Good point. 我觉得还有一个问题是，大家现在都在想办法把 AI 塞进原有的人类 workflow 里，而不是围绕 AI 重新设计 workflow。所以在 fully automated agents 出现之前，很多流程其实还是挺笨重的。 有时候最优解是直接把 workflow 重建一遍，但大公司很难做到。

"If these scaling laws continue for only a year or two longer, we are likely to get what I’ve called Powerful AI, or a country of geniuses in a datacenter." - Dario Amodei, June 2026

显然NBIS模式更好，更抗造，可以平安度过风暴。CRWV的财务杠杆在加息周期会很难受的。如果capital market关闭了，CRWV会直接破产的。NBIS是，宁愿少收点钱，至少我的资本负债表是健康的。

So a 170MW site assuming they are GB200 NVL72? Is that from Colossus 2?

人们总是习惯于在FOMO与变态挑剔中做钟摆运动而不自知。😎

技术落地大概有四个阶段： 1. 个人用新工具做旧事情 2. 公司用新工具做旧事情 3. 公司开始做以前做不到的事情 4. 整个社会为了新技术改变运行方式 Claude Code 是历史上把前两步推进得最快的产品, 没有之一。 但后面两步就没那么简单了。 公司要改流程，要改组织架构，要改考核方式；再往后还要监管、客户、供应链一起改。 human institutions是完全没准备好的。 Waymo够成熟了吧？但牌照到今天都没完全铺开。FDA 那边更不用说。

这个无损全自动书籍扫描仪后面衍生除了个开源项目 linearbookscanner.org/

Kinda funny seeing the same “PDD trading at [insert any number of high single digit multiple]” quote repeated for like six straight quarters now, with almost no acknowledgement that it’s been a value trap the entire time.

GE Vernova CEO on current constraints in FY2026 Q1 call: "[Gas power capacity] lead times are about three years, the company still has capacity in 2029/2030, and the harder constraints are EPC schedules, permitting, and fuel availability."

CRWV Management on bottleneck in 2026 Q1: "[...]not just power but labor, memory, storage, and bring-up execution."

When you see how people use Claude Code/Codex/etc it becomes clear that managing agents is really a management problem Can you specify goals? Can you provide context? Can you divide up tasks? Can you give feedback? These are teachable skills. Also UIs need to support management

THE INEVITABLE MCP BATTLE If you work in AI, you've probably heard of MCP (Model Context Protocol), an extension of tool calling that allows agents to interact with applications in a standardized way. Social has been full of glowing reviews of MCP. Standards are awesome, so why wouldn't everyone embrace it? Two simple, logical reasons that MCP will not be welcomed by the largest and most ambitious applications: (a) Customer relationships and (b) Ads (a) Any B2B or B2C app that's built deep direct customer relationships over years and decades is going to view MCP as a disintermediatory force, one that decisively makes the AI agent the owner of the relationship. The analogy to search is uncanny: MCP allows AI agents to index apps, just like Google indexed websites. And we know who ended up winning that one. (b) Any app that monetizes their audience through ads will be even more directly impacted negatively by MCP. Every user who uses an agent to access the app through MCP, instead of logging into the app directly, results in lost ads engagement and ad revenue for the app. Ads is the highest margin revenue stream for many commerce apps, one they will strongly resist any attempt to diminish this business line. Apps who've always enabled API access, should likely not have any problems with MCP. However, UX-centered apps that have never opened up API access, owe it to themselves to consider the factors above before jumping headlong into MCP. I'm curious if any of the AI agents who are touting MCP, build an MCP server that allows other AI agents to access their data. ;) This will be an interesting tug of war.

BERNSTEIN: $TSLA stock “remains high on almost every valuation metric compared to both traditional and higher-growth auto OEMs .. Our DCF now points to fair value of $93 (down from $120), primarily due to lowered estimates .. but also due to a push-out in EV adoption growth.”

Notice which of your close friends has a tendency to optimize their life, for they will cease to be your friend.

meant

brands earn their attention and standing with consumers / seems unlikely they’d succumb to becoming someone else’s “dumb api call” (for an ai agent) as Benedict Evans put it

Personalized software will be the future of consumer products. Last weekend I built a "Dad Journal" for myself that reminds me to share one memorable thing about my daughter everyday. I did this using "prompt to app" software and built much of it from my phone, while in a pool. This would have taken me weeks with a full-stack developer a year ago. Did it in a few hours. Features include: * Streaks tracker to make sure I post everyday * AI enhancement for my journal entries * Speech to text: record audio & transcribe * Generative images for journal posts where I don't have a photo * SMS to friends/family who subscribe when I post a new entry

2. The competition with other e-commerce platforms is brutal. Management knows margins are getting squeezed and growth is slowing down. 3. US investors just aren't that interested. The regulatory crackdown hit harder than most people realize. They'll need to rebuild trust over time - one meeting with business leaders isn't going to fix that overnight. 4. The governance is pretty murky, with some concerning accounting issues - no IR team and no CFO? That's a problem.

I agree those numbers are solid. But we really need to understand why the valuations are so low, and what might actually change things: 1. PDD is getting heavier on assets and moving away from the light, two-sided platform it used to be. Colin Huang talked about this shift back in his 2021 letter, and it seems to be happening even more now.

I have argued in a bunch of places that a huge part of the attraction of privates is that you can pretend they are low risk (illiquidity was considered a bug in 1985, I argue it’s now considered a feature because it makes life feel easier). If you start making an active an accurate market in illiquids you risk breaking that spell (telling people what their actual real life not make believe prices are).