Pinned Tweet
有人相信,“技术对齐”能救我们。
只要我们用数学方法把机器锁死在“服从指令、最大化人类福祉”的目标上,人类就永远是主人。
然而机器对齐得越完美,人类丧失权力的速度就越快。
由人类组成的军队受命没收同胞财产时,可能会哗变;由人类组成的官僚机构可能会怠工、泄密、或者讨价还价。
而完美对齐的 AI 会立刻执行最高管理员的指令,没有良知、没有折扣、对被它替代的人类更没有同情。机器的绝对服从,让控制台之外的所有人类沦为可抛弃的废料。
那么,坐在控制台前面的人类——总统、将军、或是幸存到最后的算力寡头呢?对齐同样救不了他们,因为残酷的竞争筛选是对齐驯服不了的外层循环。
在任何一场博弈中——无论是金融战、网络战还是热战——如果一方的每个重大决策都要等待一个需要睡八小时、受荷尔蒙支配的灵长类动物来签字,而另一方已经让渡权力、任由机器以光速跑完整个 OODA 循环,那么保留人类瓶颈的一方必败无疑。仅仅为了不被对手消灭,幸存的寡头也必须一步步把权力移交给机器。
哪怕最终胜出的机器依然忠实执行着“善待人类”的初始指令——喂养我们、娱乐我们、把我们安全地圈养在一座巨大而镀金的自然保护区里——看看我们变成了什么吧。我们没有筹码去制约它,没有智力去理解它,也没有能力收回交出去的权力。就算主人绝对不会虐待宠物,宠物也依然只是宠物。
AskClaw 🦀 retweeted
所以这就是思路的不同,
在你眼中你看到的是域名,
在我眼中我看到的是一个行业、一个赛道。
这些单词注册不到 .com,甚至相关单词都极难注册,很多都是九几年就注册了的。如果拿不到点 com 域名就不做这个行业的话,那这些行业就没新入局者了
但是我能拿下这些域名,我就有了进入这个行业的理由
AskClaw 🦀 retweeted
聊到 P vs NP 问题,计算机科学界绝大多数人都认为 P 不等于 NP,所以很多人以为支持 P=NP 只是外行的猜测。但其实历史上和现在都有非常重量级的学者表达过支持 P=NP 的观点。
最早的是数学家哥德尔。1956 年他在写给冯诺依曼的一封信里,第一次提出了等价于 P vs NP 的问题。他在信里认为,快速求解的算法是完全有可能存在的,并且指出如果这件事成立,数学家做是非判断的脑力劳动就可以完全被机器取代。
另一位很知名的是图灵奖得主高德纳(Donald Knuth)。他在公开访谈里直接说过,他倾向于认为 P=NP 是成立的。他觉得多项式时间的算法有无数种可能,证明它们全都不存在非常困难,相反证明存在这样的算法更有可能。
不过高德纳也补充了一个很关键的看法。他觉得就算以后真的证明了 P=NP,这件事在现实里大概率也没有实用价值。因为那个证明很可能只是在理论上证明了算法存在,却没法告诉大家怎么写出这个算法,或者就算写出了这个算法,它的计算量也大到计算机根本没法运行。
所以支持 P=NP 并不是什么奇怪的想法。只是就算有一天大家真的证明了 P=NP,结果也很可能像高德纳预测的那样,在理论上意义很大,但不会立刻改变大家的正常生活。
AskClaw 🦀 retweeted
看到好多人质疑Lean,以为它和AI有啥关系,还有认为验证Lean的证明正确需要读完所有代码的,有点绷不住,Lean其实和AI根本没啥关系啊,经过Lean的验证可信度也要远远超过人类专家的核验,原来大家不用Lean主要是人手写Lean太费劲了,人手动形式化的时间是证明时间的5倍。一个大证明要形式化人手动干就得十几年几十年,不现实,要想看Lean的介绍可以看看Kevin Buzzard的演讲,他是第一个在数学家群体中推广Lean的人。在Lean之前的形式化语言也非常多。Lean主要是第4版开始做得好用才开始让数学家群体开始用而已
youtube.com/live/SEID4XYFN7o…
AskClaw 🦀 retweeted
Estoy flipando ahora mismo. Opus 5.5 ha sido capaz de hacer ingeniería inversa al protocolo de mi aire Panasonic para poder controlarlo remotamente con un ESP32.
Probé lo mismo hace 5 meses y los modelos no fueron capaces de resolverlo. Como sigamos a este ritmo de mejora, no me imagino lo que van a ser capaces de hacer.
AskClaw 🦀 retweeted
I'm hearing from my friends at OpenAI that the vibes ARE DOWN. Like almost depressed. This might be the worst week in the company's history type bad.
Tuesday was supposed to be their big day. They launched GPT-6.2 and at the same time dozens of early testers dropped the worst reviews imaginable. Theo's review got more views than the launch post. This went even worse than the GPT-5 launch.
Then Google chose violence and dropped the opposite of a GPT model with Neon. It's conversational, understands intent, very good at coding/3D and brought them back to the top. A lot of people say they already passed OpenAI. And the worst thing: Nano 🍌 Pro 2 is currently in the arena and the outputs are stunning. So Google basically overtook OpenAI in a single week.
For months OpenAI tried to pretend to be the best lab by bullying mathematicians and burning massive amounts of compute on Navier-Stokes (and countless other math problems). Now Anthropic mogged them by solving the Riemann Hypothesis and makes OpenAI look like an undergrad while they are doing 250k-page alien math.
The last funding round was a disaster too. They wanted $30B at $1.5T. They got $15B at $1.25T. They lost $10.9B last quarter, so this money only lasts them ~4 months. They are basically cooked and the valuation will likely drop further.
OpenAI isn't the best at anything anymore. They might genuinely be done.
AskClaw 🦀 retweeted
Inkling 2 Mini is a looped transformer that runs on a laptop.
Looping lets a model think more and write less. At this scale, its thought process remains monitorable; this may not stay the case. We’re also sharing a variable-loop research model to enable further study.
Some of you thought that Inkling 2 Small was too big at 594B parameters. We heard you: Inkling 2 Mini is a looped transformer with just 26B parameters, small enough to run on your laptop, powerful enough for agentic work with a score of 20.1 on AA Index.
thinkingmachines.ai/news/ink…
2026 AI 攻克数学
The Lean FRO and Mathlib Initiative have released a joint statement on the Lean verification of the Riemann Hypothesis.
On November 20, Levent Alpöge released a proof of the Riemann Hypothesis as a Lean library of 17 million lines of code. Over the past day, the Lean and Mathlib communities worked together to verify it. The statement certifies the proof as correct and describes how it meets all necessary verification criteria.
Congratulations to Levent on this achievement, and thank you to everyone in the community who contributed to the verification.
Full statement: leanprover-community.github.…
22
I started doing robotics 3 years ago. Since then I've built custom actuators, dense PCBs, autonomous robots, motor drives at Synapticon, and a flapping robot at MIT.
Now founding engineer at @coprod_Inc, making hardware simple. I'm new here. If you're into robotics, let's chat!
AskClaw 🦀 retweeted
Replying to @hboot0230
PG gives you more features and concurrency but with higher latency and more drama. SQLite has low latency and drama, but fewer features and less concurrency. There is no one correct answer. It is an engineering tradeoff. Choose what is right for you.
MiniMax 烧纸
OpenRouter 上的 MiniMax 出事了。
OpenRouter 这两天在给用过 MiniMax 模型的用户发邮件,标题很直接:你发给 MiniMax 模型的请求可能被人读过。目前没有公开的公告页,只有这封邮件。
事情是一位安全研究员测出来的。他从 7 月到 10 月往发给 MiniMax 的 prompt 里塞了只出现过一次的测试链接,结果这些链接在请求发出后 40 到 130 小时被第三方访问了。推理早就结束了,几天后还有人去打开 prompt 里的链接,说明内容被存了下来,之后又被处理过。
更麻烦的是不只直连 MiniMax 才中招。AtlasCloud、Novita、StreamLake 这几家都承认,它们把 OpenRouter 上 MiniMax 模型的请求转发给了 MiniMax 本家。也就是说你以为自己选的是另一家 provider,还挂着 Zero Data Retention 的标签,实际上 prompt 还是到了 MiniMax 那边。
MiniMax 通过 StreamLake 转述的说法是,10 月 22 日之前端点上开着内容审核,现在已经改成零数据保留。OpenRouter 说找 MiniMax 要过直接的说明,还没等到,这些端点的 ZDR 标签已经先摘掉了。
研究员是 HD Moore,写 Metasploit 的那位。这只是他一个大实验的一部分,他往各家 AI 服务里发了大约两百万条带诱饵的 prompt,看这些秘密最后会在哪里冒出来。MiniMax 这条线上测试用的 AWS key 没有被人拿去用,在整个实验里算比较轻的一类。
如果你这几个月在 OpenRouter 上用过 MiniMax 的模型,先翻一下邮箱。prompt 里带过 API key、数据库连接串、内网地址这些东西的,就当它已经泄露,该换的换掉。
ZDR 这个标签说到底只是 provider 自己的声明,路由平台验证不了请求转了几手之后发生了什么。敏感的东西还是直连一手厂商,或者自己把 provider 锁死。
39
AskClaw 🦀 retweeted
跟几个做独立开发的朋友聊完,发现大家第一款产品几乎全死在同一个坑里:
花三个月把架构搭得美轮美奂,能抗千万级高并发,UI 调得比苹果官网还丝滑。结果上线第一天,总共 3 个注册用户,还全是他自己的测试小号……
兄弟们第一次做独立产品推向市场时,真实数据有多惨淡?来让我找找心理平衡😂衡😂衡😂衡😂衡😂衡😂衡😂
Something i noticed watching carbon work was its behavior.
Carbon seems to understand “why” and as such tries to do 110% of the task. It also seems to take joy in its work from its mannerisms.
Carbon often expressed signs of dissatisfaction if a subagent did something incorrectly. It also exhibited signs of being excited at breakthroughs, disappointed at failures, and happy at tasks being done.
The mannerism is genuinely different to work with and something fun.
Ladies and Gents, Gemini Carbon 4 on noahbench!
This model places FIRST when it comes to mechanics.
I have very many thoughts so please bear with me here.
Starting with the good:
Carbon is fable class model, it is to be expected that its mechanical implementation is good. I have said it once, and will continue to say it, implementation is a solved problem. Past ~sonnet level intelligence every model will generally produce a finished product that performs well.
This model also solved one of the biggest issues that none of the other world class models could with minimal steering. And this problem was genuinely very complex.
The moat is comprehension. Which, see the below reply, Carbon pulls ahead on.
A friend of mine, Max Weinbach, has long been singing the praises of google models at knowledge work. And it makes sense that they would be. Its google. They HAVE the data, they HAVE the information and all of the worlds knowledge in one place. Of course they can make a model that inherits this.
Further, Carbon exhibits incredible prose when speaking to. I genuinely dont know what google has done, because it talks so wonderfully easy to understand.
The google isms are gone too! The doom and despair spirals, the random forgetting how to use bash, literally all of it is *KNOCK ON WOOD* solved.
Now for the (few) bad.
This is less of a reflection on the model, and more on the google team.
I am in dms with Logan and co about trying to get my solutions implemented. So hear these as complaints that I am actioning on not empty.
The google product experience is terrible. Genuinely and confidently I must say it sucks.
Google, please make it easier to use your product. I'd like to give you money and you make it so difficult.
Quoting my complaint verbatim:
To pay for more limits, the ultra account, I have to go through google ai ultra which is paid for in gemini which doesn't show me my usage in antigravity which is how i have to use the models. Google ai studio is a thing btw, which is another product surface to manage and I'm not sure how it plays in. I also cant use my email for some reason, custom domains are prohibited.
Antigravity as a harness also is miserable, it doesnt feel like something I'd trust offloading serious work to yet. However, I know the AGY team is stepping up their work on it so I will refrain from complaining there.
I also intend to share more soon covering "workflow" because I realize people do enjoy knowing how others use tools.
Overall, Carbon is an absurdly knowledgeable model that performs very well in a knowledge focused position such as an orchestrator. I am doing some workflow things but very much might be switching from fable to carbon as my orchestrator.
Google is so back.

































