Pinned Tweet
MCP is USB for AI agents. Thousands of plugs, built by thousands of strangers, changing whenever they feel like it.
I wrote up why I built MCP Checkup and what it checks today.
Would you let your agent plug into a server nobody has looked at?
Predicting the next word in a sentence got us this far.
A lot of status updates were already that kind of guess.
What's the last thing you wrote that wasn't?
It is amazing that human language turned out to be so unreasonably effective at modeling the universe that training a system to predict what comes next in a sentence would get us this far.
3
Signal 宣布在 iOS 27 上默认禁止截屏,防的是手机自带的 AI 助手。
加密协议被安全专家审了十几年还站得住,Siri 扫一眼屏幕就能读到。
想截图得自己去 Signal 设置里关,它也承认文件里的数据拦不住。
聊天软件得防着自己的手机,这手机到底在给谁打工?
signal.org/blog/protecting-y…
29
微软给6万名员工发买断方案劝他们主动走人,纳德拉说是为了让大家“觉得自己有选择权”。
他举的榜样是 Cowork 团队,只有11个开发者,“造一个造产品的工厂,自己不造产品”。
他说以后每份工作都是给 agent 当经理。11个人就能管一座工厂,你凭什么觉得自己在那11个里?
techcrunch.com/2026/11/18/mi…
24
KPMG UK pointed 100+ AI agents at a FTSE 250 company's books, and its head of audit says they sometimes asked better questions than the human team.
Will the audit fee go down? KPMG hasn't said.
So what exactly is the client paying the humans for now?
europeanbusinessmagazine.com…
33
哈佛一篇论文追了718家公司:用上AI编程agent后,代码多写30%,交付的活没见多。
代码都堵在评审,PR从提交到合并多等49%,被打回改的比例快翻倍。
有个项目只有一位资深能批代码,一周看一次。AI代码闹出几次事故后,又规定必须资深批。
老板花钱买AI,到底买来了谁的效率?
complexdiscovery.com/harvard…
3
4
120
OpenAI says GPT-6.2 Astra beat Claude Fable 5.2 on Agents' Last Exam at 5% of the cost.
Its other flex: slides and docs that "read less like AI."
Do you still have to tell your boss you used it?
openai.com/index/introducing…
43
Alibaba launched Qwen4-Max with 2.4 trillion parameters, and the launch post brags it ranked 6th on Code Arena.
Same post calls that "top rankings." Five coding models just got a free compliment.
Name the five you'd actually trust with your repo.
alibabacloud.com/blog/alibab…
33
A guy plugged a $3 Arduino into his 1994 Jeep and had Claude Code and Codex hack its computer until it passed California smog.
Claude offered to run the fuel pump. He said sure. Ten seconds in, he panicked.
Mechanics, is this your problem yet?
youtube.com/watch?v=WPj3aQ-x…
9
578 STOC 2027 papers, over a third of submissions, had Google's AI review their proofs first.
Over 30% of 70 authors surveyed said it caught substantive proof bugs or significant typos.
If the AI finds the bug before the referee, what's the referee for?
research.google/blog/advanci…
4
密歇根大学测阿里 Qwen3-8B:同一套阅读题,中文回答平均 403 个 token,英文 186 个,耗电是英文的 2.1 倍。
单个 token 耗电中英文一样,多出来的全是它用中文多说的话,正确率还低了 5 个点。
按 token 付费的中文用户,这笔“啰嗦税”要交到哪天?
arxiv.org/html/2606.21869
5
Researchers checked 24,664 Lovable apps with no login and pulled records with emails from 7,403.
One app exposed 1,000+ passport records. The ads say no coding needed. The terms say the data is on you.
Name one non-coder who read the terms before shipping.
crissyfield.de/research/lova…
8











