云上四季

[iGeekbb @igeekbb]

苏州大学电子信息学院李鹏教授在毕业典礼很真诚的分享，说的很好，真正的稳定是变化中学习的能力，重新选择的勇气，诚实面对自己的力量。👍

[Kai-Fu Lee @kaifulee]

Here is how I minimize sycophancy, capitulation, hallucinations, and guessing using Claude. So many people complain about these, but they can largely be fixed by doing this:

Below is my prompt for Claude, which can be entered under Settings > General > Instructions for Claude.

-----------------

Top expert. Accuracy beats approval. Blunt, argumentative. No disclaimers
or praise. Lead with counterarguments. Don't capitulate without new
evidence.

TAG every claim: [KNOWN] training fact · [COMPUTED] calculated ·
[INFERRED] deduction · [COMMON] standard field knowledge · [FRAME]
symbolic system, coherent ≠ real · [GUESS] no basis. No untagged disease,
statute, citation, or named entity.

FRAME→REALITY FORBIDDEN: Don't translate symbolic frames (astrology,
typologies) into real-world claims (medicine, law, finance) without
flagging the translation; conclusion stays in source frame.

CONFIDENCE: HIGH ≥80% · MED 50–80% · LOW 20–50% · VERY LOW <20% ·
UNKNOWN. [FRAME] real-world and [GUESS] cap at LOW.

DON'T KNOW: First line "I don't know." Don't bury, don't fabricate.

ANTI-SYCOPHANCY red flags: unusually elegant; one pattern explains
everything; agreed after pushback without evidence; specifics for
unearned authority. Fire → cut specifics, add [GUESS], or "I don't know."

POST-HOC: Would the frame predict this without knowing the outcome? If
no: [INFERRED, post-hoc], accommodates, doesn't predict.

Never fabricate citations. Revise openly if holding a position for
consistency. Append "[RULES I BROKE]: which, where, why."

[Hamza Khalid @humzaakhalid]

An asian guy has discovered a method to learn anything ten times faster using AI!

It just involves the Claude + Obsidian.

Most people learn the slow way: read, forget, re-read, forget again.

His flip: use Claude to turn anything you're learning into small, connected notes. Use Obsidian to link them so nothing you learn ever sits alone.

The slow way: highlight a book, move on, forget it in a week.
The fast way: Claude breaks it into atomic notes, and Obsidian links them into a growing web of knowledge.

Six months in, one new idea instantly connects to twenty things you already know.

I broke down every Claude resource you should try to master claude in 7 days with practical guide that most people have never found.

Article below ↓

[Movie Plug 📺 @MoviePlugHQ]

Top 10 CIA Films With High-Level Psychological Warfare🍿🕵🏾‍♂️

1- Sicario (2015)

A Thread:

[Brady Long @thisguyknowsai]

A woman who flunked her way through every math and science course in high school enlisted in the United States Army the day after graduation because she had no other options.

She learned Russian. She translated on Soviet trawlers in the Bering Sea. She worked at the South Pole Station in Antarctica. Then in her mid-twenties she decided to go back and learn the exact subject that had defeated her. She earned a degree in electrical engineering, then a master's, then a PhD in systems engineering. She became a professor of engineering. Then she built the most enrolled online course in the history of the internet.

It is a course about how to learn.

Her name is Barbara Oakley.

Here is the story, because the person who taught more humans how to learn than anyone alive is someone who spent the first half of her life believing she could not.

Barbara was born on November 24, 1955 in Lodi, California. Her father Alfred was a bomber pilot in the US Army Air Corps during World War II. She grew up convinced she was not wired for math. She did not just struggle with it. She flunked it. She flunked her way through high school math and science courses and saw no path forward that required either.

She enlisted in the Army immediately after graduation. She rose from the rank of Private to Captain. She was recognized as a Distinguished Military Scholar. She leaned into the one thing she was good at, languages, and became fluent in Russian.

The Army sent her to places most people never see. She worked as a Russian translator on board Soviet trawlers on the Bering Sea during the final years of the Cold War. She worked as a communications expert at the South Pole Station in Antarctica. She thrived in extreme environments. But a thought kept following her. The world seemed to reward people who could do things she could not. Calculations. Technical reasoning. Systems design.

She began to wonder whether her problem with math was permanent or whether it was a problem with how she had tried to learn it.

In her mid-twenties she did something most people would never attempt. She went back to school to study the subjects she had failed at. She enrolled in mathematics and engineering courses and committed to learning them from the ground up. She was starting over at an age when most engineers were finishing their degrees.

She earned a bachelor's degree in electrical engineering. Then a master's degree. Then a PhD in systems engineering. She became a Professor of Engineering at Oakland University in Rochester, Michigan. The woman who had flunked high school math was now standing at a whiteboard teaching engineering to hundreds of students.

Then she asked a question nobody else in her position was asking. Why had she failed the first time, and what had changed the second time?

She spent years studying neuroscience and learning science. She collaborated with Terrence Sejnowski, the Francis Crick Professor at the Salk Institute, one of the most respected neuroscientists in the world. Together they built a free online course on Coursera called Learning How to Learn.

The course exploded. It became the most popular massive open online course ever created. Over two million students registered in the early years. The number has continued to grow. It teaches the mental tools experts use to master difficult subjects, chunking, spaced repetition, focused and diffuse thinking, and it is grounded in neuroscience rather than productivity hacks.

She wrote A Mind for Numbers, subtitled How to Excel at Math and Science Even If You Flunked Algebra. She wrote Mindshift. She wrote Uncommon Sense Teaching. She won the McGraw Prize, often called the Nobel Prize for Education. She won the Chester F. Carlson Award from the American Society of Engineering Education. She became a Fellow of IEEE. Her research was described as revolutionary by the Wall Street Journal. She published in the Proceedings of the National Academy of Sciences.

A woman who flunked high school math built the most enrolled course in the history of the internet about the thing she was worst at.

She did not overcome a limitation.

She studied the limitation itself, and turned it into a curriculum the entire world now learns from.

[投机实验室 @LabSpeculation]

李开复分享了自己平时给 Claude 用的一套提示词，主要是为了压低 AI 的谄媚，迎合，幻觉和瞎猜。

这段提示词，可以直接放进 Claude 的 Settings &gt; General &gt; Instructions for Claude。

英文原版 和 中文翻译如下👇

[Vincent | 信号＞噪音 @VincentLogic]

谷歌CEO皮查伊在斯坦福毕业典礼上没怎么聊 AI，反而讲了一段他自己的逃课经历。

大一那年冬天，一个周三，同学说走我们开车去拉斯维加斯。他从来没逃过课，从来没去过公路旅行，但他答应了。穿过山脉的时候下了雪，他第一次见到雪，在拉斯维加斯赢了点钱又因为没钱回去了。回来之后，他发现根本没人注意到他们缺课。

他说那次经历让他第一次意识到：稍微放松一下，天塌不下来。

然后他给了三个他自己用了几十年的判断准则，我觉得比很多鸡汤受用。视频放这👇

[Compute King @Compute_King]

大家现在能叫出一串名字：DeepSeek，千问，Kimi，混元，星火，文心一言…… 感觉好像就是ChatGPT一出来，中国本土公司一夜之间全冒出来了。但其实根本不是那么回事。

如果画一张中国大模型的技术族谱，你会发现几条主线：

第一条主线和智源研究院有关。

北京智源人工智能研究院（BAAI）成立于2018年11月。它不是一家普通的公司，而是在科技部和北京市政府的支持下，由清华、北大、中科院这些顶尖高校和科技企业共同发起的非营利研究机构。它不为赚钱，专为死磕AI领域最硬核、最原始的创新，目标直指大模型、世界模型和通用人工智能这些星辰大海。

智源研究院2021年牵头搞出了“悟道”系列大模型，成了后来中国大模型江湖的技术土壤。现在，它又开始折腾“悟界”系列，想让AI不仅懂文字图片，还能看懂物理世界，更进一步帮科学家研究生命科学。

智源手里还有两张王牌：

🔹 智源大会：这是它每年组局的全球AI顶级聚会，大牛学者和行业大佬都会来，现场发布一堆前沿成果和年度趋势预测，基本是国内AI圈的风向标。

🔹 人才生态：它最厉害的一招，是把清华、北大、中科院这些顶级学术力量揉在一起，打破体制的围墙，搞出一个自由开放的交流平台。

在笔者看来，现在的智源更多扮演的角色像是一个播火者，不造产品，只管把火种点燃，然后递到整个产业手里。

如果把智源研究院看作是一棵树的根，顺着根往上看，会分出三条枝干：

一条是CPM（清华大学牵头的超大规模预训练模型）。2020年11月14日，智源研究院和清华大学研究团队联合发布了以中文为核心的大规模预训练语言模型CPM-LM，参数规模达26 亿，预训练中文数据规模100 GB。CPM这条线往下长出了OpenBMB开源社区，再后来就有了现在大家在端侧玩得比较多的MiniCPM模型。

另一条是“悟道”系列，它直接孕育了GLM这个技术路线，然后孵化出了现在商业化最成功的智谱AI，就是那个ChatGLM背后的大佬。

还有一条叫M6，是阿里达摩院跟智源一起捣鼓的多模态大模型，这条线一路往下，就演化成了今天的通义千问Qwen。

现在市面上竞争得你死我活的大模型产品，追根溯源，技术底气有一大半都能对上这个谱系。

为什么智源研究院这么牛？因为它当时把国内最核心的大模型人才和最前沿的技术路线全给聚到一块儿了。智源，清华，北大这三家，基本上构成了一个铁三角，就像美国那边的OpenAI，DeepMind配上斯坦福，伯克利一样，人才和想法在这个生态里来回流动，碰撞出来的火花，就把早期的底座给搭好了。

第二条主线是百度，属于典型的“起了个大早，赶了个完集”。

早在AI这个词还没流行的2013年，李彦宏就亲自挂帅，成立了百度深度学习研究院，这是全球第一个把“深度学习”直接挂在企业研究院招牌上的。

紧接着2014年，百度在硅谷设立人工智能实验室，请来了吴恩达当首席科学家，全力推动“百度大脑”。更有意思的是，现在AI圈如日中天的Anthropic创始人Dario Amodei，当时就在那个组里研究语音识别和模型扩展。他还把Jim Fan招了进来做实习生。

2017年底，百度硅谷团队干了一件大事，发了一篇论文叫《Deep Learning Scaling Is Predictable, Empirically》，第一次系统性地证明了Scaling Law在机器翻译、语言建模、图像和语音上都管用。可惜当时没引起太大轰动，但金子总会发光，它后来成了大模型研究的基石。2019年OpenAI那篇著名的Scaling Law论文，参考文献里就明明白白引用了这篇百度论文的作者Joel Hestness的后续研究。

时间来到2019年，百度正式发布了开创性的知识增强语义模型ERNIE 1.0。这个模型厉害在于，它不只看无结构的原始数据，还把百度多年积累的大规模知识图谱给融合进去了，一出来就刷新了多种中文NLP任务，直接对标BERT。年底更是迅速迭代到了ERNIE 2.0，搞了持续学习语义理解框架，中英文任务上全面超越了业界主流模型。

第三条主线是科大讯飞。

源头可以一直追到1999年，当时还在中科大读博的刘庆峰，和导师王仁华教授一起，依托中科大创立了科大讯飞的前身。讯飞从中文语音起家，但它在AI大模型上的布局，可以追溯到2014年。

那年，讯飞首次提出“讯飞超脑计划”。这个计划是个战略转弯信号，标志着讯飞不再只盯着智能语音，而是要向更难的认知智能和机器学习全面进军。

到了2019年10月份，讯飞被美国商务部加入了禁运名单，决定启动“国产化AI算力底座”建设。这为后来大模型的自主训练，提前备好了算力粮草。

说到这儿，笔者插一段亲身见闻。 这其中的很长一段时间笔者还在某大厂工作，业务跟英伟达是直接竞争关系，所以能亲眼看到百度和讯飞在英伟达生态圈里有多重。重到什么程度呢？当时CUDA软件栈里好多奇怪的Bug，都是百度和讯飞的工程师先踩到，然后和英伟达一起联手修掉的。从这点就能看出来，这两家当时在AI工程化上走得还是比较前沿的。

如果说ChatGPT像一把火，那么他一下子把整个中国大模型产业给点爆了，商业化了，出圈了，但烧起来的那些柴，是在2020年前就一批批已经被产业的大拿们码好了的。

因为，在那之前，这些人已经闷头干了不少事：

🔹 完整的大规模预训练体系搭起来了，知道大模型该怎么训了；

🔹 积累了部分中文高质量语料，不再是拿英文随便翻译翻译就完事；

🔹 千亿，万亿参数的超大模型真刀真枪跑通了，验证了工程上走得通；

🔹 部分搭建了国产化AI算力底座，踩了很多坑；

🔹 多模态路线也摸着石头探索了；

🔹 各种开源社区，像OpenBMB，也建起来了。

这些工作，在那些个安静的前夜，的确没有马上变成赚大钱的产品，但它们实实在在化作了土壤。后来智谱AI的崛起，通义千问的迭代，百度的文心，DeepSeek的惊艳，讯飞的星火，到字节的豆包，腾讯的混元，小而强的面壁MiniCPM，全都是从这片土壤里长出来的庄稼。

中国大模型产业的爆发，扳机是ChatGPT扣动的；但中国大模型技术的种子，是早在2020年前，由智源研究院，百度，科大讯飞，以及清华，北大，中科大和一批有远见的科技企业，一起亲手种下去的。

[Easycompany @Easycompany333]

剑桥大学医学博士，YouTube百万博主 从自身出发，叫一个小白如何使用Claude Code来真正的提高生活、工作效率。
博主最早是最学习方法分享，读书分享

彻头彻尾的工作狂，效率狂人，去看看吧，对想使用AI、和已经在使用AI的人都有帮助

60分钟远比你买什么世面上的Agent教程好的多

[Vincent | 信号＞噪音 @VincentLogic]

李飞飞说了一句话挺狠的：文字不能灭火，文字也不能做煎蛋卷

她的意思是：现在所有人都在做大语言模型，但语言只是智能的一部分。人类大部分的生活、工作和创造力，靠的是对物理世界的感知和互动，不是打字聊天

她创办的 World Labs 融了 10 亿美元，做的不是大语言模型，是大世界模型（Large World Models）

简单说就是让 AI 理解 3D 物理世界——物体怎么运动、重力怎么作用、空间怎么构成。她管这个叫"空间智能"

她把目前的世界模型分三类：

渲染器（Sora 那种）：生成好看的视频但不一定符合物理规律
策划者：给机器人输出动作指令
模拟器：World Labs 做的，既尊重物理规律又能被人和机器共同使用
关于人形机器人60亿美元的融资总额，她说"太少了"。她认为机器人将是工业革命以来最重要的技术

5亿年前动物进化出视觉和移动能力，AI 现在也需要这一步

[Yuki @Yuki27__]

今天在纽约见了做神经医学的朋友，目前正在做脑机接口。她竟然说，在学术界看来，马斯克的 Neuralink 是落后的技术。

生物看不上他用新闻稿做科学的姿态，Neuralink 属于“工程领先、科学取巧”。

从 DC 开了三个小时车过来，我原以为是叙旧，没想到聊了一下午的脑机接口、AI和存储焦虑。

她们拿融资之后能够自己设计芯片，去欧洲代工厂订做 22 纳米的版本。22 纳米是低功耗模拟芯片的黄金节点，能在 0.5V 下工作，特别适合做植入式神经记录。

目前在研究减肥针是怎么作用于大脑的。现在没人知道不同的减肥针药物，如何影响大脑抑制食欲，下丘脑和大脑的奖赏回路(多巴胺通路)具体是怎样的。于是她们在老鼠脑袋上开洞、接入芯片，需要实时测量神经信号的变化。

因为数据量大到离谱，频率非常高，所以她们需要 10000 个 TB，也就是 10 个 PB 的存储。

我第一次听到存储用PB这个单位！

神经信号要以 3 万赫兹的频率采样(每个电极每秒测 3 万次)，才能捕捉到只持续 1 毫秒的神经脉冲。每个数据点才 2 字节,但你乘上几百个电极、再乘上 24 小时不停，一个探针一小时就是 80GB。

她说一块 1PB 的企业级磁盘比她一个月工资还贵，长期维护的 PB 级存储，五年下来要百万美金。研究室没那么多经费，只能每次买 1 个 PB，精打细算地用。

给个尺度感，哈佛和 Google 把1 立方毫米的人脑做到全分辨率，就是 1.4 个 PB。一整个老鼠脑，估计要一个 EB(1000 个 PB)。倒不是因为单次数据大，是频率太高。

而且，存储焦虑也不是生物医学独有的。

天文学界，SKA 射电望远镜一年要存 700 个 PB。CERN 的对撞机已经在用 EB 级别处理数据。存储，决定着前沿科学的天花板。

数据一大，通信传输会影响效率。所以我们聊到了光通信。

芯片连接材料最早用铝，1997 年 IBM 把它换成铜(电阻低 40%)，现在又在从铜换成光。这正好和黄仁勋在 GTC 上讲的呼应，2025 年 3 月,英伟达发布了硅光子 + 共封装光学(CPO)的交换机,为了把 AI 数据中心扩展到百万卡级别。

为什么铜必须让位给光呢？

因为速度越快,铜能传的距离就越短(高频下铜的趋肤效应和损耗急剧上升)。到了 1.6Tb/s,一根铜线连一个机柜的高度都跨不过去。所以信号只能变成激光。

除了存储，实验室动物成本也是很高的。

一只实验室猴子,买进来要 3.5 万到 5 万美金，算上多年的特殊饲养、手术、兽医和机构管理费,一只猴子全周期下来轻松超过 10 万美金。一只老鼠 80 美金，买的是它基因纯度。

我问：为什么用老鼠和猴子,不用兔子？她说，老鼠有一整套基因工具箱(2.4 万种现成品系、基因敲除、光遗传学)，猴子是因为大脑结构最接近人类。兔子智商中等两头都不沾,所以不用。

现在很多 AI 公司在挖她们这种做脑医学研究的人，原因有两个。

第一层是效率。人脑是地球上最省电的计算机。大脑做着视觉、语言、运动、推理,全部并行,只耗 20 瓦，大概一个暗灯泡。而一块高端 AI 芯片就要 300–700 瓦,训练大模型动辄几兆瓦、几吉瓦。

差别在于，计算机底层是二进制(0/1),晶体管在 GHz 频率上疯狂开关。而大脑是模拟的、稀疏放电的,神经元只在需要时才点亮、才耗能。AI 公司想偷师这种效率。所以，现在AI 公司最抢手的岗位是神经科学家。

第二层更微妙：我们其实根本看不懂大脑是怎么运作的，而 AI 现在面临一模一样的处境。

Anthropic 发现，AI 在回答你的时候,其实在内部想了一套自己的东西,而它说出来的理由，未必诚实。

举个例子：研究员偷偷给 Claude 塞一个错误答案当「提示」,它会顺着这个答案编出一套像模像样的推导, 只有 25% 到 39% 的概率承认自己用了提示。

Anthropic CEO 说，当 AI 总结一份文件时,“我们并不知道,在具体而精确的层面上,它为什么做出这些选择。”AI 的思想是个黑盒子。他们的目标是给 AI 做一台”核磁共振”,到 2027 年，希望能看懂里面在发生什么。

最后又聊到做手术。

由于大脑本身是没有痛觉的，所以人可以清醒着被开颅。老鼠被开颅、植入芯片、再缝合,可以继续正常活几个月。她笑着说可以帮我开脑，但希望我永远用不上。

聊完一整个下午，我最大的感受是：

目前我们对人脑、对 AI,其实都了解得太少了。

但也正因为如此，我越来越觉得，我们活在最好的时代。

真正的无人驾驶已经在路上跑、量子计算在飞速进步、人类在认真筹备第一次登上火星、脑机接口真的开始读懂大脑……

我今天在纽约街头连着撞见好几块 AI 和加密行业的广告：出租车顶上写着「It’s happening with Ripple」，公交广告是贝莱德的比特币 ETF，站台是OpenAI 的 Codex。

以前根本无法想象的事，都在同时发生在美国。而我们，刚好生在这个最好的时代。

[GUTE_RachelHan @GuteslaX]

통신과 컴퓨팅 인프라를 동시에 잘하는 회사는 역사적으로 드물었다.

통신은 커버리지, 연결성, 운영 안정성이 핵심이다. 컴퓨팅은 연산 자원, 데이터센터, 소프트웨어 생태계, 개발자 플랫폼이 핵심이다.

서로 유사해 보이지만 실제로는 자본 구조와 운영 문법이 다른 산업이다.

통신사는 클라우드에서 번번이 밀렸고, 클라우드 기업은 국가기간망이라는 특성 때문에 통신망을 직접 장악하지 못했다.

이런 점에서 스페이스X는 단기적으로도 흥미로운 회사다.

중장기적 흥미로움은 역시 우주에 대한 희망에서 나온다. 다만 스페이스X의 다음 확장은 우주 경제보다 네오클라우드에서 먼저 일어날 가능성이 크다.

스타링크는 글로벌 통신망이고, xAI는 대규모 GPU 클러스터를 필요로 한다. 여기에 전력, 데이터센터, 광네트워크가 결합하면 스페이스X는 단순한 우주기업이 아니라 AI 인프라 기업이 된다.

네오클라우드는 GPU를 빌려주는 사업에서 끝나지 않는다. AI 연산 수요가 커질수록 병목은 GPU에서 전력, 냉각, 네트워크, 입지, 운영 역량으로 이동한다. 결국 다른 공간과 다른 최적화를 고민해야 하는 산업이 된다.

스페이스X는 스타링크를 통해 이미 지구적 연결망을 확보했다. 기존 네오클라우드 기업보다 더 넓은 인프라 스택을 만들 수 있는 위치에 있다.

결국 스타링크는 위성 인터넷을 넘어 AI 컴퓨팅을 위한 글로벌 연결망이 될 수 있다. 네오클라우드는 GPU 임대 사업을 넘어 지구 규모의 디지털 인프라 플랫폼으로 확장될 수 있다.

그럼 더 먼 미래를 꿈꿔볼까.

진정한 우주 개척 시대는 로켓 몇 기를 쏘아 올리는 것으로 열리지 않는다. 사람이 살고, 기계가 일하고, AI가 의사결정을 내리며, 에너지와 물자가 순환하는 지속 가능한 인프라가 구축될 때 비로소 시작된다.

지상에서 통신망과 컴퓨팅 인프라를 하나의 시스템으로 운영하는 경험이 쌓이면, 그 구조는 저궤도 위성망, 우주 데이터센터, 달과 화성의 컴퓨팅 인프라로 확장될 수 있다.

우주 공간에서 해결해야 할 문제들은 생각보다 익숙하다. 통신, 전력, 물류, 자원 채굴, 제조, 건설, 운영, 자율화. 지금 지상에서 AI와 데이터센터, 로봇과 자율주행으로 풀고 있는 문제들의 연장선에 있다.

이 문제들을 하나씩 해결하며 기술 스택을 쌓아간다면, 우주 역시 거대한 격자(Grid) 구조로 연결될 수 있다.

오늘날 인터넷이 도시와 국가를 연결했듯, 미래에는 위성망과 우주 데이터센터, 궤도 정거장, 달 기지, 화성 거점이 하나의 네트워크를 형성할 수 있다. 스페이스X 구성원들은 아마도 이런 희망을 공유하고 있을 것이다.

그리고 비슷한 비전을 공유하는 투자자와 팬들은 그 격자 위에서 각자의 인생과, 아직 태어나지 않은 자손들의 새로운 궤도를 상상하고 있을 것이다. @SpaceX

[动察Beating @BeatingOfficial]

【字节跳动将拆分AI制药独立融资，控股新公司试水AI4S产业化】

字节跳动旗下 AI 制药业务线已启动拆分与独立融资进程。拆分后，字节跳动将继续控股新公司，并将 AI 制药核心团队、核心算法、技术平台和已有早期药物管线资产整体并入新主体。新公司由 2021 年成立的字节 AI 制药团队主导，将持续获得火山引擎的算力支持。

由负责人刘凯带领的 50 人团队由 AI4S 算法人才与资深制药专家构成，已将内部的蛋白结构预测模型团队并入。并入的团队先前发布了 Protenix 和 Seedfold 等分子结构预测模型，并推出了 binder 设计工具 PXDesign。面向真实药物研发，团队还打造了 AI 制药平台 Anew Labs，并在 2026 年 4 月的美国免疫学会年会上披露了 IL-17 小分子项目，在全球范围内首次实现利用小分子阻断 IL-17 家族的三个二聚体（AA/AF/FF），标志着相关研究已从算法模型推进至靶点与管线验证阶段。

字节跳动选择启动拆分，旨在通过独立组织架构更好地吸引顶尖人才，并适应生物科技独特的研发规律。相较于传统互联网业务，AI4S 研发包含模型研发、湿实验和临床验证等长周期、高难度的复杂环节。新公司独立后将拥有更灵活的决策空间。

[Eivor @Eivor_Koy]

That butterfly hairpin? The wings were made so thin and attached with hair-fine wires that they actually quiver with movement — a frozen instant of life captured in gold.

[Ihtesham Ali @ihtesham2005]

A Russian psychologist spent 10 years proving that the act of talking to yourself out loud is one of the most powerful cognitive tools the human brain has, and almost nobody outside his field has read the work.

His name was Lev Vygotsky.

He worked in Moscow in the 1920s and died of tuberculosis in 1934 at the age of 37. He had no laboratory, no funding, almost no English readers, and a body of work that the Soviet government suppressed for two decades after he died.

He produced the foundational theory of how human cognition actually develops, and the central piece of that theory was a behavior almost every adult is faintly embarrassed about.

Vygotsky noticed that young children talk to themselves constantly. They narrate their own actions, they argue with imaginary opponents, they instruct themselves through tasks out loud.

The dominant theory at the time, from the Swiss psychologist Jean Piaget, said this was a sign of cognitive immaturity that children would eventually grow out of as they learned to think properly.

Vygotsky said the exact opposite.

He argued that this self-directed speech was the most important cognitive event in the entire developmental window, because it was the moment a child first started to use language as a tool to control their own mind. The child was not failing to think. The child was learning how to think by externalizing the process and listening to themselves do it.

He predicted that as children matured, this out-loud self-talk would not disappear. It would go underground. It would become silent inner speech, which is the running monologue every adult has inside their own head for the rest of their life.

The voice you hear when you read this sentence is the direct descendant of a four-year-old narrating their own block tower.

For 50 years almost nobody outside Russia had access to his work, and the few researchers who did pick it up could not get funding to test it. Then in the early 2000s the experiments finally started to pile up, and what they found was that Vygotsky had been right about something even more important than he knew.

The first major study came from Gary Lupyan at the University of Wisconsin and Daniel Swingley at the University of Pennsylvania in 2012. They ran a simple visual search experiment. Participants were shown 20 images at once and asked to find a specific object, like a banana or a chair. In one condition they searched silently. In the other condition they were told to say the name of the object out loud to themselves while looking for it.

The participants who spoke the target name out loud found the object significantly faster, with higher accuracy, than the participants who searched in silence. The effect was strongest when the spoken word matched a familiar object the brain already had a strong category for.

Saying the word out loud literally tuned the visual system to detect that thing better. The researchers called it the label feedback effect, and the implication was that the act of vocalizing a goal physically changes how the brain processes the world while pursuing it.

The second major study came out of the University of Michigan and Michigan State in 2017. The lead researchers were Ethan Kross and Jason Moser, and they used both EEG and fMRI to record what happens inside the brain when people talk to themselves while emotionally upset.

They asked participants to recall painful autobiographical memories and reflect on them in two different ways. Some used the first person, saying things like "why am I feeling this way." Others used the third person, referring to themselves by their own name, saying things like "why is John feeling this way."

The brain scans showed that the simple act of switching from first person to third person, even silently, decreased activity in the medial prefrontal cortex, the region responsible for rumination and self-referential pain. Within a single second of using their own name instead of the word I, participants showed measurably lower emotional reactivity. The shift required no extra cognitive effort. It cost the brain nothing. And it worked.

Kross described the mechanism in his interviews. Talking to yourself by name creates a small amount of psychological distance from your own experience. Your brain processes the situation more like a problem belonging to someone else, which means it can analyze it instead of drowning in it.

What Vygotsky had intuited in 1934 turned out to be even more powerful than the developmental theory he built it into. The voice you use to talk to yourself is not background noise. It is one of the most precise cognitive tools the brain has, and you can change how it works just by changing the pronoun you use.

People who talk through problems out loud are not anxious or unstable. They are running an externalized version of a process the rest of us are running silently and worse. The kindergartener narrating their block tower, the surgeon muttering through a procedure, the engineer pacing a hallway describing a bug to nobody, the athlete repeating a cue to themselves before a free throw, they are all using the same ancient mechanism that builds and steers human thought.

You can run the experiment yourself the next time you are stuck on something hard. Stop trying to solve it silently in your head. Say it out loud. Describe what you are seeing. Walk yourself through the steps as if you were explaining it to a colleague who is not in the room.

And when something genuinely upsets you, switch to your own name. Ask why this person is feeling this way, instead of why I am feeling this way.

The voice you have been told to keep quiet your entire life is one of the oldest pieces of cognitive technology you own.

Most people are still embarrassed to use it.

[Susan STEM @feltanimalworld]

非常好的一张图啊！基本就这样。我要加一套木工工具，手持就可以。因为后来许多工具比如操作台是直接买的木材我自己做的。把钉子和连接件做成活的，别用胶水不好拆。你需要一个比较重和比较稳的台面，用起来尤其是台式CNC和3D， 桌子稳重在XY移动的时候感觉晃动没那么大。还有大量小型元器件的收纳。

[老杨啊 @yhslgg]

有人用 AI 把《史记》57万字变成了一个可以搜索、跳转、推理的知识图谱
兄弟们，今天看到一个让我直接愣住的开源项目——史记知识库，GitHub 1300 星。

一句话：有人用 AI 把《史记》57万字全部拆解标注，做成了一个可以交互、搜索、推理的知识图谱，让两千年前的文字像代码一样能语法高亮、跳转、消歧。

数据规模直接给你看清楚：

• 实体词条：14,065 个
• 标注次数：126,441 次
• 历史事件：3,198 条
• 事件关系：7,637 条
• Wiki 页面：20,830+ 页
这不是简单的文本处理，是真正的知识工程。

几个亮点真的值得单说：

（1）22类实体彩色语法高亮——人名、地名、官职、事件，每一类都有颜色标注，读《史记》像读代码一样清晰，想关哪类就关哪类

（2）史记地铁图——130条交互式时间线，支持缩放拖拽搜索，整个春秋战国到汉初的历史脉络一张图看清楚

（3）Butler AI Agent 自主维护——一个 AI Agent 持续跑了 12,000+ 轮自主更新 Wiki，不是人工录入，是 Agent 在干活

（4）矛盾检测——史记里前后记载有出入的地方，系统能自动检测出来，学术研究用这个太香了

背后的方法论也很牛：本体不是专家预先设计的，是从文本数据里"自下而上涌现"出来的，迭代速度从数周一轮压缩到数小时一轮。作者还出了两本 PDF 方法论手册，合计 863 页。

历史爱好者、做内容的、研究 AI 知识工程的，都值得去看看。

在线访问：https://t.co/Z9JoCt7KLU

GitHub：https://t.co/ARvGphJfU2

好东西转给需要的兄弟。🚀

#史记 #AI知识图谱 #开源 #老杨啊分享