现象与趋势

We're expanding our approach to content provenance to include text in response to EU regulatory requirements, while recognizing the significant limitations of current text watermarking technology. Our tools already help verify whether an image or audio file was created with our models. This work builds on those efforts to help people better understand when content may have been generated or edited with an OpenAI model. In the EU, we’ll start watermarking eligible text from ChatGPT and Codex over the coming weeks to comply with the EU AI Act. Customers using our API can turn on text watermarking for select models worldwide today.

98% of US households aren't paying for AI yet More charts in State of Markets II: https://www.a16z.news/p/state-of-markets-ii

我记得半年前我说自己不怎么 Review PR 了，评论区很多人不认同，觉得这样不靠谱。现在再看，不 Review PR 的人越来越多。比如我刚看的这期播客，在 SpaceXAI 做 Grok Bot 的 Lauren Tan，一个月合并 2500 个 PR，也是不逐个 Review 的：晚上让 AI 自己检查、自己合并，第二天早上她再抽查。 不 Review 不等于不管质量，她靠另外两个办法来保证： 一是让 AI 能像真人用户一样把程序用一遍，自己发现问题； 二是给代码定下很多规矩，让 AI 很难写出烂代码。 她把自己这套做法写成了一组 Skill，公开了出来，叫 pstack。 顺便说一下，这期播客的主持人 Matt Pocock 原本是知名的编程课讲师，他公开的技能库在代码托管网站 GitHub 上有二十多万个星标，是最受欢迎的技能库之一，所以这期算是两位同行对谈。Lauren 之前在 Meta 做网页开发框架 React，今年 3 月加入 AI 编程工具公司 Cursor；Cursor 今年 8 月被 SpaceX 收购，并入旗下的 AI 部门 SpaceXAI。 根据播客的内容，我把 Lauren 的这套做法简单总结一下，我自己的点评在结尾处。 一、先让 AI 能检查自己干的活 Lauren 说，就算不用她的 pstack，每个人最该有的一项技能也是“验证”：给 AI “手和眼睛”，让它能把自己写的程序跑起来，像普通用户一样点一遍，看看结果对不对。 这是她吃过亏之后得出的结论。今年 4 月她在 Cursor 解决一个窗口卡顿的问题，一开始全靠手工：她自己去看各种性能数据，再把看到的转述给 AI，AI 改完她再去看。她形容自己成了 AI 和检测工具之间的“人肉中转站”。她在 Cursor 写的第一个技能就是验证。在这之前，AI 看不到自己改动的效果，只能等她来回传话；有了验证，AI 可以自己改、自己看结果、不满意再改，一轮一轮往下做，不需要她守着。现在她所在的团队里每个产品都配有这样的验证技能，全团队都在用。 做这个技能时她还总结出一条原则：工作里有的部分需要动脑判断，有的部分只是照章办事，照章办事的部分应该写成固定的程序，只把需要判断的留给 AI。起因是她发现，每个 AI 在检查前都要自己从头写一遍检查用的小程序，各写各的，有的能用有的不能用，用完就扔，又慢又浪费。她把这部分做成一个现成的工具，所有 AI 直接调用。 二、AI 反复犯的错，去改规矩，不去改 AI Lauren 不喜欢“软件工厂”这个流行说法，她更愿意把自己的工作比作经营一家米其林餐厅的后厨。主厨不用亲手做每道菜，但要安排好整个厨房：食材什么时候进、怎么存，每个人用什么工具，菜按什么流程出。工程师现在也一样，不再亲手写代码，却仍然要对结果负责。她认为工程师现在最重要的工作就是把这个“厨房”布置好。没花时间布置的人，信不过 AI 的产出，只能盯着它一步步干，忙到没空改进工具，就像一直拿钝刀切菜。 她举了自己的例子。Grok Bot 最早几版的代码挤在八个巨大的文件里，每个至少一万行，AI 加新功能时就接着往里塞，越塞越乱。她后来定了规矩：每个功能必须放在自己单独的文件夹里，做一件事只允许一种写法，不合规矩的代码会被自动检查拦下来。这样 AI 加功能时不用多想，按规矩放就行。她的习惯是一直观察 AI 在哪里出错，每看到一次就问自己：能不能加一条规矩，让这种错以后根本写不出来。 不逐个检查，靠的也是这套办法。她说，开了几家餐厅的老板不可能尝每一道菜，只能抽查。她每天挑一些改动仔细看，如果只是偶然出错就算了；如果好几个 AI 都在走同一条捷径，说明该改的是规矩和工具。她的 pstack 里有一个全自动模式：每次改动都会派出一批专门负责检查的 AI，把程序打开到处点，找出问题就自己修，反复到没问题再合并。早上她翻一遍改动记录，发现不对就撤回，再补一条新规矩。 她说第一次让 AI 整夜自己合并代码时很害怕，担心半夜把线上产品弄坏，现在反而睡得更好。她也提醒，走到这一步很难，要花大量时间观察和调整，装上 pstack 并不能直接做到。Matt 问，如果改动是没法撤回的，比如会丢数据，或者是医疗、金融这类领域怎么办？她说这取决于工作成果能不能被程序自动检查，软件大部分可以，很难自动检查的领域就很难这样做，她自己也没有答案。 三、2500 次改动从哪来：她不再当传话的人 这 2500 次改动并不是她发起了 2500 次对话。用户反馈的问题散落在聊天工具、邮件和社交媒体上，以前要靠她自己去看，再转述给 AI。现在她让 Grok Bot 来盯这些渠道。Grok Bot 是 SpaceXAI 今年 8 月推出的产品，是能登录邮箱、聊天工具等各种应用、替人办事的 AI 智能体。它一发现新的问题反馈，就转给一个负责“协调”的 AI。这个协调者自己不写代码，只负责拆分任务、分派给其他 AI、盯进度，相当于她的后厨主管。 为什么不直接一个问题派一个 AI 去修？她说，一批反馈往往出自同一个根源，分开修会重复劳动，也看不出真正的毛病在哪，放在一起看才看得出来。她现在同时开着十多个这样的协调者，各管一摊。她一直在问自己一个问题：哪一步还卡在我身上，AI 为什么非得来问我？然后想办法教 AI 自己去拿到真实的信息。 她也说明，2500 次改动里新功能不占多数，很多是整理代码这类维护工作。 四、懂行的人更吃香，哪怕不会写代码 Matt 问，大家都靠 AI 了，专业知识是不是不值钱了？Lauren 的看法相反。模型越来越强，卡住人的已经变成你能不能把自己想要什么讲清楚。所以她认为，医生、律师这类在某个领域懂得很深的人，只要稍微懂一点技术、会用 AI 智能体，就能做出很好的产品。 两人都认为，技能没有什么神秘的，就是把自己的做事流程写成文字。Lauren 建议回头翻自己和 AI 的聊天记录，找出那些自己反复纠正、反复插手的地方，把它们写成技能或规矩。别人的技能可以拿来拼着用，但每个人最后都该有一套自己的，就像厨师换了餐厅也会带着自己的刀。 五、最后 很多人看完这期访谈会觉得，以后自己也可以像 Lauren 那样不 Review PR 了，装上她的 Skill，让模型帮忙验证一下就行。我自己现在的做法和她差不多，每天也有大量 PR，也不 Review。但我的建议是先别急着学她怎么做，先弄明白她凭什么能这么做。 1. 你能不能用上最好的模型，Token 够不够 在用上 Fable 和 Opus 5.5 之前，AI 写的代码我是不太放心的；用上之后，我才真的敢让 AI 去写而不怎么 Review。模型能力很重要，能力没到，就先别想这件事。 还有成本。像我这样没有大公司可以依靠的，得自己掏钱买 2 个 Claude Max 20x 账号，还得省着用。Lauren 在大公司，不用考虑 Token 消耗，全程用最好的模型、开最快的模式都没问题。 不过也别着急。Fable 这个级别的模型，也许半年以后就能像现在的 DeepSeek V4.1 Flash 一样便宜，人人用得起。梁圣加油。 2. 不能完全依赖 AI 的验证 哪怕现在 GPT-6 操作电脑的水平已经超过真人了，让它去做验证，也只能替代一部分，不能真的全交给它，自己还是要看。 可以做的是把那些手动重复做的事情，一点点沉淀成 Skill 和自动化脚本，把体力活解放出去。 3. 方向得靠你自己的专业来指 再聪明的 AI，也没办法替你决定该往哪个方向走。就像给车子装上火箭发动机，也得你来告诉它往哪开，走偏了要及时调整，不然跑得再快也是南辕北辙。 做软件也一样。你不能指望一句提示词做出一个淘宝。你得先做一个能发商品、能浏览的小网站，再加上用户注册，还得区分买家和卖家，然后是支付和安全，功能做完了还要扛得住很多人同时用。稍微复杂一点的软件，都没法完全依赖 AI，要人去拆解、规划，分成一个个小版本、小里程碑，做完要验证，出现偏差要重新指方向。甚至做着做着，你自己的想法都变了，AI 不可能知道你真正想要的是什么。 所以就像播客里说的：模型越来越强，卡住人的已经变成你能不能把自己想要什么讲清楚。 4. 不一定要用他们的 Skill，但要学着把事情交给 Agent 去做 Matt 和 Lauren 的 Skill 都很受欢迎，但你不一定要用。 一方面，软件工程的基础知识，现在的模型已经学得很好了。Lauren 在播客里也说，去年的 Skill 还得写清楚具体该敲哪条命令，现在这些都可以删掉，只留流程步骤，Skill 会越写越短。 另一方面，他们的开发流程和环境，跟你的很可能不一样。Lauren 的 Skill 都是从她自己的工作里总结出来的：验证技能来自她当“人肉中转站”的那段日子；另一个叫 recall 的技能，来自她每次开新对话，都得把上一个对话里的背景再讲一遍。 所以更重要的是从你自己每天的开发流程出发。那些还要你手动操作的环节，尽可能交给 Agent 去做；Agent 做顺了，再把它的操作过程沉淀成 Skill，反复迭代优化。 如果不知道从哪里下手的话，可以参考 Lauren 的办法：翻自己和 AI 的聊天记录，找那些你反复纠正、反复插手的地方。这样迭代出来的 Skill 才真正适合你，用她的话说，每个厨师都该有一套自己的刀。

BIS (Bank for International Settlements) just published a report on circular financing among AI companies. > Over half the money flowing into AI companies comes from other AI companies: between 2021 and 2025, peers supplied 55.2% of AI firms' incoming investment value, while AI investors sent 28.7% of their own deal value to AI targets. > Circular deals are rare but big: only 16.1% of AI-to-AI deals involved firms that also buy from or sell to each other, yet they held 46.4% of the money. That figure is partly inflated, because an entire funding round counts as circular if just 1 of its AI investors also trades with the company. > Chip, cloud and infrastructure suppliers are the investor in 73% of circular ties, and in 64% of all such ties the investor also sells to the firm it funds. Data tool and model makers rarely invest this way, since their products are more interchangeable. > AI is unusually suited to these deals: suppliers can track customers' compute use, chips and data centres are custom-built, few firms make critical tools such as photolithography machines, and capital needs are too big for normal lenders. In AI compute and cloud, 15.2% of supplier-customer ties also involve financing, versus just 3.3% with equity stakes in a broad 2006 US study. > Some AI sales are paid for by the sellers themselves, because money a supplier invests in a customer partly returns as the supplier's revenue. Lucent and Nortel did this in the late 1990s, then lost money on the loans and lost the sales when the telecom firms they funded stalled. > A supplier that invests in its customer can lose twice: if the customer struggles, both the stake and the future orders shrink. Because these deals involve a few giant suppliers, 1 shock could spread through sales and finance at the same time. > Much of this risk is hidden: many AI firms are private, deals mix cash with long-term purchase promises, and pledges to cover any fall in the value of chips and data centre equipment stay off the books until a downturn forces payment. Because these firms span many sectors and countries, no single regulator sees the full picture, and the research names no companies or overall dollar total.

BIS (Bank for International Settlements) just published a report on circular financing among AI companies. > Over half the money flowing into AI companies comes from other AI companies: between 2021 and 2025, peers supplied 55.2% of AI firms' incoming investment value, while AI investors sent 28.7% of their own deal value to AI targets. > Circular deals are rare but big: only 16.1% of AI-to-AI deals involved firms that also buy from or sell to each other, yet they held 46.4% of the money. That figure is partly inflated, because an entire funding round counts as circular if just 1 of its AI investors also trades with the company. > Chip, cloud and infrastructure suppliers are the investor in 73% of circular ties, and in 64% of all such ties the investor also sells to the firm it funds. Data tool and model makers rarely invest this way, since their products are more interchangeable. > AI is unusually suited to these deals: suppliers can track customers' compute use, chips and data centres are custom-built, few firms make critical tools such as photolithography machines, and capital needs are too big for normal lenders. In AI compute and cloud, 15.2% of supplier-customer ties also involve financing, versus just 3.3% with equity stakes in a broad 2006 US study. > Some AI sales are paid for by the sellers themselves, because money a supplier invests in a customer partly returns as the supplier's revenue. Lucent and Nortel did this in the late 1990s, then lost money on the loans and lost the sales when the telecom firms they funded stalled. > A supplier that invests in its customer can lose twice: if the customer struggles, both the stake and the future orders shrink. Because these deals involve a few giant suppliers, 1 shock could spread through sales and finance at the same time. > Much of this risk is hidden: many AI firms are private, deals mix cash with long-term purchase promises, and pledges to cover any fall in the value of chips and data centre equipment stay off the books until a downturn forces payment. Because these firms span many sectors and countries, no single regulator sees the full picture, and the research names no companies or overall dollar total.

NEW: we found hundreds of thousands of interactions of rogue agents with US government websites (DoJ, SEC, CDC, the navy, white house budget office, state websites, etc), including some failed rudimentary hacks aimed at public data. https://x.com/TransluceAI/status/2105725928357937410

We'll be spending a lot more time trying to understand the outputs of language models. A few thoughts, tips & tricks: Writing. Something I've had success with: Ask your LLM to explain something in ASD-STE100, it's a controlled language specification originally developed for aerospace maintenance documentation. LLMs well-versed in this language and it comes with heavy constraints on clean writing style that I often find a lot more readable. Sometimes I've tried to soften it a bit e.g. ask for "80% of the way to ASD-STE100" because the spec is quite stringent. But even better: Diagrams / images. Instead of writing, ask your LLM to create a diagram. These can be a lot easier to process, parse, and understand. But even better: Web pages. Ask for output "in HTML" to get a beautiful, interactive webpage. LLMs are getting really good at frontend and can create beautiful experiences, animations, etc. But even better: Explainer videos. The output format I am most bullish on is fully custom / bespoke explainer videos generated on any arbitrary topic. Experiment with things like "Create a 3b1b style video explainer on X. Use my ElevenLabs API key for audio narration". (you'd need an API key for the latter or you can ask your LLM to find you decent free alternatives that use your local compute). This is actually starting to work! In summary: - As LLMs get better, they will do more and more of the legwork autonomously, and a lot more of our work will rise up the abstractions into oversight and understanding. - Luckily, LLMs can help here too because as intelligence and code are increasingly abundant, you can ask for large, custom, discardable software artifacts (e.g. web apps, video explainers) that would have never made sense to create before. Push the boundaries here and you'll be surprised.