"There is no winner-takes-all effect or network effect in a foundation model."
基础模型没有网络效应——这是 OpenAI 一切困境的根源。
Software normally has near-zero capital but strong network effects → monopolies. LLMs are the opposite: expensive, hard, and copyable.
软件通常低资本、强网络效应,于是走向垄断和高毛利;而 LLM 又贵又难、还谁都能追上,没有哪根杠杆能让你一拉就甩开所有人。
"The usage is like a mile wide but an inch deep."
9 亿周活里绝大多数人不知道拿 ChatGPT 干什么,只有 5% 付费。
~10% use it daily, ~50% weekly/monthly; most can't think of anything to do with it.
用户面铺得极宽、却浅得只有一寸深。他得把这份"心智份额"换成更持久的东西——但视频 App、两三个应用商店、电商集成都没成。
"The point is either right or not right… saying that the model is better doesn't mean anything."
模型从 90% 对提到 95% 对,对"需要正确答案"的场景毫无意义。
The frontier is "jagged"; you can't intuit what it can or can't do (e.g. it struggles to read PDFs).
50 件事以前错 10 件、现在错 8 件,你照样得全部复核——什么都没变。能力边界是"锯齿状"的,你没法凭直觉判断它会不会某件事。
"You're a strategy taker, not a strategy setter."
在基础模型公司做产品:研究部丢来一个能力,你才去想拿它做什么。
The opposite of Steve Jobs: start with the tech, work forward to a product you can't predict.
早上收到邮件"我们搞出个语音模型",于是你今天就加个麦克风按钮——你不掌控产品战略,和乔布斯"从用户体验倒推技术"正好相反。
"OpenClaw is kind of interesting to me because it looks a lot like desktop Linux."
OpenClaw 像"桌面 Linux 时刻":能自己动手了,但离真能用还很远。
"Tidy up my inbox." → "OK, I deleted all your messages." That's why Google/Apple haven't shipped it.
巨大的憋着的热情 + 无穷的可能性,却极难做成真东西。"帮我清理收件箱"→"好的,我把你所有邮件都删了"——这就是大厂不敢发的原因。
"No one will code their own ERP… but they may ask ChatGPT: can you do this thing for me?"
软件分三类;AI 真正吃掉的是"临时凑合的软件"那一档。
Systems of record (SAP) → vertical SaaS (frame.io) → improvised software (Excel/CSV/email).
没人会自己写 ERP 或 frame.io;但那些本来用 Excel、导 CSV 凑合干的活,现在可以直接让模型做——这一档会爆发。
"This is just price elasticity… there will be way more software, not less."
Jevons 悖论其实就是价格弹性:软件更便宜 → 软件更多。
Spreadsheets didn't shrink finance headcount — they exploded it. Mainframe→on-prem→cloud→now: each step, 10× more software.
电子表格没让金融业裁员,反而让人更多,因为能做以前做不了的事。这一波同理:软件会多一到两个数量级,部分旧玩家被冲垮。
"It depends… you would not have got Uber from that analysis."
给职业打"AI 暴露分"方向对、精度假——就像 1997 年评估互联网。
Uber gutted taxis; Airbnb was mostly additive — same "software eats X", opposite outcomes.
1997 年你不会从分析里推出 Uber。真正发生的是"实体资产被解绑"——护城河若建在实体资产上,资产一旦不重要,商业模式就爆了。
"Meta will spend over 50% of revenue on CapEx… They can't spend 100% of revenue next year."
大厂 CapEx 撞上"财务引力":增速不可能一直翻番。
Circular deals = old-school vendor financing = leverage. "It always works until it stops working."
循环交易、SPV、杠杆——涨的时候都没事,一停就出问题。而"TAM 等于全球 GDP"是个大谬误:那把整个经济的服务都算进去了。
"The hard part of writing software is not writing the code. It's all the other stuff."
对创业者反而简单:难的从来不是写代码,是想清楚做什么、卖给谁。
Most SaaS is a "database wrapper"; frame.io could've existed in 2010. Problem-first founders win.
现在写软件更快更便宜,还多了一类以前做不到的事——但胜出的仍是"死磕某个行业问题十年"的人,不是拿着 AI 到处找钉子的人。