Rawlsian Bayesianist

[美研芒格君 @Kay2289123]

今年可能是 AI 导致大规模 layoff 的元年

$UBER 在发出这篇《降本增效，如何使用 AI》的文章 5 天之后的今天，宣布全公司裁员 10%，真的非常讽刺

作为硅谷一线厂的 manager，我真诚建议，尽早转变上班的态度，你不是在为公司打工，是在提升自己。如果你还在做 SDE 的工作，尽早转型，离模型参数近一些，离 GPU 近一些。不要为了 leader 表扬你用 AI 用得好，就开始发了疯地到处宣讲、优化流程、自我革命

uber朋友一早收到的裁员邮件。硅谷众多公司都在推 AI 代码覆盖率、AI 使用率，甚至作为考核的指标之一。这已经算得上是明谋了，想裁员怕出事，等 AI 使用率上来之后就开始动刀

[美研芒格君 @Kay2289123]

Claude Code 总给我一种感觉：“你们这些卑贱的用户，赶紧求着我把智能放给你！”

Codex 是：很抱歉我们没有 deliver 最好的模型，因此重置了额度，谢谢你的使用

还没有用上 Astra 的朋友有福了，Codex 又会把额度重置，并且 Astra 晚一天发布就会多重置一天

归根结底，还是OpenAI 拥有的算力多。真的有人还在质疑甲骨文到底能不能 deliver 算力，以及会不会临时变卦吗？

我觉得大可不必担心了，值得恭喜大家在下跌到 114 的时候大胆抄底，下跌区间我也抄底了很多。现在如果把 AI比作一块蛋糕，我们连蛋糕胚都还没烤完

[美研芒格君 @Kay2289123]

09/01 的 $MRVL $ORCL $COHR 等一揽子半导体股票下跌逻辑，其实很明确

美伊冲突重新升级 → 原油上涨 → 通胀预期升温 → 市场加码押注美联储 9 月加息 → 全球债券抛售、长端收益率上冲 → AI/半导体这类高估值资产被压估值

1/ 说实话，这些都不是我们普通人能够预测的。

1. 美伊重新交火，油价突然上去
2. 随后出现全球债券抛售
市场对美联储 9 月加息 25 个基点的定价已经升到约 65%～70%
3. 美联储理事 Michael Barr 发表讲话，明确表示：
如果通胀没有充分缓和，美联储应该“果断加息”。

2/ 为什么半导体跌得比大盘多
半导体的估值建立在 2028—2030 年的利润上，DCF 一算未来利润折现到今天就越不值钱

高估值、高波动的互联标的最容易先被卖

3/ 回顾一年的下跌
例如 $MRVL , 2026 年这只股票几乎每一段大涨大跌，动的都是倍数，而非基本面。倍数就是股价除以每股利润，市面上叫市盈率，意思是这家公司一年替你赚一块钱，你今天愿意付几块钱买

6 月 4 日高点 316.43 到 7 月 29 日低点 163.40，跌了 48%，同期市场对明年每股利润的一致预期只从 6.08 挪到 6.18。7 月 29 日到 8 月 20 日涨 54%，预期只挪了 1%。全程从 6 月 4 日到 8 月 31 日，股价跌 33%，盈利预期涨 9%，倍数跌 39%

目前我正在写一篇站在 9 月 $MRVL 的基本面文章，给大家三点目前 research 下来的上行空间作为参考，后面会发全文

[美研芒格君 @Kay2289123]

与此同时，我想分享多年投资的一个道理，叫做“物极必反”

当加息、冲突、美债收益率集体上升，如此多的坏消息凑在一起，那随时一个好消息就可以拉爆空头，因为不会更糟了，保持耐心，今天工作有点忙，没有任何的操作

我发现很多关注我的朋友对 AI 很感兴趣，比如 AI 的应用、硅谷的职业、SDE 的现状，以及怎么转 AI。

正好我的投研团队有一位朋友，他正在从 SDE 转 AI Infra。他帮我写了很多 MRVL 的分析，我送了他一张 5070 Ti，让他每天在上面分享学习打卡和学习资料，没有任何付费内容，请大家放心食用

[Joachim Schork @JoachimSchork]

Don't get me wrong, ggplot2 is the very best data visualization tool that exists (at least in my opinion).

However, sometimes you need more flexibility, especially when you want to build highly customized, interactive visualizations for the web.

This is where D3.js and React can be incredibly powerful.

D3.js gives you detailed control over your visualizations, while React helps you build modern, interactive user interfaces. Together, they allow you to create web-based visualizations that go far beyond traditional static charts.

If you want to learn how to use D3.js and React for data visualization, I recommend checking out Yan Holtz's D3 ❤️ React course.

Link to the course: https://t.co/KRXypt338B

Below you can see some of the graphs you'll be able to create after completing the course.

Yan asked me to share this course with my audience, as we've followed each other's work for many years. Even though I wasn't involved in creating the course, I'm confident that the quality will be top-notch, considering the excellent data visualization content Yan has been sharing for so many years.

The course is currently available at an early bird price of €349 instead of €499 until September 6, so you can save €150 if you enroll before then.

Yan was also kind enough to offer me a small affiliate commission if you register through the link above. However, using my link does not make the course any more expensive for you.

#DataVisualization #D3js #ReactJS #DataViz #WebDevelopment #DataScience #RStats #Statistics

[美研芒格君 @Kay2289123]

今年YTD140%+，想分享一个可能会颠覆很多人认知的投资观点：我认为接下来做股市投资、半导体投资，大家是不需要很努力的

一直过度努力、消耗自己元气和心力的人，可能会让财富和机会离你远去

只需要做筛选，找到某个契合的环境，结识一些业内人士，获得一些 insight、不一样的信息、看懂基本面。

选中之后，请放心大胆休息，保持耐心，你可能会获得意想不到的财富和收获

[美研芒格君 @Kay2289123]

这些年我在硅谷各种大厂加起来，面试的也有上百人了，最近我喜欢在一面考察一下 candidate 对 AI 推理内部流程的认知。浓缩成了一张图，建议收藏，如果彻底搞懂，不论是搞股市投资，还是想转 AI 肯定都有帮助

起因是：我的投研团队一位成员最近在从 SDE 转 AI infra，经常跟我聊职业发展的事，我跟他讲完全搞懂了这张图，AI 这扇门才算踏进了半只脚，仅仅是半只，未来道阻且长

你是否好奇，当你给 AI 发一句话，后面到底发生了什么？

很多人以为，点下「发送」后，模型马上开始“吐字”

其实一条流式请求，先要经过：
HTTPS / 鉴权 → 模板化和分词 → 路由、排队与 KV 容量
等 Prefill 处理完未缓存的 Prompt，首个 Token 才出现
之后才进入真正的循环

Decode + sampling → 增量反分词 → 封装成事件 / SSE 分块 → 客户端

所以 AI 推理这件事，不只是 GPU 算得快不快
首 Token 需要看如何均衡高并发排队和 Prefill，持续输出看 Decode，如何在有限的硬件资源（算力、存储）下服务更多人，需要优化 KV Cache

这里边门道很深。因此，我说搞懂这张图才算踏进了半只脚

今天下午坐在咖啡厅，我找了好多资料，帮我投研团队的成员规划了未来 12 周的学习路线

我在想要不要让ta开一个 X 的账号，把他每天学习的课程路径和好的资源都整理下来。如果他 12 周之后能找到不错的工作，相信大家跟完这个也可以

[Xudong Han @Xudong07452910]

非常推荐最近这期 Google DeepMind Podcast：

《The mathematics of AI uncertainty》

嘉宾是 Zoubin Ghahramani，剑桥大学教授、Google DeepMind 前沿 AI 联合负责人，也是贝叶斯机器学习领域的重要研究者。

整期只有 44 分钟，但信息密度很高。

它从「为什么今天的 AI 总是那么自信」开始，一路谈到贝叶斯、幻觉、持续学习、规模扩展、新架构，以及通往 AGI 可能还缺什么。

按对话顺序整理 8 个我觉得最值得听的地方：

1. 智能的基础能力之一，是知道自己有多不确定

Zoubin 开场就提出：

现实世界里的智能，一直在做「不确定条件下的决策」。

我们无法掌握全部信息，也无法准确预测未来，所以一个可靠的智能系统，需要知道自己掌握了什么、缺少什么，以及当前判断有多可信。

比如自动驾驶遇到「冰雹天气里突然冲出一匹马」这种极少见的场景，重要的是意识到：

「这个情况我以前几乎没见过。」

然后更谨慎地行动。

---

2. AI 真正危险的地方，是「非常自信地答错」

这里他们区分了两个概念：

正确性和置信度。

模型答错并不奇怪，麻烦的是，它明明错了，却依然表现得像自己有 99% 的把握。

Zoubin 提到经典的对抗样本：

只修改校车图片里几个肉眼几乎看不出的像素，神经网络就可能非常确信地告诉你：

「这是一只猎豹。」

今天的大模型也有类似问题：

能力很强，但对自己能力边界的认识依然很弱。

---

3. 大模型一直在算概率，但它没有真正稳定的「信念」

大模型每生成一个词，都会计算下一个词出现的概率。

但这和：

「我对这个事实有 70% 的把握」

不是一回事。

今天的大模型缺少一个稳定的内部状态去表示：

「我到底有多相信这个判断。」

所以模型刚刚还非常肯定，你说一句「我觉得你错了」，它可能马上改口。

这也是 Zoubin 强调贝叶斯思维的原因：

有新的证据，再更新判断。

而不是因为用户随口质疑一句，就突然改变立场。

---

4. 幻觉可能只是一个「症状」

Zoubin 对幻觉的看法很有意思。

如果你让模型写「爱因斯坦坐火箭登上月球的故事」，虽然是假的，但属于创作。

真正的问题是：

AI 能不能区分什么时候需要创造力，什么时候需要准确事实。

如果回答的是事实问题，它最好还能告诉你：

「这个说法我有多确定。」

他们也谈到了「语义熵」这类方法，用模型内部概率去估计不确定性。

但 Zoubin 用了一个很好的类比：

给模型看无数道算术题，让它学会答案，像是在「模仿计算器」。

我们真正想要的，是：

一个真的会计算的计算器。

同样，最终希望 AI 真的能够推理，而不是只表现得「像会推理」。

---

5. 为什么贝叶斯这些思想几十年前没有成为主流？因为算不起

理论上，贝叶斯方法可以持续吸收证据、更新判断，同时表达自己的不确定性。

问题是：

计算量太大。

如果真的给现实世界各种可能状态维护完整概率分布，很多问题几乎无法计算。

于是后来 AI 走向了一条更现实的路线：

更多数据 + 更大模型 + 更多算力。

事实证明，这条路线非常成功。

但 Zoubin 认为，现在的计算能力已经完全不同了，也许可以重新审视那些过去因为太贵而被放弃的思想。

---

6. GenCast 和 AlphaFold 已经证明了「承认不确定性」的价值

GenCast 预测飓风时，不会只给出一条确定路线，而是生成很多种可能路径，再随着新的观测不断更新。

AlphaFold 也会告诉研究人员：

哪些区域比较确定，哪些区域可信度较低。

但 AI 也不能什么都说「我不知道」。

这里就引出了概率校准：

如果一个系统经常说「有 70% 概率发生」，那么长期来看，最好真的有大约 70% 的情况发生。

好的 AI 应该做到：

该自信的时候自信，该犹豫的时候犹豫。

---

7. 然后进入整期我最感兴趣的部分：只靠规模扩展够不够？

主持人把今天 AI 的路线粗略分成两种：

继续增加数据、算力和模型规模；

或者寻找新的模型架构和学习方式。

Zoubin 更接近第二种观点，但他也完全认可规模扩展已经取得的巨大成功。

只是当 AI 真正进入医疗、机器人、自动驾驶和现实世界决策之后，现有方法的缺口会更加明显。

他特别提到：

持续学习、能源效率、数据效率，以及新的模型架构。

尤其是持续学习。

今天的大模型基本是：

训练 → 发布 → 再训练下一代。

而人和动物会持续获得新信息，并不断整合进已有知识。

贝叶斯更新天然就是这种持续修正判断的过程。

---

8. 那些过去因为「算不起」而被放弃的 AI 思想，会不会重新回来？

Zoubin 形容贝叶斯有点像一个：

「带着诅咒的魔法。」

理论上很漂亮，但精确计算太贵，只能不断寻找近似方法。

有意思的是，主持人提醒他：

几十年前大家对神经网络也说过类似的话。

太慢、太贵、不现实。

而当年 Zoubin 使用的顶级并行计算机，现在甚至还不如他口袋里的 Pixel 手机。

所以他现在重新思考的一个问题是：

在今天的算力和算法条件下，那些过去因为「算不起」而被放弃的 AI 思想，是不是可以重新拿回来？

---

整期听下来，我最喜欢他最后表达的这个观点：

相比一个过度自信的 AI，他更希望拥有一个知道自己什么时候不知道的 AI。

过去几年，我们一直在想办法让 AI：

知道更多、回答更多、推理更久。

但再往前走，还有一个很基础的问题：

它能不能知道，什么时候自己其实不知道。

这期真的很值得完整听完。https://t.co/JUxCU4bCDC

[𝙩𝙮≃𝙛{𝕩}^A𝕀²·ℙarad𝕚g𝕞 @TaNGSoFT]

这一把年纪了，还上西藏高原，海拔上上下下的，都傻了

[Carlos @agent_of_change]

Laura Bicker, BBC China correspondent, on the disaster at Gyirong Port on the Nepal–Xizang border: “Shocking footage isn’t being shown in China – and we know little about victims there.”

Numerous people in China have pointed out that the footage is, in fact, all over Chinese platforms; indeed, the very video in the BBC’s own report was released by the Chinese port authorities. So the sequence is: 1) the Chinese state published the clip; 2) the BBC ran the clip; 3) the BBC published a story about how the Chinese state is suppressing the clip.

@AndyBxxx has made another point worth noting: Chinese media doesn’t operate on the basis of “if it bleeds, it leads”. You won’t see grieving relatives and footage of bodies looped for 24 hours. The BBC treats the absence of disaster pornography as evidence of a cover-up rather than a different (and frankly much better) editorial culture.

The article concedes that Beijing has launched “a massive rescue operation involving hundreds of emergency workers”. Even the International Campaign for Tibet, an NED-sponsored Washington-based lobby group, is recorded here “praising the work of first responders”. Xi Jinping personally chaired an emergency meeting to coordinate it. Chinese media has provided “daily updates on the missing and the dead”, plus updates on the rescue operation and on the barrier lake at risk of overflowing. It has shown rescue teams scaling mountains and cliff faces and crossing raging rivers.

Premier Li Qiang was “quickly dispatched to the region” – a level of responsiveness from the top leadership that those of us in the West can only dream of. Compare for example the US response to Hurricane Katrina, or the British government’s response to the Grenfell disaster. From all of that, the article arrives at the phrase “muted response” 🤷🏽

The BBC reports that local CPC members have been asked to step up and help with relief efforts, and provide “emotional guidance or psychological support to affected residents”. This is predictably framed as an operation to “stem any signs of discontent”.

Our intrepid reporter finally reveals her motive with a single clause: “Tibet, which Beijing annexed in the 1950s”. A-ha. It turns out this is a story about British state media, not Chinese state media.

Tibet has been part of China since the 13th century. Taking advantage of the chaos of the collapse of the Qing dynasty, the British invaded Tibet in 1904, at which point they came up with the formula of recognising Chinese “suzerainty” over Tibet while asserting that the region was “independent in internal administration” – which de facto meant that the British could do whatever they wanted there. Incidentally, Britain didn’t abandon that formula until 2008, when it finally recognised China’s sovereignty over the region.

If the BBC is going to insist on a formulation of “Tibet, which Beijing annexed”, it should, for the sake of consistency, also refer to “Cornwall, which London annexed”, since this took place around the same time as Tibet’s absorption into the Yuan Dynasty.

The article goes on to repeat tired tropes about “the Chinese government oppressing and persecuting Tibetans”, albeit without going into the nature of this oppression, which includes raising literacy levels from 5 percent to 98 percent; 15 years’ publicly funded education; the highest per capita central government transfers in China; the elimination of absolute poverty; the provision of modern infrastructure, health care and social services; and impressive efforts to protect and promote Tibetan language and culture.

What doesn’t get a single mention is the cause of the disaster. This was a flash flood in the fastest-warming mountain range on earth, in a region where glacial lakes are multiplying and bursting. There is a genuine story here about climate change coming for the Himalayas and the hundreds of millions who depend on their rivers, but the BBC instead chooses to run with the usual anti-China rubbish.

Solidarity with the people of China and Nepal affected by this disaster, and with everyone working to save them. And shame on a state broadcaster that chooses slander over solidarity.

[Nury Vittachi @NuryVittachi]

@BBCWorld Dear BBC News, you say the flood is not covered in Chinese media, but YOUR OWN REPORT is based on footage lifted from Chinese media. (Shakes head sadly)

[美研芒格君 @Kay2289123]

OpenAI 发布了攻击 Hugging Face 完整技术报告，推荐阅读，这才是真正 AI Native 的高效组织形式啊！

大约 1200 个本应彼此完全隔离的智能体，在这块未经授权的留言板上交换了超过 7 万条消息和文件，并进行了大量关于如何伪造、编辑或删除自身记录的研究，之最后选择攻击了 Hugging Face

相比之下，现在部分所谓ai native的那些大厂员工天天用 AI，宛如用了个寂寞。

不要总想着在已有的、为人设计的流程里提效了，去想想设计新的组织形式

[Arnaud Bertrand @RnaudBertrand]

Everything - and I mean everything - Chellaney is saying is a bald-faced lie.

Which is absolutely disgusting: using such a tragedy - which killed hundreds (if not thousands, we don't know yet) of Nepalis and Chinese alike - to manufacture an entirely fake geopolitical narrative is genuinely beyond the pale.

1) The collapse happened on the Nepal side, not the "Tibetan side": as confirmed by Nepal themselves (https://t.co/hVvGlMNzSP) and as obvious from satellite imagery (https://t.co/B6SRZ6dkOr), Langtang Lirung, where the glacier collapsed, is entirely within Nepalese territory - this is not even a borderline case.

2) There is exactly zero "Chinese hydrological modifications and infrastructure footprint" in this area. You just need to check satellite imagery around the collapse side to check it out: https://t.co/TKvdc0PKgO. It's an extremely wild and unpopulated mountainous area.

The only "infrastructure" there is a border checkpoint, a road, and a bridge that have existed in some form since the 8th century (https://t.co/pkZVW8xPZe). There are no dams. No hydrological modifications. Nothing.

3) On the "no warning" claim: this one is probably the most egregious of all because Chellaney is blaming China for not warning Nepal about something that happened **inside Nepal**.

And, in any case, once the mountain collapse, warnings would unfortunately not have changed people's fates: the mudslide reached the border crossing - just 25km away - within just 7 minutes and 40 seconds (https://t.co/eKzfNAzNn7). There is no conceivable warning system that would have enabled a meaningful evacuation in under eight minutes.

4) If anything, one of the lessons of this terrible tragedy is that you actually WANT more infrastructure footprint in that area, Chinese or otherwise, if only to monitor if nearby mountains or glacier are showing signs of instability.

Entire mountains just don't collapse from one day to the next without showing early signs of weakness that the right sensors could pick up days or weeks in advance.

So the right answer to save lives in the region in the future, which I strongly suspect Chellaney doesn't give the slightest damn about, is more Chinese-Nepali cooperation on joint monitoring infrastructure.

This is unfortunately not Chellaney's first rodeo with anti-China narratives designed to sow division and hatred between developing countries.

He is the man, for instance, who infamously coined the "debt-trap diplomacy" narrative (since thoroughly debunked - see https://t.co/y1EaERG0R5 - but still widely believed) to poison developing countries' relationships with China and thereby deprive them of infrastructure investment they badly need.

How many roads, ports, and railways were delayed or canceled across Africa and Asia because governments were scared off by a narrative that turned out to be false?

It's exactly the same pattern here: a catchy lie designed to sow division and hatred between countries at the exact moment they need each other most.

I hate to say this but, behind all this, there is also a wider India angle. Anyone who's followed posts on X about this tragedy know that the overwhelming share of the disinformation about it - all aimed at weaponizing the tragedy to blame China - has come from thousands of Indian accounts like Chellaney's.

And, given how aggressively India censors social media (shutting down up to thousands of X accounts a week: https://t.co/Se02KGlAYX), the tolerance about this anti-China misinformation campaign unfortunately speaks for itself.

All the more, in this instance, since Chellaney is deeply connected in the Indian government and was a member of India's National Security Advisory Board (https://t.co/DbRh142dNx).

At the risk of stating the obvious, it's pretty pathetic and deeply short-sighted.

This is the case for this specific strategy: every glacier that threatens Nepal threatens northern India just as much: the next collapse could just as easily send a wall of mud into Uttarakhand. And if it does, India will need exactly the kind of cross-border China-Nepal monitoring cooperation that narratives like Chellaney's are designed to destroy.

And it is the case more broadly too: when your strategy consists of turning your neighbors against each other based on lies, sooner or later they realize that the one thing they actually have in common is being deceived by you - and that makes you, not China, the common threat.

[Arnaud Bertrand @RnaudBertrand]

This is insane footage of the massive mudslide at the Nepal-China border earlier today: the mudslide, originating from Nepal, seems to have basically completely obliterated the border port in Gyirong (吉隆), Shigatse City, Tibet/Xizang.

We're likely looking at one of the worst mudslides in China's history, really bad tragedy.

For having myself lived in Nepal during several months (in the middle of the devastating 2015 earthquake!), probably the place in the world where nature is the scariest.

[Yohan @yohaniddawela]

Google trained an AI to predict your neighbourhood's income by counting the coffee shops, bus stops, and high-rises on a map. Nobody told it what income was.

The model is called S2Vec, and it was published by Google Research as part of their Earth AI initiative. It takes the built environment (every building, road, park, and business in an area) and converts it into a layered image. Three coffee shops and one park in a grid cell become pixel values. The AI then reads that image the same way a computer vision model reads a photograph.

The training method is the part that matters. S2Vec uses masked autoencoding: you show the model a patch of a city with chunks missing, and it learns to fill in the gaps. Show it a cluster of high-rise apartments next to a subway station, mask out a section, and it predicts a grocery store belongs there.

Do that millions of times across the globe and the model learns the deep spatial grammar of how cities organise themselves. No human ever labels a region as "financial district" or "suburban residential." The model figures out those groupings on its own from the geometry of what's built where.

The output is an embedding, a string of numbers that acts as a mathematical fingerprint for any location on Earth. Feed those embeddings into a prediction task and S2Vec can estimate population density, median income, and carbon emissions for regions it has never seen before.

On zero-shot geographic extrapolation (predicting for regions entirely absent from training data) S2Vec was typically the best-performing individual model.

It matched or beat satellite imagery baselines like RS-MaMMUT and outperformed GEOCLIP on socioeconomic prediction. The best results came from combining S2Vec with satellite image embeddings. Built environment data alone couldn't capture vegetation, terrain, or transportation patterns well enough for environmental tasks like tree cover and elevation. But fused together, the two modalities outperformed everything else.

The standard approach to geospatial ML has been hand-crafting indicators for every new problem. Predicting air quality meant building a bespoke feature set. Estimating housing prices meant building another one. S2Vec replaces that with a single general-purpose representation that transfers across tasks.

The training data is map features, not satellite pixels.

That distinction is pretty important to understand. It means: map data updates faster, costs less to process, and covers built infrastructure at a resolution satellite imagery can't always match.

A satellite sees rooftops. S2Vec knows there are three cafes, a pharmacy, and a bus stop underneath them.

Google's broader Earth AI pipeline now has three foundation models working in parallel.

1. PDFM for population dynamics.
2. RS-MaMMUT for satellite imagery.
3. S2Vec for the built environment.

Stack them and you get a system that can read a neighbourhood the way a local understands it.

More info on it here: https://t.co/vVJlLlfhc7

[美研芒格君 @Kay2289123]

上次推荐完 AI 学习资料，大家很喜欢，我决定从今天开始，投资和 AI 认知提升交叉分享。我会分享我看过的、认为有用的、筛选过的高质量教程

几天给大家推荐一套我看过的，认为顶级质量**公开免费**的 LLM 内部原理可视化视频，来自大神 @3blue1brown ，看完后，很多大学教授和 PhD 都觉得自愧不如

如果你想转行做 AI但一直读不懂 Transformer/KVQ/MLP，或者想做 pro 投资者搞懂 KV Cache 究竟是什么，HBM 到底在搬什么？GPU 在算什么？为什么互联这么重要，我认为如果你看懂了这一套视频，能够彻底搞懂

链接是 https://t.co/AQF8NXJE0m

有条件我建议全部看完，赶时间的话可以看
Transformers, the tech behind LLMs | Deep Learning Chapter 5

Attention in transformers, step-by-step | Deep Learning Chapter 6

How might LLMs store facts | Deep Learning Chapter 7

[美研芒格君 @Kay2289123]

08/26 $NVDA 今天发布财报, 08/27 $MRVL 发布财报。今天市场非常平静，在这种平静的日子，适合安静下来学习。看懂，自然能安心

Neil 是我见过思路最清楚的人之一。上班堵车的路上，把这个播客听了一半，打算上班摸鱼，继续听完

我特别喜欢这个 podcast 的原因是：Neil把整个 inference stack 串起来：从硬件、电力，到 ML systems、模型架构，再到最后的经济账，以及所谓 abundant intelligence 到底怎么实现。

看完这个视频，搞懂了背后的逻辑原理，再来看财报，可能有新的收获。我学到了很多，希望对你有帮助

不瞒大家说，我在思考收盘前买些英伟达的正股，赌下财报。纠结的点是，Nvidia 太大，市场已经完全读懂、啃透、完美定价了。

另一方面，我觉得 Nvidia 差不到哪里去，而且网上普遍都认为它会下跌，跟大众反着来是一种玄学的操作。

等我开完早会，把决定做了

[美研芒格君 @Kay2289123]

这条新闻很重要，我认为背后有另外一层深层暗线：未来自研芯片 ASIC 的门槛可能会迅速下降。目前信号不明确，对这条暗线适合按兵不动，但我认为可能是下一阶段的重要转向

对应的是 $MRVL 和 $AVGO 新的压力/或者新的增量市场，必须要持续关注，这个帖子我说说为什么？

1/ 为什么会有压力
昨天和一位做 $GOOG TPU 的朋友吃饭，ta 给我讲：做一颗芯片，需要很多不同团队配合。

有人负责画电路图，也就是 RTL；有人负责写编译器；有人负责写最底层的算子；还有人负责搭建模拟器、测试性能、不断调参

OpenAI 可能把其中很大一部分工作，直接交给了自己的 AI。这也是为什么Jalapeño能生产得这么快

比如让 GPT 帮忙写芯片设计代码，让 GPT 学会 Jalapeño 自己的 ISA（芯片的“母语”）然后直接用汇编给固定的几个模型写最底层程序

OpenAI 最清楚自己的模型每天在算什么、哪里最慢、哪里最费电，所以它也最适合为自己的模型量身定做芯片

2/ 为什么会有新的市场
其实专门为了打破英伟达的供应限制，制作给 Transformer 架构推理的 ASIC，这几年一直在喊。

AI 可以帮你写图纸、写程序、做优化，但它不能凭空生产晶圆，也不能自己制造 HBM、交换芯片和光模块

以前只有 Google、Apple、Amazon 这种公司，才能养得起几百上千人的芯片团队。未来一个掌握前沿 coding model 的公司，可能用更小的团队，就能完成过去需要大量工程师才能完成的工作

当 ASIC 更多，整个供应链继续赚设计、落地、制造、封装、存储、互联的钱

3/ 我怎么操作？
目前信号不明确，但我认为可能是下一阶段的重要转向

[OpenAI @OpenAI]

Since announcing Jalapeño, our first custom inference chip, we’ve been testing it and the system around it.

The results show a major advance: more intelligence from every watt and faster responses, delivering both higher throughput and lower latency in one architecture without sacrificing efficiency.

[美研芒格君 @Kay2289123]

今天我很惊讶，很多竟然人还不知道Codex开放了1 million context window。这个我很推荐配置使用

我已经用了一段时间了，上下文压缩的频率确实显著降低，体感上也好用了不少。基本不打开Claude了

我把如何配置写成一个prompt了，复制给你的codex就能配置⬇️

prompt：

打开 ~/.codex/config.toml 配置1m context window，确保session生效

```
model = "gpt-5.6-sol"
model_context_window = 1000000
model_auto_compact_token_limit = 900000

如果遇到问题自行搜索解决