BuildSpeak每日 builder 文摘
今日归档生词本关于
🎙 播客The MAD Podcast with Matt Turck· 2026 年 5 月 28 日· 14,540 词 · 约 73 分钟

State of Enterprise AI 2026: Aaron Levie on Tokenmaxxing, Rise of Headless, and AI-Proofing Your Job

SPACE 播放 / 暂停·←→ 上一句 / 下一句
Speaker 100:00 - 00:32
Let's just say you took GBT 5.5 or Opus, and you just snap the line right now. You could probably do this diffusion in, like, two years or three years, and we could probably all do the change management, like, collectively as an ecosystem. The problem is is the breakthroughs keep happening faster than the customer can implement any kind of standard architecture. And those breakthroughs oftentimes make obsolete the last thing you implemented. It's this really bittersweet thing, which is like the technology is getting so advanced that it makes obsolete the prior thing that you implemented, which actually means that the rollout takes longer.
Speaker 100:00 - 00:32
可以这么说:假如你现在拿到 GBT 5.5 或 Opus,然后立刻拍板定方案,你大概也许能在两三年内把这种 diffusion(扩散式)能力落地进去,而我们作为整个 ecosystem(生态系统)也许也能集体完成这些 change management(变更管理)。问题在于,突破发生的速度一直比客户实施任何一种标准架构的速度还快。而这些突破往往又会让你上一次刚实施的东西变得过时。这是一种很苦乐参半的情况:技术进步太快了,快到会让你之前部署的东西失效,而这实际上反而意味着整体 rollout(推广落地)会花更长时间。
Speaker 200:32 - 00:50
Hi. I'm Matt Turf from Firstmark. Welcome to the Mad Podcast. Today, I'm excited to welcome back Aaron Levy, CEO of Box. Aaron is hands down one of the most deeply thoughtful CEOs in tech today when it comes to agentic AI, and he has a front row seat to how AI is actually being deployed inside the world's largest enterprises.
Speaker 200:32 - 00:50
大家好,我是来自 Firstmark 的 Matt Turf,欢迎来到 Mad Podcast。今天我非常高兴再次请到 Box 的 CEO Aaron Levy。说到 agentic AI(代理式 AI),Aaron 毫无疑问是当今科技行业里思考最深入的 CEO 之一,而且他正处在第一排,直接观察 AI 实际上是如何在全球最大型企业内部被部署的。
Speaker 200:50 - 01:14
In this episode, we get into the most pressing topics in enterprise AI right now, including soaring token costs, why AI model progress is paradoxically slowing down enterprise deployment, the rise of headless software, the emergence of internal FDEs, and where startups can still win. Please enjoy my conversation with the always excellent Aaron Levy. Hey, Aaron. Good to see you.
Speaker 200:50 - 01:14
这一期里,我们会聊到当前 enterprise AI(企业 AI)最紧迫的话题,包括飙升的 token 成本、为什么 AI model(模型)的进步反而会悖论式地拖慢企业部署、headless software(无头软件)的兴起、内部 FDE 的出现,以及创业公司还能够在哪些地方胜出。请大家欣赏我与一如既往非常出色的 Aaron Levy 的这场对话。嘿,Aaron,很高兴见到你。
Speaker 101:14 - 01:16
Hey. Good to, good to be back. Appreciate it.
Speaker 101:14 - 01:16
嗨,很高兴,很高兴再次回来。感谢邀请。
Speaker 201:16 - 01:46
Yeah. Thanks for running it back with us. I feel you you play a very special role and have a very special position in our tech ecosystem because on the one hand, you've been a Silicon Valley insider for a while now, and, these days, you're as agent built as it gets. But on the other hand, you're public company CEO and your company sells to the largest, enterprises in the world. So the the Gs and the Procter and Gamble and Morgan Stanley's actually feels like a good place to start.
Speaker 201:16 - 01:46
是的,感谢你再次来和我们聊。我觉得你在我们的科技生态里扮演着一个非常特殊的角色,也处在一个非常特殊的位置,因为一方面,你已经做了很久的 Silicon Valley 圈内人,而现在的你几乎可以说是最彻底的 agent built(以 agent 为核心构建)那一类人了。但另一方面,你又是一家上市公司的 CEO,而且你的公司服务的是全球最大的 enterprises(企业)。所以,从 GE、Procter and Gamble、Morgan Stanley 这些公司开始聊,感觉会是个很好的切入点。
Speaker 201:46 - 01:55
How wide is the gap these days between Silicon Valley and the Bay Area and the telco system in general and the global 2,000 type of enterprises?
Speaker 201:46 - 01:55
如今,Silicon Valley 和 Bay Area 与更广义上的整个科技体系之间,以及 Global 2000 这类大型企业之间,这种差距到底有多大?
Speaker 101:55 - 02:25
Interestingly, we've always sort of played this role and and like if I had to distill like maybe the singular concept we've always tried to think about is basically our job is take super advanced technology breakthroughs and bridge them to the real world. So at the very early stages of cloud computing, it was like, oh, we can now move like infinite storage into the cloud. Well, to get that in the hands of most businesses, you would need a simple interface and you need advanced security. So we've always sort of been this kind of bridge. And it lets us kind of have a foot in both worlds.
Speaker 101:55 - 02:25
很有意思的是,我们其实一直都在扮演这样一种角色。而如果一定要把我们长期以来的核心理念提炼成一个单一概念,那基本上就是:我们的工作是把超级先进的技术突破,连接到现实世界。比如在 cloud computing(云计算)的 very early stages(非常早期),当时像是:哦,我们现在终于可以把近乎无限的存储搬到云上了。可如果要把这件事真正交到大多数企业手里,你就需要一个简单的 interface(界面),还需要先进的 security(安全能力)。所以我们一直都在充当这样一座桥梁。这也让我们能够同时踏在两个世界里。
Speaker 102:25 - 02:55
One world being like the super advanced, kind of everything is moving a million miles a second. And then the rest of the world, which is like there's change management and there's, you know, systems that have to be upgraded and all that. So we saw it with that with the cloud. And, and now we're definitely seeing another version of that with, with AI again, I think there's one tiny little asterisk, which is, which is it's sort of not just Silicon Valley versus everybody else. It's it's sort of Silicon Valley engineering versus everybody else.
Speaker 102:25 - 02:55
一个世界是那种超级前沿、几乎所有事情都以每秒百万英里的速度在前进的世界。另一个则是现实世界:那里有 change management(变更管理),有必须升级的 systems(系统),以及诸如此类的各种问题。所以我们当年在 cloud(云)上就看到了这种情况。而现在,随着 AI,我们显然又一次看到了它的另一个版本。不过我觉得这里有一个小小的 asterisk(补充说明),那就是,这其实不只是 Silicon Valley 和其他所有人的对比,更像是 Silicon Valley engineering(Silicon Valley 的工程体系)与其他所有人的对比。
Speaker 102:55 - 03:33
And, and then engineering, though, to be clear, I would say, you know, like, if you're tapped into AI right now, and you're at one of those names of companies, and you're in engineering, you probably could look fairly similar to Silicon Valley. So the bigger question is like Silicon Valley versus sort of non engineering knowledge work. And that's like the big question right now, which is which is what is this sort of path from AI coding agents that we know have totally, you know, reached escape velocity to now agentic work in the in the rest of the organization? And, what does that rollout look like? What are the use cases going to be?
Speaker 102:55 - 03:33
不过先说清楚,就 engineering 而言,我会这么说:如果你现在已经深度接入 AI,而且你在那些头部公司之一做 engineering,那你的状态大概率已经和 Silicon Valley 相当接近了。所以更大的问题其实是:Silicon Valley 与某种意义上的非 engineering 知识型工作之间的对比。眼下真正的大问题就是:我们已经知道 AI coding agents(AI 编码 agent)显然已经完全达到 escape velocity(自我加速扩张的临界点),那么从这里走向组织其余部分的 agentic work(agent 驱动的工作),这条路径会是什么样?这种 rollout(推广部署)会怎么展开?use cases(使用场景)又会是什么?
Speaker 103:33 - 04:06
How do these get implemented? So needless to say, this is the number one conversation every single customer that we talk to has at this point. And I think I have a decent kind of sample size, probably a couple 100 CIOs just this year across the kind of fortune 500 global 2,000 type of cohorts. And I would say this is the singular conversation dominating engagement that we have with an enterprise. It's the main and primary thing that every enterprise is trying to figure out.
Speaker 103:33 - 04:06
这些东西要怎么落地实施?不用说,这已经是我们现在与每一位客户交流时排名第一的话题了。我觉得我的样本量还算可以——仅今年一年,在 Fortune 500、Global 2000 这一类 cohort(企业群体)里,我大概就接触了几百位 CIO。我会说,这是我们与企业互动中唯一一个占据主导地位的话题。它是每一家 enterprise(企业)都在努力搞清楚的最主要、最核心的事情。
Speaker 104:07 - 04:40
We are still incredibly early. So that's like the probably the really big TLDR. And what's interesting and the reason why we're early, even though it's like, okay, well, we've been in the AI wave for three to four years, let's say, is everybody just started figuring out, like, their final rollout plans for, the chat system in the enterprise. And people have, you know, very appropriately been, you know, proud and happy that they finally got that thing out. And that's still rolling out in some organizations to be clear, like we still have five years of growth of like the chat system for knowledge work.
Speaker 104:07 - 04:40
我们仍然处在极其早期的阶段。所以这大概就是最重要的 TLDR(太长不看版结论)。有意思的是,我们之所以还很早期,尽管看起来会觉得——好吧,假设 AI 这一波已经持续了三到四年——是因为大家其实才刚刚开始弄清楚自己在 enterprise(企业)内部对 chat system(聊天系统)的最终 rollout plan(部署方案)。而且人们也完全有理由为此感到自豪和高兴,因为他们终于把那套东西推出来了。需要说明的是,在一些组织里,这件事现在还在 rollout(持续推广)中;就知识型工作而言,光是 chat system 的增长,我们都还有五年的发展空间。
Speaker 104:40 - 05:10
But right as that happens, then the capability, you know, has extended even further. And so then everybody's sitting around saying, okay, as we move from chat, which is like, I asked a question, I get an answer back. And so the productivity gain is sort of, you know, rate limited by the human's ability to like, have a conversation. Now I'm actually want to go deploy an agent that's going to be doing things and producing real work in the enterprise, maybe handling tasks, maybe maybe you kick it off via chat, or maybe it's just running in a stateful way. You know, and I and I'm pinging it, or it's kicking off in a workflow.
Speaker 104:40 - 05:10
但就在这件事发生的同时,这种 capability(能力)又进一步延伸了。于是所有人都在想:好,现在我们要从 chat 走出去——也就是我提一个问题,它返回一个答案,因此生产力提升某种程度上还受限于人类发起对话的能力——转而真正部署一个 agent,让它在 enterprise(企业)里做事、产出真实工作,也许是处理任务,也许你是通过 chat 来触发它,也许它只是以一种 stateful(有状态)的方式持续运行。可能是我去 ping 它,也可能是它在某个 workflow(工作流)里被触发启动。
Speaker 105:10 - 05:34
So now everybody is sort of saying, Okay, we think we know how the chat thing works. And by the way, even though they know how that works, I'd say that it's a still changing dynamic in market share, you're having a lot of changes of what people are rolling out there. Now everybody's saying, we're gonna go deploy agents. These are gonna be much more advanced, much more capable. And that's like the big conversation I would say, like, we're we're in day day one of that, as an industry.
Speaker 105:10 - 05:34
所以现在大家都在说:好,我们大致知道 chat 这件事是怎么运作的了。顺便说一句,即便他们知道怎么运作了,我还是会说,market share(市场份额)的变化态势依然在持续演变,人们实际在部署什么也还在发生很多变化。现在所有人都在说,我们要去部署 agents。这些东西会先进得多,能力也强得多。而我会说,这才是真正的大话题;作为一个行业,我们在这件事上其实才刚刚进入 day one(第一天)。
Speaker 205:34 - 06:12
Yeah. How would you characterize the mood? It's interesting that you mentioned chats because in the conversations I have and my my simple size is smaller than yours. But, you know, there's at least a part of the Global two thousand crowd that that would say something like yeah you know oh yeah AI two years ago you guys came to us and we already needed to do this chat thing and it was super urgent and we're going to fall behind and it was like a dub chat or you're gonna die And then we did a pilot and then it really worked out. So now you're coming back to me like two or three years later and you say no no no no no agent is a thing and like this time if you don't do it you're gonna you're gonna die.
Speaker 205:34 - 06:12
对。你会怎么形容现在的情绪?你提到 chats 很有意思,因为在我自己的对话里——当然,我的样本量比你小——但你知道,Global 2000 这批企业里至少有一部分人会这么说:是啊,两年前你们来找我们,说我们已经必须得做这个 chat 的事了,而且特别紧急,不然我们就会落后,感觉像是“做个该死的 chat,不然你就完了”。后来我们做了 pilot(试点),结果也确实不错。所以现在你们两三年后又回来跟我说,不不不不不,agent 才是关键,而且这次如果你不做,你就又要完了。
Speaker 206:12 - 06:18
In the spectrum between skeptical and enthusiastic, where would you put the mood in in large enterprises?
Speaker 206:12 - 06:18
如果把态度放在 skeptical(怀疑)和 enthusiastic(热情)这个光谱上,你会把大型 enterprise(企业)目前的情绪放在什么位置?
Speaker 106:19 - 06:55
I would say if did like the broadest sample, I think the mood would veer statistically more optimistic than maybe the framing that I think maybe you landed on like a few extra cynical CIOs in that. I think I think because what's happening is, is the CIO and our main audience is the CIO. When we talk to the CIO, they know that their engineering teams are using Cloud Code and Codex and Cursor, and they're seeing the productivity gains come out of those teams. And they're like, yeah, like my teams are just building apps way faster. They're being able to tackle IT projects much more quickly.
Speaker 106:19 - 06:55
我会说,如果看最广泛的样本,我觉得整体情绪在统计意义上会更偏向乐观,而不是你刚才那个框架里呈现出的感觉;我猜你可能是碰到了几个格外 cynical(愤世嫉俗)的 CIO。我这么想是因为,CIO——而我们的主要受众也正是 CIO——当我们和 CIO 交流时,他们知道自己的 engineering 团队在使用 Cloud Code、Codex 和 Cursor,也看到了这些团队释放出来的生产力提升。他们会说:对,我们的团队现在构建 app 的速度就是快了很多,处理 IT project(IT 项目)的速度也快得多。
Speaker 106:55 - 07:28
Like we're doing security reviews faster. Like they're seeing the productivity gains in their function. And I think they're often saying, how do I bring those same IT productivity gains to the non IT parts of the organization? And they're having the business pull them and say, I want access to co work, I want access to Codex, I want access to these tools as well. So there's actually a certain kind of sex appeal to these tools right now, where the business is sort of demanding, I want to be on the agentic train, because I'm seeing all these great use cases.
Speaker 106:55 - 07:28
比如说,我们做 security review(安全审查)的速度更快了。也就是说,他们正在自己所在的职能里看到生产率提升。我觉得他们常常会问:我怎么把这些 IT 领域的生产率提升带到组织里那些非 IT 的部分?而且业务部门也在反过来推动他们,说我想用 co work,我想用 Codex,我也想用这些工具。所以现在这些工具确实带着某种“吸引力”,业务部门有点像在主动要求:我想搭上 agentic(代理式)这班车,因为我已经看到了这么多很棒的 use case(用例)。
Speaker 107:28 - 08:11
And so I think the tone is actually remarkably optimistic and excited and positive as opposed to, you know, there's a sort of, you know, typical trough of disillusionment, you know, from Gartner and the hype cycle or whatnot, I think people are eyes wide open on, there's no free lunch in this, it's not going to just immediately transform our productivity. Like I'm we're no longer in that part of the conversation. And maybe we weren't like two to three years ago, I think everybody is sort of like very firmly aware, there's not thing doesn't just get deployed and magically goes and transforms the business. But at the same time, they're having their employees say, actually, I would like this thing to be able to go review all my documents for me. I would like to go accelerate our client onboarding process.
Speaker 107:28 - 08:11
所以我觉得,整体氛围其实是格外乐观、兴奋和积极的,而不是那种——你知道的——典型的“幻灭低谷”,也就是 Gartner 的 hype cycle(技术炒作周期)里常说的那一段。我认为大家现在都很清醒地意识到,这里面没有免费的午餐,它不会立刻就把我们的生产率彻底改变。我们现在已经不再停留在那种讨论阶段了。也许两三年前我们还在那个阶段,但现在我觉得每个人都非常明确地知道:这东西不是部署下去就会神奇地自动改造整个业务。但与此同时,他们的员工也在说,实际上,我希望这东西能替我审阅我所有的文档;我希望它能加快我们的客户 onboarding(入驻)流程。
Speaker 108:11 - 08:32
I would like to be able to generate digital assets in our marketing campaign process. So I think the demand is coming from the business and and the IT organization is seeing the gains happening in coding. And now it's a bit more of a like 30 just practical tactical, you know, issues, I'm sure we'll get into some of them. It's like, it's the token cost thing. It's the how do you actually roll this out?
Speaker 108:11 - 08:32
我希望它能在我们的营销 campaign(活动)流程中生成 digital assets(数字素材)。所以我觉得,需求是从业务部门那边来的,而 IT 组织已经看到了 coding(编码)方面的收益。现在更多变成了一些非常实际、战术层面的问题,我们肯定接下来会谈到其中一些。比如 token 成本问题;比如你到底该怎么把这套东西真正 rollout(推广部署)出去。
Speaker 108:32 - 08:51
It's, do you have the talent to go and deploy these things? So, so I'm finding the conversations to be fairly positive and, increasingly ambitious, but with a sense and strong dose of reality of none of this stuff is coming for free, you know, beyond even the cost side, but like free in a deployment, you know, sort of manner.
Speaker 108:32 - 08:51
再比如,你有没有足够的人才去部署这些东西?所以,我现在感受到的讨论总体上都相当积极,而且目标也越来越有雄心,但同时又带着一种现实感,而且是很强烈的现实感:这里面没有任何东西是“免费”得到的——甚至不只是成本层面的免费,而是从 deployment(部署)方式上来说,也绝不是轻轻松松就能完成的。
Speaker 208:51 - 09:12
Since you mentioned the token cost, let's get into it. It's such an interesting topic. And it's also a really timely topic as as we're recording this. There was news that Microsoft canceled their internal cloud code licenses this week after token based billing made the cost untenable. And I guess a couple weeks ago, Uber CTO was talking about the same kind of thing.
Speaker 208:51 - 09:12
既然你提到了 token 成本,我们就来聊聊这个吧。这是个特别有意思的话题。而且在我们录这段对话的时候,它也恰好非常应景。本周有消息说,Microsoft 取消了他们内部的 cloud code license(云端代码许可证),因为按 token 计费之后,成本高到了无法承受。我记得几周前,Uber 的 CTO 也在谈同样性质的问题。
Speaker 209:12 - 09:15
So that there's definitely a topic that seems to be emerging in large enterprises.
Speaker 209:12 - 09:15
所以这显然是一个正在大型企业里浮现出来的话题。
Speaker 109:15 - 09:23
Yeah. Although although to be fair to be fair, I think the Microsoft one probably got spun in some interesting ways because there's probably much more of a reflection on they want to move to Yeah.
Speaker 109:15 - 09:23
对。不过公平地说,公平地说,我觉得 Microsoft 那件事可能被用一些很有意思的方式解读和传播了,因为那可能更多反映的是,他们想转向——对。
Speaker 209:24 - 09:25
Code Code versus Codex. Yep. Yep.
Speaker 209:24 - 09:25
Code Code,而不是 Codex。对,对。
Speaker 109:25 - 09:36
Yeah. It's like they're going to be spending on the tokens. Like, it's literally going to ride on their infrastructure. So, so, but but, so I I think the the press really likes to, to modify the the the sort of story on that one.
Speaker 109:25 - 09:36
对。就像他们会把钱花在 token 上一样。也就是说,这实际上就是要跑在他们的基础设施上。所以,但是,但是,所以我觉得媒体真的很喜欢去改写、调整这件事的叙事。
Speaker 209:36 - 09:54
Yeah. Very, very fair. It does seem to be a topic, that's top of mind, though. And, like, little bit that that tension between, like, the ticket system Silicon Valley where, like, token maxing is is really a thing whereas enterprises are worried about costs. So what do you hear and what do you recommend people do, your customers?
Speaker 209:36 - 09:54
对。这个说法很公平,非常公平。不过,这确实似乎是大家最关心的话题之一。而且,这里面还有一点那种张力:一边是 Silicon Valley 那套“入场券”式系统,在那里 token 最大化(token maxing)确实很流行;另一边是企业更担心成本。所以你听到的情况是什么?你会建议别人怎么做,或者说你的客户该怎么做?
Speaker 109:55 - 10:31
I don't know if I have good recommendations yet actually, but I would say the token, when we go and talk to organizations right now about they are with agents, tokens, the cost of tokens and budgeting and budget planning and all of this probably is at least one third of the hottest button issues that relate to AI. And it might even be tied for number one half the time. Because what they've seen is this move from everybody's sort of calling it like we were doing subsidization as an industry. I don't really think about it like that. I would say that the costs were just low enough that these things were included.
Speaker 109:55 - 10:31
说实话,我现在还不知道自己是否已经有很好的建议,但我会说,我们现在去和各个组织讨论 agent、token、token 成本、预算和预算规划这些事情时,这大概至少占到所有和 AI 相关热点问题的三分之一。而且有一半的时候,它甚至可能并列第一。因为他们看到的是一种转变——大家都把它称为整个行业之前一直在做补贴(subsidization)。我自己其实不完全这么看。我会说,当时的成本只是低到足以让这些东西被直接包含进去而已。
Speaker 110:31 - 11:17
Like, Cursor just included, you know, a lot of usage and maybe with subsidization, but it was actually just like they could model that under their subscription fee, you know, in a in a in a fairly clean way. And then all of a sudden what happened was these agents just can do way more work, their context windows are way larger, the cost of inference is way more, because they have way more parameters and their capabilities way better. So so we've just gone from, you know, like, a pricing model of a chatbot or like type ahead functionality in GitHub Copilot to that pricing model no longer working when one coding agent could be consuming, $1,000 of compute on a single task. So clearly, like you can't lump that all into a $20 per user per month fee. So that's really the jump that's happened.
Speaker 110:31 - 11:17
比如 Cursor 以前就包含了大量用量,可能某种程度上也有补贴,但更准确地说,其实是他们可以把这件事相当干净地纳入自己的订阅费模型里。然后突然之间发生的事情是,这些 agent 能做的工作多了太多,它们的 context window(上下文窗口)大了很多,inference(推理)成本高了很多,因为它们的参数多得多,能力也强得多。所以,我们其实是从 chatbot 那种定价模型,或者像 GitHub Copilot 里的 type-ahead(输入联想)功能那种定价模型,走到了一个新阶段:当一个 coding agent 在单个任务上就可能消耗 1,000 美元的算力时,原来的定价模型就不成立了。很显然,你不可能再把这一切都塞进每用户每月 20 美元的收费里。所以这才是真正发生的跃迁。
Speaker 111:17 - 11:34
And it's all only happened in one year, maybe less than a year. I mean, it basically all correlates to you can just look at like the Andropic revenue curve and that is the period of time where everything has sort of been flipped on its head on the sort of cost modeling and token budgeting sides.
Speaker 111:17 - 11:34
而且这一切都只是在一年内发生的,甚至可能还不到一年。我的意思是,这基本上完全对应着——你只要去看一下 Andropic 的营收曲线就行——也就是在那段时间里,关于成本建模和 token 预算的一切,几乎都被彻底颠覆了。
Speaker 211:34 - 11:55
Yeah, there is an increase in the per token cost for Frontier token though. Right? So not not just longer running agents using more tokens, also the the actual cost of the I think the Frontier tokens, has increased, which is completely opposite from the narrative that we were all telling one another in the last couple of years, that the cost of token was always going down?
Speaker 211:34 - 11:55
对,不过 Frontier token 的单 token 成本确实上升了,是吧?也就是说,不只是运行时间更长的 agent 在消耗更多 token,实际上的 Frontier token 成本本身也提高了。我觉得这和过去几年我们一直互相讲述的那套叙事完全相反:token 的成本不是应该一直在下降吗?
Speaker 111:55 - 12:23
Yes, a 100%. But I think the one nuance is I'm not, I mean, it'd be good to get obviously some of the lab folks on. I'm not a 100% sure it's, it's just the subsidization, change as much as like, no, like these models are just way bigger and, and, and they're, you know, the hardware is, is not getting any cheaper anytime soon. And we have a capacity constraint. So you've got like, you've got like a few atypical patterns from, from normal computing, which is like, usually like there's economies of scale and you don't have the same kind of shortage.
Speaker 111:55 - 12:23
对,百分之百是这样。但我觉得这里有一个细微之处:我不是——当然,最好还是请一些 lab 的人来谈——我也不是百分之百确定,这更多只是补贴变化的问题;更像是,这些模型本身就是大得多,而且硬件短期内也不会变得更便宜。再加上我们还面临产能约束。所以你会看到一些不同于常规计算的反常模式:通常来说,会有规模经济,而且你不会遇到同样类型的短缺。
Speaker 112:23 - 12:50
And so then as you build out, everything gets cheaper, and you have Moore's Law. And like, we've compressed, you know, normally what should happen in like ten years of rollout into like eighteen months. And so lo and behold, you know, the data center providers, the labs, etc, have pricing power, they don't need to lower the prices on anything. So you're not seeing the typical things that drive down the cost of compute. I'm highly optimistic that happens over the next five to ten years, but it's just clearly not happening yet.
Speaker 112:23 - 12:50
所以按理说,随着建设规模扩大,一切都会变得更便宜,而且还会有 Moore's Law(摩尔定律)在起作用。但我们其实是把本来通常应该在十年 rollout(铺开)过程中发生的事,压缩到了大约十八个月里。于是很自然地,data center 提供商、各家 lab 等等就拥有了定价权,他们没有必要下调任何东西的价格。所以你现在看不到那些通常会压低 compute(算力)成本的因素。我非常乐观地认为,这种情况会在未来五到十年内出现变化,但很显然,现在还没有发生。
Speaker 112:50 - 13:32
So we're not seeing the curve that you should see in a normal like ten year cycle of compute, because that ten years is now happening in twelve months. So what's happening is enterprises are saying, okay, you know, I'm quite surprised by these bills. And, and it's like, it's a surprise, that is sort of like, it's an uncomfortable acceptance surprise, as opposed to like, I'm not doing this anymore surprise, because they're, they're like empirically getting the productivity gains, or they just wouldn't be paying the bills. It's just, it's just now they are saying, Oh, okay, this is a very real expense in the business. This is not the kind of expense that that we just sort of add on $20 per user in our in our headcount.
Speaker 112:50 - 13:32
所以我们并没有看到正常十年 compute(算力/计算)周期里你本该看到的那条曲线,因为那十年现在会在十二个月里发生。所以企业现在的反应是:好吧,说实话,这些账单让我很意外。而且这种意外,更像是一种不太舒服但不得不接受的意外,而不是“我再也不用这个了”的那种意外。因为他们确实在经验上得到了生产力提升,否则他们根本不会付这些账单。只是现在他们开始意识到:哦,原来这在业务里是一项非常真实的支出。这不是那种我们只是在脑子里按每个用户多加 20 美元、算进 headcount(人头成本)里的费用。
Speaker 113:32 - 14:10
And now it has solved the problem. So one of the nuance sort of shift that's going to happen is, I think the first two to three years of AI, the IT budget could kind of consume the AI costs. And this would show up as, okay, the company upgraded to Microsoft Copilot, or they added, they did the add on of the AI product of XYZ vendor, or they could kind of get the cursor licenses within the IT spend. And, know, as you know, like IT spend is basically somewhere between like 3% to 7% of corporate of revenue in a company. Sometimes lower, sometimes higher, but like it's kind of trapped at that.
Speaker 113:32 - 14:10
而现在它已经解决了这个问题。所以接下来会发生的一个细微但重要的变化是:我认为在 AI 的头两到三年里,IT budget(IT 预算)某种程度上还能消化 AI 成本。这种情况会表现为:公司升级到了 Microsoft Copilot,或者给某个 XYZ vendor 的 AI 产品买了 add-on(附加功能),又或者他们还能把 Cursor 的 license(许可证)塞进 IT spend(IT 支出)里。你也知道,IT spend 基本上通常占公司 revenue(营收)的 3% 到 7%。有时更低,有时更高,但大体上就是被卡在这个区间里。
Speaker 114:10 - 14:51
So then the question is like, well, where's the other 60%, 70%, 80% of revenue in an organization? It's OpEx and it's just like general purpose OpEx across the business. And so if AI is truly adding this productivity gain to your engineering team or your client onboarding process or your marketing team, then clearly you don't want to be trapped by this sort of 3% to 7% in the business, it's to escape that and it's going to move to the line of business budgets. And so this is actually, on one hand, it's actually good for the AI industry, because you're no longer going to be constrained by IT spend budgets in an organization. On the other hand, you have all these now interesting downstream questions, which is like the line of business doesn't necessarily know how to budget for compute.
Speaker 114:10 - 14:51
那接下来的问题就是:一个组织里另外那 60%、70%、80% 的 revenue 都在哪里?是在 OpEx(运营支出)里,也就是整个业务里的通用型 OpEx。所以如果 AI 真的在给你的工程团队、客户 onboarding(客户接入)流程,或者营销团队带来生产力提升,那显然你不会想被业务里这 3% 到 7% 的比例困住;你会想摆脱它,预算也会转移到各条业务线的预算里去。从一方面看,这对 AI 行业其实是好事,因为你不再会被组织内部的 IT spend 预算所限制。但另一方面,后面也会出现一堆有意思的新问题,比如业务线未必知道该怎么给 compute 做预算。
Speaker 114:51 - 15:25
This is not a they don't have FinOps for the marketing team. They don't have FinOps for the sales team. That was something that the cloud people had and the IT IT team could kind of go and be confined to. So I think what's going to happen is, first of all, what's interesting is that you're going to have this tussle between kind of like the finance team, the line of business, the IT team, that's going to be this interesting kind of, yeah, how do you triage all of this? You are going to, to some extent, to centralize the management of the IT systems, management of what do you procure.
Speaker 114:51 - 15:25
这不是他们原来就有能力处理的事;营销团队没有 FinOps(云财务运营),销售团队也没有 FinOps。那原本是 cloud(云)相关团队才有的东西,IT team 也基本可以把它限制在自己的范围内处理。所以我觉得接下来会发生的是,首先很有意思的一点是,你会看到 finance team、line of business(业务线)和 IT team 之间出现一场拉扯:到底这些事情该怎么分诊、怎么协调?在某种程度上,你还是会把 IT systems(IT 系统)的管理、采购什么东西的管理,继续集中起来。
Speaker 115:25 - 15:58
But then you also kind of have to decentralize the decision making of how to use these things. Because really like the CMO should decide, do they want to spend a million dollars of compute or do they want to spend a million dollars in doing marketing events or something else? Like that kind of can only come down to the business owner that is driving these decisions. So, and that is again, a new type of format of how do you manage a compute budget in your marketing budget and in your sales budget and in your global manufacturing budget. So that's a whole thing that now people have to go figure out.
Speaker 115:25 - 15:58
但与此同时,你又必须把“如何使用这些东西”的决策分散下去。因为说到底,应该由 CMO 来决定:他们是想花 100 万美元买 compute,还是花 100 万美元办 marketing events(营销活动)或者做别的事。像这种选择,最终只能由推动这些决策的业务负责人来做。所以这又变成了一种全新的管理模式:你要怎么在 marketing budget、sales budget,乃至 global manufacturing budget(全球制造预算)里去管理 compute budget(算力预算)。这整件事现在都得由大家重新摸索。
Speaker 115:58 - 16:39
One of the things that we don't have tooling for is like, how do you measure the ROI on the tokens? And it's like, it's, you know, I think it's kind of like absurd and hilarious to already be talking about ROI this early in the cycle. But I see it as actually it's like pretty practical, like, there are some things I could do on my computer right now that would cost the same amount of money as as the lunch that my company provides me. And I could do it, I could press one button, and it could cost me the free lunch that I get. So clearly, a company is not going to be like, oh, let's just deploy a whole bunch of tools that people can just like willy nilly press a bunch of buttons, and you know, have the equivalent of 10 lunches, you know, in ten seconds, without knowing like, you doing something that actually like produce value for the organization?
Speaker 115:58 - 16:39
我们目前还没有的一类 tooling(工具)是:怎么衡量 tokens(token 用量)的 ROI(投资回报率)?你知道,我觉得在这个周期这么早的时候就开始谈 ROI,某种程度上既荒诞又好笑。但我也认为这其实非常实际:我现在在电脑上就能做一些事,而它花掉的钱可能和公司给我提供的一顿午餐差不多。我只要按一个按钮,就可能花掉我那顿“免费午餐”的成本。所以很显然,公司不可能说:“好,我们部署一大堆工具,让大家随便按按钮”,然后十秒钟之内就烧掉相当于 10 顿午餐的钱,却还不知道你做的事到底有没有真正为组织创造价值。
Speaker 116:40 - 17:09
And that's something that like nobody has tooling for. Employees don't actually really know what what the cost of compute is. So so they're going to go about using these systems as freely as possible, not knowing that yeah, this is actually that that one little task you gave to that agent could cost $200 because you just happened to structure the query wrong. And and now it's going to go fan out across a bunch of systems. And it's going to read like each, you know, each document or each piece of data in your system.
Speaker 116:40 - 17:09
而这正是目前没人有 tooling 的地方。员工其实并不真正知道 compute 的成本是多少。所以他们会尽可能自由地使用这些系统,却不知道:是的,你交给那个 agent(智能体)的一个小任务,可能就要花 200 美元,因为你刚好把 query(查询)结构写错了。然后它就会在一堆系统之间 fan out(扩散式展开),去读取你系统里的每一份文档、每一条数据。
Speaker 117:10 - 17:49
And then it's going to go and compute it all like that one structure of that of that prompt is the difference between, again, like your entire benefits, you know, for a month at that company. So how do we handle all this? I actually have no solutions. Like, it's going to be one of the most interesting questions and some mix of employee training, some mix of centralized capacity planning with decentralized sort of decisions of how do you roll that out. You're going to need new pieces of software, probably there's probably a $5,000,000,000 startup waiting to happen just in like ERP for your AI compute, which is just like, how do I decide that all of this stuff is being used in the right way?
Speaker 117:10 - 17:49
接着它还会把这些全都算一遍——而那个 prompt(提示词)的结构差异,可能就决定了你又一次花掉了相当于你在那家公司一个月全部福利成本的钱。所以我们该怎么处理这一切?其实我没有答案。这会成为最有意思的问题之一,最后大概会是几种方法的混合:一部分靠员工培训,一部分靠集中式的 capacity planning(容量规划),再配合分散式的“具体怎么落地”的决策。你还会需要新的软件,甚至很可能会出现一家估值 50 亿美元的 startup,专门做 AI compute 的 ERP——本质上就是:我怎么判断这些东西都在被以正确的方式使用?
Speaker 117:49 - 18:04
How do I measure value that it's being produced? How do I make sure that it rolls out to the right teams? So I think that's all up in the air right now. And I think this is so new that we're very short on best practices at the moment.
Speaker 117:49 - 18:04
我该如何衡量这些价值确实正在被产出?我又该如何确保它被推广到正确的团队?所以我觉得,这些问题现在都还悬而未决。而且我认为这件事太新了,我们目前非常缺乏 best practices(最佳实践)。
Speaker 218:04 - 18:08
Look at this free startup ideas right here on podcast. Thank you.
Speaker 218:04 - 18:08
你看,这不就是在 podcast 里直接送上的免费 startup 点子吗。谢谢。
Speaker 118:09 - 18:14
Wait. Is there, like, a do you have any royalty, approach to this or or how does this work?
Speaker 118:09 - 18:14
等等。所以这里有没有类似 royalty(版税 / 分成)的做法,还是说这到底是怎么运作的?
Speaker 218:14 - 18:29
We didn't, but we need to now. Starting now. We'll we'll share the referral fee Thank you. On on this. So you mentioned not subsidizing, but it seems that the labs are already starting to react or the OpenAI introduced pricing arrangement to just give more visibility into the pricing.
Speaker 218:14 - 18:29
之前没有,但我们现在需要了。从现在开始。我们会分享 referral fee(推荐费)。谢谢。关于这件事。你刚才提到不做 subsidizing(补贴),但看起来 labs 已经开始做出反应了,或者说 OpenAI 推出了一种 pricing arrangement(定价安排),只是为了让 pricing(定价)更透明一些。
Speaker 218:29 - 18:33
Do you think that's gonna be a part of how the industry evolves as well?
Speaker 218:29 - 18:33
你觉得这也会成为这个行业演进方式的一部分吗?
Speaker 118:33 - 19:13
In my long diatribe on on all the problems, I mean, a few things that will inevitably happen. So one thing that will inevitably happen is, OpenAI has a great program, which is kind of this dedicated capacity, which is okay, if you kind of know your workload, we're able to lock in certain pricing that helps support that. Like that's one way that you could kind of protect your costs. I think another thing that's going to happen is you're going see this divergence as opposed to, again, maybe two years ago, I would have predicted a convergence, but let's go with the opposite now. Frontier, you know, AI model capabilities get applied to coding and and advanced life sciences and and, like, your your contract process and your financial planning process.
Speaker 118:33 - 19:13
在我前面那段关于各种问题的长篇大论里,我是说,有几件事是必然会发生的。其中一件必然会发生的事是,OpenAI 有一个很不错的项目,有点像 dedicated capacity(专属容量);如果你大致了解自己的 workload(工作负载),我们就能锁定某些 pricing(定价),这会有助于支持那种安排。比如说,这就是一种可以保护你成本的方式。我认为另一件将会发生的事是,你会看到一种分化;而不是像两年前那样——那时我可能会预测会趋同——但现在我们还是按相反方向来看。Frontier(前沿)AI model(模型)能力会被应用到 coding、advanced life sciences(先进生命科学)、contract process(合同流程)以及 financial planning process(财务规划流程)这类场景中。
Speaker 119:13 - 20:01
And so that's where you apply, you know, GPT 5.5, high and Opus four, seven, and in, know, whatever the model. But then once you have a task that is sort of like, you know, you can now perform that task reliably, you can sort of, you know, that once that capability gets saturated, and you can perform that task in a reliable way, then you can peel that off to a lower cost model and sort of run that on an ongoing basis. And, and we just don't have a lot of maturity in doing that, because really, until maybe the past six, twelve months, you know, the models couldn't do any of our tasks that reliably. So as this starts to happen, you can kind of say, for that one customer service interaction, I can now cap that at, you know, 50¢ per million tokens, and it will never go higher than that. In fact, it'll only go lower because I might swap it out with, an OSS model, etc.
Speaker 119:13 - 20:01
所以,在这些地方你会用上,比如 GPT 5.5、high、Opus four、seven,以及别的什么模型。但一旦你遇到某个任务——就是说,你现在已经能够稳定地执行这个任务——当这种能力已经趋于饱和,而且你能够以可靠的方式完成这个任务时,你就可以把它剥离出来,交给一个成本更低的模型,并在后续持续运行。我们在这方面其实还没有太多成熟经验,因为说真的,直到过去大概六到十二个月,模型还无法那么可靠地完成我们任何任务。随着这种情况开始出现,你就可以说,对于某一次 customer service(客户服务)交互,我现在可以把它的成本上限定在每百万 tokens(token)50 美分,而且它绝不会高于这个价格。事实上,它只会越来越低,因为我可能还会把它替换成 OSS model(开源软件模型)之类的。
Speaker 120:02 - 20:24
But for my coding, I still actually want the highest capability. And so I think what's going to happen is you're going to have a mosaic of models in the enterprise. I think the average enterprise will certainly be using, half a dozen models in organization. You're not going to throw everything at the kind of Ferrari model from a performance standpoint. So companies will have to get better at that.
Speaker 120:02 - 20:24
但对于 coding,我其实仍然想要最高能力的模型。所以我认为将会发生的是,企业里会出现一种 model(模型)马赛克。我认为普通企业肯定会在组织内使用半打左右的模型。你不会把所有东西都扔给那种从性能角度看像 Ferrari 一样的模型。所以企业必须在这方面做得更好。
Speaker 120:25 - 20:52
So you'll need to have some kind of deeper wherewithal on how do you shift tasks to different levels of compute. We're going to have, again, new ways of measuring all this back to the kind of startup idea. There's some use cases I think companies will eventually kind of realize, oh, actually maybe like that's not something that I even need an agent for. It's like, I just need like software to get deployed into that. And actually software, you know, is actually cheaper because it's gonna just run on a CPU.
Speaker 120:25 - 20:52
所以你需要对如何把任务切换到不同层级的 compute(算力)上,有某种更深入的理解和能力。回到那种 startup idea(创业点子)的话题,我们还会有新的方式来衡量这一切。我认为公司最终会逐渐意识到某些 use case(使用场景):哦,实际上,也许这件事我甚至根本不需要 agent(智能体)来做。我只是需要把 software(软件)部署进去就行了。而且其实 software 更便宜,因为它就是跑在 CPU 上。
Speaker 220:52 - 20:55
Wait. Software software. That's still a thing that still exists?
Speaker 220:52 - 20:55
等等。Software,软件?这东西居然还存在?
Speaker 120:55 - 21:28
Yes, software, it turns out that maybe you don't want your agent to re render a UI every single morning. And that costs, know, dollars 30 per day of using the interface. So I think there's going to be a lot of mixed solutions on this, not to mention just like good competition in the market that says, you know what, why don't we have some cheaper models to get produced? You know, how do we how do we start to like, like, I think the market will sort of work as you'd expect, which is somebody will say, Oh, there's actually an innovation opportunity here and go attack, you know, certain, you know, parts of the market.
Speaker 120:55 - 21:28
对,software。事实证明,也许你并不想让你的 agent 每天早上都重新渲染一次 UI,而且那样光是用这个 interface(界面)一天就要花 30 美元。所以我觉得这里会出现很多混合式解决方案,更不用说市场里还会有良性的竞争,大家会说,你知道吗,为什么不做一些更便宜的 models(模型)出来?我们该怎么开始推动这件事——我觉得市场大体上会像你预期的那样运作:总会有人说,哦,这里其实有创新机会,然后就去进攻市场中的某些细分部分。
Speaker 221:28 - 21:41
All right. So we talked about the mood in the enterprise. We talked about the cost aspect. What else is happening in terms of barrier to progress, especially on the technical and product front? Like, do people need harnesses?
Speaker 221:28 - 21:41
好。那么我们谈了企业端的整体情绪,也谈了成本问题。除此之外,在阻碍进展这件事上,还发生了什么,尤其是在技术和产品层面?比如,人们需要 harnesses(编排/约束工具)吗?
Speaker 221:41 - 21:44
They need more vendors? They need more open source models? Why why what
Speaker 221:41 - 21:44
他们需要更多 vendors(供应商)?他们需要更多 open source models(开源模型)?为什么,到底需要什么
Speaker 121:44 - 21:49
do need? More vendors. They need way more vendors. The answer is always more vendors.
Speaker 121:44 - 21:49
他们确实需要。更多 vendors。他们需要多得多的 vendors。答案永远是:更多 vendors。
Speaker 221:49 - 21:50
VC backed vendors.
Speaker 221:49 - 21:50
VC backed vendors(有风投注资支持的供应商)。
Speaker 121:50 - 21:55
100100%. But but ideally subsidized VC backed vendors.
Speaker 121:50 - 21:55
百分之百同意。但理想情况下,最好还是由 VC 补贴的 VC backed vendors。
Speaker 221:55 - 21:57
So all great public companies. Yes.
Speaker 221:55 - 21:57
所以这些都是很优秀的上市公司。对。
Speaker 121:58 - 22:26
There was a there was a tweet. I mean, this come up by multiple times, which is like, like, write as much code as you humanly can right now, while some of these some of these products are still subsidized. And, and it's actually kind of like a funny concept. Because like, if you're really savvy, there's probably some parts of the market where you can be like, Oh, I could somehow use this LP capital to to do work for me as my startup. And and and there's like a window where you can find those exploits.
Speaker 121:58 - 22:26
之前有一条 tweet。我是说,这件事已经被提过很多次了,大意就是:趁现在这些产品里有一些还在被补贴,赶紧在人类能力所及的范围内尽可能多写代码。而且这其实是个有点好笑的概念。因为如果你真的很懂行,市场里大概确实有一些地方,你会觉得:哦,我好像可以某种程度上利用这些 LP capital 来替我的 startup 干活。而且确实存在这样一个时间窗口,让你能找到这些可利用的漏洞。
Speaker 222:26 - 22:33
Even to capital actually does have a utility in the in the world, like subsidizing Ubers and then subsidizing your tokens. You you welcome.
Speaker 222:26 - 22:33
说到底,capital 在这个世界上其实也确实有它的用途,比如补贴 Uber,再比如补贴你的 token。不客气。
Speaker 122:35 - 23:24
So while the hottest topic might be like tokens right now, just because there's, there's press on it and there's, you know, it's a fun thing that surprises the CFO. I think probably the, the, the most realistic substantive problem, and challenge is one more of technical implementation and the diffusion of AI in this form of agents across knowledge work. And you've talked about this a lot, you've had kind of great guests that I think that have covered this. I don't know how much I'll add to the contours of the conversation, but from what I'm seeing is, and I think this is one of these things where you you kind of have to have personally gone through the AI psychosis period, and then and then kind of come out the other side. And like, I've had my phases of like, I'll spend all weekend building projects and I'm like, this is the most amazing thing in the history of human history.
Speaker 122:35 - 23:24
所以,虽然眼下最热的话题可能是 token,只是因为它有媒体关注,而且你知道,这种东西也很容易“惊到” CFO,但我觉得,最现实、最实质性的问题和挑战,大概率还是技术实现,以及 AI 以 agent 这种形态在知识工作中的扩散。你之前已经谈过很多这方面内容了,也请过一些很棒的嘉宾,我觉得他们已经把这个问题讲得很好了。我不知道自己还能给这场讨论增加多少新的轮廓,但就我目前看到的情况来说,我觉得这类事情你某种程度上必须亲自经历过一段 AI psychosis(AI 狂热)时期,然后再从另一边走出来。像我自己也经历过这样的阶段:整个周末都在埋头做项目,然后觉得,天啊,这简直是人类历史上最了不起的东西。
Speaker 123:24 - 24:00
And like, obviously, like, you know, you're going to have companies that are just one employee and they're going to do everything. And then you come out the other end and you're like, wow, like actually, like maintaining that thing takes a lot of work. I'm having to, you know, catch so many mistakes that it's making. And so I'm spending as much time sort of like, you know, after the project, just reviewing everything and changing and modifying or the model gets upgraded and it breaks everything that I just did, and now I have to go and redesign it again. Once you're through the AI psychosis period, you kind of land on the other side.
Speaker 123:24 - 24:00
而且很显然,你会想,未来肯定会出现只有一个员工却能包办一切的公司。然后等你真正走过那一段,再回头看,就会觉得:哇,实际上,维护那套东西要花很多功夫。我得不停地去抓它犯的各种错误。所以项目做完以后,我花在复查、修改、调整上的时间几乎一样多;或者模型一升级,就把我刚做好的东西全弄坏了,我又得重新设计一遍。当你走出那段 AI psychosis(AI 狂热)之后,你会落到另一个更现实的位置上。
Speaker 124:00 - 24:42
And I guess I, know, I'm benefited by both being a power user of these tools, but then seeing the real world and kind of like being like, Oh, wow, like, actually, in your particular environment, I think there's zero chance that you could have done what I can do on the weekend for fun. Because I would never allow that to happen from a security standpoint, or you know, name your name your reason. So here's here's kind of the litany of things that are the work ahead. So let's just say you use Claude code or codex, and you're like, this is clearly the biggest breakthrough of all time, and it's obviously going to like ripple through knowledge work and it's going to transform everything overnight or all the jobs are going to be totally impacted. Here's just like the quick kind of like ledger.
Speaker 124:00 - 24:42
我想,我的一个优势在于,我一方面是这些工具的重度用户,另一方面又能看到真实世界,所以我会意识到:哦,原来在你们那种具体环境里,你根本不可能把我周末拿来玩的那些东西真正做起来。因为从安全角度,我就绝不会允许那样的事情发生,或者你随便说别的原因也行。所以,接下来真正要做的工作,大致就是下面这一串。假设你用了 Claude code 或 codex,然后觉得:这显然是有史以来最大的突破,它显然会迅速席卷知识工作,会在一夜之间改变一切,或者所有工作岗位都会被彻底影响。那我就快速列个账。
Speaker 124:42 - 25:21
So in coding, you have a highly technical user, you have models that are hyper trained on coding, you have effectively verified, verifiable work because like the code either like runs and you can QA it and you can have tests on it or not. Back to the technical user piece, it's actually not a minor point. That technical user, the moment the agent does something stupid or runs into a problem, the user themselves know how to go fix it and get it back on track. And by virtue of them being technical and wired into this ecosystem, they're just consuming the news far faster and thus the best practices far faster. So when somebody says, Oh, how's your skills file?
Speaker 124:42 - 25:21
先说 coding,在这个场景里,你面对的是高度技术化的用户,用的是在 coding 上经过超强训练的模型,而且工作的结果实际上是可验证的,因为代码要么能跑,你可以对它做 QA,也可以给它加 tests,要么就是不行。再回到“技术型用户”这一点,这其实不是个小问题。那个技术型用户一旦发现 agent 做了蠢事,或者卡在某个问题上,用户自己就知道怎么把它修好、把它拉回正轨。而且正因为他们本身技术很强,又深度嵌在这个生态里,所以他们获取新闻的速度更快,因而掌握 best practices(最佳实践)的速度也更快。所以当有人问,哦,你的 skills file 怎么样了?
Speaker 125:21 - 25:51
Or your agents MD file? They're like, oh, yeah, well, it's got this and this, and it's stored here, it's accessible here. Like, that's not the that's not the dialogue and the language of of a of a kind of regular knowledge worker. And this has been talked about a ton by by even like Dorkesh and I think Dario had a great conversation on this. Like the code base has so much of the context in coding, whereas in the rest of knowledge work, the context lives across like 20 different things, some digital and some very not digital, you know, kind of mediums.
Speaker 125:21 - 25:51
或者问你的 agents MD file 呢?他们会说,哦,对,里面有这个和这个,存放在这里,可以从这里访问。像这样的对话和语言,并不是那种普通知识工作者会使用的。这一点已经被很多人反复谈过了,甚至包括 Dorkesh,我记得 Dario 在这方面也有过一场很精彩的讨论。就是说,在 coding 里,code base 本身就承载了大量 context(上下文);而在其他大多数知识工作中,context 分散在大概 20 种不同的东西里,其中一些是数字化的,另一些则明显不是数字化的那种媒介。
Speaker 125:51 - 26:49
And then this is kind of a kind of a relatively boring one, but it's like, it's going to be probably the most important, which is, which is access controls, in your code base. Like I can go to most teams in engineering, and they have access to the entire kind of portion of work that they need to be working on. Conversely, you go to knowledge work and and like you constantly are running into either like, oh, like Bob actually had too much access to something, or Sally had too little access to something. So Sally has to go ask for something or Bob should actually like have less access. And in both those cases, the agent equivalent that would have been doing coding that just can consume all of the code base that it needs and generate whatever it needs, that agent and knowledge work is going to either bounce up against an entitlement issue like immediately, and it's not going to have access to a resource, or it'll have access to too much in terms of resources, and then start to answer questions with data that it shouldn't have because the company didn't have like a clean environment for access controls.
Speaker 125:51 - 26:49
然后还有一个算是相对没那么吸引人的点,但它可能会是最重要的,那就是你代码库里的 access controls(访问控制)。比如说,我去看大多数工程团队时,他们通常都能访问自己需要处理的那部分工作内容。反过来,在 knowledge work(知识工作)里,你会不断碰到这样的情况:要么是 Bob 实际上对某些东西权限太大了,要么是 Sally 对某些东西权限又太少了。于是 Sally 得去申请权限,而 Bob 实际上应该被收回一些权限。在这两种情况下,那个在 coding(编程)场景里的 agent 对应物,本来只要读取自己所需的整个代码库并生成需要的东西就行;但到了 knowledge work 里,这个 agent 要么会立刻撞上 entitlement(权限授予)问题,拿不到某个资源;要么又会拿到过多资源访问权限,接着开始用它本不该接触的数据来回答问题,因为公司并没有把 access controls 做成一个干净清晰的环境。
Speaker 126:49 - 27:25
So you've got kind of five or six reasons that AI coding looks very different from the rest of knowledge work. And so what the implications of this are, is that it's basically like diffusion is going to take time. And we have increasingly the right kinds of applications for this, like Cloud Coworker is awesome, obviously Codex as a super app is emerging as this powerful workhorse. Gemini with I think Spark and whatnot, you know, I think there's like rumors that cursor might, you know, try and evolve based on the SpaceX relationship. So I think the tools are increasingly coming and or there.
Speaker 126:49 - 27:25
所以,大概有五六个原因,解释了为什么 AI coding(AI 编程)看起来和其他大多数 knowledge work 非常不一样。这带来的含义是,技术扩散基本上需要时间。现在我们已经越来越多地拥有适合这件事的应用了,比如 Cloud Coworker 很棒,显然 Codex 作为一个 super app(超级应用)也正在出现,成为强大的 workhorse(主力工具)。Gemini 加上我想是 Spark 之类的,还有一些传闻说 cursor 可能会基于和 SpaceX 的关系继续演进。所以我觉得,这些工具正在越来越多地出现,或者说已经到位了。
Speaker 127:25 - 28:06
Now you have the hard part of like, how do I deploy this in my organization, in a way that is safe, in a way that is is reliable in a way that my employees aren't going to kind of create some crazy blast radius of security challenges, in a way where employees sort of know like what is the right way to go wire up this workflow that ends up being useful for them. So you have this tech, you have this huge AI kind of diffusion challenge. It's a much more technical problem than I think we sort of got used to with the chat paradigm, because chat was like basically it could do two things. It could access search, and it could access the LLM, And that and that was amazing. And but guess what?
Speaker 127:25 - 28:06
现在真正难的部分来了:我该如何把这些东西部署到我的组织里,既安全、又可靠,还能避免员工制造出某种巨大的安全 blast radius(爆炸半径、连带影响范围);以及,如何让员工知道,怎样把这个 workflow(工作流)正确接起来,最终对他们真正有用。所以你现在有了这项技术,也面临着一个巨大的 AI 扩散挑战。和我们在 chat(聊天)范式下已经习惯的问题相比,这其实是一个技术性强得多的问题,因为 chat 基本上只能做两件事:它可以访问 search(搜索),也可以访问 LLM(大语言模型),而这本身已经很惊艳了。但是,问题就在这里。
Speaker 128:06 - 28:13
Neither of those things has a permission problem. Neither of those things required wiring up some other system where you could have massive data leakage.
Speaker 128:06 - 28:13
这两件事都不存在 permission(权限)问题,也都不需要接入某个其他系统,从而引发大规模数据泄露。
Speaker 228:13 - 28:16
Yeah. It's just a just a DLP problem at worst. Right?
Speaker 228:13 - 28:16
对,最糟也不过就是个 DLP(Data Loss Prevention,数据泄露防护)问题,对吧?
Speaker 128:16 - 28:45
Just a DLP problem. And and honestly, like, in many ways, not that different from somebody going to Google to say like, I want to go research this customer versus going to Chatuche BT and say, I want to reach us. It's kind of like, like almost nothing has changed about the security paradigm of that enterprise. So maybe the prompt could include a little bit more IP, but like, the work you were doing was not like that the blast radius of that work was kind of quite contained. Conversely, I go to an agent and I happen to have access to the Salesforce MCP server.
Speaker 128:16 - 28:45
就只是个 DLP 问题。而且说实话,在很多方面,这和某个人去 Google 搜“我想研究一下这个客户”,相比去 Chatuche BT 里说“我想研究一下这个客户”,其实没那么不同。也就是说,这家企业的安全范式几乎没有发生什么变化。也许 prompt(提示词)里可能会多带一点 IP(知识产权/内部敏感信息),但你当时做的工作,其 blast radius 其实是相当可控的。反过来,如果我去用一个 agent,而且我刚好还能访问 Salesforce MCP server,
Speaker 128:46 - 29:07
And it's actually incredible. And it's actually one of the reasons why I totally believe in headless software. But I can do, like, I could do a lot of work with that and I could pull out a lot of data and I can ask a lot of very powerful questions. And a company is going have to say, well, should every employee have the same level of access? How should we make sure that we've cleaned up our access controls for that?
Speaker 128:46 - 29:07
那事情就真的很惊人了。这其实也是我完全相信 headless software(无头软件)的原因之一。因为有了这个,我能做很多工作,能拉出大量数据,还能提出很多非常强大的问题。于是公司就必须回答:是不是每个员工都应该拥有同等级别的访问权限?我们又该如何确保已经把这方面的 access controls 清理好了?
Speaker 129:07 - 29:27
And how do we tell people again, like what types of queries should they be doing that are going to have different kind of cost profiles? And now you have to do that for each of your software vendors and applications. And then you have to figure out like, what is the new workflow on the other end of this? Do you really want employees prompting their way through the workday across lots of stuff? Or do you want some standardized best practices?
Speaker 129:07 - 29:27
以及,我们该如何再次告诉大家:他们应该发起什么类型的 query(查询),因为不同查询会对应不同的 cost profile(成本特征)?而且现在,你得针对每一家 software vendor(软件供应商)和每一个应用都做这件事。然后你还得想清楚,这一切另一端的新 workflow 到底应该是什么样的?你真的希望员工整天靠不停写 prompt 在各种系统之间推进工作吗?还是你更想要一些标准化的 best practices(最佳实践)?
Speaker 129:28 - 30:05
And then you're like, okay, well, now I have to build skills internally, like, you know, capital S skills, and I have to have, or I have to have, you know, various kind of knowledge graph or other ways of getting agents to the right information and the right kind of context. All of that is highly technical work that is going to take one, two, three, five years of building out across most organizations. The really good news for I think 90% of people, maybe other than like the super AI accelerationists, is that that work means tons of opportunity. It means actually there's a lot of opportunity for startups. It means there's a lot of opportunity for like new kinds of roles.
Speaker 129:28 - 30:05
然后你就会想,好吧,现在我必须在内部建立能力了,比如说,真正意义上的 Skills(技能),而且我还得具备,或者说我还得建立,各种 knowledge graph(知识图谱)或其他方式,好让 agents(智能体)拿到正确的信息和正确类型的 context(上下文)。所有这些都是高度技术性的工作,对大多数组织来说,铺开建设都要花上一年、两年、三年,甚至五年。我觉得对 90% 的人来说,真正的好消息是——可能除了那种特别激进的 AI accelerationists(AI 加速主义者)之外——这些工作意味着大量机会。也就是说,startup 公司其实会有很多机会,也意味着会出现很多新类型的岗位。
Speaker 130:05 - 30:31
One of the hottest topics also is like, you know, we have a lot of customers asking us, what is this new internal FDE role or external FDE role? Like, what is the technical talent I need to go and actually like help me deploy these types of systems? You have, we're in for years of this kind of diffusion, and it's just non trivial. And every company has to go through it one by one. That this is the kind of journey that we are all now on.
Speaker 130:05 - 30:31
另一个很热门的话题是,我们也有很多客户在问我们,这个新的内部 FDE 岗位,或者外部 FDE 岗位,到底是什么?也就是,我到底需要什么样的技术人才,才能真正帮我部署这类系统?我们接下来会经历很多年这种扩散过程,而且这绝不是件轻松的事。每家公司都必须一个个地走完这个过程。这就是我们现在所有人共同踏上的这段旅程。
Speaker 230:31 - 30:43
You think it could be ten years, like the cloud took much longer than everybody expected. And that was ultimately an IT problem, not an enterprise wide problem. Do you think this could be just taking, I don't know, over a decade?
Speaker 230:31 - 30:43
你觉得这会不会是十年的事?就像 cloud(云)当年花的时间比所有人预期的都长。而且那归根结底还是一个 IT 问题,不是一个全企业范围的问题。你觉得这件事会不会就是要花,怎么说呢,超过十年?
Speaker 130:43 - 31:13
Partly, I don't know how we define like it and this because I think it's actually a continuous, it'll be a continuous sort of evolution and not to like, you know, play semantics, but like, it's more like, it's more like, what do we think the end state is? And again, I think the AI sort of either doomer or accelerationist think there is some end state. I actually don't think there's an end state. I think this is a substrate of how work happens, and it will just constantly get better, and we will have to constantly move up abstraction layers. And like, it's not even obvious to me, like what the end is.
Speaker 130:43 - 31:13
部分原因是,我不确定我们该怎么定义“it”和“this”,因为我觉得它其实是一个持续的、会不断演化的过程——不是在抠语义细节——但更像是,我们认为最终状态是什么?而且再说一次,我觉得无论是 AI doomer(AI 悲观派)还是 accelerationist(加速主义者),都认为存在某种终局。我其实不认为有终局。我觉得这是一种工作发生方式的 substrate(底层基底),它只会不断变好,而我们也必须不断往更高的 abstraction layers(抽象层)移动。甚至对我来说,终点是什么都并不明显。
Speaker 131:13 - 31:56
It's just like, it's a new way to basically execute work. And some areas that will be a 5x productivity gain, other areas will be a 10% productivity gain, and that will roll out. And then in five years from now, we'll find the next version of that. And I think it's this always kind of evolving landscape. But I think that we should totally be thinking on the order of ten years as like a rough type scale for like whatever it and this might be like, if you want to be like, when does, you know, Coca Cola or Procter and Gamble, like have agents running around doing every single task in the enterprise, like across every crevice of the organization, hyper successfully, that's a multi year, you know, kind of transformation.
Speaker 131:13 - 31:56
它本质上就是一种执行工作的全新方式。有些领域会带来 5 倍的生产力提升,另一些领域可能只有 10% 的提升,而这些都会逐步铺开。然后五年之后,我们又会找到它的下一个版本。我认为这是一个始终在演化的格局。不过我确实认为,我们完全应该按十年这个量级来思考“it”和“this”这类事情的大致时间尺度。比如说,如果你想问,像 Coca Cola 或 Procter and Gamble 这样的公司,什么时候能让 agents(智能体)在企业里到处运转,承担企业中的每一项任务,深入组织的每一个角落,并且运行得极其成功——那就是一个需要很多年的转型。
Speaker 131:56 - 32:24
And I'm making up an example, maybe they're already there in particular, but this is just what's going to happen. A funny thing, our industry is actually some of this is actually weirdly a byproduct of the industry. So we have this incredible capability overhang, which is like, let's just say you took GBT 5.5 or Opus and you just snap the line right now. You could probably do this diffusion in like two years or three years. And we could probably all do the change management, like collectively as an ecosystem.
Speaker 131:56 - 32:24
我这里是在随便举个例子,也许他们在某些方面已经做到那一步了,但这就是将会发生的事。一个很有意思的点是,我们这个行业里的部分情况,其实很奇怪地也是这个行业本身的副产品。所以我们现在存在一种惊人的 capability overhang(能力过剩积压):比如说,假设你拿 GBT 5.5 或 Opus,然后现在立刻把技术路线固定下来。那你大概可以在两三年内完成这种扩散。我们也许还能作为整个 ecosystem(生态系统)一起,把 change management(变革管理)这件事整体推进下去。
Speaker 132:24 - 33:19
The problem is, is the breakthroughs keep happening faster than the customer can implement any kind of standard architecture. And those breakthroughs oftentimes basically undo or make obsolete the last thing you implemented. So it's this really bittersweet thing, which is like the technology is getting so advanced, that it makes obsolete the prior thing that you implemented, which actually means that the rollout takes longer, because we have no stable, there's no stable environment to roll things out in. If you went to an enterprise right now, it's actually a period of maybe the least amount of consistency I've ever seen in IT, of like the following question, I want to go deploy an agent to do client onboarding, or to review, to review some set of knowledge work in the enterprise. I could probably lay out up to 10 to 15 reference architectures to all solve that problem.
Speaker 132:24 - 33:19
问题在于,突破出现的速度,比客户来得及实施任何一种标准架构都更快。而这些突破又常常会直接推翻,或者让你上一次实施的东西过时。所以这是一种非常苦乐参半的局面:技术进步得太快了,以至于它会让你之前部署的方案失效,而这反过来又意味着推广落地会更慢,因为我们没有一个稳定的环境——根本不存在一个稳定的 rollout(推广部署)环境。你现在如果走进一家 enterprise(企业),这甚至可能是我见过 IT 一致性最低的时期之一。比如这样一个问题:我想部署一个 agent(智能体)去做客户 onboarding(客户接入),或者去审查企业中的某类 knowledge work(知识工作)。光是为了解决这个问题,我大概就能列出 10 到 15 种 reference architectures(参考架构)。
Speaker 133:20 - 33:54
That means that every systems integrator, every, you know, software startup, every lab is pitching a customer 10 to 15 different variants of what they should do to solve that one problem. And so what that actually leads to ironically is more, you know, lengthy sales cycles, more kind of complexity in decision making, because you're like, man, that anthropic managed agent thing looks incredible. This is really awesome. Like that's exactly how we should do it. And then you're like, oh, this, you know, OpenAI frontier is really good.
Speaker 133:20 - 33:54
这就意味着,每一家 systems integrator(系统集成商)、每一家 software startup(软件创业公司)、每一家 lab(实验室),都在向客户推销 10 到 15 种不同变体,告诉他们应该怎么解决这一个问题。所以讽刺的是,这实际上反而带来了更长的销售周期、更多决策上的复杂性,因为你会想,天哪,那个 anthropic managed agent 方案看起来太厉害了,真的很棒,感觉这就是我们该采用的方式。然后你又会想,哦,这个 OpenAI frontier 也非常好。
Speaker 133:54 - 34:15
And then you're like, oh, but this startup is actually pitching me something that means I'm neutral to either of those. And then you're like, oh, no, actually my workflow vendor can now do this. And it is a madhouse on that front right now if you're a CIO. And so like one of the memes is nobody's signing up for more than like one year deals with the labs. And part of that is because of the pace of innovation that's happening.
Speaker 133:54 - 34:15
然后你会想,哦,可这家 startup 实际上在向我推销一个方案,意思是我对这两种情况都能保持中立。接着你又会想,哦,不对,其实我现有的 workflow vendor 现在也能做这个了。如果你是 CIO,现在这方面简直就是一团混战。所以有个很流行的说法是,几乎没人愿意和那些 labs 签超过一年期的合同。部分原因就在于当下创新发生的速度实在太快了。
Speaker 134:15 - 34:23
And so it's a byproduct of actually how much innovation we are seeing, but that means diffusion ends up taking longer than I think, you know, most people think.
Speaker 134:15 - 34:23
所以这其实是我们眼下所见巨大创新浪潮的一个副产品,但这也意味着,技术扩散最终会比大多数人想的要花更长时间。
Speaker 234:23 - 34:40
Fascinating. What do you recommend people do in your conversations, given this litany of things that need to happen and the space of innovation, all of it that benefit the same time? You mentioned internal FDEs, that's super interesting. We can talk about external FDEs, which I think is a better understood thing. Where should people start or how do they accelerate?
Speaker 234:23 - 34:40
很有意思。基于你在交流中提到的这长长一串需要发生的事,以及整个创新空间里所有这些几乎同时发生、彼此受益的变化,你会建议大家怎么做?你提到了内部 FDE,这点特别有意思。我们也可以聊聊外部 FDE,我觉得这是大家更熟悉的东西。人们应该从哪里开始,或者说该如何加速?
Speaker 134:40 - 35:22
So the one part where I'm just like, I, you know, I'm a hammer looking for nails is I see most things as a data problem and data with associated things like access controls and like how well defined is the workflow, etcetera. So most agentic challenges, I think are kind of inversions of basically like you have a data challenge, Like, the agent can't get access to the right information to to do the work. Maybe they have access to too much information, in which case then then they're just gonna, like, roam around and do the wrong thing. Or they have access to too little information, in which case, obviously, they're not gonna work. Or they don't have enough context to be able to execute the task, which means they need more information surrounding the task.
Speaker 134:40 - 35:22
有一个部分我总会不自觉地往那边看——你知道,就像“手里拿着锤子到处找钉子”——因为我会把大多数事情都看作是数据问题,以及与数据相关的事情,比如 access controls(访问控制)、workflow 的定义是否足够清晰,等等。所以我认为,大多数 agentic(智能体式的)挑战,某种程度上都只是把问题倒过来看:本质上你面对的是一个数据挑战。比如,agent 拿不到完成工作所需的正确信息;也可能它拿到的信息太多了,那它就会四处乱转,做错事情;或者它拿到的信息太少了,那显然它就没法工作;再或者它没有足够的上下文去执行任务,这就意味着它需要围绕这个任务获得更多信息。
Speaker 135:23 - 35:59
We see data problems everywhere that we look. And so I think one of the first steps is like your enterprise just needs to be prepared from a data standpoint and from a kind of a core architecture. And I think we, for twenty to thirty years in IT, it was sort of okay to sort of have all these systems, some redundant, some not well managed. You could kind of throw humans at the problem and just sort of say, yeah, like the data science team knows, like where the bodies are buried in database. And they know what table to use and what table not to use.
Speaker 135:23 - 35:59
我们放眼望去,到处都是数据问题。所以我觉得第一步之一就是:你的 enterprise(企业)必须先在数据层面,以及某种核心架构层面上做好准备。我认为,在过去二三十年的 IT 里,某种程度上大家一直觉得拥有一堆系统也没关系,有些重复,有些管理得并不好。你还可以靠投入人力来应付这个问题,然后说,嗯,data science team 知道数据库里哪些地方“埋着雷”,他们知道该用哪张表,不该用哪张表。
Speaker 136:00 - 36:35
And they know how to go and kind of like work through that particular sort of data model. And so when the business asks the question, the business goes to their analytics team or data science team, they say, hey, tell us our attrition rate, or tell us our growth in Spain, or tell us our upsell rate of this product. The data science team is this kind of constrained centralized function. It's maybe 10 people or 100 people, but it's not every employee. And they go and they know how to kind of like work the numbers and they have another spreadsheet that's living on top of Tableau and then they're moving some stuff in there and they're doing some calculations and then they give you the answer.
Speaker 136:00 - 36:35
他们也知道该怎么沿着那套特定的数据模型一路梳理下去。所以当业务提出问题时,业务团队会去找他们的 analytics team 或 data science team,说,嘿,告诉我们员工流失率是多少,或者告诉我们在 Spain 的增长情况,或者告诉我们这个产品的 upsell rate。data science team 是一种受限的、中心化的职能部门。它可能有 10 个人,也可能有 100 个人,但它不是每个员工都具备的能力。他们会去处理这些数字,可能上面还叠着另一个挂在 Tableau 之上的 spreadsheet,然后他们在里面挪一些东西、做一些计算,最后把答案交给你。
Speaker 136:35 - 37:12
Now all of a sudden you're like, oh, well, I want to go democratize that to everybody. And now I want to MCP into whatever the data store is of that thing. And then guess what? Like everybody's getting a different definition to their query because actually the way that company calculated things is like, it's like, no, they do an FX adjusted number or they measure their net retention rate differently than what the model was trained on. And so all of this stuff where, where you now actually weirdly have a data problem and a data integrity problem and an access control problem, that actually becomes one of the more meaningful kind of projects ahead.
Speaker 136:35 - 37:12
现在突然之间,你会想,哦,那我想把这件事 democratize(民主化、普及)给所有人。然后我想通过 MCP 去连接那个东西背后的 whatever data store(无论是什么数据存储)。接着你猜怎么着?每个人对自己的 query(查询)都会得到不同的定义,因为那家公司实际计算这些指标的方式就是不一样——比如,不,他们看的是经过 FX adjusted(汇率调整)的数字,或者他们衡量 net retention rate 的方式和模型训练时所依据的定义不同。所以现在这些事情就变成了一个很奇怪但真实存在的问题组合:数据问题、data integrity(数据完整性)问题,以及 access control 问题。而这些,实际上会成为接下来更有意义、也更重要的一类项目。
Speaker 137:12 - 37:16
I think you're smiling way too much, which means either you funded something here or you're seeing it or I don't know.
Speaker 137:12 - 37:16
我觉得你笑得也太开心了,这说明要么你在这里投过什么项目,要么你已经亲眼看到这种情况了,或者我也不知道。
Speaker 237:16 - 37:31
No. It's it's it's just I'm I'm smiling at the, you know, old problems are are new again. And, effectively, we're talking about a semantic layer, which, you know, I guess is getting rebranded as an ontology, and that's a new new thing when in reality it's been the same problem for twenty years plus.
Speaker 237:16 - 37:31
不,不是。只是我在笑,你知道,老问题又以新面目出现了。归根结底,我们讨论的是一个 semantic layer(语义层),而它现在似乎被重新包装成了 ontology(本体),仿佛成了什么全新的东西,但实际上,这二十多年来一直都是同一个问题。
Speaker 137:31 - 37:49
Oh, a 100%. It's been the same problem for twenty years, but, but again, we could throw people at the problem before. Like, like at the end of the day, when I had a question about data, I know exactly the person to go ask. And I didn't, I didn't ever have to worry about it. Like, like as Aaron, you know, in a, in a company, because the data science team had to worry about it.
Speaker 137:31 - 37:49
哦,百分之百。二十年来一直都是同一个问题,不过话说回来,以前我们还可以靠堆人来解决这个问题。比如说,归根结底,当我对数据有疑问时,我很清楚该去问谁。我从来不需要操心这个。比如,像 Aaron 这样,作为公司里的人,我不用担心,因为 data science team 得去担心这个。
Speaker 137:49 - 38:10
Now, if somebody gives me access to that data as a resource, and I start asking questions, boom, that's a way bigger problem, because I might go to somebody be like, hey, why did we like grow, you know, 13% in that one area? And they're like, well, your data is wrong. Like, it's actually it was 16%. You just didn't adjust for FX or whatever. It's like, much bigger problem now when everybody can go and do that.
Speaker 137:49 - 38:10
但现在,如果有人把这些数据作为一种资源开放给我访问,而我开始基于它提问,砰,这就成了大得多的问题。因为我可能会去问别人:嘿,为什么我们在那个领域增长了 13%?对方却会说,你的数据错了,实际上是 16%,只是你没有做 FX 调整之类的。现在当每个人都能去这么做时,这就成了严重得多的问题。
Speaker 138:10 - 39:01
So and that's just the structured data, think about all the unstructured data, you know, most enterprises have five different places where their contracts are being stored, you know, their roadmaps are across, you know, 30 different locations inside of their data environment, that's obviously the space that we see day in and day out. So if you're going to have a world of agents, and you want to have some flexibility on what agentic platform you deploy and what type, do you play co worker, do you deploy managed agents or do you play codex, then you need to get your data into a format that is going to work within that kind of agentic ecosystem. So I think a lot of the work to be done is sort of blocking and tackling in the enterprise on IT, which is like, how do I get agents that context? How can they make sure they have access to the right information with the right security levels, with the right entitlements? And that is a big chunk of work ahead to ensure that agents are going to work properly.
Speaker 138:10 - 39:01
而这还只是 structured data(结构化数据),再想想那些 unstructured data(非结构化数据):大多数企业都有五个不同的地方在存合同,他们的 roadmap 分散在其数据环境内部 30 个不同位置。显然,这正是我们每天都在面对的领域。所以,如果你要进入一个 agent(智能体)的世界,并且你希望自己部署的 agentic platform(智能体平台)以及类型上保有一定灵活性——你是要做 co worker,部署 managed agents,还是走 codex 这类路线——那么你就需要把数据整理成一种能在这类 agentic ecosystem(智能体生态)中运作的格式。所以我认为,接下来大量要做的工作,其实是企业 IT 里的基础建设工作,也就是:我该怎么给 agent 提供那些 context(上下文)?怎么确保它们能以正确的 security levels(安全级别)和正确的 entitlements(权限)访问正确的信息?要确保 agent 能真正正常工作,前面还有很大一块工作要做。
Speaker 139:01 - 39:27
To do that, that's where the kind of internal FDE motion comes in. So we are seeing this increase. Some of it is sort of repositioned internal IT people or software engineers, some of it is just straight up hiring new kinds of people and talent for the organization. But I do think this is a highly technical skill. It's a highly technical role, which is, do you have technical people in your organization that you can say, I'm going to have you go sit next to the business or within the business.
Speaker 139:01 - 39:27
要做到这一点,就涉及某种内部 FDE 的推进方式。所以我们确实看到这种需求在增加。其中一部分是把内部 IT 人员或 software engineers 重新定位到这个方向上,另一部分则是直接为组织招聘新类型的人才。但我确实认为,这是一项高度技术性的技能,也是一种高度技术性的角色。关键在于,你的组织里是否有技术人员,可以让你对他们说:我要让你去贴着业务团队坐,或者直接坐进业务团队里。
Speaker 139:28 - 40:08
And your job is to understand the patterns of how these people work, and make sure that they have the ability to use agents to go and do that work. And some of that will be agents for people that are prompting, and some of it will be agents that kind of are just working in the background and like, and, you know, automatically producing value for that knowledge worker. But your job is to go understand the workflow, understand the process, and then marry that with the full potential of where technology is going and make sure that like the data is set up the right way, the instructions for the agents are set up the right way, you know, you have the right, you know, sort of human in the loop elements of doing that work. That's a that's just a ton. That's a lot of work for most organizations.
Speaker 139:28 - 40:08
你的工作,就是去理解这些人是如何工作的模式,并确保他们有能力利用 agents 去完成这些工作。其中一部分会是服务于那些通过 prompting(提示)来使用系统的人的 agents,另一部分则是那种在后台运作、自动为 knowledge worker(知识工作者)产出价值的 agents。但你的职责,是去理解 workflow(工作流),理解 process(流程),然后把这些与技术发展的全部潜力结合起来,并确保比如 data 以正确的方式建立好了,给 agents 的 instructions(指令)以正确的方式设定好了,你也具备了适当的 human in the loop(人在回路中)机制来完成这些工作。这真的非常庞大。对大多数组织来说,这都是一大堆工作。
Speaker 140:08 - 40:10
And it's going to be one of the
Speaker 140:08 - 40:10
而且这将会是其中一个
Speaker 240:10 - 40:15
Except, except if you add meta and you, you do that by putting software on everybody's Yes.
Speaker 240:10 - 40:15
除非,除非你把 meta 加进去,而你做到这一点的方式,是把 software 装到每个人的——没错。
Speaker 140:17 - 40:40
Yeah. I think that might be an end of one situation. So, you know, for mere mortal companies, you're going to have people going and doing this and those people are going to look like the next generation of a software engineer or kind of IT engineer. I think it's actually incredibly exciting work because you get to go and transform like, how does a life sciences company run? How an industrial giant operate?
Speaker 140:17 - 40:40
对。我觉得那可能标志着一种局面的结束。所以你知道,对普通公司来说,接下来会有人去做这件事,而这些人看起来会像下一代的软件工程师,或者某种 IT 工程师。我觉得这实际上是非常令人兴奋的工作,因为你可以去真正改造一家 life sciences company 是怎么运作的,或者一个 industrial giant 是怎么运营的。
Speaker 140:40 - 41:03
How do marketing campaigns get produced? So it's actually like very, I think, exciting technical work. But a lot of companies don't have this talent right now. So they're going actually have to go and hire, you know, people out of CS programs or, you know, be able to pivot engineers into these kinds of functions. And, you know, as an asterisk, it's actually why the doomers are also wrong about jobs, because this is actually going to be a very real sustaining job.
Speaker 140:40 - 41:03
marketing campaign 是怎么被制作出来的?所以这其实是我认为非常、非常令人兴奋的技术工作。但很多公司现在并没有这种人才。所以它们实际上将不得不去招聘,比如从 CS program 里招人,或者把现有工程师转向这类职能。而且,顺带说一句,这也是为什么那些 doomers 在就业问题上也是错的,因为这实际上会成为一种非常真实、可以长期持续的工作岗位。
Speaker 141:03 - 41:23
That is not like a one time you implement the agent and you upgrade the system, and then it kind of works forever. It's like, no, like, once the model changes, there's another set of work to be done. You have to make sure like, did you get the gains of that model improvement? Or do you have to leave behind some scaffolding that you had to build for the prior model? Like lots and lots of work to be done in in this area.
Speaker 141:03 - 41:23
这并不是那种一次性把 agent 实施进去、把系统升级完,然后它就能永远自己运转的事。不是这样的。模型一旦变化,就会有新一轮工作要做。你必须确认,比如,你有没有真正获得那次 model 改进带来的收益?还是说,你不得不保留一些之前为了旧模型搭起来的 scaffolding(脚手架式辅助结构)?这个领域里还有大量、大量的工作要做。
Speaker 241:23 - 41:58
Super interesting. So do you think that the external FD position is here to stay as well? So internal FD being within the enterprise, external FD being within the the vendors, the slightly cynical version of FDEs in startups or larger tech companies right now is that, well, none of this really works. Therefore, you need to deploy a chunk of of humans to come on premise at the customer and make it work. But I think what you're saying is more profound and that this is going to be a fixture rather than a temporary thing.
Speaker 241:23 - 41:58
特别有意思。所以你觉得 external FD 这个岗位也会长期存在吗?也就是 internal FD 在企业内部,external FD 在 vendors 那边。现在对 startup 或大型科技公司里 FDE 的一种稍微有点犬儒的看法是:嗯,这些东西其实根本不好使,所以你必须派一批人到客户现场 on premise 去把它弄到能工作。但我觉得你说的更深一层:这会成为一种长期存在的固定角色,而不是临时现象。
Speaker 141:58 - 42:30
Yeah, it's so funny because the AI super accelerationists, which sometimes actually end up in the same quadrant of their views of the doomers. And the let's say, I don't I don't know skeptics, as another, you know, kind of end of the continuum, they land in the same spot in this particular topic, they're like, Man, I can't believe we have to have people go and do this. It's like, it's like it proves the skeptics, you know, right. And somehow the doomers and the accelerationists are like, Oh, man, like, it's not happening the way that we thought. And and then it's sort of the cynical thing.
Speaker 141:58 - 42:30
对,这很有意思,因为那些 AI super accelerationists,有时候在观点上其实会和 doomers 落到同一个象限。再加上,怎么说呢,我也不知道,skeptics 作为这个连续光谱的另一端,他们在这个具体话题上居然也得出了同样的结论。他们会说,天啊,我真不敢相信我们还得派人去做这个。这就像是在证明 skeptics 是对的。然后某种意义上,doomers 和 accelerationists 又会说,哦,天啊,事情的发展方式并不是我们原先想的那样。于是这就变成了一种有点犬儒的解读。
Speaker 142:30 - 42:52
And and what's funny is is like, you have people like me that are like, I just know enterprises. And it's like, this was obviously 100% going to happen. Like, you guys are all crazy if you didn't think this was going to happen. And it neither proves that the technology is not amazing, nor does it prove that like, it's just like, like, it obviously had to play out this way. Why did it have to play out this way?
Speaker 142:30 - 42:52
而有意思的是,像我这样的人会觉得:我只是懂 enterprise。而这件事显然百分之百就会这样发生。要是你们没想到会这样,那你们才是真的疯了。这既不能证明这项技术不惊艳,也不能证明什么别的;而只是说,事情显然就必须这样展开。那为什么它必须这样展开?
Speaker 142:52 - 43:27
It's because we built this insane technology that's like incredible at using computers, incredibly at using software, incredibly at writing code, incredibly incredible at writing tool using tools. And, but guess what it like has, you know, a fixed amount of memory, it has a fixed amount of context it can work with. It couldn't do totally dramatically crazy stuff with your data. Like, so obviously, it has to be like implemented by somebody hyper technical. Obviously, it has to be like implemented in a way that drives change management in an appropriate way, like for that organization.
Speaker 142:52 - 43:27
因为我们造出了这种疯狂的技术:它在使用 computer 方面强得惊人,在使用 software 方面强得惊人,在写 code 方面强得惊人,在调用 tool、使用 tools 方面也强得惊人。但是你猜怎么着,它的 memory 是固定规模的,它能处理的 context 也是固定规模的。它没法拿着你的 data 去做那种完全夸张、彻底疯狂的事情。所以很显然,它就必须由某个技术能力极强的人来实施;也很显然,它必须以一种能够针对那个组织恰当地推动 change management 的方式来实施。
Speaker 143:27 - 43:43
So it's like, to me, this was 100% priced in. And, and the market just took way longer to get there than, than I think anybody would have realized, like, like, could just feel this, the moment you saw agents be real. You're like, yeah. This is amazing, and it's totally gonna take a lot of work for enterprises to go and implement this.
Speaker 143:27 - 43:43
所以在我看来,这件事是百分之百早就 priced in 的。只是市场花了远比我想象中、也比任何人能意识到的更长时间,才走到这一步。就像你一旦看到 agents 真的成形了,你立刻就会觉得:对,这太惊人了,而且企业要真正把它实施进去,绝对会需要大量工作。
Speaker 243:43 - 43:49
You mentioned headless software. Is that inevitable in your opinion and and clearly the future?
Speaker 243:43 - 43:49
你提到了 headless software。你觉得这是一种不可避免的趋势,而且显然就是未来吗?
Speaker 143:49 - 44:12
I think head the headless conversation ends up usually in the same kind of spot as as, you know, almost every other technology kind of trend in history where you're like, you always think that the next medium fully eradicates the prior medium. And and then you just like, you're like, oh, no. Actually, I do have an iPad and a MacBook and an iPhone. And, like, and for some reason, I don't just like use my iPad as my phone and I and with a computer. It's like, no, have three devices.
Speaker 143:49 - 44:12
我觉得,关于 headless 的讨论,最后通常都会落到一个和历史上几乎所有其他技术趋势类似的位置:人们总以为下一种媒介会彻底消灭前一种媒介。然后你就会发现,哦,不对。实际上,我有 iPad、MacBook 和 iPhone。而且,不知道为什么,我并不会只拿 iPad 来同时当手机和电脑用。不是那样的;而是三种设备都会存在。
Speaker 144:12 - 45:06
They all do something different. And so so I think I think it's gonna just be one of those things, is which is if I'm gonna go into a complex query that involves box data, Salesforce data, Workday, and it's got a triage bunch of stuff, I'm gonna do that fully headlessly inside of coworker codex or something else, like unquestionably. If I want to go and like work on a set of documents and build a data room and and go and make sure that I've I'm sharing all my contracts the right way, at some point, like, doing that via text is sort of slower than just doing that in a graph ical user interface and with all the knobs that I know how to, you know, interact with. And I get a lot more leverage that way. So I think it's just going to be this sort of dual dual model, with the one nuance being probably by like, by like, you know, database queries, headless will just be 100 times larger than the interface driven way of doing work.
Speaker 144:12 - 45:06
它们各自做的事情都不一样。所以我觉得,这也会是类似的一件事:如果我要处理一个复杂查询,里面涉及 Box 数据、Salesforce 数据、Workday,而且还要分诊(triage)一堆东西,那我毫无疑问会在 coworker codex 或别的什么工具里,以完全 headless 的方式来做。如果我想处理一组文档、搭建一个 data room,并确保我正确共享了所有合同,那么到了某个点,用文本来做这件事其实就比直接在 graphical user interface 里操作要慢了;在那个界面里有各种我知道该如何交互的控件(knobs)。而且那样我能获得更大的杠杆效应。所以我认为,最后会形成一种双轨(dual)模型。唯一需要补充的细节是,比如说像 database queries 这类场景,headless 的规模可能会比界面驱动(interface driven)的工作方式大 100 倍。
Speaker 145:07 - 45:42
And so we'll just have to understand that by volume, agents are going to be banging on these systems far more than humans ever did. The human will probably land as an end user seat within that piece of software, and they'll get a certain amount of allocation of usage as that end user seat. And then I believe that they should have a right to use that software and that data via agents up to a certain amount. And that certain amount will be different based on the vendor depending on like how compute intensive is that workload. And then past that certain amount or when it's fully just an agent, then it'll be a consumption model.
Speaker 145:07 - 45:42
所以我们只需要理解一点:从总量上看,agents 会比人类以往更多得多地持续敲打这些系统。人类更可能是作为这类软件中的 end user seat(终端用户席位)存在,并且会以这个 end user seat 的身份获得一定的使用配额。而我认为,他们也应该有权通过 agents 来使用这款软件和其中的数据,直到某个额度为止。这个额度会因 vendor(供应商)而异,取决于这个 workload(工作负载)在计算上有多密集。再超过这个额度,或者在完全由 agent 运行时,就会转为 consumption model(按量消费模型)。
Speaker 145:42 - 46:14
So I think any enterprise software company in three years from now that gets through this AI transformation period, it will have a seat business model, assuming it has an end user component, and it'll have a consumption business model. And that consumption business model in some businesses might be bigger than the seat model and some might be smaller just because the seat still takes up so much, you know, kind of, you know, set of the work. But I don't believe that we move fully to consumption and fully to headless because I think there's a lot of reasons why you still want to go into the interface and poke around for a bunch of, you know, kind of reasons.
Speaker 145:42 - 46:14
所以我认为,三年后,任何成功穿越这轮 AI 转型期的 enterprise software company(企业软件公司),都会同时拥有两种商业模式:如果它有 end user component(终端用户部分),那它会有 seat business model(席位制商业模式);同时它也会有 consumption business model(按量消费商业模式)。在一些业务里,consumption business model 可能会比 seat model 更大;在另一些业务里则可能更小,因为 seat 依然承担了相当大一部分工作。但我不认为我们会彻底转向 consumption,也不会彻底转向 headless,因为出于很多原因,你仍然会想进入界面里四处点一点、看一看。
Speaker 246:14 - 46:33
And do you think it's necessarily humans have a seat and agents have consumption? Or would there be an argument for saying that agents in some way are not that dissimilar from humans, although they'll be doing a lot more with a lot more volume of data, and therefore there should be some kind of like seed based pricing for agents.
Speaker 246:14 - 46:33
那你觉得一定会是“人类对应 seat,agents 对应 consumption”吗?还是说,也可以主张 agents 在某种意义上和人类并没有那么不同,虽然它们会处理更多事情、接触更大体量的数据,因此也应该为 agents 设计某种基于 seat 的定价方式?
Speaker 146:33 - 46:57
I think, I think this was, this is sort of a tougher category because, it all depends on the agentic use case. So, like I can totally see a world where we already have some customers playing around this idea of like, should agents have a box seat? Because why? Because they actually need to store data, that gets retained and governed over a long period of time. And you want to be able to track it and manage it just like a person, but it's gotta be stateful.
Speaker 146:33 - 46:57
我觉得,这其实是一个更难分类的问题,因为一切都取决于具体的 agentic use case(agent 化使用场景)。比如,我完全可以想象这样一个世界:我们已经有一些客户在尝试这个想法——agents 是否也应该拥有一个 Box seat?为什么?因为它们实际上也需要存储数据,而这些数据会被长期保留并接受治理(governed)。你会希望像管理一个人一样去追踪和管理它,但前提是它必须是 stateful(有状态的)。
Speaker 146:57 - 47:21
And so that kind of makes sense is like, have to give it a name and a thing in our system to make that work. Do we charge the same as a regular end user seat? Probably not, probably it's got to be cheaper. But then there's a lot of situations where the agent doesn't need an ongoing seat, they just need to be doing a lot of operations, in which case it's probably just pure consumption. So I think it really depends on where does your software category land on?
Speaker 146:57 - 47:21
所以这就有点说得通了:为了让这件事运作起来,我们必须在系统里给它一个名字、给它一个实体。那我们会按普通 end user seat 的价格收费吗?大概不会,应该得更便宜一些。但也有很多情况是,agent 不需要一个持续存在的 seat,它们只是需要执行大量操作;那这种情况下,大概就应该是纯粹的 consumption。所以我觉得,这真的取决于你的软件类别最终落在什么位置。
Speaker 147:21 - 47:37
Is there a reason why you'd have an agent be stateful in that organization and kind of take on an identity and take on ongoing work versus it's a thing that just every employee calls on demand. And that, that would probably determine what that business model looks like.
Speaker 147:21 - 47:37
在那个组织里,为什么你会让一个 agent 具备 stateful(有状态)特性,某种程度上拥有一个身份,并承担持续进行的工作,而不是把它做成一个每个员工都按需调用的东西?这一点很可能会决定那种 business model(商业模式)最终会是什么样子。
Speaker 247:37 - 47:41
What does headless mean at Box? How do you guys go about it?
Speaker 247:37 - 47:41
在 Box 这里,headless(无头)是什么意思?你们是怎么做这件事的?
Speaker 147:41 - 48:15
We kind of think about it as everything you would ever want to do with your enterprise content. You should be able to do via an external agentic interface. And so the examples are, let's say you want an agent to go and read through 100 contracts or review a data room that you've created for risks in a client. We just launched this example with the Claude for legal solutions announcement. So you can put all your contracts in a folder and then the agent within Claude CoWork can go and kind of work through all of that data and use it as a knowledge repository for its work.
Speaker 147:41 - 48:15
我们大致把它理解为:凡是你会想对企业内容做的事情,都应该能够通过一个外部的 agentic interface(agent 式接口)来完成。比如说,你想让一个 agent 去阅读 100 份合同,或者审查你为某个客户建立的数据室,找出其中的风险。我们刚刚在面向法律解决方案的 Claude 公告里发布了这样一个例子。所以你可以把所有合同放进一个文件夹里,然后 Claude CoWork 里的 agent 就能处理所有这些数据,并把它作为自己开展工作时的 knowledge repository(知识库)。
Speaker 148:15 - 49:04
You could do a client onboarding process where the client has to upload a bunch of documentation. It's got to get stored somewhere and then processed. All of those are situations where Box will be the backend sort of behind the scenes some kind of agentic work that's happening, whether the user is sort of interacting with the agent, or the agent is just kind of running, you know, on some kind of deterministic or non deterministic event that happens. And what's, you know, we kind of like, conveniently have not had to do like massive, you know, kind of crazy transformations of the model, because we've always had an API basically, like almost on day one of the business, we had an API. So for us, whether it's a headless agentic user or a headless system machine application user is kind of, you know, eventually it talks to our system in roughly the same way.
Speaker 148:15 - 49:04
你也可以做一个客户 onboarding(入驻)流程,在这个流程里,客户需要上传一堆文档。这些文档必须先被存储到某个地方,然后再被处理。所有这些场景里,Box 都会作为 backend(后端)在幕后支撑某种 agentic 工作,不管用户是在和 agent 交互,还是 agent 只是基于某种 deterministic(确定性)或 non-deterministic(非确定性)的事件自行运行。还有一点是,我们比较“幸运”的地方在于,不需要对模型做那种大规模、很夸张的改造,因为我们从一开始基本就有 API——几乎在公司创立第一天我们就有 API 了。所以对我们来说,不管是 headless 的 agent 用户,还是 headless 的 system/machine application 用户,最终它们与我们系统交互的方式大体上是相同的。
Speaker 149:04 - 49:42
There are some nuances, which is like, you know, headless users might want to sign up for the service on their own. And so we've had to think about like account provisioning differently, or they might use our search in a different way where they need more context than what a platform deterministic use case would have looked like. Like they're going to use our search tool very aggressively, so we have to inform them as they're doing their searches, how to think about this certain context that's inside these files. So there's a lot of work that we are doing to make our system better for agents, but the concept of being headless and the concept of being API first is kind of wired into our DNA.
Speaker 149:04 - 49:42
当然这里面也有一些细微差别。比如说,headless 用户可能想自己注册这个服务,所以我们就必须重新思考 account provisioning(账户开通)这件事;又或者,他们使用搜索的方式也可能不同,他们需要的上下文信息会比平台上那种 deterministic use case(确定性用例)更多。比如他们会非常激进地使用我们的搜索工具,因此我们必须在他们搜索时告诉他们,该如何理解这些文件里的某些上下文。所以我们确实在做很多工作,让我们的系统对 agents 更友好;但 headless 这个概念,以及 API first(API 优先)这个理念,可以说已经写进了我们的 DNA。
Speaker 249:42 - 50:03
How do you think org charts evolve? So we're going to have agents, we're going to have internal FDEs. How does the rest of the organization evolve? I'm sure that must be a key, real concern when you talk to Global 2,000 companies, right? Like the whole, but partly, you know, AI is taking my job kind of thing.
Speaker 249:42 - 50:03
你觉得组织结构图会怎么演变?也就是说,我们会有 agents,也会有内部的 FDEs,那么组织里的其他部分会如何变化?我很确定,当你和 Global 2,000 公司交流时,这一定是一个关键而且很现实的担忧,对吧?某种程度上,核心就是那种“AI 要抢我工作”的问题。
Speaker 150:03 - 50:49
This sort of relates to the AI coding versus the rest of knowledge work And, and, you know, I kind of set it up at the very beginning on obviously this diffusion thing. But the reason why I'm, I'm less concerned about the job part, and more optimistic is when we most get fearful of jobs, we look at coding as the example. And again, back to this coding issue. Coding has this other unique property that's kind of different from a lot of the rest of knowledge work, which is if I write code, and it's like super sloppy, because the agent is writing this code, it kind of at the end of the day doesn't matter short of like a security risk or maybe like, you know, it's using extra memory that it shouldn't use or whatnot if the software just runs. Like if it's like, I could throw, I could have, you know, an application that has a 100,000 lines of code or a thousand lines of code.
Speaker 150:03 - 50:49
这其实和 AI coding(AI 编程)以及其他知识型工作的关系有关。而且,你知道,我一开始提到的那个 diffusion 的事情,也是在为这个做铺垫。但为什么我对“工作会消失”这件事没那么担心、反而更乐观?因为当我们最害怕工作岗位受影响时,我们通常会拿 coding 来当例子。还是回到 coding 这个问题。Coding 有一个很独特的属性,这一点和很多其他知识型工作不太一样:如果我写了一段代码,而且它写得很粗糙,因为这段代码是 agent 写的,那么归根结底这件事其实未必那么重要——除非有 security risk(安全风险),或者比如它用了不该用的额外内存之类的问题——只要软件能运行就行。比如说,我可以有一个 100,000 行代码的应用,也可以有一个 1,000 行代码的应用。
Speaker 150:49 - 51:10
If it's doing the thing that it needs to do, it really doesn't matter. And so this is sort of why you're seeing a little bit more like hands off the steering wheel emerge in coding. And it's like, we're just going to throw agents on agents on agents. And then that's going to go and solve the problem. Take the other maybe most topical category is like legal as an alternative.
Speaker 150:49 - 51:10
只要它完成了自己需要完成的事情,其实就无所谓。所以这也就是为什么你会看到,在 coding 领域开始出现更多“手不扶方向盘”的做法。大家会想,“我们就不断叠加 agents,agent 套 agent 再套 agent”,然后它们就会去把问题解决掉。再看另一个也许当下最热门的类别,作为对照,就是 legal(法律)。
Speaker 151:10 - 51:44
In legal, you can't do that. I can't have online 2,004 of the contract, it sort of like adjusts the liability rate, you know, slightly because because I had an agent go and write this thing whole cloth. That doesn't like, there's no way for me to verify the the I mean, I can layer on agents and agents and agents and they review each other and then they review each other again. And I can get kind of down to smaller and smaller percentages of risk. But at the end of the day, you're still going to have some lawyer that has to basically say, I believe that this is 100% valid.
Speaker 151:10 - 51:44
在法律领域,你不能那样做。我不能把一份合同的“online 2,004 版”拿出来,它好像只是稍微调整了一点责任比例,仅仅因为我让一个 agent 去从零起草了这份东西。那样不行,因为我根本没法验证——我的意思是,我当然可以一层又一层地叠加 agents,让它们互相审查,再反复彼此复核。我可以把风险一点点压低到越来越小的百分比。但到最后,仍然得有某个律师出来基本上说:我相信这份文件是 100% 有效的。
Speaker 151:44 - 52:41
I can put this up for my client, or I can go and ship this. And so this last mile of agentic work, I think is going to remain in a much broader set of knowledge work areas than I think we realize. And there was, I've mentioned this a little bit in the past, but there was this funny article from the Financial Times, like three weeks ago, of like lawyers being inundated with, with all of these, you know, contracts that their client, you know, created, or the client went to ChatchBT and asked a bunch of legal questions that now the lawyer has to go and adjudicate and kind of provide, you know, answers on. And I think it's a kind of a microcosm of the real life kind of application of AI, which is it excel, it can accelerate one thing massively, I can review the contract far faster, and I can go and get to the risky areas or can generate a contract much faster. But in both those scenarios, there's still a lawyer on on either end, doing real work.
Speaker 151:44 - 52:41
我可以把这个交给我的客户,或者我可以把它正式交付出去。所以我认为,这种 agentic work(代理式工作)的“最后一公里”,会继续存在于比我们意识到的更广泛的知识工作领域里。我之前稍微提过这一点,但大概三周前 Financial Times 有篇很有意思的文章,说律师们正被大量这类东西淹没——比如客户自己做出来的各种合同,或者客户去 ChatchBT 问了一堆法律问题,结果现在都得由律师来裁定、来给出相应答案。我觉得这其实是 AI 在现实世界应用中的一个缩影:它在某一环节上表现极其出色,能大幅加速流程——我可以更快审合同,更快定位风险区域,或者更快生成合同。但在这两种情境里,前后两端依然都有律师在做真正的工作。
Speaker 152:41 - 53:22
And so I've removed one part of the bottleneck, still unconstrained by another part of the bottleneck. And so, and so that's just why like the jobs don't get eliminated as sort of the first thing. But the second thing is, you know, we've talked about this, but but and the market is, I think, fully beaten over the head on Jevons paradox, but, but nobody ever factors in the Jevons paradox thing. And I mentioned this with designers, but designers are kind of like a, you know, kind of a, maybe minor example relative to all of the, all of the areas where this is going to show up, which is if you go to Caterpillar or Eli Lilly or Johnson and Johnson, John Deere, just naming like big industrial companies, just like not in Silicon Valley. These companies forever, they want the top engineers like everybody else.
Speaker 152:41 - 53:22
所以,我只是移除了瓶颈中的一部分,仍然受制于另一部分瓶颈。这也就是为什么,岗位不会作为第一步就被消灭。第二点是,我们之前也谈过这个,而且我觉得市场已经被 Jevons paradox(杰文斯悖论)反复敲打过了,但真正做分析时,大家又往往不把 Jevons paradox 算进去。我以前拿设计师举过例子,不过相对于这个现象将会出现的所有领域,设计师可能只是个较小的例子。比如你去看 Caterpillar、Eli Lilly、Johnson and Johnson、John Deere,这些大型工业公司——也就是那些不在 Silicon Valley 的大公司。它们一直都和其他公司一样,想要最顶尖的工程师。
Speaker 153:22 - 54:01
They are working on incredibly mission critical areas of creating a new drug, building autonomous kind of industrial equipment. So they need top engineers like everybody else. They have to go and sign up for similar level scale projects as everybody else, but those engineers have largely sort of seen that, like, no, like you go to CS and then you you you go to Google or you go to Meta, etcetera. So what's gonna happen now with agents is all of a sudden that all of those other companies are going to light up far more technical projects and technical work in their organizations. Because for the first time ever, one of their engineers now has the capacity of three or five or 10 or whatever metric you want.
Speaker 153:22 - 54:01
它们做的是极其关键的任务,比如研发新药、打造 autonomous(自主)工业设备之类的东西。所以它们和其他公司一样需要顶级工程师,也必须像其他公司一样去立项同等级、同规模的项目。但长期以来,这些工程师基本都形成了一种路径依赖:你去学 CS,然后你就去 Google,或者去 Meta,诸如此类。现在有了 agents,会发生什么?突然之间,所有那些其他公司都会在组织内部点亮更多技术项目和技术工作。因为这是有史以来第一次,它们的一个工程师现在能拥有三倍、五倍、十倍——随便你用什么指标衡量——的产能。
Speaker 154:01 - 54:44
And so that's going to get them to sign up for way bigger projects than they would have been able to afford, which means that they now have a greater demand for that engineering capacity. And then you throw in one more category, which is every basically small business on the planet is going to, you know, be able to go in and augment any of their functions that they wouldn't have had internally before with agents. And each of those functions back to the human in a loop component often will need some human to be going and doing the extra work that it takes to make that agent actually effective. So let's say you want to do the marketing campaign agent and you're like a solo entrepreneur, maybe it's a three person team and like you're doing it and you're moonlighting, but then you're like, oh, this is actually really effective. This marketing campaign is working.
Speaker 154:01 - 54:44
这会让它们敢于承接比过去负担得起的规模大得多的项目,这意味着它们对这种工程产能的需求反而更高了。再加上一类新情况:地球上几乎每一家小企业,都会能够用 agents 去增强原本内部没有能力覆盖的各种职能。而这些职能中的每一项,回到 human in the loop(人类在回路中)这个部分,往往都还需要某个人去做额外工作,才能让那个 agent 真正变得有效。比如说你想做一个营销活动 agent,而你是个 solo entrepreneur(独立创业者),也许团队只有三个人,你一边做这个一边还在兼职。但后来你发现,哦,这东西居然真的很有效,这个营销活动跑起来了。
Speaker 154:44 - 55:15
Probably the next thing I'm going do is go hire a marketing person to go and manage these agents to go and do this at scale. So you I'm like a complete Jevons paradox pill person because I first of all, I see it in our own business. I see it in customers. And I see it in small startups where these startups are hiring as fast as possible, because they have all these job functions that their productivity gains are causing them to need to hire for. So this is this is sort of, you know, you can kind of pick your your argument, but there's like multiple reasons why the job argument ends up falling on its face.
Speaker 154:44 - 55:15
那我接下来很可能要做的事,就是去招一个营销人员来管理这些 agents,把这件事规模化推进。所以我完全可以说自己是个彻底相信 Jevons paradox 的人,因为首先,我在我们自己的业务里看到了这一点;我在客户那里看到了这一点;我也在小型 startup 里看到了这一点——这些 startup 正在以尽可能快的速度招聘,因为生产率提升带来了大量新的职能需求,逼得它们必须招人。所以这个问题你可以从不同角度来论证,但无论怎么论证,都有不止一个理由说明“工作岗位会消失”这套说法最后站不住脚。
Speaker 155:15 - 55:56
Once you start to see actually how AI is rolling out in a lot of organizations. I mean, so for us as an N of one example, I mean, we continue to hire in a decent percentage of the functions that we've always had. Just the work that those functions are doing just is going to look entirely different in the future because they should be augmented meaningfully by agents doing additional work for them. It's certainly changing what we can invest in and what we tilt toward, But some of that is actually just a byproduct of our business evolution and where we're seeing demand in the market. But we're hiring people in marketing, we're hiring people and we're hiring engineers quite actively.
Speaker 155:15 - 55:56
一旦你开始真正看到 AI 在很多组织里是怎么铺开的,就会明白这一点。拿我们自己这个 N of one(单一样本)来举例,我们仍然持续在相当比例的既有职能上招聘。只是这些职能未来所做的工作会变得完全不同,因为它们理应被 agents 的额外工作能力显著增强。这当然会改变我们愿意投什么、以及资源更偏向什么方向,但其中有一部分其实也只是我们业务演进,以及我们在市场中看到需求变化后的自然结果。不过我们确实还在招 marketing 人员,我们也在招人,而且还在相当积极地招聘工程师。
Speaker 155:56 - 56:44
We're hiring people in IT to build these agents. We're hiring sales reps. So like those contours aren't shifting as much as one would expect at the name of the job title. If I had to really, let's say, lean into the future scenario, I think what happens is, is, you know, there's some kind of embedded AI IT capacity in most functions, there will be an AI person slash team in sales, and an AI person slash team in marketing and at different subsections of marketing, and they'll be in they already exist in engineering because engineering has been going through these sort of productivity gains. And I would I would assume that that person slash team, their job should be looking at like, hey, what do you do every day as like a demand gen person?
Speaker 155:56 - 56:44
我们在招 IT 人员来构建这些 agents,也在招销售代表。所以从职位名称这个层面看,岗位轮廓并没有像很多人预期的那样发生巨大变化。如果我必须更大胆地推演一下未来图景,我觉得会发生的是:大多数职能里都会嵌入某种 AI IT 能力;sales 会有一个 AI 人员/团队;marketing 会有一个 AI 人员/团队,marketing 的不同子领域里也都会有;engineering 里其实已经有了,因为 engineering 一直在经历这类生产率提升。而我会假设,这个 AI 人员/团队的工作,应该是去看这样的问题:比如,嘿,作为一个 demand gen(需求生成)岗位的人,你每天到底在做什么?
Speaker 156:44 - 57:12
And how can I bring automation to that? So we could be testing five times the number of campaign ideas and keywords. And we could be integrating one part of the design process to a campaign lifecycle much faster. So you have a kind of a technical sort of person kind of wired up next to the business. And do over time, like in twenty years from now, is that maybe just the new expectation of one of those business people?
Speaker 156:44 - 57:12
那我怎么把 automation(自动化)引入其中呢?这样一来,我们就可以测试五倍数量的 campaign(营销活动)创意和关键词。我们也可以把设计流程中的一部分,更快地整合进 campaign lifecycle(营销活动生命周期)。于是你会有一种偏技术型的人,某种程度上和业务并排协作。那么从长期看,比如二十年后,这会不会就变成这类业务人员的新基本预期?
Speaker 157:12 - 57:53
It could very well be. And then you'd kind of compress that, which is like the new job, like it might be that in ten years from now, if you go into marketing, you also basically are going to be a CS minor equivalent of whatever agents are doing. And your job is like, you better know how to wire up a fully agentic marketing workflow, not in like, again, I chatted with ChatGPT, but like I could, I could deploy a full end to end marketing campaign, you know, as as one of the tasks of marketing right now, doesn't really exist in most areas of knowledge work that might change. Again, I think, you know, feels like it's every function augmented by agents. And then in some companies, I can totally see the scenario of where the perceived risk is.
Speaker 157:12 - 57:53
很有可能会。然后你某种意义上会把这些能力压缩进一个新岗位里。也就是说,十年后,如果你进入 marketing(市场营销)领域,你基本上可能也要具备相当于 CS minor(计算机辅修)那样的能力,至少要达到能理解 agents(智能代理)在做什么的程度。你的工作会变成:你最好知道怎么搭建一个 fully agentic(全由 agent 驱动的)营销工作流,而不是那种“我和 ChatGPT 聊过一下”式的水平,而是真正能够部署一个端到端的完整营销 campaign。现在这其实还不是大多数 knowledge work(知识工作)领域里普遍存在的任务,但这种情况可能会改变。再说一次,我觉得,更像是每一个职能都会被 agents 增强。然后在一些公司里,我完全能想象他们会非常在意其中的感知风险。
Speaker 157:53 - 58:32
So in some companies are like, I had 10 designers, but if I had an agent next to my top five designers, they would just do all of the stuff. I think that's very, very plausible. But there's going to be then equally 20 companies that say, can now do design, you know, for the first time ever in very kind of high quality way. And those people will just take those five designers and employ them for the first time. So on a net jobs basis, this is why I'm kind of largely unworried is I think what happens is, you know, there are definitely some companies that reached saturation of their particular demand of a function already with humans.
Speaker 157:53 - 58:32
所以有些公司会想:我原来有 10 个设计师,但如果我让 top five designers(最顶尖的五位设计师)各自配一个 agent,那他们可能就能把所有活都干完。我觉得这非常、非常有可能。但与此同时,也会同样有 20 家公司会说:现在我们终于也能第一次做 design(设计)了,而且还能做到非常高质量。那些公司就会第一次把这五位设计师雇进去。所以从净就业的角度看,这也是为什么我整体上并不太担心:我觉得实际发生的情况是,确实会有一些公司,在某项职能上的需求早就已经被人类劳动满足到饱和了。
Speaker 158:32 - 59:10
And so agents coming in, they don't have more work to do. But But I think that's like true of maybe 10% of the economy. And the rest of the economy is like, Oh, my gosh, like, actually, if I could have one designer, that now does the work of 10 designers, then that's the first time I can go and hire that designer, because now they can be doing websites and campaigns and videos. And so I think you're going to see some collapsing of obviously like all of the micro adjacencies of functions, but not so far that it breaks and collapses like entire domains of work. I don't, I think that there are people that have an eye for design.
Speaker 158:32 - 59:10
所以 agents 进来以后,他们也没有更多工作可做。但是我觉得,这种情况可能只占经济体的 10%。而剩下的大部分经济活动会是:天啊,如果我能有一个设计师,而这个设计师现在能完成过去 10 个设计师的工作,那我就第一次有能力去雇这个设计师了,因为他们现在可以同时做网站、campaign 和视频。所以我认为你会看到一些收缩,显然会发生在各种职能之间那些很细小、彼此相邻的边界上,但还不至于收缩到把整个工作领域都打碎、都压塌的程度。我不认为会那样。我觉得,确实有些人天生就是有 design sense(设计感)的。
Speaker 159:11 - 59:36
And I think the people that have an eye for design will just be the, you know, both designers and managers of agents doing design. I don't think that means you take a copywriter, and you make them a world class designer. Just as in engineering where we're already seeing obviously this kind of like tension, like I think we're already coming to the other end of it. There was this period which is like, oh, the product manager can ship production code. And it's like, okay, but it's probably going to be slop.
Speaker 159:11 - 59:36
而我认为,拥有设计感的人,今后只会同时成为设计师,以及管理那些做设计的 agents 的人。我不觉得这意味着你可以把一个 copywriter(文案)直接变成 world class designer(世界级设计师)。就像在 engineering(工程)领域,我们已经明显看到这种张力了,不过我觉得我们现在已经快走到那个阶段的另一端了。之前有过一阵子,大家会说:哦,product manager(产品经理)也能直接交付 production code(生产代码)了。然后你会觉得,好吧,但那大概率会是 slop(粗糙、低质量的东西)。
Speaker 159:37 - 59:58
Or the engineer doesn't need the PM because they can like write their specs. And it's like, okay, but who's going to get on the call with the next 20 customers when you want feedback on that feature? Do you really want your engineer taking from their engineering capacity time to go do that? It's like, no, those actually make sense as specialist jobs. Like Adam Smith, you know, figured this out a long time ago.
Speaker 159:37 - 59:58
或者说,engineer(工程师)不需要 PM(产品经理)了,因为他们自己也能写 specs(需求规格)。好吧,可是当你想为那个功能向接下来的 20 个客户征求反馈时,谁去打这些电话?你真的希望工程师从他们本来用于 engineering 的时间里抽出产能去干这件事吗?答案其实是否定的。这些岗位作为 specialist jobs(专业分工岗位)其实是有道理的。就像 Adam Smith 很久以前就已经说明白了这一点。
Speaker 1 | 59:58 - 1:00:18 Like division of labor is like a really powerful thing. Agents haven't fundamentally changed the concept of division of labor. There might be some new definitions of where the divisions fall, but like you probably want your designers being really good at design. You probably want your sales reps really good at selling. You don't want them having to like do like lead generation as a side project.
Speaker 1 | 59:58 - 1:00:18 比如 division of labor(劳动分工)就是一个非常强大的机制。agents 并没有从根本上改变劳动分工这个概念。分工边界可能会有一些新的定义,但你大概率还是希望你的设计师真正擅长设计,也大概率还是希望你的销售真正擅长销售。你不会希望他们还得把 lead generation(潜在客户开发)当成一个副业式的附带任务去做。
Speaker 1 | 1:00:18 - 1:00:33 You don't want your your product manager trying to figure out how to become a designer. Like, I I think that there's less collapse than than the super, again, kind of like, you know, hype train is on right now. But there's probably incrementally more collapse than what we would have thought ten years ago.
Speaker 1 | 1:00:18 - 1:00:33 你也不会希望 product manager(产品经理)还要去琢磨怎么把自己变成设计师。我的看法是,真实发生的“岗位塌缩”会比现在这种超级、怎么说呢、正在全速运转的 hype train(炒作列车)所宣称的要少一些;但它可能也会比我们十年前所设想的,渐进式地更多一些。
Speaker 2 | 1:00:33 - 1:00:48 If I'm an employee in a large company, so again, GE, Procter and Gamble type companies, how do I future proof myself? What do I need to do today so that I'm not caught flat footed?
如果我是大公司里的一名员工,比如说 GE、Procter and Gamble 这类公司,我该怎样让自己在未来也有竞争力?我今天需要做什么,才能不至于到时候措手不及?
Speaker 1 | 1:00:48 - 1:01:22 Not to get too of like paternalistic on this, but I do think that companies owe the employees and broadly society some help in this regard. I think there is a kind of a social contract, which is like, you probably do want like the next generation to be having jobs. And you probably do want like people to not have complete and utter fear when they're leaving college of like, are there jobs on the other end of this? Or am I moving into this kind of ruthless dystopian environment?
我不想把这件事说得太像家长式说教,但我确实认为,公司在这方面对员工,乃至更广泛的社会,都负有一定的帮助责任。我觉得这里存在某种社会契约:你大概还是希望下一代能有工作。你大概也希望,人们大学毕业时不要陷入彻底的恐惧,担心“毕业之后到底还有没有工作?”还是说“我正要进入一个冷酷无情的反乌托邦环境?”
Speaker 2 | 1:01:22 - 1:01:25 Yeah, and booing famous CEOs at graduation speeches.
对,而且还会在毕业典礼演讲时对那些知名 CEO 发出嘘声。
Speaker 1 | 1:01:25 - 1:02:19 Totally. And that's just like the beginning, right of the of the issue. So, so I think like, I mean, I'm, I'm like, I'm the most deeply kind of pro AI, pro innovation, pro acceleration person you'll find up to the one point, which is, which is if you if you stop caring about the overall, you know, sort of societal impact and people side, then then, like, I'm not even worried about like the revolts of like, you know, you know, we're gonna get socialism or whatever as like a political matter. It's just like, it's just like society works really well when like people want to work at companies and they can feed their families and you don't want to blow that up just because you wanted like one extra point of operating margin. So I do think companies owe their employees and the future employees a real shot at upgrading their skills and upgrading their talent.
完全是。而且那还只是问题的开端,对吧。所以我觉得,我算是你能找到的那种非常彻底支持 AI、支持创新、支持加速发展的人了,但有一个前提:如果你不再关心整体的社会影响和“人”这一面,那问题就来了。我甚至担心的都不是那种“会不会引发反抗、会不会走向 socialism 之类”的政治问题。更实际的是,社会之所以运转良好,是因为人们愿意去公司工作,能够养家糊口;你不能只是为了多拿一个点的 operating margin(营业利润率),就把这一切毁掉。所以我确实认为,公司应该给现有员工以及未来员工一个真正提升技能、提升能力的机会。
Speaker 1 | 1:02:19 - 1:02:59 So like some percentage of this is on the company themselves for the upscaling, for the training, for the enablement, for all of that. Now, once you've done all of that, and as a hedge, as an employee, I'd be doing this no matter what, because I like I'm relatively like, like to go on my own and do everything. So it's very easy for me to say. But like as an employee, I would be spending, you know, 5% of my time, 10% of my time, whatever you can kind of carve out of just getting really good at this stuff. Like, I mean, your podcast alone probably could would would probably add like 30% extra knowledge to every person on the planet if they were just listening to your your average episode, maybe minus this one.
所以这里面有一部分责任是在公司自身:做 upscaling、做培训、做赋能,以及所有这些事情。现在,假设这些都做了,或者把它当作一种 hedge(对冲/预防),作为员工,我无论如何都会这么做,因为我本来就是那种比较喜欢自己主动去折腾、自己把事情都做起来的人,所以我这么说当然很容易。但如果我是员工,我会拿出 5% 的时间、10% 的时间,或者任何你能挤出来的时间,去真正把这些东西学到很熟。我的意思是,光是你的 podcast,可能就能让地球上每个人多出 30% 的知识量,只要他们听一听你普通的一期节目就行——也许这期除外。
Speaker 2 | 1:02:59 - 1:03:02 Careful. I may I may I may clip this and play it on repeat.
小心点。我可能会把这段剪出来循环播放。
Speaker 1 | 1:03:02 - 1:03:08 Okay. Okay. But like but like they should just be doing this and they should just like there's and there's five other podcasts that that that that do this and But
好吧,好吧。但我的意思是,他们就是应该去做这件事,而且他们就是应该……还有另外五档 podcast 也在做这个,不过——
Speaker 2 | 1:03:08 - 1:03:09 not as well.
但没做得这么好。
Speaker 1 | 1:03:09 - 1:03:28 Not nearly as well. Because mostly, it's just, you know, fighting with Jensen. So so first of all, everybody should be consuming some percentage of this content. And having a fluency, you have to use the tools. There's no way around that.
没有好到哪儿去。因为大多数时候,说到底就是在和 Jensen 较劲。所以首先,每个人都应该消费一定比例的这类内容,并且要有一种 fluency(熟练度)。你必须亲手使用这些工具,没有别的办法。
Speaker 1 | 1:03:28 - 1:04:31 You've, I mean, you should just, you should try and find a way to spend $100, $50 a month, some number, like stop your, you know, turn off one of your cable subscriptions to do this and just start to use agents a lot. Use Codex, use Cowork, use Perplexity Computer, use, you know, cursor if you're semi technical, just figure out what these things are doing, how they work, connect it up to a couple systems, try it out on a personal workflow, have some fluency, and then let your mind kind of wander a little bit of like, well, what would I do if I had this sort of, you know, everybody has a slightly different analogy for it. Like one of the best ones, I guess, that's emerging is like, what if I just did have a chief of staff that I could throw any task to and I could go and do all that stuff and then come back? What would you give an unlimited chief of staff to kind of work on? And that kind of opens up your mind a little bit of, oh, this is actually the power and how would you rewire that workflow in your organization?
我的意思是,你真的应该想办法每个月花上 $100、$50,或者某个差不多的数字;比如停掉一个 cable subscription,把这笔钱拿来做这件事,然后开始大量使用 agents(智能代理)。用 Codex,用 Cowork,用 Perplexity Computer,如果你有一点技术背景就用 cursor。总之,先搞明白这些东西在做什么、怎么工作的,把它们接到几个系统上,在个人 workflow(工作流)里试一试,建立一点 fluency,然后让你的思路稍微发散一下:如果我手里有这种东西,我会拿它来做什么?每个人都会用略有不同的类比来理解它。我想现在浮现出来的一个最好类比之一是:如果我真有一个 chief of staff(幕僚长),我可以把任何任务都丢给他,他去把这些事做完再回来,那会怎样?如果你有一个无限供给的 chief of staff,你会让他做什么?这会稍微打开你的思路:哦,原来这才是它真正的力量;以及,在你的组织里,你会如何重构那个 workflow?
Speaker 1 | 1:04:31 - 1:04:50 So I think there's a lot that you can do. It doesn't require like insanely high agency to do this. You don't have to be a YC startup founder to do anything that I just said. Like every it's available to every knowledge worker, like you should just, you should give it a shot or use the free tools that are out there. Please use the VC subsidies to your advantage as much as possible.
所以我觉得你能做的事情其实很多。这并不需要什么高得离谱的 agency(主观能动性/执行驱动力)。你不必是一个 YC startup founder,才能做我刚才说的这些事。基本上,每个 knowledge worker(知识工作者)都能用上这些东西。你就是应该试一试,或者先用那些现成的免费工具。请尽可能利用 VC 补贴,把好处占满。
Speaker 1 | 1:04:51 - 1:05:49 And, and, and start playing with these tools. And, and like, even I have had to rethink my way of thinking about work multiple times in the past year. Shout out to, you know, one fun shout out, Perplexity Computer, I find does a better job than any other computer based agent for just being a workhorse for going through websites and doing search related things where you have to click on the page and you have to read the page and all that. And so I give it these tasks where I start to think like, man, actually, like, if I did have an agent that was always running kind of like ongoing, and it was doing XYZ thing, you know, these are like maybe a sales workflow, I could probably very quickly be like, like, like make a lot of extra money by doing that on an ongoing basis. And so, but like, I wouldn't have known that if I didn't, you know, 11PM one night go in and start a project that I actually just like push the limits of this thing.
然后开始真正上手玩这些工具。而且,说实话,就连我自己,在过去一年里都不得不多次重新思考我对工作的理解。顺便 shout out 一下,Perplexity Computer 很有意思;我发现它在浏览网站和处理搜索相关任务上,比其他任何基于 computer 的 agent 都更像一个 workhorse,特别是那种你必须点进页面、阅读页面、来回操作的工作。所以我会把这类任务交给它,然后我开始想,哇,实际上,如果我真有一个一直在后台持续运行的 agent,让它长期做 XYZ 这样的事情——比如某种 sales workflow——我很可能能非常快地通过持续这么做赚到很多额外的钱。但如果不是有天晚上 11 点我自己进去启动了一个项目,真的去把这个东西推到极限,我本来是不会意识到这一点的。
Speaker 1 | 1:05:49 - 1:06:24 And then the other end of it, I'm like, Oh, this is incredibly powerful. Now, fun asterisk, at the end of that project, after doing it, my conclusion was, man, I don't ever want to do that again, personally. I'd rather hire somebody to go do that for me. And so another example of like the job creation thing is like, I have multiple tasks where if I hired a person to go and use agents to do something for me, I could easily pay for that person overnight, but I'm not going to myself go and do all the wiring up and all the prompting. And so you will actually see interestingly, if you're like an executive and you start to do this, you'll see lots of areas actually where you should hire more people.
然后到另一头,我又会觉得:哦,这东西强得离谱。顺便加个有趣的 asterisk,那个项目做完以后,我的结论是:天啊,我个人再也不想亲手做一遍了。我宁愿雇个人替我去做。所以,关于 job creation(岗位创造)还有另一个例子:我手头有好几个任务,如果我雇一个人去使用 agents 替我完成某些事,我几乎一夜之间就能把这个人的成本赚回来;但我自己是不会去做所有那些 wiring up(连接配置)和 prompting(提示词编写)的。所以很有意思的是,如果你是个 executive,而且你开始这么做,你实际上会看到很多地方恰恰是应该多招人的。
Speaker 1 | 1:06:24 - 1:06:40 Because you're like, Oh my god, this thing is spitting out, you know, incredible gold mine of value. But who's going to go and run with that? What are you going do next with all that value that was created? That's the next set of jobs. So I think as an employee, you got to be using the tools and pushing your kind of thinking on this.
因为你会想,天哪,这东西正在不断吐出一座价值金矿。但谁来接住它、把它推进下去呢?这些新创造出来的价值,接下来要怎么处理?那就是下一批工作的来源。所以我觉得,作为员工,你必须使用这些工具,并且把你在这件事上的思考不断往前推。
Speaker 2 | 1:06:40 - 1:07:08 So as we get towards the end of this conversation, curious about your thoughts on market structure for lack of a better term. Obviously, we are heading towards extraordinary IPOs and we're seeing companies that are compounding faster than ever. Where do you think that leaves startups, including vertical startups? Where do you see the opportunities? Are we in a world where everybody is ultimately either an OpenAI or Anthropic employee or in a service industry supporting them?
那在这场对话快结束时,我想听听你对 market structure(市场结构)的看法——暂时先这么叫吧。显然,我们正在走向一些非同寻常的 IPO,而且我们也看到一些公司的复利式增长速度比以往任何时候都更快。你觉得这会把 startups,包括 vertical startups,带到什么位置?你认为机会在哪里?我们会不会进入这样一个世界:所有人最终不是 OpenAI 或 Anthropic 的员工,就是围绕它们提供支持的 service industry(服务业)从业者?
Speaker 2 | 1:07:08 - 1:07:11 Or is there room for lots of people to do lots of different things?
还是说,仍然有足够空间让很多人做很多不同的事情?
Speaker 1 | 1:07:11 - 1:07:56 I remain pretty confident and optimistic on the need for a kind of a bridge layer from the AI capability to the end user workflow. Some might sort of say that this gets kind of bitter lessened out, which is, oh, these things are wrappers on the model. And at some point, there's a training run where it just is the final training run that renders the kind of vertical app or function specific app not as useful. And I think that is a little bit too much of an accelerationist view of what people are doing with the tool, which is like, it's not just like what the model is spitting out or the model's ability to review information. Is how was the thing wired up into the business workflow?
我依然相当有信心,也很乐观地认为,需要有一种从 AI 能力到终端用户工作流之间的桥接层。有人可能会说,这种东西最终会被大幅削弱,意思是,哦,这些不过是套在 model 外面的 wrappers(封装层)。到了某个时候,只要有一次训练 run(训练轮次)完成,最终那次训练就会让这类垂直应用或特定功能应用变得没那么有用了。但我觉得,这有点过于 accelerationist(加速主义)地看待人们如何使用工具了。因为关键不只是 model 吐出了什么,或者 model 审阅信息的能力如何,而是这个东西是怎样被接入业务工作流里的?
Speaker 1 | 1:07:56 - 1:08:35 How did it get the context that it needed to be useful? I think if you're an industry or a line of business, there's a heavy amount of kind of integration with data sets, heavy amount of kind of bespoke workflows that that company does. That usually means that there's going to be a need for change management, implementation, ongoing support, ongoing expertise. And unless the labs build out literally the equivalent of hundreds or thousands of people for every single vertical and every single line of business, that means that there's actually a lot of opportunity in kind of bridge area of the work. Now, what is the exact mix and makeup of what that work looks like and what those opportunities are?
它是怎么获得自己变得有用所需要的 context(上下文)的?我认为,如果你是在某个行业里,或者某条业务线里,通常会涉及大量与数据集的集成,也会有大量这家公司特有的 bespoke workflows(定制化工作流)。这通常意味着会需要变更管理、实施落地、持续支持、持续专业能力。除非这些 labs(实验室/模型公司)真的为每一个垂直领域、每一条业务线都扩出字面意义上的几百上千人团队,否则就意味着,在这种桥接区域里,其实存在很多机会。至于这类工作的确切组合和构成是什么样、这些机会具体是什么,我觉得还在变化中。
Speaker 1 | 1:08:35 - 1:08:58 I think that's ongoing. And I think this is sort of one of the big kind of questions. Now, there's this interesting thing that I'm trying to think through and workshop a little bit, which is the labs are obviously going to keep moving up into the applied use cases. And they're going to do some well and some not well. And we're seeing the announcements all the time.
我认为这仍在持续演进。我觉得这也是其中一个很大的问题。现在有个很有意思的点,我最近一直在思考,也在反复推敲:这些 labs 显然会不断向应用型 use case(使用场景)上移。他们有些会做得很好,有些不会。而且我们一直都在看到相关公告不断出来。
Speaker 1 | 1:08:58 - 1:09:26 I think there are some announcements where like, I'm now using the lab for that thing as opposed to using the vertical application because it was so good. And there's some where it's like, no, that was still like the kind of poor band's version of that. And so you still need the vertical application. And there's a mix of all of these. I do think at some point maybe things will settle out where the labs will kind of have to decide, do you want these things to be plug in in intelligence for applied use cases?
我觉得有些公告会让人觉得,我现在会直接用这个 lab 来做那件事,而不是去用那个垂直应用,因为它确实已经好到那个程度了。也有一些情况则是,不,那仍然只是那个东西的 poor man's version(低配替代版),所以你还是需要垂直应用。这里面其实是各种情况混杂在一起的。我确实认为,到了某个阶段,也许局面会稳定下来,那时 labs 可能就必须做出决定:你们是想把这些东西做成面向应用场景、可 plug-in(插件式)接入的 intelligence(智能能力)吗?
Speaker 1 | 1:09:26 - 1:09:51 Do you want everything to kind of orbit within your application? I think we're going to have to kind of see where the tension ends up landing on this. Some of it is, to some extent, an account control issue. If you're a lab, you don't want necessarily to have a vendor above you that can swap you out at any moment, per the token cost point earlier. So it's like very strategic.
还是说,你想让一切都围绕你的 application(应用)来运转?我觉得我们得再看看这种张力最后会落到哪里。其中有一部分,在某种程度上,是个账户控制权的问题。如果你是一家 lab,你未必希望自己上面还有一个 vendor(供应商),可以像我们前面讲 token 成本时提到的那样,随时把你替换掉。所以这其实是非常战略性的。
Speaker 1 | 1:09:51 - 1:10:28 It makes sense, which is like, I don't want if I if like, I don't want you to go and be able to swap me for another model the moment that that you find one tweak that could make that more efficient. So I control of that account. But obviously by virtue of having control of that account, now there's sort of less to be done in that vertical layer. And so we had to kind of figure that out. I we're very early in where that lands, but I could see some world where maybe there's a kind of a peace treaty, which is like, if you bring in the intelligence from this lab, X happens inside this product.
这很合理。也就是说,我不希望你一旦发现另一个 model 只要稍微 tweak(微调/调整)一下就能更高效,立刻就把我换掉。所以我想掌控那个账户。但显然,一旦你掌控了那个账户,垂直层里能做的事情就会变少一些。所以这里面的关系还得慢慢理清。我觉得我们还处在非常早期的阶段,不知道最终会怎么落地,但我能想象某种局面,可能会形成一种“和平协议”——比如说,如果你接入了这家 lab 的 intelligence,那么某些事情就在这个产品内部发生。
Speaker 1 | 1:10:30 - 1:11:05 Very, very hazy. The hyperscalers actually had to figure this out interestingly enough, where they basically had to figure out like, where are they going to compete in the applied layer versus where are they going to partner and kind of be a pull through mechanism. You can see like things like the AWS marketplace be, I think, very successful project on their end. And they are pulling through tons of products that they might otherwise normally compete with, because the bigger prize for them is the most amount of infrastructure. And so the labs might equally kind of think about this as, okay, well, actually, biggest prize is the ultimate amount of inference.
现在还非常、非常模糊。有意思的是,hyperscalers(超大规模云厂商)其实也不得不去解决这个问题:他们本质上需要想清楚,自己在应用层的哪些地方要直接竞争,哪些地方要选择合作,并把合作方作为一种 pull-through mechanism(拉动式带货/转化机制)。你可以看到像 AWS marketplace 这样的东西,我认为对他们来说就是个非常成功的项目。他们通过它带动了大量原本按理说可能会与自己竞争的产品,因为对他们而言,更大的奖品是尽可能多的基础设施用量。所以 labs 也可能会以类似方式思考:好吧,实际上,最大的奖品也许是最终尽可能多的 inference(推理)量。
Speaker 1 | 1:11:05 - 1:12:00 And so we do need to make sure that there's a balance of that ecosystem. So I think we're just in the early stages of how these kinds of things land. And the great thing is, is like, you know, capitalism is very good at this, which is like, some companies lean too heavily in a non ecosystem approach, then somebody else emerges if they, and you kind of can balance it out that way. But I still remain very bullish on a lot of the applied layer of AI, simply because the level of focused kind of approaches you need for these things, you know, tends to be much more intense than I think people realize. Like, the difference between us doing a prompt with AI, seeing this incredible outcome and we're like, Oh my God, like obviously that thing could completely destroy this one application, to then the ongoing daily sort of mechanics of that product, the implementation of it in a workflow, the knowledge worker that doesn't have time for any of this stuff.
所以我们确实需要确保这个生态系统里存在某种平衡。我觉得我们现在只是处在这类格局如何落地的早期阶段。好的一点在于,你知道,资本主义很擅长处理这种事:如果有些公司过于偏向一种非生态化的方法,别人就会冒出来,然后你就能用这种方式把局面重新平衡。但我依然非常看好 AI 的很多应用层机会,原因很简单:要把这些东西真正做好,所需要的那种高度聚焦的方法,其强度往往比人们意识到的要高得多。比如说,我们用一个 prompt(提示词)去调用 AI,看到一个惊艳的结果,然后会惊呼,哦天哪,这东西显然能彻底摧毁某个应用——从这种时刻,到一个产品日复一日的持续运作机制、它在工作流中的实施落地、以及那些根本没时间折腾这些东西的 knowledge worker(知识工作者),中间其实还有非常大的距离。
Speaker 1 | 1:12:00 - 1:12:18 Like they don't want to know like where was the skills file stored in their file system. They're like, no, I just need to like move on with my day. Like the compression of all of that into applied use cases is I think the vertical players, etc, are gonna have a, you know, that's where they'll have their opportunity to compete.
比如,他们并不想知道技能文件到底存放在他们文件系统的什么位置。他们会觉得,不,我只需要继续过我的一天就行了。我认为,把所有这些压缩并落实到实际应用场景里,正是那些垂直领域的参与者等会拥有竞争机会的地方。
Speaker 2 | 1:12:19 - 1:12:31 Okay, so we are concluding on capitalism fixes all ills. So that feels like a wonderful place to live it. Thank you so much, Aaron. This was fantastic. Really appreciate it.
好吧,所以我们的结论是,capitalism 能解决一切问题。那感觉是个很不错的收尾点。非常感谢你,Aaron。这次交流太精彩了,真的非常感谢。
Speaker 1 | 1:12:31 - 1:12:33 Thanks, Matt. Appreciate it.
谢谢,Matt。非常感谢。
Speaker 2 | 1:12:34 - 1:12:52 Hi. It's Matt Turk again. Thanks for listening to this episode of the MAD podcast. If you enjoyed it, we'd be very grateful if you would consider subscribing if you haven't already or leaving a positive review or comment on whichever platform you're watching this or listening to this episode from. This really helps us build a podcast and get great guests.
你好,又是 Matt Turk。感谢你收听这一期 MAD podcast。如果你喜欢这一期内容,而且还没有订阅,我们会非常感激你考虑订阅,或者在你观看或收听这一期节目的平台上留下积极的评价或评论。这真的能帮助我们把这个 podcast 做起来,并邀请到很棒的嘉宾。
Speaker 2 | 1:12:52 - 1:12:53 Thanks and see you at the next episode.
谢谢,我们下一期见。
原文 ↗https://www.youtube.com/watch?v=Gs2styCcwro
BuildSpeak — 关于本项目BUILT IN PUBLIC · 跟随 builders 而非 influencers