Building And Structuring An AI Native Company

频道: YC Root Access
视频: https://www.youtube.com/watch?v=Z3JyAqh4ixg
原文语言: en
统计: 共 29 轮 · 提问者 1 · Tom 26

[0:03] 提问者

How's your day going so far?

今天到现在为止过得怎么样？

[0:05] Tom

Good.

挺好的。

[0:06] Tom

Great. Um So, I've got about 25 minutes or something. Um and I'm going to try and talk about building AI-native companies. This is based on a talk I gave about a month, maybe 6 weeks ago. It's on YouTube. Um I've updated it. I've added some thoughts. I think the main caveat I want to start with is no one knows how to do this. And so, this is um very theoretical. It's based on uh hundreds of YC companies that we're working with right now. Um lots and lots of people are trying things. And I think if anyone tells you they've got it all figured out, they are probably lying. I think internally at YC, we have an amazing software team. We've been pushing the boundaries of I think what LLMs can do for a couple of years now. I think we have a bunch of really interesting use cases. Um which I'll I'll give examples of those as we go through. But, I I don't think we've got it I don't think we've yet got it fully figured out, but I expect um by the end of 2026 or possibly by the first batch in 2027, we it will be possible, I don't know if we'll take the PR risk of it, it will be possible to from end to end have an AI fully read all the applications, pick the companies to interview, conduct the interview, select the companies to fund, give them advice throughout the batch on how to do like run their companies more effectively, introduce them to investors, review their pitch decks, and then debug their pitch meetings.

太好了。嗯，我大概有 25 分钟左右，想跟大家聊聊怎么打造 AI-native 的公司。这次的内容是基于我大概一个月、也许六周前做的一场演讲，那场在 YouTube 上能找到，这次我做了更新，加了一些新的想法。开场我最想强调的一个前提是：没有人知道这事儿该怎么做。所以接下来讲的东西都还非常理论化，素材来自我们现在正在合作的几百家 YC 公司——非常非常多的人在做各种尝试。我觉得如果有谁跟你说他已经全想明白了，那多半是在吹牛。YC 内部有一个非常出色的软件团队，这两年我们一直在往 LLM 能力的边界上推，也攒下了一批挺有意思的用例，后面我会一个个举例。但我不觉得我们已经彻底想明白了。不过我预计，到 2026 年底、最晚到 2027 年第一批 batch 的时候，有件事在技术上是能做到的（至于我们敢不敢承担那个公关风险就不好说了）：从头到尾，让 AI 读完所有申请、挑出要面试的公司、亲自做面试、决定投哪几家、在整个 batch 期间给他们怎么把公司跑得更高效的建议、把他们介绍给投资人、审他们的 pitch deck，最后再复盘他们的融资会议哪里出了问题。

[1:41] Tom

I think an AI will be able to do all of that end to end. Probably by the end of this year, if not by the first 3 months of next year. We've got quite a lot of it working already. We're exposing a lot to our founders. I'll tell you more about that, but that's sort of where this talk has come from. So, what is an AI native company? Well, the first thing to start with is what is not an AI native company. This is how the Roman legions were were structured. The smallest unit was the contubernium. So, eight soldiers who share a tent, equipment, and a mule, and there's a decanus, and then 10 of these things called a contubernia form a century, which is actually 80 men, not 100, which is very confusing. But, you see this kind of hierarchical um way of organizing humans, right? The Roman legions used this to project power across North Africa, across um Near Asia, up into the UK, up to Hadrian's Wall. This was the way that the Roman Empire projected power. And they would pass information down and receive reports up. And a human being was responsible for being that conduit the whole way. And bizarrely, like 2,000 years later, we're still basically running the same kind of thing. And then like they look slightly different. This is a Carpathia tweet, uh which I thought was very good.

我相信 AI 端到端跑完这一整套是能做到的。大概今年年底，最晚不会超过明年前三个月。其实相当一部分我们已经跑通了，也开放给创始人在用了。这些后面细讲，这场演讲基本就是从这儿来的。

那么，什么是 AI-native 公司？先从反面说起——什么不是 AI-native 公司。这是罗马军团的组织方式。最小单位叫 contubernium（八人帐篷组）：八个士兵共用一顶帐篷、一套装备和一头骡子，配一名 decanus（十夫长）；十个 contubernia（八人帐篷组）组成一个 century（百人队），而百人队实际上是 80 个人、不是 100 个，这点特别容易搞混。但你能看到，这就是一种把人一层层组织起来的科层结构，对吧？罗马军团靠它把力量投射到北非、近东，一路往北推到英国、推到哈德良长城。这就是罗马帝国投射权力的方式。指令一层层往下传，报告一层层往上收，而全程负责当这个传导管道的是人。诡异的是，两千年后的今天，我们基本还在跑同一套东西。只是长得稍微不太一样。这是 Karpathy 的一条推文，我觉得画得非常到位。

[3:00] Tom

But, um like they're all organized slightly differently, but the same principle is true, which is that humans are the conduit for information going up and down um up and down the rungs, right? And so, there's a Jack Dorsey tweet tweet which kicked this whole thing off about 2 months ago. There's an underlying assumption that organizations have to be hierarchically organized with humans as a coordinating mechanism. And I think basically AI breaks that apart. So, you don't have to have humans as a coordinating mechanism any longer. And I basically think most companies have mis- Misunderstood is the wrong word. Most companies sort of experienced ChatGPT as a Q&A bot. They put put ans- They put questions in, they get answers out. And then maybe they've sort of moved to longer running perhaps what you might call an agentic system or something. But basically, you tell you ask the agent to go and do something, it goes away until it gets stuck, and it comes back to you, and you as a human are like the gating mechanism. And so, while you are asleep, the the thing can't work, right? Like if it stops halfway through at 3:00 in the morning, like it's waiting on input from you, it can't go anymore. You as a human are the gating mechanism.

虽然每家公司的组织方式略有差别，但底层原理是一样的：信息在层级之间上上下下，靠的是人来当传导管道。大概两个月前 Jack Dorsey 发了一条推文，把这个话题彻底点着了。我们心里都有个默认假设：组织必须按层级来搭，人是那个协调机制。而 AI 基本上把这个假设拆掉了——你不再需要用人来当协调机制。

我觉得大多数公司都误——「误解」这个词不太准确。大多数公司对 ChatGPT 的体验，基本停留在「问答机器人」：丢进去一个问题，吐出来一个答案。后来可能进化到运行时间更长一点的东西，也就是大家说的 agentic 系统之类的。但本质上还是：你让 agent 去干件事，它一路干到卡住，然后回来找你，而你——这个人类——就是那道闸门。所以你睡着的时候，这东西是转不动的。凌晨三点它跑到一半停下来等你输入，就再也走不动一步。你这个人类，就是那道闸门。

[4:17] Tom

And so, you you know, you can do things like making engineers 20% more productive, you can add co-pilots for lawyers to make their existing workflows get a little better, you can ship more software with AI, but really all of these things um the men in the suits are coming for me. I don't know what to I don't know. um uh really it's dependent on humans being the coordinating mechanism in all of these places. And if you think about it not just as like increasing productivity, but in terms of new capabilities, Huh, that's interesting. Oh, there we go. um W- We're basically a single person can do way more work than previously thousands of people. Your entire company becomes queryable, and you're not just shipping more software, but the the software you ship is like agent native. I These are a lot of buzzwords. I'm going to go into more detail about what I mean. AI At the moment, AI A lot of companies like bolting AI on onto the side of the company, right? In all the ways we've spoken about. The Q- question answer bots, they're sort of very um lightweight agents that can call tools effectively like a web search or something like that. um But if you think about the company as a series of self-improving AI loops from the ground up, you actually get to a very very different conclusion. Okay?

于是你能做的事情就是：让工程师效率提升 20%，给律师加个 copilot 让现有工作流好一点，用 AI 多发一点软件。但说真的，这些东西——（穿西装的人要来抓我了，我也不知道该说啥）——归根到底，它们全都依赖人在每个环节充当协调机制。

但如果你换个角度，不把它看成「提升生产力」，而是看成「解锁新能力」——嗯，这就有意思了。哦，出来了。基本上，一个人能干的活可以顶过去几千人；你的整家公司都变得「可查询」；而且你不只是多发了些软件，你发出去的软件本身就是 agent-native 的。我知道这一串全是热词，后面我会展开讲清楚具体指什么。

现在很多公司的做法，是把 AI 「挂」在公司边上，就是我们刚才说的那些形态——问答机器人，或者能调调工具（比如网页搜索）的轻量级 agent。但如果你从地基开始，把公司理解成一连串自我改进的 AI 闭环，你会得到一个非常非常不同的结论。对吧？

[5:37] Tom

Previously, AI is like make each person 20% more productive or 30% more productive. What happens if we reimagine the company as a series of AI loops? What is an AI loop? It's basically like um So, the top you've got um product telemetry inbound messages. This is like data from the real world, right? Like billing signals, support tickets, code changes, stuff that gives you information in. And then you go down to the perhaps a policy layer, which is like what rules can it um what rules constrain the AI? What things it must ask approval for? What does it have to log? Um you have a tool layer where it can call perhaps internal AI APIs, um send emails, uh update the billing, um and you might have APIs for this or MCP or whatever. And then you have a series of quality gates, which it which could be a human, and I'd argue probably should not be a human in in all but the most extreme cases, but the quality gate can often be a second adversarial LLM. It might be something that's that's looking at the output saying, are we suffering from prompt injection here? Or in a case of a bank, are we giving financial advice here, yes or no? And if we are, we can't, you know, like constrain it or a really easy example for engineers is like um the quality gate might simply be a second model doing code review.

以前 AI 的作用是让每个人效率提升 20%、30%。那如果我们把公司重新想象成一串 AI 闭环呢？什么叫 AI 闭环？大概是这样：最上面是产品埋点数据、进来的各种消息——也就是来自真实世界的数据，比如计费信号、客服工单、代码变更，这些给你输入信息的东西。

再往下是策略层：有哪些规则在约束这个 AI？哪些动作必须先申请审批？哪些必须留日志？

然后是工具层，它可以调用内部的 AI API、发邮件、更新账单，走 API 也行、走 MCP 也行，随你。

再然后是一串质量闸门。闸门可以是人，但我认为除了最极端的情况，都不该是人。质量闸门往往可以是第二个「对抗性」LLM：它盯着输出问，我们是不是中了 prompt injection？或者在银行的场景里问，我们这是不是在给投资建议？是还是不是？如果是，那就不行，得给它套上约束。对工程师来说最容易理解的例子是：质量闸门就是第二个模型来做 code review。

[6:56] Tom

Um and then a learning mechanism, which is basically you loop back around, right? You deploy the thing, you make the change, and you see how it impacts the real world. If you can do this entire loop without a human, your product starts improving itself when you're sleeping. Can you scroll down a little bit to the bottom of this, please? Yeah, all the way to the bottom. Thank you very much. So, um the first time I saw these things, it like broke my brain a little bit. Um internally at YC, uh we've got a lot we've got a data on 7,000 companies, 20,000 founders, like many hundreds of thousands of applications. So, a year or two ago, we built an AI agent that it basically human it's like English to SQL, right? Like it's a simple data querying agent. You can be like, what was a split of European versus uh American founders this batch?" And it would go and translate that human query into SQL and run it and come back with the answer. It's like, yeah, pretty pretty magical. But that feels like a tool that kind of makes a data analyst 20% or 30% more productive. For me, actually makes me as a partner 10% or 20% more productive and I don't have to hire a data analyst. It was fine. And it would keep hitting edge cases or like not be able to do certain things.

再往后是学习机制，说白了就是绕回来把环闭上：你把东西部署上去、做了改动，然后看它对真实世界产生了什么影响。如果这一整圈能在没有人参与的情况下跑完，那你睡觉的时候，你的产品在自己变好。

能帮我往下滚一点吗？对，一直滚到最底下。非常感谢。

我第一次看到这类东西的时候，脑子有点炸。YC 内部有 7000 家公司、20000 名创始人、几十万份申请的数据。一两年前我们做了一个 AI agent，本质上就是「英语转 SQL」——一个简单的数据查询 agent。你可以问它：「这一批里欧洲创始人和美国创始人的比例是多少？」它就把这句人话翻译成 SQL、跑一遍、把答案给你。挺神奇的。但那种东西感觉只是让数据分析师效率提升 20%、30% 的工具；对我来说，是让我这个合伙人效率提升 10%、20%，而且我不用去招一个数据分析师。也就还行吧。它老是撞上各种边界情况，有些事就是干不了。

[8:08] Tom

I'd be like, "Oh, that's annoying." And then we shipped something about two or three months ago, which for me was like the head explosion moment, which was a second agent, which uh sort of lives on top of this um data query agent. And the second agent runs overnight and it looks for all the queries that any human has done during the day and it looks for success and failure. So, the success is like, yeah, I copy it, I send the email, I do whatever. The failure is like, ah this isn't work or it I hit a permission issue or whatever. Like, the database index doesn't work or it has it requests some permission it doesn't have. And then that second AI agent overnight goes and puts in pull requests to fix all of the problems from the day before. And so, if you go back as a human on the second day and run the same query as yesterday, it now works. And so, this process of like evaluating what the system is doing and then proposing changes to the system and updating the system makes the whole thing self-improving. And that for me was like a head explosion moment. I think if you were here earlier for um James's talk from PostHog, they basically have the same thing for products now. They take in all your product telemetry and they see where it's breaking and then they put in pull requests and eventually auto merge those pull requests. So, your entire product surface will become self-improving. It will just dynamically You'll have a a bunch of headline metrics you want to optimize for. You'll have a bunch of product like vision documents, which says, "This is the kind of thing we want to build. Hit This is in scope and out of scope. And the AI will basically look at your product telemetry, come up with ideas, deploy those ideas, test them with humans, and see if it moved the needle or not, and just self-improve.

我就想，唉，真烦。然后大概两三个月前我们上线了一个东西，对我来说是「脑子炸掉」的时刻——第二个 agent，架在那个数据查询 agent 之上。这第二个 agent 在夜里跑，把白天所有人跑过的查询全翻一遍，看哪些成功、哪些失败。成功就是：我把结果复制走了、发了邮件、干了点什么。失败就是：这玩意儿没跑通、撞上权限问题、数据库索引不对、或者它申请了一个自己没有的权限。然后这第二个 AI agent 通宵去提 pull request，把前一天的问题全修掉。于是第二天你这个人类回来，跑昨天那条一模一样的查询，它现在能跑通了。

这个「评估系统在干什么 → 提出对系统的改动 → 更新系统」的过程，让整套东西开始自我改进。对我来说这就是脑子炸掉的时刻。

如果你早些时候听了 PostHog 的 James 那场演讲，他们现在对产品做的基本是同一件事：把你所有的产品埋点数据吃进去，看哪里坏了，然后提 pull request，最后自动合并。于是你的整个产品面都会变成自我改进的。它会自己动态调整——你给它一组想优化的核心指标，再给它一堆产品愿景文档，写清楚我们想做什么样的东西、什么在范围内、什么在范围外。然后 AI 就去看你的产品埋点、想出点子、把点子部署上线、拿真人测一测，看指标动没动，然后自己迭代自己。

[9:48] Tom

Copacetic um uh had a great tweet about 3 weeks ago now on like the auto research thing. Basically applying this to machine learning. Just having the AI I think it was on GPT-2 equivalent or something like that. But coming up with research ideas, testing it, running overnight, and just hill climbing. All this is just hill climbing. If you have a an end measurable outcome, you can just come up with ideas, you can test it. Did we go up the hill or down the hill? If down hill, discard. If up, like keep. And just keep hill climbing. And computer systems are really really good at this kind of hill climbing work because they never stop. They will just run indefinitely indefinitely until you reach local maximum. It's kind of cool cuz you they don't have to sleep. Um Another example from YC, we started recording office hours about 6 months ago. Um and I was given the job of rewriting a section of We have a user manual internally. Um which is written over the last 15 years. And it's like 500 pages long. And a lot of it was very very good like 5 years ago. And then AI happened. And like honestly a bunch of the advice is just not relevant anymore. And I had the job of rewriting a bunch of this stuff. And I like procrastinated and procrastinated and procrastinated.

大概三周前，Copacetic 发过一条很棒的推文，讲的是「自动做研究」这件事，基本就是把这套方法用到机器学习上：让 AI 自己——我记得是在 GPT-2 量级的模型上——自己想研究点子、自己测、通宵跑，就这么一路爬山。

其实这一切都是爬山（hill climbing）。只要你有一个可量化的最终结果，就可以：想点子、测一测，我们是往山上走了还是往山下走了？往下就丢掉，往上就留着，然后接着爬。计算机系统特别特别擅长这种爬山的活，因为它们从不停下来，会一直跑一直跑，直到爬到局部最优。挺酷的，因为它们不用睡觉。

再举个 YC 的例子。大概六个月前，我们开始把 office hours 录下来。当时我被派了个活：重写我们内部那本「使用手册」的一部分。这份手册是过去 15 年里一点点写出来的，大概 500 页。里面很多内容放在五年前非常非常好，然后 AI 来了，说实话，一大堆建议现在已经不适用了。我的任务就是重写其中一大块。然后我一拖再拖、拖了又拖。

[11:04] Tom

And we started recording the office hours 6 months ago. Now we have like I don't know 3 or 4,000 hours of office hours recorded. And someone on the team, Harj I think it was, had the idea of like what if we just transcribe and then mine all of the advice we actually give in office hours, and then use that to rewrite the user manual. And so

六个月前我们开始录 office hours。现在大概攒了三四千小时的录音。团队里有人——我记得是 Harj——提了个想法：要不我们把这些录音全转写出来，把我们在 office hours 里真正给出的建议全挖出来，然后拿这些去重写使用手册？于是——

[11:23] Tom

[snorts]

[笑声]

[11:23] Tom

when the advice changes, the AI will see what the humans are saying, extract the insight, and just rewrite the user manual, and push it. So it becomes this like living guide of how we actually advise in companies. And then obviously once you've got it in a user manual, you just make the entire thing queryable to an advice agent, right? Like if a if a founder asks us a question in Slack and we get a lot of questions that are pretty similar, honestly, like how should I price my first customer or my third customer is is about to churn, what should I do or this customer has asked for this esoteric feature, what I'm just like we've answered versions of this question like 500 times. And so now the AI can give not just Tom's answer, but like here's how Niccolo would answer it or here's how Harj would answer it. I think you can actually get superhuman advice from this thing because it's able to recall perfectly and access the intelligence of 16 partners, not just one with fallible memory. And so you can kind of see how this these like loops can run in the background and get better and better and better and better. And then I think you give them a virtual machine. So, you give your agent a virtual machine with things like tools, so search the web, crawl your internal company directory, search your Slack history, persistent file storage, so it can write it can come up with a plan and write the plan to disk so if it fails halfway through it can resume.

当建议发生变化时，AI 会看到人类在说什么，把洞见提取出来，直接重写使用手册并推上去。于是它变成一份「活的」指南，反映我们实际是怎么给公司提建议的。

显然，一旦内容进了使用手册，你就可以把整份东西做成一个「建议 agent」能查询的知识源。比如创始人在 Slack 里问我们一个问题——而我们收到的问题其实高度雷同，说真的，像「第一个客户该怎么定价」、「我第三个客户快流失了怎么办」、「这个客户提了个很偏门的需求怎么办」，这类问题的各种变体我们已经回答过大概 500 遍了。所以现在 AI 不只能给出 Tom 的答案，还能告诉你「Niccolo 会怎么回」、「Harj 会怎么回」。

我认为这东西真的能给出超越人类水平的建议，因为它的记忆是完美的，而且能同时调用 16 位合伙人的智慧，而不是只有一个记性还不靠谱的人。

所以你大概能看出来，这些闭环可以在后台一直跑，然后越来越好、越来越好、越来越好。

再往下，我觉得你要给它们一台虚拟机。给你的 agent 一台 VM，配上工具：搜网页、爬公司内部通讯录、搜 Slack 历史记录，还有持久化的文件存储——这样它可以先想出一个计划、把计划写到磁盘上，万一中途挂了还能接着往下跑。

[12:56] Tom

And then it if it needs to, it can write code and execute the code to solve a problem. And then see how that worked against the plan and adjust and repeat. And now you've basically got an AI employee. And I we're at the very, very early stage of this with something like Open Claw or Hermes is a basically VMs with an agent living inside on a on basically on a repeating loop.

如果有需要，它还能写代码、执行代码来解决问题，然后对照计划看效果如何、做调整、再来一遍。到这一步，你基本就有了一个 AI 员工。

我们现在还在非常非常早期的阶段，像 OpenClaw 或者 Hermes 这类东西，本质上就是一台 VM，里面住着一个 agent，在一个不断重复的循环里跑。

[13:20] Tom

[gasps]

[惊叹声]

[13:22] Tom

So, where we are now, right? I think YC has a few of these loops internally. I think the next thing that happens like it occasionally a human has to intervene, right? We're like, "Well, the loop's going wrong." Whatever. Like, "Let's take some insight from there put it over there." I think the thing that happens afterwards is the loops start talking to each other or the AI employees start communicating with each other to share ideas and re-plan and assess and and whatever. And you end up with this like I mean, company brain is is like an overused term. Now, it got very very popular about 6 weeks ago and as of like 2 weeks ago it's I see people in the audience with company brain written on a t-shirt now. There he is.

那我们现在走到哪一步了？我觉得 YC 内部已经有几个这样的闭环了。接下来会发生的是——偶尔还是得有人介入一下，比如「这个环跑歪了」，或者「把那边的洞见挪到这边来」。

但我认为再往后会发生的是：这些闭环开始互相对话，或者说这些 AI 员工开始互相沟通，交换想法、重新规划、互相评估等等。最后你就得到了一个……我是说，company brain（公司大脑）这个词现在已经被用滥了。它大概六周前突然火起来，而到两周前——我在台下已经能看到有人穿着印了 company brain 的 T 恤了。喏，就是那位。

[14:05]

[laughter]

[笑声]

[14:06] Tom

Um but what So, what is this as an idea? It's the Mhm. I'm going to steal some of Jack Dorsey's words, but in a conventional company intelligence is spread throughout the people and the hierarchy roots it. So, in this new model intelligence lives in the system. What I mean by that basically is that um lots and lots of companies try and write down standard operating procedures, right? Like, the way things should be done. But then when you actually observe people, how they actually do the job they're like, "Oh, well, in in this edge case it's really I should do this other thing." And there's a lot of stuff in people's heads that if it weren't in people's heads and it was actually written down you'd get a lot of value out of it. And so, this [snorts] idea of a company brain I think is the combination of like basically taking all of the data in your company and that's things like YC's application data, it's it's transcripts of every call with with every meeting with every founder, it's the advice we give um the way we pick companies. And then you record that and make it legible. You make You make it so that the AI can access all of it. And then you have a bunch of these reinforcing loops that run indefinitely and then are able to talk to each other, and the entire thing becomes a self-improving company.

那么，作为一个概念，这到底是什么？我要「借用」一下 Jack Dorsey 的话：在传统公司里，智能分散在每个人身上，而层级结构把它钉死在那儿。在这个新模型里，智能存在于系统之中。

我的意思大概是：非常非常多的公司都试图把标准作业流程（SOP）写下来——事情应该怎么做。但你真去观察员工实际怎么干活，他们会说：「哦，遇到这种边界情况，其实我该那么做才对。」有大量东西留在人的脑子里，而这些东西如果不待在脑子里、而是真被写下来，你能从中榨出巨大的价值。

所以 company brain（公司大脑）这个想法，我觉得本质上是：把你公司里所有的数据都拿出来——对 YC 来说就是申请数据、和每位创始人每次会议的通话转写、我们给出的建议、我们挑公司的方式——把它们记录下来，并且做成「可读的」，做到 AI 能全部访问。然后你搭一堆这样的强化闭环，让它们无限期地跑下去，还能互相对话。最后，整家公司就变成了一家自我改进的公司。

[15:25] Tom

I think that's basically where we're going. That you just pump tokens in, you get access to all of your data, and the entire thing becomes self-improving.

我觉得大方向基本就是这样：你不停地往里灌 token，模型能拿到你全部的数据，整套系统就开始自我改进。

[15:33] Tom

[snorts]

[笑声]

[15:33] Tom

And people live at the edge. Right? So, you have this kind of company brain, this data, these self-reinforcing loops, and then the human beings live around the edge where this intelligence makes contact with reality. So, I think Jack Dorsey said, "People reach into places the model can't yet go. They sense things the model can't perceive. Intuition, opinionated direction, cultural context, trust dynamic, the feeling in a room." Um they make the calls the model shouldn't make on its own, especially ethical decisions, novel situations, and high-stakes moment where the cost of being wrong is existential. A world model, this company brain that can't touch the real world, is just a database. And so, these humans, what what he's saying, I think, is the humans uh in the sales calls, right? They're like visiting the client office to reassure people, to convince the CEO, to pitch the investor. It's all of the emotional uh kind of interpersonal stuff that the humans are there for. But, the humans are no longer necessary to do the information routing. Right? Like, all of the information is routed automatically to the person who needs it. They don't need to go through layers and layers of management to get approved.

而人待在边缘，对吧？你有这么一个「公司大脑」，有数据，有这些自我强化的循环，人则分布在边缘——也就是这套智能跟现实世界发生接触的地方。Jack Dorsey 好像说过：人能伸进模型还去不了的地方，能感知模型察觉不到的东西——直觉、有主见的方向感、文化语境、信任关系、一个房间里的气氛。人来做那些不该由模型自己拍板的决定，尤其是伦理判断、从没遇过的新情况，还有一旦错了代价就是生死存亡的高风险时刻。一个碰不到真实世界的世界模型——这个所谓的公司大脑——那不过是个数据库罢了。所以这些人的位置在哪儿？我理解他的意思是，人出现在销售会谈里，对吧？跑到客户办公室去安抚对方、去说服 CEO、去向投资人 pitch。人留下来做的，全是这类带情绪的、人跟人打交道的事。但信息传递这件事不再需要人了。对吧？所有信息会自动路由到需要它的那个人手上，不用再一层一层往上走审批。

[16:46] Tom

Um And so, that's basically what I think companies are going to be. They're going to be much, much smaller. They're going to be centered around a kind of company brain, and the people will interface with the real world, and then feed all that information back into into the brain. A reasonable question is, "Sounds great in theory. How do like this seems like sci-fi kind of. Like, will it happen? Who knows. What can we actually do now to make it happen?" Um and so, practically, these are the tips we're giving founders right now. The first is I'd burn tokens, not headcount. So, we're seeing founders get to demo day with a million dollars of revenue, and some getting to series A um with 10 million of revenue with like a fraction of the the people they had before. You can argue about whether it's a thousand X or whatever, but it's multiple times more, right? Like someone a polyglot who really knows how to use these tools is worth many, many times uh the people before. Um and you can just do this As a result, the headcount of the whole company is going to be less. In particular, it means no middle management. We just don't think you need layers and layers of like directors and VPs anymore.

所以我认为未来的公司基本就长这个样子：规模会小非常非常多；核心围着一个公司大脑转；人负责跟真实世界打交道，再把拿回来的信息全部喂回大脑里。有个很合理的疑问是：理论上听着挺美，可这不有点像科幻吗？真会发生吗？谁知道呢。那我们现在具体能做点什么，让它真的发生？所以从实操层面讲，下面是我们现在给创始人的几条建议。第一条：烧 token，别烧人头。我们看到有创始人走到 Demo Day 时已经做出一百万美元营收，还有的拿 A 轮时营收一千万美元，而团队人数只有过去的零头。到底是不是一千倍可以争，但至少是好几倍。一个真正会用这些工具的多面手，抵得上过去好多个人。结果就是整家公司的人数都会更少。尤其意味着一件事：不要中层管理。我们真心觉得，你不再需要一层又一层的总监和 VP 了。

[17:56] Tom

The key roles for me are the like the IC. Everyone has to be a individual contributor actually doing the work, coming to meetings with working prototypes, not decks. And DRI like DRIs is for me actually the biggest thing a lot of companies are missing, which is like a single person whose head is on the block to make something happen. As soon as you have a committee, I worked at a bank and we had so many committees for everything, and it grinds things to a halt. So, having a single person who's directly responsible for the outcome, I think these are the most two important the two most important roles. And I think middle management is gone. Practically speaking, um I would try to make This is, by the way, easier when you're starting out. I would make everything in your organization legible to Legible's like a really popular word in Silicon Valley right now. It means you can read it. Make all your data readable to AI. That means recording everything, transcribing literally every meeting. Um if you can't get access to Slack DMs, I would ban the Slack DMs. Make everything um in public channels the AI can access. Every action needs to create an artifact. It needs to create a uh written or recorded version, otherwise it basically didn't happen to the AI.

在我看来关键的角色就两个。一个是 IC（individual contributor，一线执行者）——每个人都得亲手干活，开会时带来的是能跑的原型，不是一堆 PPT。另一个是 DRI（直接责任人）。DRI 恰恰是很多公司最缺的东西：有一个人把脑袋押上去，负责让这件事成。一旦变成委员会制就完了——我在银行待过，什么事都有一堆委员会，最后就是一切都停摆。所以必须有一个人对结果直接负责。这两个角色我认为是最重要的。而中层管理，我觉得会消失。具体怎么做——顺带说一句，越早期越好办——我会把组织里的一切都做成「可读的（legible）」。legible 这个词现在在硅谷特别火，意思就是能被读取。把你所有的数据都变成 AI 读得懂的。这意味着什么都要录下来，每一个会都要转写成文字。如果你没法读取 Slack 私聊，那我干脆就禁掉 Slack 私聊，所有沟通都放在 AI 能访问的公开频道里。每一个动作都必须留下产物——一份写下来的或录下来的版本，否则对 AI 来说，这件事根本就没发生过。

[19:05] Tom

Ob- obviously, I think on-demand internal software for all internal in uh all company operations is um absolutely necessary. A really cool thing we've just started doing, which I don't think we've spoken about yet, um, is, uh, and it's inevitable it comes out anyway, so, um, obviously you record all your, um, sales meetings, right? And you can feed those sales meetings back into an AI to be like, how did I do with my sales meeting? What if you also recorded all your meetings with investors? And you took those investor calls and you had access to several hundred or thousand companies that are all meeting all the same investors every 3 months. And the first thing you can do is tell founders, here's what's going wrong with your pitch. You spent 90% of the time talking or you really fumbled this question or you need to be much more concise on your go-to-market. Very, very easy to do and many of our founders are taking 40 or 50 or 60 investor calls. You just pump them in and the AI will, if you're recording them, tell you what's going well and what's going badly. But we also get data on the other side of that conversation. These founders are meeting the same investors. So, investor X will definitely, here's how like, we could simulate the investment call, right? We could say, here's exactly how Andreessen Horowitz does every single investment call. Whereas, Founders Fund will ask you this question, this question or this particular partner will always say, what's your wedge or what is like, what's your motive Anthropic comes after this?

另外很显然，公司所有内部运营都用「随取随做」的内部软件来跑，这是绝对必要的。我们最近刚开始做一件挺酷的事，好像还没对外讲过——反正早晚也会传出去。你肯定会把所有销售会谈录下来，对吧？然后你可以把这些销售录音喂回给 AI，问它：我这场销售谈得怎么样？那如果你把跟投资人的会也全录下来呢？再把这些投资人通话汇总起来——而你手上有几百上千家公司，他们每三个月都在见同一批投资人。第一件能做的事，是告诉创始人：你 pitch 的问题出在这儿。你 90% 的时间都在自己讲；或者这个问题你答砸了；或者你的 go-to-market 得说得精炼得多。这事非常非常容易做，而我们很多创始人一轮要打 40、50、60 场投资人电话。只要你录了音，全丢进去，AI 就能告诉你哪些地方做得好、哪些地方做得糟。但我们同时还拿到了这场对话另一侧的数据。这些创始人见的是同一批投资人。所以投资人 X 一定会怎样怎样——我们甚至可以模拟一场投资会谈，对吧？我们可以说：Andreessen Horowitz 的每一场投资会谈到底怎么走，一清二楚；而 Founders Fund 会问你这个问题、那个问题；某位特定的合伙人一定会问：你的切入点（wedge）是什么？或者——如果 Anthropic 也下场来做这件事，你的护城河在哪？

[20:32] Tom

And here's a great way to answer that. And I'm sure I'm I don't feel bad about doing this at all because I absolutely know that those investors are also recording this and pumping into their AI.

然后再告诉你：这个问题这么答最漂亮。我对这么干一点心理负担都没有，因为我百分之百确定，那些投资人也在录音，也在把内容灌进他们自己的 AI 里。

[20:41] Tom

[snorts]

[笑声]

[20:42] Tom

And so, anyway, basically, if you can record it, if you can make it legible, then the AI can start to comprehend it and give you superpowers. I will finish on this. If you are starting a company today, would you build it like this? And for, this was a talk to very early stage YC founders. Most of you are even earlier stage and you're small enough to just build it right from the start. So, anyway, I'll finish there. I think we have, oh, no, we'll just do questions with all of us, I think, rather than me taking questions on my own. Why did we do that? Thank you.

总之，基本逻辑就是：只要你能把它录下来、能让它变成可读的，AI 就能开始理解它，然后给你超能力。最后我用这句话收尾：如果你今天要创业，你会照这个方式来搭公司吗？这场分享原本是讲给非常早期的 YC 创始人听的，而你们大多数人甚至更早期，团队小到完全可以从第一天就把它搭对。好，我就讲到这儿。我记得我们还有——哦不对，等会儿是我们几个人一起答问，而不是我一个人接问题。当时为什么这么安排来着？谢谢大家。

[21:15]

[applause]

[掌声]

[21:19] Tom

Thank you.

谢谢。