangzzz/status/2051531322502283671) - @HiTw93: 31 条;重点:Today was the most productive coding day I’ve had in 10 years. I spent 1.6k out of the 10k credits Cursor gave me, mainly using Opus 4.7 1M Max and GP [链接](https://x.com/HiTw93/status/2051597128980746407) - @Khazix0918: 6 条;重点:上一条不是我发的!!!已经删掉了 被盗号了,莫名其妙发了什么诡异的币圈的东西。 五一在家没怎么刷X,朋友问我怎么玩币了我才看到感觉天都塌了,立刻开了X的二次验证。 再说一遍,那个不是我发的!!!不是我!不是我!是被盗号了!!! [链接](https://x.com/Khazix0918/status/2051578264263737436) - @nash_su: 10 条;重点:最近在研究个人和企业 AI 知识库,这个系列的文章可以看看,我汇总成了一个中文版: https://t.co/5VleOxEqcp [链接](https://x.com/nash_su/status/2051548211844894761) ## 重点推文 - @gregisenberg: Coinbase is now testing 1 person teams + AI agents and announced laying off 700 employees. Other companies doing this (layoffs + AI): 1. Shopify: No new headcount unless you prove (赞 1381 / 转 109 / 回 152 / 看 250874) [链接](https://x.com/gregisenberg/status/2051683558486392890) - @AnthropicAI: As AI takes on work humans can't fully check, a capable model could deliberately hold back—and we'd never know. New Anthropic Fellows research finds that such a model can be traine (赞 1148 / 转 104 / 回 110 / 看 148108) [链接](https://x.com/AnthropicAI/status/2051718308702081047) - @AnthropicAI: New Anthropic Fellows research: Model Spec Midtraining (MSM). Standard alignment methods train AIs on examples of desired behavior. But this can fail to generalize to new situation (赞 907 / 转 84 / 回 67 / 看 86604) [链接](https://x.com/AnthropicAI/status/2051758528562364902) - @lidangzzz: 如果让我设计一款中国大学: 1. 只有一个计算机专业; 2. 微积分、线性、概率、复变、大学物理(上下)、基础编程语言、基础数据结构,全部下放到高中,AP修完入读; 3. 大一开学一顶门,直接上算法、操作系统、体系架构、编译器四门课,大二大三大四全部专业课,删掉所有政治课; 4. machine learning、deep learning、LLM、AI A (赞 788 / 转 119 / 回 127 / 看 124480) [链接](https://x.com/lidangzzz/status/2051531322502283671) - @gregisenberg: cringe is just the entrance fee to doing anything interesting cringe mountain looks different for everyone. for some it's posting. for some it's cold DMing. for some it's charging (赞 744 / 转 52 / 回 92 / 看 42421) [链接](https://x.com/gregisenberg/status/2051637194314436964) - @lidangzzz: 1. 本科没学习,大学四年一门课都没修明白 2. 所有业余时间全部在打游戏刷抖音 3. 没有任何实习,没做过任何开源项目 4. 简历白纸一张,除了学校专业以外没有任何专业技术 5. 整个大学四年写代码不超过2000行 6. 偏远省份或者艺术升体育生入学 7. 数学太差,考本校专硕都考不上 基本就是这一类的画像 (赞 708 / 转 35 / 回 137 / 看 203343) [链接](https://x.com/lidangzzz/status/2051549070959714631) - @gdb: Major ChatGPT upgrade rolling out now, in the form of GPT-5.5 Instant: (赞 726 / 转 32 / 回 62 / 看 44001) [链接](https://x.com/gdb/status/2051759920467312940) - @garrytan: What makes gbrain genuinely different? It's not a memory layer OR a code tool OR a search engine. It's all three unified under one graph with one query interface. Nobody else has t (赞 637 / 转 20 / 回 115 / 看 56961) [链接](https://x.com/garrytan/status/2051525161380364315) - @signulll: great marketing should feel more like a drum beat than a complex melody. like deep house, it should be one beat, repeated, until the message stops sounding like marketing & sta (赞 526 / 转 22 / 回 33 / 看 20767) [链接](https://x.com/signulll/status/2051481457986842635) - @lidangzzz: 全世界大学制度估计近50年要崩溃毁灭一次,现有的漏洞实在是太多了。 1. 绝大多数教授热衷于做research,把>50%的时间用在发垃圾paper上,并没有办法给本科生认认真真地上课,而给本科生上课是最基本要求; 2. 一个方向火了,学校开始大量招faculty,开设新专业,新的program,人越来越多, 而一个方向死了,学校根本不去砍垃圾program (赞 430 / 转 59 / 回 62 / 看 75985) [链接](https://x.com/lidangzzz/status/2051503615777402906) - @trq212: See you all soon! We've got some fun announcements ahead. I'll also be doing a workshop on "how we Claude Code" with some workflows I'm excited to share. Don't worry if you're not (赞 453 / 转 20 / 回 45 / 看 31093) [链接](https://x.com/trq212/status/2051808772423819507) - @signulll: today i met eight ppl from openai incl the ceo. what a fun day. (赞 472 / 转 1 / 回 41 / 看 28721) [链接](https://x.com/signulll/status/2051788928232182023) ## 标签分布 - china: 96 条 - ai: 83 条 - agent: 69 条 - ai-coding: 61 条 - tools: 61 条 - product: 42 条 - engineering: 32 条 - startup: 19 条 - yc: 9 条 - model: 7 条 - anthropic: 7 条 - claude: 7 条 ## 一句话判断 - 2026-05-05 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_2026-05-05.json` - 信息源:`/Users/bytedance/Documents/info-bu/信息源/信息源.md`