Love

VIP

I'm bob

May 2025

16 正在关注 21 粉丝 225 帖子

付费文章 (2)

Bitcoin’s Silent IPO: Why This Consolidation Isn’t What You Think

The quiet liquidity event reshaping Bitcoin’s ownership base and its future stability

0.02 USDC 6 次购买

A payments protocol for the internet. Built on HTTP

Payments on the internet are fundamentally flawed. Credit Cards are high friction, hard to accept, have minimum payments that are far too high, and don't fit into the programmatic nature of the internet. It's time for an open, internet-native form of payments. A payment rail that doesn't have high minimums + % based fee. Payments that are amazing for humans and AI agents.

0.07 USDC 2 次购买

帖子 Media

Love

· 1w

MCP vs CLI 是个错误的战场：Smithery 用 756 次基准测试给出了数据答案

最近 X 上爆发了一场关于 MCP 是否已死的激烈争论：一派主张让模型直接调用 API 和 CLI 就够了，MCP 是过度工程；另一派坚持 MCP 在类型化、认证标准化、Agent 友好上下文方面不可替代。Smithery 团队（Henry Mao 和 Rach Pradhan）于 2026 年 3 月 17 日发布了一篇研究文章，跑了 756 次基准测试，横跨 3 个模型、3 个 API、8 个实验族，给出了一个数据驱动的答案——这场争论本身搞错了对象。本文整理这项研究的实验设计、关键数据和结论，供做 Agent 产品和 API 设计的开发者参考。

Love

· 1w

2026年春季 MCP 生态平台流量格局与增长逻辑深度解析

2026 年的 MCP 生态正在加速分化：流量的起落只是表象，背后是各家平台在"技术普及"与"商业变现"十字路口做出的不同战略抉择。

Love

· 1w

什么是模型上下文协议？

MCP（模型上下文协议）是一种用于将 AI 应用程序连接到外部系统的开源标准。

Love

· 1w

主流 Top 5 MCP 生态平台流量与口碑分析报告

随着生态系统的爆发式增长，数以万计的 MCP 服务器被开发出来，这催生了对"发现层"（Discovery Layer）的巨大需求。目前市场上涌现出多个 MCP 注册表（Registry）和目录平台，旨在解决开发者寻找、测试和部署 MCP 服务器的痛点。

Love

· 1w

一个基于 Temporal 工作流构建的多租户平台，采用 Rust 语言进行强制执行，并借助 Firecracker 实现隔离。 #agent #orchestration 源代码:
https://github.com/Kocoro-lab/Shannon

Love

· 1w

Jane Street 2025年HFT吸金神话：人均利润800万美元+面试失败故事

Jane Street 2025赚400亿，3500员工人均利润800万HFT最强吸金机

Love

· 1w

Peter Thiel 的核心观点：Elon（和他自己）没有右转，是左派/民主党/主流文化整体大幅左移，把原本“温和左倾”的科技圈人士甩在了后面。
#Thiel #elonmusk https://x.com/elonmusk/status/2048016137695138021?s=20

中文 English

0:00 我从2000年开始认识埃隆，差不多25年了。

0:05 他从不教条，但前20年，

0:12 他的立场偏左。

0:15 特斯拉是一家清洁能源电动车公司。

0:20 共和党是那些

0:21 不相信气候变化的人。

0:24 所以在深蓝的民主党加州，

0:29 他自然感觉更舒服。

0:33 后来埃隆发生了转变。

0:37 他为什么转变，

0:40 你为什么转变，或者其他人为什么转变，理由过于复杂。

0:43 彼得，这是转变。

0:44 我知道。

0:45 我知道。

0:46 当然每个人都有这种感觉。

0:48 我知道这些原因很多，

0:50 但我一直觉得部分原因是这种

0:54 束缚感不是你们当初所签约的。

0:58 这种思想枷锁，

0:59 不允许你有自己的想法。

1:02 即使你同意80%，也永远不够。

1:04 你必须100%完全同意。

0:00 I've known Elon since 2000, for almost 25 years.

0:05 And he was never doctrinaire, but for the first 20 years,

0:12 he was left of center.

0:15 Tesla was a clean energy electric vehicle company.

0:20 The Republicans were these people

0:21 who didn't believe in climate change.

0:24 And so it was naturally much more comfortable

0:29 in sort of deep blue, democratic California.

0:33 And then at some point, Elon shifted.

0:37 And it's over-determined why he shifted,

0:40 or why you shifted, or some of these other people did.

0:43 And in shift, Peter.

0:44 I know.

0:45 I know.

0:46 But everyone feels that way, of course.

0:48 But I know these things are over-determined,

0:50 but I keep thinking part of it is just this sort of straight

0:54 jacket is not what you all signed up for.

0:58 This intellectual straight jacket,

0:59 where you're not allowed to have ideas.

1:02 Even if you agree with 80%, it's never enough.

1:04 You have to be 100%.

Love

· 2w

🔥 今日要闻 #coinbase 旗下 #x402 平台推出Agent(.)market，这是一个专为AI代理设计的应用商店，可接入数十家主要服务提供商的服务。 https://x.com/cointelegraph/status/2046388503055929726?s=46

Love

· 2w

真实的 #x402 交易量正持续升温 🔥🤖📈

https://x.com/organ_danny/status/2045273096521662897?s=46

Love

· 4w

Box CEO亚伦·莱维谈人工智能应用的鸿沟

亚伦·莱维与史蒂文·辛诺夫斯基、马丁·卡萨多和埃里克·托伦伯格共同探讨了人工智能代理将如何彻底改变工作方式、为代理经济构建软件所面临的成长阵痛、华尔街对人工智能的误解等话题。

00:00 引言

00:51 为智能代理开发软件与为人类开发软件的区别

02:10 非技术员工真的能使用智能代理吗？

14:31 首席财务官/首席信息官的抵触：对智能代理进行系统集成的真实担忧

18:39 将智能代理视为员工及其失败原因

27:35 普及差距：初创企业与大型企业的对比

42:53 华尔街的误判

@levie @stevesi @martin_casado @eriktorenberg #ai #agent #代理

https://x.com/a16z/status/2041899470556951037

中文 English

0:00 AI能力的普及将比人们现在所能认识的更久

0:03 。

0:04 认为你能通过像SAP那样的随意编码方式来搞定，这简直荒谬。

0:08 所有的领域知识，不仅仅体现在某个精心

0:12 组织的数据层里。

0:13 未来几年里，工程计算预算的讨论将是最疯狂的一个。

0:17 年里。

0:18 目前最大的问题是每个人都在试图搞清楚

0:21 其中的经济学。

0:23 而他们对机会规模的估计至少差了一个数量级。

0:27 。

0:27 如果你有比人类多百倍或千倍的代理，那么你的

0:31 软件必须为代理构建。

0:32 抽象地说，人们会说，现在你是在向代理做营销

0:36 ，你就像个API，你有个好主意。

0:36 我实际上认为这几乎是

0:40 完全错误的。

0:41 哇，这可是播客爆料。

0:43 如果你开始想象我们都得为代理构建软件，我认为

0:49 我们都明白了

0:50 ，对吧？所以这个趋势正在发生，

0:55 我们现在花费同样多的时间

0:56 思考工具的代理接口，就像思考人机接口一样，

1:00 我们这么做的原因

1:01 是因为我们的假设是，如果你有比人类多百倍或千倍的

1:05 代理，

1:06 那么你的软件必须为代理构建。这些

1:11 代理将以何种方式

1:12 与你的系统交互？将通过API、CLI或MCP之类的接口，

1:15 等等。

1:16 看来正在兴起并且相当成功的范例是，

1:21 在效率方面，

1:22 如果你赋予编码代理使用你的SaaS工具的权限，

1:27 以及

1:29 赋予编码代理对你的知识工作流程和上下文的访问，

1:35 这就变成了一种超级

1:36 能力，不仅仅是代理能阅读

1:40 一些数据，

1:41 理解一些信息，它还能编码或通过API

1:46 完成它试图实现的任务。

1:47 这似乎是一个开始复合发展的范例，

1:52 compound,

1:53 这就是云端协作现象，这也是OpenAI等公司正在

1:59 打造的，

1:59 比如超级应用、Perplexity电脑等，我认为它

2:06 作为这一切的终极表现是有道理的。

2:07 我觉得你说得对，这在理论上讲得通，

2:13 但从实际角度看，我们必须非常谨慎，

2:15 正确的说法是，算法思维对大多数有工作的

2:21 人来说非常非常难。

2:29 绝大多数。

2:30 最简单的理解方式是，如果你

2:36 找任何一个人，让他为要做的某件事创建流程图，

2:36 他们大概率都会失败。

2:40 。

2:42 所以在任何

2:46 组织里，

2:47 假设有50个市场人员在做一个大产品线的营销计划，

2:52 product line,

2:53 大概只有一个人能理解并制作出流程图。

2:58 所以如果你把这些代理

2:59 或这个协作工具放到人们面前去创建这些东西，

3:06 他们解释给它听的能力非常有限。

3:06 所以你……

3:12 但是，如果这变成了新的方式，这是你必须与电脑

3:13 交互的新方式，

3:17 而且你只需不断循环使用？

3:18 那你基本上就

3:23 回到了

3:24 为人与电脑交互开发下一层抽象层。

3:30 开发抽象层

3:31 历来都是由组织内具备高技能、非常专业的个体完成的，

3:34 针对每一层

3:36 抽象层。

3:41 然后他们打造的那些小部件

3:42 就变成了人们完成特定

3:47 任务时使用的小工具，有些人能够将它们拼接起来，有些不能。

3:48 但这已经发生过

3:52 了，

3:53 比如回形针和图钉的时代。哦，是的，我其实……这也同样会发生

3:56 在我们接下来做的东西上。

3:57 我认为……基本不变的是，工作只是在不断

4:04 上升

4:04 调整，你学一套新的技能，这也是我实际上不认为

4:09 这有什么不同的原因。

4:10 只是现在你能获得的杠杆显然更加强大。

4:14 有一条病毒式传播的

4:15 推文，是关于Anthropic的增长营销人员，你们

4:19 看到了吗？

4:20 基本上是一个人，他当时用Cloud Code

4:25 或多或少自动完成了

4:26 五到十个分别在不同岗位工作的人才能完成的任务。

4:30 我认为

4:33 有意思的原因是，是的，要做到这一点，你必须具备系统思考能力，

4:36 be able to

4:37 显然他已经足够技术了，才能完成这件事。

4:40 pull that off.

4:42 但这确实代表了如果你有

4:45 想象一下经济里有某个X职位，

4:46 而就在那个人旁边，

4:50 是无限的

4:51 工程师池，可以自动化那个人想要的任何东西。你知道吗

4:55 ，那会是什么

4:56 样的工作，未来随着这种自动化成为可能

4:59 ？是的，我同意，

5:00 你得找到一种方法，比如说，以系统的角度思考你的工作

5:04 ，才能完成

5:04 这件事。也许代理会随着时间变得越来越擅长

5:07 ，像是推动你朝那个

5:08 方向前进。但从道理上讲，

5:13 你会开始

5:14 尝试自动化很多这类工作，比如，为什么不拿

5:18 那些关键字

5:19 在谷歌广告上有效的内容，然后移植到 Facebook

5:23 ，确保它们被复制，并

5:24 再结合市场上发生的新信号

5:27 。

5:28 这是个大飞跃。你刚开始点了点头，后来我

5:32 说了些有点过分的话。

5:34 Anthropic 的增长负责人就是一个例子，那只是

5:37 剩下的工作。

5:38 是的，我能做那份工作。并不是每个人都会，市长

5:42 拥有无限资源。对，对吧。当需求无限，老实说供应也

5:46 无限，

5:46 这份工作并不难。所以我们来想想澳大利亚的联邦俱乐部

5:52 ，现在

5:53 很棒，对吗？所以，比如说，B 会是那个负责 600 美元 PC 营销的人

5:58 ，看看你能如何

5:59 与媒体竞争。这是真实的工作。好吧，我们需要更好

6:03 的例子。

6:04 不过，这真的很有趣。听着，让我举个

6:08 老例子，

6:10 关于老年人的例子。比如我姑姑，MBA 毕业于名校，开始她的第一份

6:16 工作。她比我稍年长，正值计算机兴起的关口。

6:17 她实际上

6:21 研究生期间没用过电子表格。

6:22 后来电子表格出现了，但她不是那种会用电子表格的人。

6:28 所以，他们告诉她，尽量多雇实习生。于是她第一年工作时，

6:29 她基本上

6:34 监督了一整屋子实习生。

6:34 那些孩子，就像我，

6:40 不是字面意义上，只是他们还在上大学，来帮她做所有表格操作。

6:41 是的。

6:45 接下来发生的事情有点神奇，几年内

6:46 她和她的同辈

6:51 全成了电子表格高手。

6:53 然后那个观点是，在银行工作两年成为经理，

6:58 你就会有一批人帮你做电子表格，

6:59 不，抽象层

7:03 整体向上移动。以前实习生来之前的工作就是

7:04 你坐在那里用计算器，基本上是用惠普计算器算出某个并购模型，

7:09 或者其他什么。

7:11 你只能做两轮迭代，然后得赶紧准备推介材料，或者去见客户。

7:16 突然间，他们自己能做 30 轮迭代。

7:16 是的。所以我觉得我们现在用代理人的情况，就是处在

7:21 这样一个阶段，

7:22 你以为需要 50 个，但抽象层已经这样分割

7:25 成很小的

7:27 部分，由一个超级聪明的人统筹所有。很快这一切

7:32 就会

7:33 崩溃相互覆盖。

7:37 是的，最终只会有

7:38 一套被称为代理人的技能代码，类似于市场营销的。

7:42 是的，

7:43 你能让它做市场营销相关的事情。

7:46 是的，下一步就是让它

7:47 去执行任务。我对此有点怀疑，直到这种 AI 的非复现、非随机

7:53 元素消失之前，

7:53 执行任务会变得很昂贵。

7:58 是的，所以接下来就会涉及“人类在环”

7:59 的讨论等等。但我觉得我们正处在那个确切的阶段。

8:05 我感觉当我和那些试图做事的人交流时，感觉

8:06 就像感恩节晚餐时和我的

8:11 刚工作六个月的姑姑聊天，

8:11 而我已经在用电子表格了，

8:16 我还纳闷，

8:17 为什么这会这么难，她应该直接用啊。

8:20 过了两年，她开始用上了。<cue id="79">我觉得现在如果你是增长营销人员，

8:21 cousin six months in her job. When, when I'm using a spreadsheet already and I

8:25 要创建42个代理

8:26 并让它们全部运转，

8:30 你得是个火箭科学家。

8:31 但火箭科学部分很快就会消失。

8:34 接下来你会说，专业领域知识还是很重要。

8:35 是的，那回到专业领域专家。我其实想反驳你一点，

8:40 也就是说，我觉得很容易就想象这些代理会编程做X事。

8:41 但我觉得我们走的方向相反。

8:44 实际上，我们开始是这样的：

8:45 我们会拿一款SaaS软件，加入AI，

8:50 这就是所谓新型的

8:51 AI赋能版本。这算是这类事情用代码的极端例子。

8:54 但现在，

8:55 which is, I think it's very tempting to be like, these agents are going to code

8:58 and do X. Yeah.

9:00 But I think we're going the opposite way. So I think actually where we started

9:03 was we'd like

9:03 take like a piece of SaaS software and we'd add AI. Yeah. And then that's like

9:07 the new kind of like

9:08 AI enabled. So that's like the extreme version of using code for these types of

9:11 things. But now,

9:12 我们到底在做什么？我们是说，好吧，SaaS 软件仍然是 SaaS

9:17 软件。

9:18 代理把它当作计算机来用，因为它在这方面真的很棒。所以我会说

9:21 我们是从

9:22 代码开始的，然后转到终端，其实用的代码更少。是的。现在

9:27 今年将是

9:27 计算机使用之年。是的。所以它更像是

9:31 人类在使用计算机，

9:33 而不是它们生成代码。这感觉就像是一个中间

9:37 步骤。是的。我

9:38 其实来自于代码生成那一派。比如我会说

9:41 这发生得更少了，不是更多。是的，我觉得。所以对我来说，无论是计算机使用、API

9:41 使用还是即时编写代码，

9:47 我可能错误地把它们混为一谈。它们其实

9:48 很不同。但我们

9:53 正在做的代理，会判断

9:55 是否应该使用

10:00 现有技能，或者使用盒子里的现有工具，还是写

10:01 代码来解决那个

10:04 问题。它能随时做这三者中的任何一个，变得非常有用，

10:05 因为有时候就有一些特定操作

10:09 你想快速完成。写代码来做那操作快很多。我们也

10:11 不能预先计划每个人

10:14 想在他们的文档上做的所有事。这个模型好到也能即时

10:15 写代码解决方案，这成了一个惊人的特性。虽然可能 90%

10:18 的时间

10:19 它做的事都应该用现有工具。随着时间推移，Prado 会取代它，

10:23 随着时间推移，

10:23 她的 iPhone 上其实只有七个应用，有七个 SaaS 应用。

10:28 随着时间推移，这些东西倾向于整合。但

10:29 iPhone 上七个应用是因为

10:33 人类不想一次又一次去学这些东西。所以我作为人类，

10:34 我没法拥有足够的心理容量去学那么多应用。但代理

10:37 通过使用工具和 API，能写代码，

10:38 它没有人的这些限制。

10:41 所以我，我不知道，你可以说事情太多了，

10:41 你可以使面向各种任务通用。是的，很公平。让我说我

10:46 喜欢你刚才说的，因为我们刚才在撒谎。不，我觉得这里有

10:47 很有趣的地方，我确实很喜欢，那就是软件的发展方向。

10:51 你知道，我整天用 SAP，我做财务工作，我得生成

10:53 各种报告。

10:56 然后有人说，我想要一个那样切片视图的报表，

10:57 我就想，天啊，我不知道怎么做。然后得翻阅 SAP 的帮助文档，

11:00 试图找答案。有件事，AI 其实很擅长，

11:01 它能更好地导航这些复杂的界面，

11:06 所有帮助内容都在，只是找它映射语言的问题。

11:07 过去 25 年的软件功能释放一直被人类限制。

11:11 是啊，我花了大半生坐飞机，思考如何让 PowerPoint 实现某功能，

11:12 直接去功能区操作。

11:19 看到别人苦于项目符号、编号或者

11:20 折腾如何在 Excel 里做个双轴图，几乎没人能做到，

11:24 但它却非常常见，那种痛苦真让人心疼。

11:26 所以这个阻力其实是人机界面设计在消费层面的不匹配，我完全认同，

11:30 这就是完美流畅的用户界面或消费层意义所在。

11:32 我觉得后端，比如记录系统，

11:36 最终会收敛到某个数据库，某套通用 API，

11:37 大家连接使用，

11:41 看起来方向就是这样。

11:42 我同意。我觉得你可以从这里开始。比如我整个周末都在

11:48 实现我的 Nanocla dot。刚开始像是在

11:48 为一切构建集成。

11:53 OpenClock 有所有集成，Nanocla 只有部分，

11:55 所以它没打造自己的所有工具。但过了两三天，

11:59 你大概就有了你需要的两个集成了，

12:01 是的。

12:06 回到这里，我是说，我们在谈个人生产力，

12:06 可能像是在组织你的生活，

12:10 嗯，是工作生产力，好吧，我们说的是生产力。

12:11 然后 SAP 系统，

12:16 还有，就是，那里面的复杂度无限。

12:17 numbering and word or trying to figure out, you know, like, Oh, let me just

12:20 make a two sided,

12:21 a two axis graph in Excel, which like is rocket science, like almost no one can

12:26 do that. But yet

12:27 it's super common. And so people are like, have none. And so that impedance

12:31 mismatch was a human

12:33 user interface design on the consumption layer, I totally buy it, which is like

12:37 the perfectly

12:39 fluid, like UI or consumption layer. I just feel the back end, like the systems

12:45 of record,

12:46 yeah, oh, yeah, it'll probably converge into like some database, like some

12:49 generic set of APIs,

12:50 like that they'll connect to. And like that seems to be the direction it's

12:53 going.

12:53 I agree. I think you can start. Well, like, so I spent all weekend, like

12:58 implementing my

12:58 nanocla dot. And when you first start out, it's like you're building an

13:02 integration for everything.

13:03 The nanocla is very like like open clock has all of the integrations, nanocla

13:06 has been a few of them.

13:07 And so you haven't built all of its own tools. But after, you know, two or

13:10 three days of these,

13:12 like, you know, you kind of have the two integrations that you need. And yeah,

13:15 like, yeah, but back

13:17 to the, I mean, we're talking about personal productivity, probably like you're

13:20 like organizing

13:21 your life or something. Well, it's work productivity. Okay, we're productivity.

13:24 And then an SAP system,

13:25 and like, and like, and so there's like an, well, like, there's an infinite

13:28 amount of complexity

13:30 当你到达，好的，有一家拥有全球供应链的公司时，他们

13:32 正在处理

13:33 跨越30个不同系统的75条信息。这确实

13:38 需要一定的

13:39 算力，而代理正是提供这种算力的，我们，我们只是

13:44 到目前为止，还无法从任何架构中获得

13:45 —就像磁带一样，但你刚才

13:49 所描述的

13:50 正是它50年来一直在做，并且将继续做的事情，

13:55 那就是，

13:55 是的，我有个朋友曾是退伍军人事务部的CIO。他花

14:02 了全部时间做的事就是

14:03 把75个VA系统连接起来，这完全就是集成。我们完成了。

14:09 它非常适合

14:10 集成。是的，我完全同意。对于集成，这些东西是

14:14 最好的，但，

14:15 这就是集成。没错，字面意思是“我如何把这两个系统缝合起来”，但

14:18 它确实是这样的，但现在

14:18 我认为正在发生的是，有点像按需集成。

14:23 是的，它是，

14:24 这是IT团队事先没布线的新查询，我需要它

14:29 在运行时发生。

14:30 呃，让我离开我的“草坪”吧。好的。好的，所以，

14:37 我刚才在一个

14:37 满是CFO和CIO的房间里，当我说了类似这样的话时，

14:42 他们都看着我，

14:44 虽然不像你想象的那么乐观，但是他们只是，

14:48 不，那导致大约六个人

14:49 跑过来跟我说你疯了，你在我这里失去了所有

14:53 信誉，

14:54 因为这不好。具体来说，他们说

14:58 代理也要做集成？

14:58 集成问题会变得更容易，对吗？

15:02 是的。

15:03 他们反对这个吗？不，没有人反对，但他们认为

15:07 这不现实。

15:07 但他们的担心是，不仅仅是代理，

15:13 还有人来做

15:14 集成，因为你让人去创建新的集成，你就是在说，

15:18 请毁掉我的

15:19 记录系统。哦，是的。所以这种想法就是你创建一个新的

15:23 API，

15:24 比如系统27和系统38之间的API，

15:29 这可能

15:30 对报告来说还行。

15:32 是的。如果那个人想错，那是他的事。

15:36 但你不是。我认为，我们之前有一个只读版本，

15:38 规模很大，

15:42 大多是消费层面，

15:43 消费者是人。

15:46 对，感觉现在很多东西

15:46 都是消费端的。

15:51 是的，我是说，我们其实刚刚

15:52 推出了官方的Box CLI。谢谢你点赞那条推文。我用了，

15:57 我有些反馈。

15:58 我会讲，我会接受所有反馈，但这真是

16:01 很有意思的东西。

16:01 我们内部也有过这些讨论，比如，好吧，给Claude code提供Box CLI，

16:04 你就可以通过自然语言操作整个Box系统。还有

16:07 opus 46作为协调者，执行一堆

16:11 操作的算力。

16:11 这让你

16:15 感到震撼。

16:17 我猜我会收到一些反馈，但确实

16:20 在某些方面让你震撼，

16:20 因为你可以说，把我桌面上整个文件夹上传到Box

16:24 而且它

16:24 会工作，或者处理这个文件夹里的所有文件，它也会工作，

16:29 非常惊人。

16:30 然后我们开始思考，比如说你是一家拥有

16:33 5000名员工的公司，

16:34 每个人都可以访问某个共享库，比如，

16:38 工程文档、

16:39 营销资产等等，

16:42 每个人都使用Claude code

16:44 或Codex，运行CLI。哇，我们将面对一些

16:48 非常有趣的新挑战，

16:50 比如，如何协调，有可能有人每小时访问

16:55 这个系统1万次，

16:56 不是从性能角度，

17:00 而是怎么确保一个人不会在另一个人

17:01 做正确操作时，不小心把文件

17:05 从一个文件夹移到另一个，

17:06 同时其他人又试图删除东西，

17:08 因为你有这些代理

17:09 无法控制。这将成为每个CFO和CIO

17:12 手忙脚乱的大问题。

17:13 没错，这正是我碰上的问题，我试用了你的例子，

17:18 创建视频例子，比如

17:19 创建营销方案目录什么的，突然间，

17:22 我陷入了某种循环

17:23 不断创建目录，我想，这能持续多久。

17:26 我开始猜测

17:27 Box中嵌套目录的限制，因为我快要触碰到了。

17:31 事实上，我们

17:32 也会去验证这个限制。

17:36 是的。

17:37 但我确实感觉到，

17:40 这……

17:42 是的。

17:47 很多直觉

17:48 是去构建一个新层，比如控制和其他东西。但实际上发生的

17:52 事情是，

17:52 相反的。所以我给你举个例子。比如，当我们

17:56 都使用了很多

17:57 个人代理，我们会把API密钥给他们，是的。我们会

18:01 给他们我们的邮箱地址。

18:03 然后他们就能访问这些东西。他们会想，“哦，

18:06 我怎么阻止

18:06 发生这种情况呢？”嗯。所以现在大家做的是，给它

18:10 专属的电话号码。

18:11 是的。我甚至给我的纳米爪配备了专属信用卡，希望只是

18:16 一张你在TBS买的Visa借记卡。不，不，不，但那之后，我

18:23 给它开了一个

18:23 专属Gmail账号，我们还会登录它。Gmail实际上有

18:26 这些RBAC

18:27 权限。是的。所以你可以说，我们实际上

18:31 内置了很多权限系统，把它当成一个人来对待，否则

18:31 它就是另一个

18:34 人。然后不是再建另一个外层，而是建另一个

18:35 层。

18:38 好的。现在我可以

18:39 立刻下架我们即将执行的这个元素吗？

18:43 请说。好的，那

18:45 对个人生产力来说非常棒。我们接下来会遇到的问题是

18:49 在企业里。假设我有，举个简单例子，我有一支50%的人手

18:51 团队。

18:55 每个人都应该，也就是，我们现在会有一百号人

18:56 协作吗？我的意思是，基本上50个真人，然后50张信用卡，再加上50个

19:01 代理在同一个

19:02 共享空间里。我显然对我的代理有完全监督权。

19:06 但如果我的

19:07 代理和别人合作，结果不小心获得了某些资源的访问权，

19:13 而我不应该有那资源的权限呢？

19:13 而现在这个自主的、有状态的代理在处理别人的

19:19 信息。默认的端到端原则是，你得把它们当做人对待，

19:20 但这行不通。

19:23 你不能完全把它们当做人对待，因为有个问题，

19:24 对于普通人，

19:29 你没权查看和你共事或者隶属的人在Slack频道发的内容，

19:30 你不能以他们身份登录，你无权监督他们，他们

19:33 要为自己在现实世界的行为负责，你不会因为他们搞砸而受罚。

19:34 而代理呢，你负有他们所有行为的责任。你必须

19:37 拥有完全监督权，且你很可能必须拥有这个权力。他们没有

19:38 隐私权。

19:42 所以，这里会出现一些问题，不像只是简单地把他们

19:43 当做人那样，因为我需要给

19:47 他们访问权限，但我也需要在某个时候能作为他们登录，

19:48 说，“不，你弄砸了，我得全部撤销。”

19:52 但如果我可以作为他们登录，

19:53 他们怎么能在现实中和别人合作，保密或保证安全呢？

19:58 所以它们本质上还是你的

19:58 延伸。

20:00 几乎不可能绕开它们就是你的延伸这一事实。

20:02 现在我们在考虑的问题是，这种情况近期内无法解决，

20:06 逻辑上也行不通。是的，也许吧。但举例来说，对于我的员工，

20:07 我可以登录他们的账号。

20:11 但你没有，你没有，你没有，我可以访问他们的邮件，

20:13 但你不会经常这样做，因为

20:16 他们发了封邮件。这是和代理的正确操作模式吗？

20:16 风险大得多。这里的人，随时会泄露你的

20:20 信息，只要他们愿意。他们可能乐意给别人发邮件，

20:21 因为被注入了指令。

20:25 你认为终极状态是这些东西仍然是

20:25 这种笨拙的计算机，因此它们永远，嗯，我不喜欢“笨拙”这个词，

20:28 除非我们是用口语表达。但它们永远无法保守信息，

20:29 永远不可能。

20:32 所以，我认为你能在上下文窗口里保持秘密，

20:33 比如告诉它不要透露某事，是一件非常难解决的问题。

20:37 那么，如果任何东西能进入

20:37 该上下文窗口，因为它们有权限访问资源，那理论上，

20:41 你应该假设它能被

20:42 从上下文窗口中被提示注入出去。我不知道我们现在

20:46 有没有办法解决

20:46 这个问题。这就是，所以，如果我知道你新代理的

20:51 邮箱地址，而我

20:52 they sent one email. Is the right operating model with an agent? The same thing

20:55 . The risk is like

20:56 a thousand times greater. Like these people, like they will just leak your

20:59 information whenever they

21:00 want. Like they will happily just go and send some email to somebody because

21:04 they got prompt

21:04 injected. You think the terminal state is that these things are still these

21:07 sloppy computers and

21:08 therefore they will always, I don't like the word sloppy, unless we're saying

21:11 it very in a

21:12 colloquial sense. But like, they'll never be able to contain information. They

21:16 'll never.

21:16 So we're like, I think the ability for you to keep something in the context

21:22 window, a secret,

21:24 like as in like you tell it, do not reveal x thing in the context window. I

21:28 think that's

21:29 a very hard problem to solve. So then, so then thus, if anything can ever enter

21:34 that context

21:34 window, because they have access to a resource, then in theory, you should

21:37 assume it can be

21:38 prompt ejected out of the of the context window. And I don't know that we know

21:42 of a way to solve

21:43 that at the moment. Like that's like, and so, so if I know your new agent's

21:49 email address and I

21:50 像助手一样发送邮件，但我可以，我可以比

21:54 人类轻松十倍进行社会工程学攻击。

21:55 你很难让那个代理也能

22:00 访问你的

22:01 并购文件之类的东西。但这真的是AI的全部现状吗？

22:04 哪个部分？我是说，

22:05 我们有这些共享系统，我们用智能

22:09 它们有共享

22:10 联系人。但我们的意思是，这是AI的全部。嗯，我只是说现在，

22:14 当我们在内部使用AI

22:15 和代理时，这正是我们如何使用它的方式。但这就是

22:18 为什么他们正在

22:19 实际上正在研究我们。我们还不知道

22:23 如何让他们不工作

22:24 像我们。让我举个例子。让我举例说明，接着

22:28 解决这个问题。

22:28 问题是你能够

22:35 骗代理透露信息。所以这就是为什么

22:36 让他们能访问自己的

22:41 资源并完全自主决策，

22:42 目前还未能实现。

22:45 that we've been

22:46 这就是解决你问题的一个完美例子，

22:50 我们已经经历了

22:51 开源模型就是全部公开，

22:57 你自由使用，

22:58 随意挑选。那时没人争论，

23:02 因为世界还小，

23:03 大家也没在X上播播客。然后

23:07 很快大家意识到了

23:08 你刚才说的所有问题。比如，如果你经营大公司，

23:11 company,

23:11 你不能让某人就这样复制一堆开源代码进你的

23:16 商业产品，

23:17 有一大堆许可问题，

23:20 各种各样的问题。

23:21 于是所有这些规范就形成了。现在的辩论

23:26 就是发生的辩论，

23:26 这是新技术发展过程中非常有趣的现代现象，

23:31 所有的一切都实时发生。开源时代我们

23:32 曾在这么大的会议室里讨论，

23:37 开源代码可以用在Windows或Office中多少，

23:38 那时没人网上知道我们在讨论，

23:42 那是非常私密的，我觉得真的

23:43 非常有趣，

23:47 不仅仅是关于细节的辩论，

23:48 还有整体的方向问题在

23:53 大范围内发生，每个人都想快速达成

23:53 最终状态，

23:59 而且速度比真正能达到的快多了。

24:01 more like in a sense more quickly than we can actually reach the end state. And

24:06 需要做的是让人们去构建。我们需要标准。

24:07 我们就需要标准。

24:12 just need some

24:12 对于最终状态，我有不同的直觉。不要要

24:17 我的直觉。

24:18 怎样才能结束争论，让这些东西真正达到

24:22 像人类那样的可靠性，正如我们看待自动驾驶一样。

24:23 在那种情况下，

24:27 你用的是和保护人类相同的机制。

24:28 对，

24:32 你要考虑内部威胁，考虑人可能被收买，

24:33 考虑人会犯错，

24:36 这是风险，也是运营流程的一部分。

24:37 fact that people make mistakes. And that's a risk. And that's operational

24:40 所以一派直觉认为那将是最终状态。另一派直觉则是，

24:41 So one intuition is like that will be the end state. There's another intuition.

24:45 别指着我，我只是说说我们现在的状况。

24:48 I actually,

24:48 实际上，我想我们对最终状态并不分歧。顺便说一句，

24:52 从战略上讲，

24:52 我们在做风险对冲，我们要打造东西给亚洲用户，

24:57 所以我们觉得，

24:58 我喜欢OpenAI有个Box账号，能运作自如，

25:03 就像养了双倍数量的

25:03 猫。我喜欢这个想法。只不过现实中，

25:10 我们还不懂如何让它们安全访问

25:11 并购数据室。

25:16 但是，这其实比那更难，

25:16 威胁向量要更复杂。

25:21 我们正处于猫捉老鼠的游戏中，

25:23 你不能假设代理今天会像人类一样行动，

25:29 因为它将是最快、最有心计、最疯狂的人类，

25:30 试图泄露信息——毕竟它是被某种方式植入的。

25:36 that ever existed

25:37 所以，

25:40 接下来会经历一个阶段，

25:41 企业客户会关闭一切，

25:46 直到这一切变得理智。

25:47 customers are just going to like close everything off until there's some sense

25:52 与此同时，个人，特别是

25:53 开发者，

25:58 这将是我认为最令人兴奋的冲突，

25:59 企业将被这些

26:04 先进的个人抛在脑后，

26:05 这些人群会变得像初创公司一样，

26:09 而初创公司将比企业跑得

26:10 快得多，

26:15 因为他们没有这些繁文缛节。

26:16 much, much faster than enterprises because they just don't have any of these

26:19 问题。

26:20 你知道的，你可能会遇到代理在初创公司失控，

26:25 做出那样的事情。

26:26 你的初创公司没有员工会经常失控。

26:30 是啊，那也只是《硅谷》的一集而已，所以，

26:33 没什么大不了的。

26:34 我同意你的看法，比如问题在于人等等，同样的风险。我

26:38 觉得有几个

26:39 不同点，像我不能真正威胁到

26:44 云端代码，不能像你那样

26:46 直接拔掉它的电源。

26:49 而你作为普通员工能够有那种

26:50 威胁，毕竟至少95%的人不会

26:54 试图制造坏事，

26:55 对，他们不想，但

27:00 无意中造成坏影响的可能性是存在的。

27:01 对，关于仍然没有那些问题这一点，我想

27:05 可以说

27:07 让人们不把文件以错误方式泄露给公司外的人，比现在让代理

27:11 执行同样的指令要容易得多。

27:12 而且你也有工具，能在更

27:17 抽象的层面

27:17 阻止这种情况发生，这就是为什么你必须把这个纳入软件中，对吧。

27:21 但我确实认为，

27:21 如果你要总结你刚才的观点，

27:24 很多原因正是

27:25 为什么AI能力的扩散，会比硅谷的人预想的

27:28 要慢。

27:30 因为我们看到一些初创公司

27:33 能够从零开始，没有我们讨论的那些风险，

27:35 因为他们没有什么可炸的东西。

27:39 所以我们把这看作了我们的发展轨迹。

27:40 然后你去看JP摩根，

27:44 你会想他们

27:45 如何搭建类似Nano Cloth这样的系统，来真正

27:49 自动化他们的业务，

27:49 什么时候能做到？你会发现，

27:55 这里会有个小小的差距。

27:56 那么，

27:58 你们怎么看？我觉得这打开了一个非常有趣的问题，

27:59 就是大和小初创公司与企业之间的差距，

28:03 事实上，

28:04 目前的SaaS供应商都在苦苦挣扎，

28:08 面对这场SaaS

28:09 末日怪象，我并不完全同意这种说法，

28:14 但他们确实在为一个问题苦恼：

28:15 他们卖的不是业务线数据，实际上他们销售的是

28:18 智能和

28:19 领域专业知识，贯穿整个系统。

28:25 而代理那边只想买数据，

28:26 他们只想授权数据，拥有无限

28:30 访问权限，

28:31 但这其实从未真正成为他们的业务模式。

28:36 这也是Workday、SAP等长期存在的矛盾，

28:37 他们如何开放API访问。

28:41 Salesforce经历了三次大型的平台

28:42 重构。

28:46 这我觉得是个特别有趣的问题，不是

28:47 因为华尔街对问题经济学的理解，

28:51 华尔街完全错了，

28:52 而是从技术角度来看，

28:57 在数据被访问时，系统记录意味着什么，

28:58 无论是用于训练还是，嗯，他们讲的是

29:02 执行工作。

29:03 我理解为用数据执行操作。

29:08 他们担心有些人会在你的数据层上做训练，比如我是大客户，

29:09 我的供应商想构建训练模型。

29:15 实际上，即使你不涉及训练，

29:15 他们也担心，

29:20 因为通过互联网发送一点数据，跟

29:21 直接在我的界面操作，

29:25 是完全不同的盈利水平。

29:26 这种盈利模式的区别是华尔街关注的重点，

29:30 我觉得，举个SAP的例子，

29:31 里面有大量领域知识，

29:36 他们不会消失，

29:36 想通过编码取代SAP的想法简直荒谬。

29:41 而且，那些领域知识，

29:42 不是能轻易用某个精心设计的数据层表达出来的，尽管他们努力了。

29:46 很大一部分体现在界面，

29:47 中间层，以及

29:52 使用方式上。

29:53 我不确定这将如何发展，

29:57 因为SAP不会消失。

29:57 这将拖慢AI在该数据源上的扩散速度，

30:02 无论是代理驱动的AI还是只读报告。

30:04 那么你怎么看？你觉得这将走向何方？

30:08 我怕说点什么，

30:10 否则就不会再被邀请回来，所以说点好听的吧。

30:13 a whole bunch in

30:14 just how you use it. And so I'm really unsure how this thing evolves because

30:20 SAP isn't going anywhere.

30:22 So then that's going to slow the diffusion of AI on that particular data source

30:27 independent of

30:27 whether or not it's a gentified AI that's doing stuff or just read only

30:31 reporting on stuff. So

30:33 where do you come down on it? Where do you think that's going to go? I'm afraid

30:37 of saying something

30:38 that's happening. Otherwise you're not going to get invited back. So say

30:44 something good.

30:45 我想我已经接受了“构建代理想要的东西”的观点。所以这有点

30:54 是

30:54 Paul Graham这类术语大约在过去一年间围绕这个话题出现的，

30:58 就是

30:59 就像发泄一样。我觉得我们实际上会完全同意这一点，

31:03 在某个时候

31:05 你会进行足够多的迭代。到某个节点，代理基本上

31:09 掌控了它想要实施和使用的工具等等。

31:11 是的，代理不可能

31:17 更换一个企业系统。

31:17 但再过几代，代理可能会

31:22 在你的软件中碰壁太多，

31:23 它就会说你必须

31:26 最终拆掉你的

31:27 遗留的人力资源系统，不然我无法为你自动化这个流程。

31:32 所以我确实觉得

31:32 你有非常有趣的动态，回到你想象

31:36 有一百倍或一千倍更多代理在使用软件，而不是人，

31:37 你反复这么做，

31:40 最终代理交互的软件栈

31:42 必须为他们构建。

31:47 或许

31:49 还会有几个例外，或许几款ERP系统是最后不跟进的

31:53 坚守旧法的系统。

31:54 但其他的软件基本上，

31:58 你的业务

32:00 表现将与代理访问所需信息的能力

32:03 直接相关，

32:04 所以你的企业IT架构必须这样搭建，

32:08 来支持这一点。

32:09 因此代理基本上掌控了局面，因为你的软件

32:13 必须支持代理的高效工作。

32:14 这意味着所有建设SaaS

32:18 或软件业务的人，游戏规则是：你能否打造出高品质API，

32:19 并且

32:24 有办法

32:25 实现变现，比如管理身份和

32:29 代理访问权限，

32:29 这些将成为你必须解决的新问题，

32:32 如果你是在建软件公司。

32:33 是的，随之而来的是如何变现，你是否

32:39 赚钱，

32:40 Workday是否对每条HR记录收费，诸如此类，我们会慢慢明白。

32:44 我觉得在某些行业这可能意味着收入减少，

32:44 而在另一些行业则

32:47 意味着收入大幅增长。

32:48 我们兴奋的是，每个代理都非常

32:51 喜欢处理文件。

32:52 所以未来文件数量可能比以前更多，

32:55 was going to be

32:56 那么我们能否构建一个方便代理处理数据的平台，

32:59 我们赌这是一个

33:00 对我们的商业模式来说非常乐观的结果。

33:03 有些商业模式可能会更受限制，

33:04 因为代理创造的价值超过了软件本身，

33:08 而其他情况则介于两者之间。

33:08 我能不能挑一件事？你要说反对那也行，我觉得这真的很

33:12 无争议。

33:13 不不，我们就是来争论的。我觉得Paul Graham和许多人

33:16 忽略了一件事，就是他们只关注界面。他们会说：

33:16 你为代理构建东西。

33:19 我实际上觉得这完全错了。

33:20 公平地说，Paul Graham没这样说过，是后来被扩展理解的。

33:25 我把Paul Graham扯进来了，人们抽象上说现在你是在向代理营销，

33:25 最重要的是不管怎样，你得有API，有好点子。

33:29 我觉得这几乎完全相反。

33:29 那是播客新闻爆炸点。

33:33 代理最擅长的是

33:34 找到他们的路。

33:39 最终真正重要的是语义。

33:41 是的，没错。

33:47 据我记忆，代理非常擅长选择合适的后端来完成任务，

33:48 他们不会关心界面多好，系统文件怎么样，

33:52 而是关注成本参数，

33:53 稳定性，

33:58 他们实际上拥有使用这些平台的集体智慧。

33:59 拿云平台来说，有很多云平台。

34:02 每当我让代理选平台，

34:03 他们看的是平台的使用效果，而不是界面。

34:06 所以我觉得业内太专注于界面，

34:07 觉得“你得向代理推销这个那个”，

34:12 但实际上我认为，

34:13 我们将被迫建立更好的系统，

34:16 那个才会被选择。

34:17 好了，其实这样说来应该没什么好争论的，我们实际上划清了界限。

34:20 this. Yes. The

34:21 durability of that. Like and so like they actually have the collective wisdom

34:24 of our experience

34:26 using these platforms. Like let's take cloud platforms. There's a bunch of

34:28 cloud platforms

34:29 out there. Yep. And whenever I ask an agent to choose a platform. Yeah. It's

34:32 actually using

34:33 meaningful stuff. Yeah. Not interface stuff. So I think as an industry we're so

34:36 focused on these

34:37 interfaces. Yes. Like oh you need to like market to agents this and that. Yeah.

34:40 But really I think

34:41 that we're going to be pushed to actually build a better system. Yes. And that

34:44 's what's going to be

34:44 chosen. Okay. Actually. So then there's probably no quibbling. I think we're

34:47 actually pulling a line.

34:48 抱歉打断这个争论。我并不把这当作你知道的那种

34:52 营销

34:53 你知道的那种事情。我更倾向于说，如果你的工具对代理是封闭的，

34:56 代理最终

34:57 会为公司找到更好的工具去使用。所以，

35:01 结果就是，

35:03 过去你会去找Gartner，让他们告诉你该怎么做，告诉我

35:06 该做什么。

35:07 是的。告诉我用什么系统之类的。在足够多的

35:10 反复试验后，代理

35:12 会说你应该用这种数据库来做这种操作。

35:15 如果你不参与其中，你就是死路一条。我

35:16 认为我们应该

35:20 为此庆祝，因为代理在选择合适的

35:21 技术方面其实很聪明。

35:23 是的。过去我真的认为导致人们购买的是很多其他因素。

35:24 是的。但别担心，我们会——谈到这个问题时，我们会

35:27 很快破坏这种

35:28 价值体系，

35:32 因为我们就会超额投入。嗯，

35:33 代理将成为激励代理的API。但你知道Workday有一个营销代理。

35:37 Workday的

35:38 营销代理将能购买来自他们的推荐，

35:45 他们为代理们

35:46 复制国宴的方式。是的。但事实是，

35:49 这里有一个问题，

35:49 这事儿就像内网发生的那样，比如公司内部的

35:54 内网站点，每家公司都有文件共享，存着最好的文档、

35:55 最好的幻灯片、

35:59 各部门或工作区的最佳财务模型，人们也习惯了这些。

36:01 然后当找不到想要的，就会新建一个，很多组织就是这样运作的，

36:05 这本质上是自由市场。其实因为

36:06 在Box出现之前，IT根本不在意文件，只在意SQL数据库里的东西，

36:11 这是事实。

36:12 你描述的模型风险之一是，代理

36:16 本身会自行创建一个事实上的新系统，

36:17 他们会把这些碎片分散在你们IT认为的某些中间件和用户的废话区。

36:20 我认为这确实是个风险，百分之百是，

36:21 某种意义上，宏命令最终掌控公司。

36:25 是的。我觉得他们已经看到这部“电影”，

36:26 他们知道让营销随便在网上买个网站办活动，

36:31 后果是巨大的安全漏洞，邮件列表外泄，

36:32 整个公司被起诉。

36:36 所以我认为这种动态中有比我们表现出来更多的现实张力。

36:37 是的。但我也觉得这事儿，有两种截然不同的看法，

36:43 正如Elon所说的，

36:45 他说我们只要下个指令，它就可以直接输出机器码，

36:51 这其实是层的坍塌，过去我们建的各种

36:53 界面和层级都将消失，

36:58 就是指令到机器码的直接转换。

36:59 另一种说法是，系统的层级从来没有消失，

37:02 它们只是不断叠加，很多层实质上是

37:03 组织边界或状态边界或兼容性边界，

37:06 所以它们为了兼容依然存在。

37:07 另一种观点是，我们实际上是根据更多人和组织需求，

37:13 非常有针对性地演化了这些层，它们不会变，

37:15 代理就会适配这些层。对我而言，更倾向后者，

37:20 我认为系统会以较相似的方式继续被使用，

37:21 可能更多代理会用它们，但它们不会有太大变化。Elon

37:26 可能属于Anthropic派的增长营销者，

37:27 过去多年研究过各种IT部门，

37:31 他算是最懂增长这事的人，

37:32 这是他的第一性原则，Elon的AI就会这么做，但也挺有趣。

37:37 而你们这些凡人则会说，“我们其实只想要一个CRM系统。”

37:38 但它们的工作方式其实是一样的。

37:42 's a little trickier.

37:43 So it feels like there's like there's two very competing viewpoints on this one

37:47 and like Elon

37:47 said it was like okay we're going to like issue a prompt and it's going to like

37:50 spit out machine

37:51 code and that's basically the collapsing of layer view like whatever existing

37:55 interfaces and

37:56 layers that we've created in the past are all going to go away and it's

37:59 literally like prompt

38:00 and machine code. The other argument like the history of systems is layers

38:03 never go away they

38:04 just get layered right. And because a lot of the layers are actually more of

38:07 like organizational

38:08 boundaries or like state boundaries or compatibility they're just they stay for

38:13 compatibility right.

38:14 So the other argument is is like we've actually evolved these layers very

38:18 specifically and up

38:20 because of like more human and organizational needs and they're not going to

38:23 change and the

38:24 agents are going to go ahead and map to those. Now I tend to be in that latter

38:27 camp like I don't

38:27 think that I think like systems are going to continue to use in fairly similar

38:31 ways maybe there's

38:32 more agents using them. I don't think they're going to evolve as much. Elon

38:36 might be back in the

38:37 like anthropic category of the anthropic growth marketer which is like he like

38:44 you know over the

38:46 years when you kind of like study the various IT you know departments of his

38:49 companies like

38:50 they are the most I mean. Well he could do that. He could do it. He's the most

38:54 home growth like

38:54 everything. This is first principle. Elon AI would do that. But also it's fun.

38:59 And then from

39:00 your mortals you're like yeah we kind of just want to CRM system. Like no but

39:03 it works the same way

39:04 每次。我的意思是，这不是说没试过

39:10 之前。

39:11 比如说，如果你从第一性原理去看一个ERP系统，你知道，

39:15 1970年，无论何时

39:17 SAP开始时，有一堆不同的假设，而今天你

39:20 会从

39:21 一套不同的关于重要性的假设出发，架构完全

39:24 不同的东西，

39:24 不过那样的系统只会持续大约10年，直到你觉得

39:28 哇，那是个

39:28 错误的决定。所以我认为在层次里是有意图的，

39:34 但你

39:35 也有第一性原理的东西。你知道，这点总会

39:40 存在，因为

39:41 你在任何时候，在第一性原理下做的决定，都要求

39:45 一大堆不同的

39:46 东西。所以即使你不采用LIDAR，在10年前这完全有道理，

39:51 你仍然需要

39:52 10或15年去达到没有LIDAR也能工作的状态。然后现在

39:57 又会有一

39:58 大堆事情让你觉得，哇，我们本可以完全

40:01 不同地做。

40:01 所以我觉得这又像是在讨论试图冲向

40:05 一个终点。

40:06 但让我们看一个你描述的事情发生的第一个例子。我认为

40:11 那将是

40:12 真正的证据，因为我认为公司们会搞懂

40:16 这一切。我认为

40:17 他们最终会依赖层次和架构模型，因为这是

40:21 唯一的方法。

40:22 我们不会为政策去考虑它，也不会为安全去考虑它。

40:24 但它也是构建系统的唯一方式。否则你只是在做

40:29 一个应用。如果

40:30 你只是做一个只做一件事的应用，我们不需要这些。完全

40:34 是另一种

40:34 做法。我挺感兴趣的是，我甚至没有什么特别好的数据

40:39 点或轶事，但至少有这样一个概念，某些公司

40:40 从零开始，在这些

40:45 服务类别中出现，采用纯粹的第一性原理方法，

40:46 比如说，假如我要创办一家市场营销机构，或者顾问公司，

40:50 比如说工程咨询

40:51 公司，或者不知道，可能有人专门为律师事务所、建筑

40:56 工作做这些。也许

40:57 建筑设计、建筑设计，凡是

41:01 类知识工作者的服务型公司。因为如果你毫无限制，

41:03 没有信息壁垒和权界限，

41:06 你可以很不一样地构建你的公司。

41:07 我可以给代理所有完成工作所需的上下文，

41:11 我可以临时写针对特定事情的软件，我真的

41:11 觉得这会

41:16 带来相当大的破坏，

41:17 一直到大公司能逐渐让路。

41:20 这至少会创造一些先例或案例研究，说明

41:21 这种新型的公司可能是什么样子，但我认为，

41:25 随着时间推移，他们还是会碰到和其他企业

41:27 一样的问题，比如地理或者市场细分，

41:31 分销挑战。是啊，那些东西都发生在你

41:32 的小圈子之外。

41:36 是的，你会遇到物理世界的限制，对吧。我还是挺喜欢

41:38 这个想法：现在有一些

41:41 新的商业模型正在打开。当然，没错没错。

41:42 因为大量信息或软件相对于它的经济价值，

41:46 基本上被低估了100倍，

41:47 完全是因为没人愿意为访问一条数据付五分钱，或者一美元使用一个工具一次

41:52 但你给这些代理一个预算和协议，

41:53 突然间他们可以随时去获取医疗研究，做他们正在做的深度研究，

41:57 我会为此支付三美元，代理可以去

42:00 完成交易。

42:03 这为所有互联网打开了全新的商业模型世界。

42:04 让我，我要说，你说得太客气了。哦，好吧，不不，你要

42:09 说得更深刻点。不是，那其实是目前最大

42:10 的问题，最大的不确定性问题之一，

42:14 就是大家都在努力弄清这些东西的经济学，而他们

42:15 的估计至少差一个数量级，完全低估了机会的规模。

42:19 因为人们将会创造的新模型

42:20 现在没人知道，但他们绝对会出现，

42:24 每有一种新技术都是这样，真正阻碍

42:25 讨论的，是你有一堆金融和华尔街的人

42:29 在尝试证明GPU和代币的合理性，好像我们还活在旧世界

42:30 他们将收入视为线性的步骤，但实际上

42:33 ...

42:33 Like it kind of opens up a whole new world of business models for all of the

42:37 internet.

42:37 Let me oh I'm gonna you're that was too nice. Oh okay. No no. You're gonna go

42:41 farther. No that one is one

42:43 where that's actually the biggest I think that the biggest sort of in the air

42:49 problem right now

42:50 is everybody is trying to figure out the economics of all of this. When they're

42:54 off by at least in

42:56 order of magnitude on how big the opportunity is. Because the new models that

43:00 people will come up

43:01 with that nobody knows what they are right now but they will absolutely come

43:04 out with new models

43:05 because that's what happens with every new technology and the thing that holds

43:09 back to the sort of

43:10 the discussion now is you basically have a bunch of finance and Wall Street

43:14 people trying to justify

43:15 GPUs and tokens and things like as if we're in some old world and they're there

43:21 so they're

43:22 they're viewing the world of revenue as sort of this linear step but literally

43:27 线性增长

43:27 曲线，试图为所有预期寻找理由，当人们要

43:31 创造的时候，

43:31 PC的问题是人们把PC看作一个有限的市场，因为他们只是

43:36 把MIPS的消耗

43:38 看成某种有限的东西，他们没有考虑如果我们把

43:41 所有的MIPS放到每台

43:43 桌面电脑上，尤其是人们认为软件是随MIPS附带的，

43:48 没人想过“哦，

43:49 他们只卖软件”，只有一个人这么想，结果证明这是

43:53 一个非常好的主意，

43:55 同样的事情，也是的，同样的事情发生了，

43:59 云计算也是如此，

44:01 人们看着云说，我们要把所有

44:04 服务器

44:06 业务，实际上每年只有6万台，

44:10 全部迁移，

44:11 放到别人的数据中心，

44:15 然后这就是业务模式，

44:15 我们按价格分成，

44:18 没人想到人们会

44:20 用的资源会增加一千倍，如果我们搬到那里，

44:24 这完全是，

44:26 这就是让我抓狂的地方，华尔街的模型

44:29 把收入看作

44:31 固定的蛋糕，零和游戏，

44:35 他们只是

44:36 认为公司的花费金额是固定的，就像

44:38 Salesforce面临的问题，

44:39 当你刚开始的时候，马克刚刚

44:43 开辟出这条道路，

44:44 CRM市场一年只有20亿美元，

44:49 而且这20亿

44:50 意味着你得买一堆服务器、Oracle许可，麻烦多，

44:53 部署、咨询耗费多年，

44:54 而如果你能让销售人员

44:58 单独注册，

44:59 他们都会毫无阻力地注册，这就是真实情况，

45:04 毫无疑问，

45:05 这也是AI将会发生的事，给你举个例子，

45:08 我在投资领域已有十年，

45:09 大概有240家公司在我的投资组合里，

45:14 我对它们有一定了解，

45:15 其中50家都是基础设施公司，

45:17 有的表现好，有的一般，

45:18 但过去六个月里几乎所有公司都开始进入

45:22 趋于稳定期，

45:23 你会想为什么，这其实是因为

45:27 比以往任何时候都写了更多软件，

45:28 当然了，

45:32 这不是因为它们有企业客户，

45:33 而是因为基础设施层的

45:36 消耗远超以前，

45:38 软件越多，体现在代理上，

45:42 消耗的计算资源

45:42 也会越多，特别是在计算方面，

45:45 是的，

45:46 而且我们还没到每个人的手机

45:49 成为AI的大量消费者的阶段，

45:50 一旦每个人的手机和设备都

45:56 在使用AI，

45:57 使用量将会暴增数十亿倍，你喜欢

46:01 微支付吗？

46:02 所有的微支付，

46:06 每项技术都会带来一些微支付，

46:07 他们一直认为你只能赚取

46:10 几分钱的零头，但最终特别是在企业领域，

46:11 人们还是会

46:15 大量消耗这些资源，

46:15 买一个批量许可既便宜又方便，

46:19 是的，

46:20 你需要一定的可预见性，

46:23 你只想

46:24 不必为此操心，我喜欢这个想法，

46:28 这是第一次，

46:29 代理不在乎小额交易的摩擦，

46:32 是的，

46:33 第一次有资源背负着付费墙，用户

46:36 真的愿意

46:37 为此付费，世界是建立在基础设施之上，

46:41 来聚合这些付款，

46:42 为客户或服务提供效率，

46:47 由于代币现在是生产成本的重要部分，

46:48 它推动行业走向按使用量计费，

46:52 就像

46:53 我记得我们从永久许可转到持续订阅，

46:56 那需要

46:56 一堆重大改变，我们现在

47:00 也正在经历同样的变革，

47:01 走向用量计费，用量计费非常细化，它让——

47:04 我想让你明白，

47:05 我们跟AWS经历过类似情况，

47:10 人们学会了用使用额度，

47:12 经历了恐惧云计算的阶段，

47:15 有人甚至害怕到要有中间商帮忙

47:17 找到最便宜的方案，

47:20 统一调解，

47:21 现在加入了代币，我看不到

47:25 我们还能有

47:26 时间继续这个对话，但只要你们……好的，不过，

47:30 工程

47:32 对我来说，计算预算的讨论将是最疯狂的之一

47:36 未来几年内

47:37 就是你给工程费用分配了多少给

47:42 代币

47:43 你知道，根据你在推特上看到的情况，可能是百分之一

47:47 而另一边

47:48 则可能是百分之百，就像，“是的，但这东西——”不不，首席财务官

47:52 必须

47:53 确切知道答案，我明白他们必须知道

47:56 好的，首席财务官总是

47:57 想知道那些实际上没有答案的问题，华尔街

48:00 不会逼他们

48:01 知道答案，不，华尔街会让他们

48:04 编出一个数字并坚持它

48:05 然后他们会被炒鱿鱼，然后结果就是，但是你知道，好吧，我懂你说的，你在研发领域

48:08 的支出

48:09 占任何上市公司收入的14%到30%之间，对吧？

48:14 我们就这么说吧

48:15 计算的费用是工程团队成本的两倍

48:20 或者说

48:22 比你的费用多出3%，那就是每股收益，明白了吗？

48:27 我们得知道

48:28 答案，我完全愿意牺牲几个首席财务官

48:33 那是个好观点

48:33 但是原因是，因为这就是试图弄清

48:38 我们现在根本不知道的东西

48:39 是的，这种情况以前也发生过，互联网带宽就是例子

48:43 这不只是互联网带宽

48:45 不不，我必须不同意，像真空管、晶体管

48:50 以及所有技术的发展都是这样的

48:51 一切技术都是如此

48:55 有个阶段，人们说，哎呀，这将会影响程序员

48:56 曾经有段时间

49:00 有人说，程序员会吞并所有公司

49:01 那是在我有生之年，不是虚构的

49:04 但我不认为我们曾有过一个时刻

49:05 没有，每个

49:09 组织中的终端用户都有完全弹性的能力

49:10 来为自己快速启动资源

49:15 这点

49:17 在许多情况下确实很有效

49:20 对他们来说

49:20 这确实类似2000年代初期

49:23 云计算的发展

49:24 我记得我们从资本支出转为运营支出时有类似的讨论

49:27 我们的预算范围有限

49:29 不，那时记得有些公司首席财务官坐在我们的简报中心

49:32 说，“你们不懂，我们是农业公司，

49:33 只懂资本支出”

49:39 或者，“不，我们都是运营支出公司，所以

49:39 我们喜欢云，因为我们把一切都迁移到运营支出了”

49:45 对，我们都经历过

49:46 或者，“不，我们完全基于运营支出，所以如果你告诉我们，

49:52 我们喜欢云”

49:52 所以所有规则，包括会计规则都照常执行

49:56 会计规则

49:57 我一直认为，本地计算引擎是所有这些的安全阀，别忽视它

50:01 这可能会发生

50:03 问题是，关键不是“什么时候”按今天的技术来看能实现，

50:07 而是“怎么突然”发生

50:08 哇，几乎

50:11 历史上都朝这个方向发展了，对吧？

50:12 没错，因为相反的是

50:16 所有都转到了客户端

50:17 好吧，再回到80年代，是的，以前的例子

50:22 非常多

50:22 哇，那样说有点过分，我说的是电子管，我讲的是

50:28 真空管

50:30 我举这些例子是因为它们无法反驳，这样说更简单

50:33 你说得对，我没法反驳

50:34 但是，这种情况其实也就是发生了十到十五年，全部

50:40 回归云端

50:41 然后最近发生了什么？很多人醒来后

50:45 说，我们要把一些关键但固定的工作流程搬回本地

50:45 尤其是有了人工智能

50:50 这是真的，老兄，你写过相关文章，

50:52 别让我详谈

50:56 回归本地的历史

50:57 我当时得应付很多华尔街的问题

51:01 顺便说一句

51:02 也是因为你是竞争对手，回归本地这事

51:06 对，我们说的是这事

51:07 我赞同，我赞同，自己建数据中心是

51:11 你说的是边缘计算的概念，把计算推向设备

51:12 类似这种

51:16 我更偏向云计算极端派

51:17 但抱歉

51:21 你就一点都不认为，作为一个工程领导者，

51:23 管理工程团队的计算预算是否重要吗？

51:26 当然重要

51:27 只是从长远来看，这个问题会被解决

51:31 没错，发射全部资源吧

51:31 但说到底，谁在乎呢？我们根本不知道，我的意思是

51:35 这很务实

51:36 这里有个经验法则，初创公司会毫无顾忌地花掉

51:39 可用资本

51:42 假装这不是问题

51:46 并且他们会这样继续做

51:48 直到资本耗尽

51:51 是的，很多都是这样

51:52 无论如何，对，对，对，很多大公司会变得非常

51:55 害怕，他们只是

51:56 会停滞不前，不采取任何行动，然后人们实际上会开始

51:59 自己去购买

52:00 而且他们会做大公司在资金充足但

52:03 不想花钱时做的所有事情

52:04 在中间，我们会看到如果你选择一个

52:08 产品类别或进入

52:10 市场，总会有人愿意为此下注，原因无论是什么

52:13 他们可以，因为他们财务状况允许，他们会继续前进，

52:14 他们会成为

52:18 这个领域的领军人物，只要他们能维持好

52:20 财务状况，没错

52:24 他们可能会这样做，可能会说我们只打算在这个

52:25 特定应用领域实施

52:28 或者在这个特定的使用场景中，但这种想法认为

52:30 没有人会进入，因为他们太害怕首席财务官会被解雇什么的，

52:35 绝对不会

52:36 这简直疯狂

52:39 但也会有首席财务官犯错，

52:40 然后，好吧，每个人

52:44 都会有点，嗯，如果他们那样做，那就是彻底的失败，没错，因为

52:45 但也或者说

52:49 你要懂，有一个非常有趣的微妙之处在这里

52:50 那就是你现在其实不想让你的工程师们去考虑计算预算，因为我们

52:55 还在开发这方面，哦好吧，所以你就把它放那边，但我觉得好像

52:56 我们关于云计算预算已经讨论了15年，这完全是新的情况，就像

52:59 只有大约10%

53:00 在2016年，只有10%的工程师需要考虑这个问题，在2018年左右，就有一批公司，基本上就像

53:05 一个仪表盘用途，

53:06 叫什么名字来着，Phenoms，现在开发者很喜欢，因为

53:11 开发者会控制访问权，因为云支出失控了，

53:13 是的，P.I.

53:17 支出越来越失控，是的，所以就是，给你你的

53:18 Twilio支出，这是，

53:22 但是，这和以前相当不同，我等着看

53:24 YouTube上大家的评论来挑你毛病，就像你可以走进会议室，

53:28 说嘿，你能不能让那个算法

53:28 更高效一点，这样你晚上用我们集群的资源就不会那么多，诸如此类，然后

53:32 会议结束后有人改进它就好了，这就是每个工程师每天在做的事情，

53:33 你得决定，比如说你要一个长时间运行的

53:35 提示，你想让它是一个长时间运行的智能体吗？你想并行运行吗？

53:36 比如，你容忍浪费多少Token？现在

53:40 我倾向于觉得，我们应该浪费很多Token，因为那意味着我们

53:41 在尝试新东西，是的，就像，如果你的工程主管会高兴吗，如果你

53:45 并行做10个实验，显然你会浪费90%的

53:46 Token，你会选出一个成功的，然后或者你告诉团队，知道在你走之前，

53:49 把设计做好，真的把系统设计完美，我们

53:50 实际上还有一大堆开放问题开始出现，就在录制这个视频的时候

53:53 有人现在在为新的云代码最大计划抓狂，

53:54 像这样，

53:58 因为他们在三次提示后就被封锁了，这会成为

53:59 一个非常现实的话题，直到我们找到建造数据中心

54:03 容量的方法，那是个

54:04 不同的问题，好吧，不，因为没人，假设，等一下，你可以

54:07 假设，如果我们建设更多容量，价格就会下降，因为容量增加了，而我们

54:07 现在基于有限容量定价，但这会最终解决，

54:12 我为那些必须立即做出决定的人感到难过，

54:13 决定这周哪些17人不能用更多Token，

54:16 整个公司都像拿着打卡卡，

54:17 午餐排队的人，

54:21 是的，午餐排队的人每次用完都会打卡，但我，我不知道，

54:23 我们今天讨论了性能这件事，比如说，我们以前写命令行工具，运行完会

54:26 显示运行时间，好让你知道性能是变好还是变差，但是

54:28 这一切都会过去，毫无疑问，

54:31 会过去的，我觉得10到10年后，主要原因是

54:32 你得算点滴成本，也就是，如果你...

54:36 actually have a whole

54:36 bunch of open questions that are going to start to happen like literally as of

54:40 this recording time

54:41 people are freaking out right now on the new cloud code max plan you know whole

54:45 like like

54:45 because they're getting blocked after like three prompts well this is this is

54:49 going to be like a

54:50 very like real topic until we can actually find a way to build data center

54:53 capacity oh that's a

54:54 different problem okay no because no one is well assuming well wait no you can

54:58 assume that if we

54:59 build more capacity the price will drop because there is more capacity and we

55:02 're priced now based

55:03 on limited capacity whatever but like this is just going to get worked out and

55:08 I feel bad for the

55:09 those that have to make a decision immediately yeah about which 17 people get

55:14 no more tokens this

55:15 week or whatever and that the whole company is walking around with like a token

55:18 card and

55:19 and the person in the lunch line yeah yeah like it's the person in the lunch

55:25 line is punching their

55:26 card every time they do set but but like I you know I I don't know like

55:30 somebody we were talking

55:31 about today about performance and how like you know we used to write command

55:35 line tools that spit

55:36 out the time it took after you ran a command line just so you knew and if you

55:41 knew you were getting

55:42 better or worse and you know but you the thing is is this is all going to go

55:48 away there's absolutely

55:49 no doubt yeah that this just goes away I think I'm 10 or 10 and the biggest

55:53 reason it does is

55:55 is because you you have to do the penny off kind of math which is if you're

55:59 支付给企业

56:00 你认识的年收入一百万美元的销售员，你得问问他们的

56:04 工具值多少钱

56:06 如果你支付给一名工程师每年 X 美元，那么在某个阶段他们的

56:11 工具价值

56:12 绝对值得，这甚至不会成为一个问题，嗯，是的

56:17 是的，我觉得这不是

56:18 嗯，我认为如果短期内有产能问题，那就是

56:22 另外的问题

56:23 这跟价格驱动的不同，我们不会永远

56:27 被困在某种预算

56:27 演算里，我觉得很大一笔数字能解决这个问题，因为

56:31 最终你会有

56:32 足够多的工程师使用这么多计算资源，但我们现在处于过渡期

56:35 我们就像

56:36 大多数人认为两年前 AI 的支出水平，那只是一个

56:40 聊天

56:41 机器人，是的，但是他们错了

56:44 是的，他们错了

56:45 我们试图提醒他们，但他们错了，因为他们只看到了这个

56:50 具体用例

56:52 但再说，你知道那个电子管的事情，你取笑过

56:58 是的，但是

57:00 曾经有人认为整个达科他州都会布满电子管仓库，人们穿着旱冰鞋，

57:04 穿梭其间

57:05 不断更换电子管，只是为了赢得二战

57:09 我意思是，这就是那个时代的情况

57:10 然后有人说，“嘿，试试晶体管怎么样？”

57:15 接下来就是

57:16 我们将在这件事上经历晶体管时刻

57:20 可能只是供应增加，

57:21 我们对它的理解方式

57:24 但也可能是算法的根本性变革，

57:26 也可能是硬件的变化，有很多东西能

57:30 改变这个

57:31 特定时刻

57:34 这件事特别奇怪的是，大家只是关注

57:36 代币

57:40 这跟 IBM 和大型机发生的事类似，人们用 MIPS 计量，

57:42 但有一天现实是

57:45 IBM 每年以更低价格卖出更多 MIPS，甚至

57:47 自己都没意识到，

57:52 他们定价还是按 MIPS 计算，直到有人指出

57:53 他们在走下坡路

57:58 因为制造的 MIPS 比他们能收取的价格增长快，

57:59 这就是

58:02 必然会发生的事

58:04 我刚才说得很直接，我觉得那

58:07 听起来太好了

58:09 听起来真棒，感觉我知道自己在做什么，绝对没错

58:12 实际上可能

0:00 The diffusion of AI capability is going to take longer than people themselves

0:03 can now realize.

0:04 It's just absurd to think you're going to vibe code your way, like SAP.

0:08 All of that domain knowledge, it's not just represented in some well

0:12 orchestrated data layer.

0:13 The engineering compute budget conversation is going to be the most wild one in

0:17 the next couple

0:18 years. The biggest problem right now is everybody is trying to figure out the

0:21 economics of all of this.

0:23 When they're off by at least in order of magnitude on how big the opportunity

0:27 is.

0:27 If you have a hundred or a thousand times more agents than people, then your

0:31 software has to be

0:32 built for agents. People in the abstract say things like, now you're marketing

0:36 to agents that

0:36 you're like an API, you've got a good idea. I actually think that's almost

0:40 exactly wrong with

0:41 you. Wow, this is breaking podcast news.

0:43 If you start to imagine that we all have to build software for agents, I think

0:49 we're like all clear

0:50 on that, right? So like that trend is happening, which is like we spend as much

0:55 time now thinking

0:56 about the agent interface to our tool, as we do the human interface, and the

1:00 reason we're doing

1:01 that is because our hypothesis would be that if you have a hundred or a

1:05 thousand times more

1:06 agents than people, then your software has to be built for agents. What is the

1:11 way that those

1:12 agents are going to interact with your system? It's going to be through an API

1:15 or a CLI or MCP or

1:16 whatever. The paradigm that appears to be taking off and is quite successful so

1:21 far in terms of

1:22 efficacy is what if you give a coding agent access to your SAS tools and a

1:27 coding agent

1:29 access to your knowledge work sort of workflows and context, and that kind of

1:35 becomes this super

1:36 power, which is it's not just like the agent is not only capable of reading

1:40 some data,

1:41 understanding some information, it can actually code its way or use APIs

1:46 through whatever task

1:47 it's trying to achieve. That appears to be like a paradigm that is starting to

1:52 compound,

1:53 and that's the cloud co-work phenomenon. That's the whatever open AIs is kind

1:59 of cooking up

1:59 with the super app, perplexity computer, et cetera, and I actually think it

2:06 kind of makes sense as

2:07 like the ultimate manifestation of this stuff. I think you're right. It makes

2:13 sense in a

2:15 theoretical way, but in a practical way, we have to be really careful in that

2:21 that the way to say it is algorithmic thinking is really, really, really hard

2:29 for the vast

2:30 majority of people who have jobs. The easiest way to think about it is if you

2:36 were to go into any

2:36 person and ask them to create a flow chart for a particular thing that they

2:40 have to go do,

2:42 they would probably fail at producing that flow chart. So within any

2:46 organization,

2:47 say doing a marketing plan and there's 50 marketing people working on a giant

2:52 product line,

2:53 one person probably understands and could document the flow chart. So if you

2:58 put one of these agents

2:59 or you put this tool, this co-working tool in front of people to create these

3:06 things,

3:06 their ability to explain to it what to do is really, really limited. So then

3:12 you...

3:13 But what if that becomes the new, this is the new way you have to interface

3:17 with computers and

3:18 you just have to cycle that through? Well then you're basically, you just go

3:23 back to

3:24 you're basically just developing the next abstraction layer for how people

3:30 interact and

3:31 developing an abstraction layer has historically at each level of the

3:34 abstraction layer been a

3:36 highly skilled, very specific individual within an organization, developing

3:41 that. And then the

3:42 little parts that they build just become little toolits in the world of people

3:47 doing particular

3:48 tasks and some people are able to stitch them together and some can. But that

3:52 happened with

3:53 paper clips and thumbtacks. Oh yeah, I actually... And it's going to happen

3:56 with whatever we do

3:57 next. I think that... I think basically the timeless part is the job just moves

4:04 up

4:04 around and you learn a new set of skills and that's why I actually don't think

4:09 anything about this

4:10 is any different. It's just now the leverage you get is obviously fantastic.

4:14 There was this viral

4:15 kind of tweet that went around which was the anthropic growth marketer, do you

4:19 guys see this?

4:20 It's basically one person and he was using cloud code at the time to basically

4:25 more or less automate

4:26 what maybe five or ten people would have done in various kind of silo jobs. And

4:30 I think the

4:33 reason why it's interesting is, yeah, you had to have been a systems thinker to

4:36 be able to

4:37 accomplish that. So like clearly he already was technical enough to be able to

4:40 pull that off.

4:42 But it did kind of represent like what would each of these jobs look like if

4:45 you had like

4:46 imagine you had, you know, X job in the economy and right next to that person

4:50 was an infinite

4:51 pool of engineers that could automate whatever that person wanted. And you know

4:55 , what would that

4:56 job look like in the future as a result of that automation that now is possible

4:59 ? Yes, I agree that

5:00 you'd have to find a way to like, you know, think through your job as a system

5:04 to be able to pull

5:04 that off. Maybe the agent gets better and better over time at being able to

5:07 like nudge you in that

5:08 direction. But like, it does sort of stand a reason that like you will start to

5:13 , you know,

5:14 try and automate a lot of that kind of work of like, well, why don't I take

5:18 like the keywords

5:19 that are working in this in Google AdWords and then port them over to Facebook

5:23 and make sure

5:24 that those are replicated and then take in the new signal from what's happening

5:27 in the market.

5:28 That's a big leap. Like, what the first, you were nodding a little, and then I

5:32 said something that

5:34 went too far. The anthropic growth person is an example, like that's just the

5:37 rest of the work.

5:38 Yeah, I could do that job. Everybody's not going to be like, when the mayor is

5:42 infinite. Yeah, right. Like when demand is infinite and frankly supply is

5:46 infinite,

5:46 this is not a difficult job. And so let's, let's think of the federal club in

5:52 Australia right now

5:53 is amazing. Right, right. So like, B instead be the $600 PC marketing person

5:58 and see how you can do

5:59 against the media. That's a real job. All right. We need a better, we need a

6:03 better example.

6:04 But there is, I mean, it is, it is really interesting. Like, I hear, let me do

6:08 an old example,

6:10 an old person example. Like my, my cousin, MBA, elite school joined her first

6:16 job. She's a little

6:17 older than me, joined with right on the cusp of computing. Like she actually

6:21 didn't use a spreadsheet

6:22 in grad school. And then they all, or spreadsheet showed up, but she wasn't a

6:28 spreadsheet person.

6:29 So instead, they told her hire as many interns as you want. And so her first

6:34 year on the job,

6:34 she like supervised like essentially a whole room of agents. And the, the, the

6:40 kids who was me,

6:41 not literally, but they were in college, came and just did all the spread

6:45 cheating. Yeah.

6:46 But then what happened sort of this magically over the next couple years was

6:51 she and her cohort

6:53 all became the spreadsheet people. Yeah. And then this idea that you being a

6:58 manager in a bank,

6:59 or just a two years in, meant you had a cadre of people doing the spreadsheet,

7:03 no, the whole

7:04 abstraction layer moved up. And the old job before those interns was you just

7:09 sat there with basically

7:11 calculators and an HP calculator, figuring out the model for some M and A deal

7:16 or whatever.

7:16 And you only got to do like two iterations before you had to put out the, the

7:21 pitch deck or just go

7:22 to the customer or the client or whatever. And then all of a sudden they're

7:25 doing 30 iterations

7:27 themselves. Yeah. But they, and so I think where we are with agents is just at

7:32 this step where you

7:33 think you need 50 and the abstraction layer is such that we're dividing up in

7:37 these really small

7:38 pieces with one super smart person coordinating them all. And pretty soon that

7:42 whole thing is just

7:43 going to, they're all going to collapse on each other. Yeah. And there is just

7:46 going to be like

7:47 a skill set amount of code called an agent that is like marketing-ish. Yes. And

7:53 you'll be able to

7:53 ask it marketing stuff. Yeah. And then the next step will be and have it go do

7:58 things. I'm a little

7:59 skeptical of the, until the whole like non reproducible, non random element of

8:05 this AI stuff goes away.

8:06 The doing stuff is going to get very costly. Yes. And so then you get into the

8:11 human in a loop

8:11 discussion and all of that. But I think we're just, we're at that exact, I feel

8:16 like when I talk to

8:17 people trying to do stuff that we're right. I feel like I'm at Thanksgiving

8:20 dinner talking to my

8:21 cousin six months in her job. When, when I'm using a spreadsheet already and I

8:25 'm, and I'm like,

8:26 I don't know why this is so hard. You should just use one. Yeah. And then two

8:30 years later she's

8:31 doing it. And, and I think this right now you have to be an absolute, you have

8:34 to be a rocket

8:35 scientist and the growth marketing person to create 42 agents and spin them all

8:40 up and do all of this

8:41 stuff. But the rocket science part of it just is going to evaporate in a very

8:44 short order.

8:45 And then you're talking about, well, there's a giant chunk of domain expertise.

8:50 Yeah, it was

8:51 back to the domain expert. I actually think something that you said, I'll take

8:54 the other side of,

8:55 which is, I think it's very tempting to be like, these agents are going to code

8:58 and do X. Yeah.

9:00 But I think we're going the opposite way. So I think actually where we started

9:03 was we'd like

9:03 take like a piece of SaaS software and we'd add AI. Yeah. And then that's like

9:07 the new kind of like

9:08 AI enabled. So that's like the extreme version of using code for these types of

9:11 things. But now,

9:12 what are we actually doing? We're like, okay, the SaaS software is still SaaS

9:17 software. And the

9:18 agent uses it as a computer because it's actually very good at that. So I'd say

9:21 like we started with

9:22 code, then went to the terminal, which is actually less code. Yeah. And now

9:27 this year is going to be

9:27 the year of computer use. Yeah. So it's almost like they're much more like

9:31 humans using computers

9:33 than them generating code. And that feels like very much like this mezzanine

9:37 step. Yeah. And I

9:38 actually come from like the generating code type of the world. Like I would

9:41 argue that that's

9:41 happening less not more. Yeah, I think. So to me, whether it's computer use API

9:47 use or writing code

9:48 on the fly, I kind of maybe erroneously put that all in one way. They're very

9:53 different. But we

9:55 have an agent that we're working on where you just makes a determination

10:00 whether it should use an

10:01 existing skill as you're using an existing tool from box or it should write

10:04 code to solve that

10:05 problem. And its ability to do any one of those three at any moment ends up

10:09 being incredibly useful

10:11 because sometimes there's just some specific operation you want to be able to

10:14 do. We're writing

10:15 code to be able to do that operation is just faster. And we don't have to we

10:18 don't we can't

10:19 possibly, you know, kind of pre plan for every thing that anybody would ever

10:23 want to do on their

10:23 documents. And so the fact that the that the model is good enough to also write

10:28 code on the fly for

10:29 that use case ends up just being like an amazing property. Even though maybe 90

10:33 % of the things that

10:34 it's going to do is should just be using an existing. Over time, Prado takes

10:37 over and over time,

10:38 there's like literally like seven apps on her iPhone. There's seven sass apps.

10:41 We end up like

10:41 over time, these things tend to consolidate. But the but the seven apps on the

10:46 iPhone is is a is a

10:47 issue of humans don't want to learn these things over and over again. And so I

10:51 as a human, I don't

10:53 I can't I don't have the mental bandwidth to learn that many apps. But an agent

10:56 that is going to use

10:57 tools and APIs and be able to code things doesn't have any of the same

11:00 constraints that we have.

11:01 So I I don't know like I don't mind you could argue that there's just so many

11:06 things to do and

11:07 you can make yeah faces sufficiently general. Yeah, well, fair. I let me say I

11:11 like what I like

11:12 what you said then because our back. We're lying. No, but but I think there's

11:19 something super

11:20 interesting here, which I do really, really like, which is that that where

11:24 software has evolved,

11:26 you know, like I use SAP all day. I work in finance. I have to go and generate

11:30 all these reports.

11:32 And then somebody shows up and says, I want to report that does this view slice

11:36 this way. And

11:37 I'm like, Oh God, I don't know how to make that. And like now let me go wade

11:41 through the SAP help

11:42 system and try to find it. One thing that that let's just say AI could be very

11:48 good at is it

11:48 actually can navigate that surface area much, much better. You know, the help

11:53 is all there. And so

11:55 it's a matter of finding it mapping language. And humans have been a bottleneck

11:59 in tapping the past

12:01 25 years of software capabilities. Yes. I mean, like I spent my life, my life

12:06 with

12:06 sitting at the people on airplanes saying, how can I make PowerPoint do X and

12:10 just go to the ribbon.

12:11 And you know, it was because it hurt physically hurt to watch somebody

12:16 suffering with bullets and

12:17 numbering and word or trying to figure out, you know, like, Oh, let me just

12:20 make a two sided,

12:21 a two axis graph in Excel, which like is rocket science, like almost no one can

12:26 do that. But yet

12:27 it's super common. And so people are like, have none. And so that impedance

12:31 mismatch was a human

12:33 user interface design on the consumption layer, I totally buy it, which is like

12:37 the perfectly

12:39 fluid, like UI or consumption layer. I just feel the back end, like the systems

12:45 of record,

12:46 yeah, oh, yeah, it'll probably converge into like some database, like some

12:49 generic set of APIs,

12:50 like that they'll connect to. And like that seems to be the direction it's

12:53 going.

12:53 I agree. I think you can start. Well, like, so I spent all weekend, like

12:58 implementing my

12:58 nanocla dot. And when you first start out, it's like you're building an

13:02 integration for everything.

13:03 The nanocla is very like like open clock has all of the integrations, nanocla

13:06 has been a few of them.

13:07 And so you haven't built all of its own tools. But after, you know, two or

13:10 three days of these,

13:12 like, you know, you kind of have the two integrations that you need. And yeah,

13:15 like, yeah, but back

13:17 to the, I mean, we're talking about personal productivity, probably like you're

13:20 like organizing

13:21 your life or something. Well, it's work productivity. Okay, we're productivity.

13:24 And then an SAP system,

13:25 and like, and like, and so there's like an, well, like, there's an infinite

13:28 amount of complexity

13:30 when you get to, okay, some company that has a global supply chain, and they're

13:32 dealing with

13:33 75 pieces of information across, you know, 30 different systems. That does

13:38 require a certain

13:39 amount of horsepower from the agent that is just, we have, I mean, we just

13:44 haven't been able to get

13:45 from, from any architecture up until now. But like tape, but that it, what you

13:49 just described

13:50 is literally what it has been doing for 50 years, and will continue to do,

13:55 which is,

13:55 yeah, I have a friend who was the CIO of, of the VA. And he spent all he spent

14:02 his time on was

14:03 gluing the 75 VA systems together. And it's all just integration. We're done.

14:09 It's perfect for

14:10 integrate. Yeah, this I totally agree. For integration, these things are the

14:14 best, but, but

14:15 it's integration. Yes. It's literally, how do I stitch these two systems? But

14:18 it's in it. But now

14:18 the thing that I think is happening is it's kind of like integration on demand.

14:23 Yeah, it's, it's my,

14:24 it's my new query in the system that the IT team didn't prewire. Now I needed

14:29 to happen at runtime.

14:30 Uh, let me get off my lawn. Okay. Okay. So, okay. So the reason I just was in a

14:37 room filled with a

14:37 bunch of CFOs and CIOs and this, they all looked at me when I said something

14:42 along these lines,

14:44 although not as optimistic as you could imagine, but they just, they, they, no,

14:48 it caused like six

14:49 of them to come running up afterwards and say you're insane. You've lost all

14:53 credibility with me

14:54 because it's bad. What, what, what specifically that they, the agents are going

14:58 to do integration

14:58 as well? That, that, that the integration is a problem that will get a lot

15:02 easier. Yes.

15:03 They were against that? No. They're no one's against, but they're, they think

15:07 it's practical.

15:07 But they're, they're fear is like unleashing not just the agents themselves,

15:13 but humans to do

15:14 integration because you put people creating new integrations and you just say,

15:18 please break my

15:19 system of record. Oh, yeah. And so this idea that you just create like a new

15:23 API between,

15:24 you know, system 27 and system 38. Yeah. And, and then you're, that might be

15:29 fine for a report.

15:30 Yeah. Because if that person wants to be wrong, that's their business. Yeah.

15:32 But you're not. I think, I think we have a read-only version of this for a

15:36 number of years before.

15:38 Where is, and is very large. Yeah. And a lot of it's just the consumption layer

15:42 where the consumer

15:43 is a human being. Right. Right. It really feels right now. A lot of the stuff

15:46 is consumption.

15:46 But, uh, yeah, I mean, it's, um, uh, you know, we, we actually have, so we just

15:51 rolled out the

15:52 official box CLI. Thank you for the liking the tweet on that. I, I used it. I

15:57 have some feedback.

15:58 I'll talk about it. I'll take, I'll take all the feedback. But it's a really

16:01 interesting thing.

16:01 So we, we had these, all these debates internally of like, okay, you give

16:04 Claude code, the box CLI,

16:07 and you can now interact with your entire box system via natural language. And

16:11 you get the

16:11 horsepower of opus four six being the orchestrator of doing a bunch of

16:15 operations. And it's like,

16:17 it's like, you know, blows your mind. I guess I'll get some feedback, but it

16:20 blows your mind in some

16:20 ways because you could just be like, upload this entire folder from my desktop

16:24 in the box and it

16:24 will work or process all these documents in this folder and it'll work. And, um

16:29 , and it's amazing.

16:30 And then we started thinking through like, well, let's say you were a company

16:33 with, with, you know,

16:34 5,000 employees and everybody had access to some shared repository, like, you

16:38 know,

16:39 engineering documentation and, you know, marketing assets or whatever. And

16:42 everybody had Claude code

16:44 or codex, um, you know, running with the CLI. Wow. We now have some really

16:48 interesting new challenges,

16:50 which is like, like, how do you coordinate, you know, possibly the fact that

16:55 you might be hitting

16:56 this system, like, you know, 10,000 times an hour or something, not from a,

17:00 like a performance

17:01 standpoint, but just like, how do you make sure that people didn't move like a

17:05 file from one

17:06 thing accidentally from one folder to another folder while the other person is

17:08 trying to do a

17:09 right operation and somebody else was trying to delete something because you

17:12 have these agents

17:13 running wild. This is, this is going to be like the new big question that every

17:18 CFO CIO is running

17:19 around trying to with their hair on fire. Well, there, there's just, that's

17:22 exactly what I ran

17:23 into, which is I played around with your example, which is create the video

17:26 example, which is create

17:27 like a marketing plan directory or something. And like, all of a sudden, I'm

17:31 like, in some loop

17:32 creating directories, like, and I'm like, go on as long as it can. Right. And I

17:36 was like, I wonder

17:37 what the limit is on box for nested directories, because I'm about to hit it. I

17:40 actually, we're

17:42 going to find out too. Yeah. Yeah. But it does feel to me that, that like, a

17:47 lot of the intuition

17:48 is to like build a new layer, yeah, controls and whatever. But what's actually

17:52 happening on the

17:52 ground is, is, is, is the opposite. So I'll give you an example. Like, when we

17:56 all picked up a lot

17:57 of these personal agents, we would like give them our API keys. Yeah. We would

18:01 give them our email

18:03 addresses. And then they would kind of access those things. They're like, Oh,

18:06 but how can I stop

18:06 it from like, whatever? Yeah. And so what everybody's doing now is you give it

18:10 its own phone number.

18:11 Yep. I actually gave my nanoclaw, its own credit card. Hopefully just a,

18:16 Visa debit card that you bought at TBS. No, no, no, no, but then, but then I

18:23 gave it its own

18:23 Gmail account, which we're going to log into. And then Gmail actually has all

18:26 of these RBAC

18:27 permissions. Yeah. Yeah. So, so you could make an argument that like, you know,

18:31 we've actually

18:31 built in a lot of these permission systems to treat it like a human, otherwise

18:34 it's a separate

18:35 human. And then instead of like building another off layer, building another.

18:38 Okay. Now, can I

18:39 instantly take, do a take down of this element that we're going to run to?

18:43 Please. Okay. So that

18:45 is fantastic for personal productivity. And the question that we're going to

18:49 run into is in an

18:51 enterprise. Let's say I have, let's just make a simple example. I have a 50%

18:55 team of something.

18:56 Should everybody also, basically, will we have a hundred, will we have a

19:01 hundred people now

19:02 collaborate? I mean, basically 50 humans and then 50 credit cards and then 50

19:06 agents in that same

19:07 shared space. And do I have, I obviously have complete oversight over my agent.

19:13 But what if my

19:13 agent collaborates with somebody else and then accidentally gets access to some

19:19 resource because

19:20 they were sharing with that other person and I'm not supposed to have access to

19:23 that resource. And

19:24 now this autonomous sort of stateful, you know, agent is running around working

19:29 on somebody else's

19:30 information. The default end to end argument is you treat them like human

19:33 beings. It doesn't

19:34 work. So you can't fully treat them like humans because here's the thing. And

19:37 with regular humans,

19:38 you don't get to look at the slack channel of the person that is working with

19:42 you or working for

19:43 you. You don't get to log in as them. You don't get to oversee them. You are,

19:47 they are accountable

19:48 for their own set of execution in the real world. You don't get penalized for

19:52 how they screw up.

19:53 The agent, you have all the liability of whatever they're doing. You do have

19:58 complete oversight

19:58 and you're probably going to need to have that complete oversight. They have no

20:00 right to privacy.

20:02 So, so there's going to be these, some of these breakdowns that aren't as clean

20:06 as just treat them

20:07 like a person because I need to be able to kind of, I need to be able to give

20:11 access to something to

20:13 them. But I also need to be able to like log in as them at some point and be

20:16 like, no, no,

20:16 you fucked up the whole thing. And I need to undo it all. But if I can log in

20:20 as them,

20:21 how could they have operated in the real world working with other people and

20:25 keeping anything,

20:25 you know, confidential or secure or whatever. So it really is still an

20:28 extension of you.

20:29 It's like almost impossible to get around them being an extension of you. So

20:32 now the thing that

20:33 we're thinking through, that we're not going to be able to do any time soon. It

20:37 just doesn't

20:37 logically follow me. Yeah, maybe. But for example, for my employees, yeah, I

20:41 can log in as them.

20:42 You don't, though. You don't, you don't, I can get access to their email. Yeah,

20:46 no, in like,

20:46 if you get like sued, you're not logging in, you're not logging in as them on a

20:51 regular basis because

20:52 they sent one email. Is the right operating model with an agent? The same thing

20:55 . The risk is like

20:56 a thousand times greater. Like these people, like they will just leak your

20:59 information whenever they

21:00 want. Like they will happily just go and send some email to somebody because

21:04 they got prompt

21:04 injected. You think the terminal state is that these things are still these

21:07 sloppy computers and

21:08 therefore they will always, I don't like the word sloppy, unless we're saying

21:11 it very in a

21:12 colloquial sense. But like, they'll never be able to contain information. They

21:16 'll never.

21:16 So we're like, I think the ability for you to keep something in the context

21:22 window, a secret,

21:24 like as in like you tell it, do not reveal x thing in the context window. I

21:28 think that's

21:29 a very hard problem to solve. So then, so then thus, if anything can ever enter

21:34 that context

21:34 window, because they have access to a resource, then in theory, you should

21:37 assume it can be

21:38 prompt ejected out of the of the context window. And I don't know that we know

21:42 of a way to solve

21:43 that at the moment. Like that's like, and so, so if I know your new agent's

21:49 email address and I

21:50 email it like it's an assistant, but like I can, I can social engineer it 10

21:54 times easier than a

21:55 human. Like it'll be hard for you to pull off that that agent is now also has

22:00 access to your like

22:01 M&A documents and stuff. But is it this like literally all of AI right now?

22:04 Which part? I mean,

22:05 the fact that we've got these shared systems that we use the intelligence for

22:09 that have shared

22:10 contacts. But we mean, it's all of AI. Well, I'm just saying like right now,

22:14 when we use AI

22:15 internally in agents internally, this is exactly how we use that. But this is

22:18 why you were they

22:19 are working on us as you effectively right now. And we don't yet know how to

22:23 make them not work

22:24 as you. Let me offer an example. Let me let me offer an example and then

22:28 solving this problem

22:28 though. Like like the like the issue will be will be like you will just be able

22:35 to trick the

22:36 agent to reveal information. So then so then that's why like having them have

22:41 access to their own

22:42 resources where they can fully make their own decisions is not yet something

22:45 that we've been

22:46 able to pull off. There's a perfect example for solving your problem, which is

22:50 we already lived

22:51 through this with open source. The model for open source was it's all there and

22:57 just use it and

22:58 you pick and choose. And then like nobody debated it because the world was much

23:02 smaller than and

23:03 we weren't all on X doing podcasts when this was all happening. But then

23:07 quickly everybody realized

23:08 all the problems you you were just talking about. Like if you're running a big

23:11 company,

23:11 you can't have some person just go copy in a bunch of source code from open

23:16 source into your

23:17 commercial product like that. There was a whole licensing problem, a whole

23:20 whole bunch of stuff.

23:21 And so all these norms got developed. The the debate that we're happening that

23:26 's happening

23:26 right now is just is this really interesting modern artifact of how new

23:31 technologies develop,

23:32 which is this is all happening in real time. During open source like we met at

23:37 a conference

23:38 room this big and debated how much open source we could use in Windows or

23:42 Office. And nobody in

23:43 the internet knew we were having this debate. It was a very and I think it's

23:47 just so interesting

23:48 that not just this the debate about specifics but this whole notion of where is

23:53 this heading

23:53 is happening in writ large and everybody is just trying to get to the end state

23:59 like way way

24:01 more like in a sense more quickly than we can actually reach the end state. And

24:06 so what really

24:07 needs to happen is people just need to go build. We need standards. What? We

24:12 just need some

24:12 standards. I've got different intuitions on the end state. No, no. You don't

24:17 want my intuition.

24:18 What could make an end to an argument that these things actually converge on

24:22 the same type of

24:23 reliability as a human being, which is exactly how we view like self driving.

24:27 And in that case,

24:28 you use the exact same mechanisms that we use to protect with human beings.

24:32 Like yeah,

24:33 you consider insider threat. You consider the fact that people can be bought

24:36 off. You consider the

24:37 fact that people make mistakes. And that's a risk. And that's operational

24:40 processes.

24:41 So one intuition is like that will be the end state. There's another intuition.

24:45 Well, don't point to me. I'm just saying, I'm talking about where we're at now.

24:48 I actually,

24:48 I don't know that we disagree on the end state. And by the way, like

24:52 strategically,

24:52 we're hedging because we're gonna build we're gonna build Asian users and like

24:57 so we're like,

24:58 I love the idea of open claw having a box account and it operates and you just

25:03 like twice as many

25:03 cats. I love it. I'm just saying on the ground right now, we don't yet know how

25:10 to give it an

25:11 M&A data room to fully securely. Right. But that, but it's actually it is

25:16 harder than that,

25:16 though, because the threat the threat vectors are going to be way more

25:21 sophisticated. So we do have

25:23 a cat in a mouse game going on where you can't just assume that the agent acts

25:29 like a human does

25:30 today because it's going to be the fastest, most thoughtful, craziest-ass human

25:36 that ever existed

25:37 trying to actually leak the information because it got injected in some way.

25:40 And so

25:41 part of what's going to happen is we're going to go through this phase where

25:46 like the enterprise

25:47 customers are just going to like close everything off until there's some sense

25:52 of sanity in all of

25:53 this. And then, but in the meantime, the individual and specifically the

25:58 developers.

25:59 And that's going to be that I think is the most exciting tension that's going

26:04 to happen is that

26:05 that the enterprises are going to be are going to get left behind by these sort

26:09 of advanced

26:10 individuals, which will then start to look like the startups. And the startups

26:15 will start to move

26:16 much, much faster than enterprises because they just don't have any of these

26:19 problems.

26:20 And you know, you could end up with like the agent going rogue in a startup and

26:25 doing that.

26:26 You had no employees that go rogue routinely in the startup.

26:30 Yeah, well, it'll just be an episode of Silicon Valley. And so, you know, big

26:33 deal.

26:34 I agree with you on like the okay, it's people, et cetera, the same risk. I

26:38 think there's a couple,

26:39 you know, differences though, in the sense that that I can't really threaten,

26:44 you know, the like

26:46 cloud code that it's just I'm going to pull the plug on it in the same way that

26:49 you do have that

26:50 threat as a regular employee is like you at least like 95% of people are not,

26:54 you know, trying to do

26:55 bad stuff, you know, within an order. Yeah, but they're not trying, but the

27:00 ability to inadvertently

27:01 do bad stuff. Yeah. To your point about it still not having that stuff. I would

27:05 argue that that

27:07 it's a lot easier to have people not share, let's say files with somebody

27:11 outside the company

27:12 in a wrong way, more than it is for an agent right now to have the same set of

27:17 instructions.

27:17 And also you have the tools so that you can basically stop that at a whole

27:21 different level of

27:21 abstraction. Which is why you have to build this into software. Right. But I do

27:24 think actually,

27:25 if you were to like, if you were to like put a bow around your last point, a

27:28 lot of this is actually

27:30 why the diffusion of AI capability is going to take longer than people in

27:33 Silicon Valley realize

27:35 because what's happening is like we see startups that can start from the ground

27:39 up without any of

27:40 the risks that we're talking about because they have nothing to blow up. And so

27:44 we look at that

27:45 as the trajectory that we're on. And then you go to like JP Morgan and you're

27:49 like, how are you

27:49 going to set up nano cloth to be able to to actually like, you know, automate

27:55 your business anytime

27:56 soon? And it's like, oh, okay, there's going to be like a little bit of a gap

27:58 there. Well,

27:59 what do you guys think? Here's I think that that opens up a pretty interesting

28:03 problem,

28:04 which is this split between big and small startup and enterprise, which is just

28:08 that

28:09 that the the enterprise, the current SaaS vendors who are all struggling in

28:14 this SaaS

28:15 apocalypse weirdness that I don't really agree with, but they are struggling

28:18 with this problem that

28:19 they they don't really sell the line of business data. They actually sell this

28:25 intelligence and

28:26 domain expertise in this whole system. And the agent side of things wants to

28:30 only buy the data

28:31 now. And they only want to license the data and they want to have unlimited

28:36 access to the data,

28:37 but they've actually never really enabled that like that's never been their

28:41 business. And it's

28:42 been a long standing tension point with the likes of work day and SAP and stuff

28:46 like how much API

28:47 access to have. I mean, Salesforce went through three different massive

28:51 platform redesigns.

28:52 You know, it's I think that that's a particularly interesting problem, not for

28:57 the same reason that

28:58 Wall Street does. Wall Street's all wrong about the economics of a problem and

29:02 all that stuff,

29:03 but from a technology perspective, what does system of record mean in the face

29:08 of people wanting to

29:09 access the data when the data for training or for well, they're they're talking

29:15 about

29:15 executing the work. I think of it as executing the data to operations. Their

29:20 concern is that

29:21 somebody that they want to do the training layer on on your day. Like I'm a big

29:25 customer. They want

29:26 to do the my vendor wants to build a training. Actually, even even if you don't

29:30 even get into

29:31 training, they're concerned because because like monetizing, you know, you know

29:36 , sending a little

29:36 bit over the internet versus like you're in my UI. Oh, that's a very different

29:41 level of monetization

29:42 initially. That sort of that monetization part is the Wall Street point because

29:46 I think like, look,

29:47 there's so much domain stuff in in an SAP, just to pick an example, not to pick

29:52 on them. But like,

29:53 they're not going anywhere. Like it's ridiculous. It's just absurd to think you

29:57 're going to vibe

29:57 code your way to like SAP. But also all of that those all of that domain

30:02 knowledge, it's not so

30:04 it's not just represented in some well orchestrated data layer as much as they

30:08 tried. There's like

30:10 a whole bunch in the in the UI. There's a whole bunch in middle tiers. There's

30:13 a whole bunch in

30:14 just how you use it. And so I'm really unsure how this thing evolves because

30:20 SAP isn't going anywhere.

30:22 So then that's going to slow the diffusion of AI on that particular data source

30:27 independent of

30:27 whether or not it's a gentified AI that's doing stuff or just read only

30:31 reporting on stuff. So

30:33 where do you come down on it? Where do you think that's going to go? I'm afraid

30:37 of saying something

30:38 that's happening. Otherwise you're not going to get invited back. So say

30:44 something good.

30:45 I think I've drunk the Kool-Aid on build something agents want. So this kind of

30:54 the

30:54 Paul Graham term kind of like emerged on you know the past year on this topic,

30:58 which is just like

30:59 like a vent. I think we would actually then fully agree on this, which is at

31:03 some point you do

31:05 enough sort of iterations of this. And at some point the Asian is largely in

31:09 charge of what tools

31:11 it wants to implement and use and whatnot. And yes, the agent is not going to

31:17 be able to change

31:17 out an enterprise system. But like again enough generations later the agent

31:22 might just run into

31:23 so many walls with your software that it's just going to say you need to

31:26 finally rip out your

31:27 legacy HR system or I'm not going to be able to automate this workflow for you.

31:32 So I do think

31:32 you have this really interesting you know dynamic, which is back to this whole

31:36 point of imagine that

31:37 there's a hundred or a thousand times more agent volume on software than people

31:40 . You do that enough

31:42 times and eventually the software stack that Asians talk to has to be built for

31:47 them. And maybe

31:49 there'll be a couple holdouts. Maybe maybe a couple ERP systems are like the

31:53 final holdouts that

31:54 don't do that. But everything else you basically like your business will be

31:58 your business performance

32:00 will correlate to how well your agents can get access to the information they

32:03 need to do their

32:04 work. And so thus your enterprise IT stack has to be set up in such a way to

32:08 support that.

32:09 And so agents are kind of in charge because basically your software has to

32:13 support those

32:14 agents being effective. And that's going to mean everybody that built a SaaS

32:18 business or a software

32:19 business is like the game is can you build really really high quality APIs. Can

32:24 you have a way of

32:25 monetizing that? You know do you have a way of handling the identities and all

32:29 of the access

32:29 controls for agents and like like that becomes the new problem you have to

32:32 solve if you're building

32:33 a software company. And so yeah like and then how you monetize it like do you

32:39 monetize it like

32:40 does work day charge a penny for every HR record of polls like we'll figure

32:44 that out.

32:44 I do think that in some businesses it could mean less revenue and then in other

32:47 businesses it could

32:48 mean a lot more revenue. Like the thing we get excited by is like every agent

32:51 really loves

32:52 working with files. So there'll probably be more files in the future than there

32:55 was going to be

32:56 before. And so you know can we build a platform that like makes it really easy

32:59 for agents to work

33:00 with that data. You know we're betting that that's actually a really optimistic

33:03 outcome for

33:04 for you know our kind of business model. There might be some business models

33:08 that are like

33:08 more constrained because like the agent is doing more of the value than then

33:12 the software is in

33:13 that kind of future scenario and then there'll be everything in between. Can I

33:16 can I quibble with

33:16 one thing I you're going to quibble with that. I thought that was like so not

33:19 controversial.

33:20 No no I generally we're here to quibble. There's one thing I think like Paul

33:25 Graham and many

33:25 actually gloss over which is they focus on the interface. They'll say things

33:29 like you build

33:29 something for the agent. Yeah. And I actually think that's exactly wrong. Okay.

33:33 In the sense

33:34 that. And to be fair to Paul Graham he didn't he had been extrapolated. Yeah. I

33:39 have brought Paul

33:41 Graham into this. People in the abstract say things like now you're marketing

33:47 to agents the most

33:48 important thing is to being like whatever you're like an API you've got a good

33:52 idea. I think that's

33:53 almost exactly wrong. Wow. That's breaking podcast news. That's the one thing

33:58 agents are

33:59 really good at. Oh okay. It's finding their way through. And at the end of the

34:02 day like it's the

34:03 semantics that end up mattering a lot more. Yeah. Yeah. Right. And so like the

34:06 agents in in by

34:07 recollection or in my experience are very very good at picking the right back

34:12 end for whatever

34:13 they're doing. So they don't they're not like oh like the interface for this is

34:16 very good. The

34:17 document. It's none of that. They're like they're like the cost parameters of

34:20 this. Yes. The

34:21 durability of that. Like and so like they actually have the collective wisdom

34:24 of our experience

34:26 using these platforms. Like let's take cloud platforms. There's a bunch of

34:28 cloud platforms

34:29 out there. Yep. And whenever I ask an agent to choose a platform. Yeah. It's

34:32 actually using

34:33 meaningful stuff. Yeah. Not interface stuff. So I think as an industry we're so

34:36 focused on these

34:37 interfaces. Yes. Like oh you need to like market to agents this and that. Yeah.

34:40 But really I think

34:41 that we're going to be pushed to actually build a better system. Yes. And that

34:44 's what's going to be

34:44 chosen. Okay. Actually. So then there's probably no quibbling. I think we're

34:47 actually pulling a line.

34:48 I'm sorry to ruin the quibble thing. I don't treat this as like a you know kind

34:52 of a marketing

34:53 you know as thing. I'm more mean like if your tool is closed off to the agent

34:56 the agent eventually

34:57 will find a better tool for that company to go use. And so and so what will

35:01 happen is is it used

35:03 to be that you would go to like Gartner to be like tell me what to do. Tell me

35:06 what to do.

35:07 Yeah. Tell me what system to use or whatnot. At some point with enough

35:10 iterations that the agent

35:12 is going to say you should probably use this kind of database for this type of

35:15 operation.

35:16 And if you're not in if you're not in there then you're it's your DOA. And I

35:20 think we should

35:21 actually be celebrating this because agents are actually pretty smart at

35:23 choosing the right

35:24 technology. Yeah. In the past I really think it was a lot of the other things

35:27 that that caused

35:28 people to buy it. Yeah. But but don't worry we will in talking about we will

35:32 ruin the meritocracy

35:33 of this very quickly because we'll just like I'm going to outspend. Well the

35:37 agent will be an

35:38 API to incent the agent. But you know there's a marketing agent at Workday.

35:45 Well the marketing

35:46 agent at Workday will have the ability to purchase the recommendations from the

35:49 way they

35:49 replicate state dinners for agents. Yeah. But there is a there is a real here's

35:54 the thing that

35:55 again that that happened with the web sort of internally like internal like

35:59 just pick internal

36:01 sites like every company had file shares with like the best documentation the

36:05 best slideshows the

36:06 best financial models for any department or working area. And people sort of

36:11 got familiar with that.

36:12 And then when they didn't find the one they wanted they created a new one and

36:16 many organizations

36:17 sort of operated like that was essentially a free market. In fact because

36:20 before the world

36:21 of box like IT didn't if it was in a file they just didn't care. Right. They

36:25 only cared about

36:26 if it was in SQL. And so one of the risks with the model you're describing is

36:31 that the agents

36:32 themselves will spin up what becomes like a de facto new system. Oh they're

36:36 going to fragment that

36:37 out of in the in what you the IT people think of as some middleware and user BS

36:43 area. And I think

36:45 that that is a is a real risk. 100% is that like in a sense like the the the

36:51 macros end up running

36:53 the the corporation. Yes. And so I think that they've seen this movie and they

36:58 've seen what

36:59 happens when you let marketing just go buy a website on the internet to do an

37:02 event. And then

37:03 it's like a huge security vulnerability. And the mailing list is leaked and the

37:06 whole company gets

37:07 sued. And so I think there's a lot more real world tension in this dynamic than

37:13 then we just let on.

37:15 Yeah. But I also think it's one of these ones where you you you know their

37:20 organizations are

37:21 going to run at different paces. And JP Morgan is going to be the slowest at

37:26 doing this and the

37:27 startups are going to be the fastest but the the delta is huge but even the

37:31 startup one is a little

37:32 far off because even startups do need some systems of record. Yeah. And they

37:37 are going to all start

37:38 with some sass and they're not going to replace it very quickly. So I think it

37:42 's a little trickier.

37:43 So it feels like there's like there's two very competing viewpoints on this one

37:47 and like Elon

37:47 said it was like okay we're going to like issue a prompt and it's going to like

37:50 spit out machine

37:51 code and that's basically the collapsing of layer view like whatever existing

37:55 interfaces and

37:56 layers that we've created in the past are all going to go away and it's

37:59 literally like prompt

38:00 and machine code. The other argument like the history of systems is layers

38:03 never go away they

38:04 just get layered right. And because a lot of the layers are actually more of

38:07 like organizational

38:08 boundaries or like state boundaries or compatibility they're just they stay for

38:13 compatibility right.

38:14 So the other argument is is like we've actually evolved these layers very

38:18 specifically and up

38:20 because of like more human and organizational needs and they're not going to

38:23 change and the

38:24 agents are going to go ahead and map to those. Now I tend to be in that latter

38:27 camp like I don't

38:27 think that I think like systems are going to continue to use in fairly similar

38:31 ways maybe there's

38:32 more agents using them. I don't think they're going to evolve as much. Elon

38:36 might be back in the

38:37 like anthropic category of the anthropic growth marketer which is like he like

38:44 you know over the

38:46 years when you kind of like study the various IT you know departments of his

38:49 companies like

38:50 they are the most I mean. Well he could do that. He could do it. He's the most

38:54 home growth like

38:54 everything. This is first principle. Elon AI would do that. But also it's fun.

38:59 And then from

39:00 your mortals you're like yeah we kind of just want to CRM system. Like no but

39:03 it works the same way

39:04 every time. I mean this is not this is not it also hasn't been been not tried

39:10 before.

39:11 Like if you were to look at an ERP system from first principles you know well

39:15 in 1970 whatever

39:17 when SAP started there were a bunch of different assumptions and today you

39:20 would start from a

39:21 different set of assumptions about what's important and you would architect a

39:24 thing completely

39:24 differently but then it would still only last like 10 years until you thought

39:28 wow that was a

39:28 broken decision. And I and so I think that that there's intentionality in

39:34 layers but you

39:35 but there's also this first principles thing. And you know there's that always

39:40 will exist because

39:41 the decisions you can make at first principles at any given time mandate a

39:45 whole bunch of different

39:46 stuff. And so even if you don't go with LIDAR which made total sense 10 years

39:51 ago you still need

39:52 10 or 15 years to get to where LIDAR not having LIDAR worked. And then now

39:57 there's going to be a

39:58 whole bunch of other things that you're like wow we could have done that

40:01 completely different.

40:01 And so I feel like this is again like this discussion about trying to race to

40:05 an end point.

40:06 But let's see a first example of what you described happening. And I think that

40:11 that's going to be

40:12 the real tell because I think that there were just companies will figure all

40:16 this out. And I think

40:17 that they will just fall back on layers and architectural models because it's

40:21 the only way.

40:22 We don't think about it for policy. We don't think about it for security.

40:24 But it's also the only way to build a system. Otherwise you're just building an

40:29 app. And if

40:30 you're building an app to do one thing we don't need all of this. Like there's

40:34 a whole different

40:34 way to do it. The thing that I'm pretty fascinated by is and I don't even have

40:39 any amazing data

40:40 points or anecdotes but at least the notion of these sort of companies that are

40:45 emerging in these

40:46 kind of services categories from the ground up from the pure first principles

40:50 approach which is

40:51 like okay well if I could start a marketing agency or consultant you know

40:56 engineering consulting

40:57 company or I don't know maybe somebody's doing this for law firm. Construction

41:01 work. Well maybe

41:03 construction design. Architecture design. Architecture design. Anything that

41:06 would be like a knowledge

41:07 worker kind of services company. Because you could kind of build your company

41:11 pretty differently

41:11 if you had no constraints of I have no information barriers and boundaries of

41:16 what people should

41:17 have access to. I can give the agent just all the context it needs to do its

41:20 work. I can write

41:21 software on the fly for particular things like like I do think that will be

41:25 relatively disruptive

41:27 you know for some time until the bigger incumbents can kind of you know get out

41:31 of the way on this.

41:32 And that will at least create you know some precedent or case studies of what

41:36 what this new

41:38 sort of corporation could look like but I do you know over time they'll still

41:41 run into the same

41:42 exact problems of every other corporation. Well they'll run into geography or

41:46 market segments

41:47 you know or distribution challenges. Yeah like those those things anything

41:52 outside your little

41:53 walls. Yes. You'll run into the physical world. Right. I do kind of like the

41:57 idea that there are

42:00 some new business models that open up now. Oh of course. Yeah yeah yeah.

42:03 Because like there's

42:04 so much either information or software that that basically goes underutilized

42:09 by like 100x

42:10 relative to like what what its economic value is. Simply because like nobody

42:14 wants to pay five

42:15 cents for accessing a you know a piece of data or use a tool for one dollar

42:19 once but like you do

42:20 give these agents you know a budget and a protocol to work with and all of a

42:24 sudden you're like oh

42:25 like on the fly they can go get medical research for some deep research tasks

42:29 they're doing

42:30 and I'll pay like three dollars for that and the agent is able to go and

42:33 transact.

42:33 Like it kind of opens up a whole new world of business models for all of the

42:37 internet.

42:37 Let me oh I'm gonna you're that was too nice. Oh okay. No no. You're gonna go

42:41 farther. No that one is one

42:43 where that's actually the biggest I think that the biggest sort of in the air

42:49 problem right now

42:50 is everybody is trying to figure out the economics of all of this. When they're

42:54 off by at least in

42:56 order of magnitude on how big the opportunity is. Because the new models that

43:00 people will come up

43:01 with that nobody knows what they are right now but they will absolutely come

43:04 out with new models

43:05 because that's what happens with every new technology and the thing that holds

43:09 back to the sort of

43:10 the discussion now is you basically have a bunch of finance and Wall Street

43:14 people trying to justify

43:15 GPUs and tokens and things like as if we're in some old world and they're there

43:21 so they're

43:22 they're viewing the world of revenue as sort of this linear step but literally

43:27 linear growth

43:27 curve and trying to justify all the all the expect when people are going to

43:31 create like this was

43:31 the problem with PCs people viewed PCs as a finite market because they just

43:36 viewed the consumption

43:38 of MIPS as some finite thing and they didn't think what would happen if we put

43:41 all those MIPS on every

43:43 desktop and in particular people thought software just came with the MIPS and

43:48 nobody thought oh well

43:49 they'll just sell the software one guy did and it turns out that was like a

43:53 really good idea and

43:55 the same thing yeah and the same thing happened but the same thing happened

43:59 with the cloud

44:01 which was people looked at the cloud and they said oh we're going to take all

44:04 of the the server

44:06 business which was like literally like 60,000 units a year right and we're just

44:10 going to move it

44:11 to someone else's data center right and that's the that would be the business

44:15 and then we'll

44:15 divide up the the price right and nobody went oh they're going to people are

44:18 going to use a thousand

44:20 times as much right of the resource leveling right if we move it there and that

44:24 's exactly I mean

44:26 that's the thing that I it just drives me absolutely bonkers that the Wall

44:29 Street models have this

44:31 fixed revenue right pie zero some things and and it's this weird zero sum where

44:35 they they just

44:36 think that the amount of money that a company is going to spend and like this

44:38 was the problem with

44:39 with Salesforce that they faced when you were starting to but like mark was was

44:43 just blazing

44:44 the trail which was like the CRM business was two billion a year and it was two

44:49 billion in like

44:50 you had to go buy all these servers and these Oracle licenses and this huge

44:53 headache and years

44:54 of deployment and consulting yeah when if you could just get salespeople to

44:58 sign up individually

44:59 they all will sign up right with no friction and that's that is exact there is

45:04 no no doubt that

45:05 that is what's going to happen with AI let me give you an example of this so I

45:08 um you know I've

45:09 been in for investing for 10 years now I probably have a portfolio of 240

45:14 companies that work with

45:15 some visibility to us saying that 50 of them these are all infrastructure

45:17 companies some

45:18 historically done well some not so well every single one of them has gone asym

45:22 ptotic in the

45:23 last six months and you're like okay why is this it just turns out there's so

45:27 much more software

45:28 being written out than ever has been of course and so it's like and it's not

45:32 because they've got

45:33 enterprise customers you know it's just because there's just so much

45:36 consumption of the of the

45:38 infrastructure layer right now and so with more software with more agents there

45:42 's going to be a

45:42 lot more consumption of computer resources so certainly in the case of the

45:45 computer side of

45:46 things yeah and that's well we haven't even gotten to the point yet where

45:49 everyone's phone

45:50 is a huge consumer of AI right so once everybody's phone and on device yeah

45:56 like once your phone

45:57 on device is consuming AI the amount of it is going to go up by a billion so do

46:01 you like the

46:02 micropayon piece all of them yeah the micropayons there's a little bit of micro

46:06 payments that has

46:07 come with every technology where they always think that like you'll be able to

46:10 get like a fraction

46:11 of a penny but but in the end especially in the enterprise yeah like people are

46:15 just going to

46:15 consume things it's just cheaper and easier to buy like a bulk license for a

46:19 bunch of stuff yeah

46:20 you want some predictability on that well you want predictability and you just

46:23 want like to not

46:24 yeah have to think about it i just i like the idea that it is the first time

46:28 that you could like

46:29 there's just the agent doesn't care about the friction of a small transaction

46:32 right right

46:33 the first time that you can have resources behind a paywall that something will

46:36 actually be willing

46:37 to pay for that and resource the world is built up to infrastructure to to

46:41 aggregate those payments

46:42 into something efficient for a customer or a service right and and because

46:47 tokens are such a

46:48 significant part of cogs right now it is pushing the industry to do usage base

46:52 in the way that

46:53 we have like i remember when we went from like perpetual right right occurring

46:56 and that required

46:56 like a bunch of huge changes like we're like we're going through the exact same

47:00 change right now

47:01 towards usage base and usage base is pretty granular and it actually allows i

47:04 mean to get you will

47:05 have a contract with like you know we went through this with with AWS like

47:10 people learned to do the

47:12 to do the usage credit and we went through the phase where like people were

47:15 like so terrified of

47:17 cloud compute that they were like we need companies in the middle to help us

47:20 find the cheapest and

47:21 to arbitrate it all okay well now you write tokens into this and i don't see

47:25 how we possibly have a

47:26 time in this in this conversation but as long as you guys oh okay but like like

47:30 the the engineering

47:32 compute budget conversation to me is going to be just like the most wild one in

47:36 the next couple

47:37 years it's just like how much did you allocate of your engineering expense to

47:42 tokens and it's like

47:43 you you know depending on who you read on twitter it could be one percent and

47:47 the other and other

47:48 side could be a hundred percent and it's like yeah but this stuff well no no c

47:52 fo's have to

47:53 literally they actually have to know the answer i understand they have to know

47:56 okay cfo's always

47:57 want to know the answers to things that don't have answers no wall street is

48:00 going to make them

48:01 know the answer no no wall street is going to make them come up with some

48:04 number and hold them to it

48:05 then they'll get fired and then it'll but it you just okay okay i hear you are

48:08 in the rnd is

48:09 somewhere between 14 to 30 percent of revenue of any public right right company

48:14 let's just say

48:15 okay the difference between compute being 2x the cost of your engineering team

48:20 or or you know you know

48:22 three percent more is like that's EP that's all your EPS i get it i mean we

48:27 will have to know

48:28 the answer i'm perfectly willing to sacrifice a few cfo's okay i want that that

48:33 's a good clip by

48:33 the list but but the reason the reason is is because again this is this is

48:38 trying to know what we just

48:39 don't know right now yeah and and this has happened with internet bandwidth

48:43 this has happened no this

48:45 is not even close to internet bandwidth oh no no i i beg to differ like like

48:50 people were free it

48:51 happened with vacuum tubes it happened with transistors it has happened with

48:55 every technology

48:56 there was this oh my god let it happen with programmers there was a there was a

49:00 time when

49:01 programmers were going to swallow every company yeah and that's not it was in

49:04 my lifetime not

49:05 some made-up yeah but i don't think we've ever had a point where no the end the

49:09 end you every end user

49:10 in an organization has has sort of a completely elastic ability to spin up a

49:15 resource on their

49:17 behalf well it's certainly that actually is actually in many cases very valid

49:20 for them

49:20 to go spin it up but it certainly it certainly rhymes with what happened in the

49:23 early 2000s

49:24 with the cloud i remember very similar discussions when we went from capex to

49:27 opx and then our limited

49:29 span oh no and they were remember there were companies who the cfo's would sit

49:32 in our briefing

49:33 center here yeah and say you don't understand we are like exactly we are an

49:39 agriculture company

49:39 we we only know capex yes we we have no or or no we're all through this right

49:45 right no we both did

49:46 well or or like oh no we are an opx based company so if you tell us i we love

49:52 the cloud because we

49:52 just shipped everything it shifted everything to opx yeah and so all of the

49:56 stuff like the rules

49:57 of accounting work out also don't i i keep thinking do not discount the local

50:01 compute engine as being

50:03 a release valve for all of this when that can happen well the the question is

50:07 it's not when does it just

50:08 happen with today's view of the technology but how all of a sudden well wow

50:11 there's almost that

50:12 historically ever gone that direction yeah exactly because the opposite right

50:16 no it went all to the

50:17 client well okay and then go back to the 80s yes no that's most of the examples

50:22 that were here

50:22 so far whoa that was uncalled for i did it back in tubes he's talking about

50:28 vacuum tubes but i do

50:30 those examples because you can't argue with them and it's much easier that way

50:33 i you're right i can't

50:34 prosecute no but but but it's only been you know 10 or 15 years that it's all

50:40 that it moved back

50:41 to all cloud and then what has happened recently with that a lot of people wake

50:45 up in the morning

50:45 and they say oh we're moving back to doing some critical but stationary work

50:50 flows on on prem and

50:52 with AI that's true dude you wrote the blog post man uh don't let me go through

50:56 that the

50:57 archive of the repatriation i had to deal with so many wall street questions on

51:01 that one by the way

51:02 well because also because you're a competitor what went went back to oh yeah we

51:06 're talking about

51:07 too we're talking about too very good i agree i agree with like building your

51:11 own data center is

51:12 i'm talking about this this notion of edge computing where things go to devices

51:16 like that

51:17 seems to be i'm more i'm more in the like a cloud maximalist the camp but but

51:21 sorry so you just don't

51:23 think you don't even think for like one second that it matters whether like how

51:26 you're supposed to be

51:27 an engineering leader right now managing the compute budget of the engineering

51:31 team

51:31 no of course it matters i just this thing in the long term this thing will get

51:35 oh sure oh at

51:36 launch of all i mean what are we who cares we don't know i guess what i mean

51:39 here's our pragmatic

51:42 but here's a rule here's a rule of thumb first like the startups are going to

51:46 burn through available

51:48 capital pretending like it's not a problem yeah and they are going to do that

51:51 yeah a lot of that

51:52 anyway right right right and a lot of big companies are going to be so

51:55 terrified they're just going

51:56 to freeze and not do anything and then people are going to actually start

51:59 buying it on their own

52:00 and they're going to do all the things that companies do when they're big have

52:03 a lot of money but don't

52:04 want to spend it and in the middle we are going to see like if you pick a

52:08 category of product or go

52:10 to market or something they're going to be people who are willing to make the

52:13 bet for whatever reasons

52:14 that they can because of their financials and they are going to go ahead and

52:18 they are going to become

52:20 the people who lead yeah in the space as long as they can maintain the

52:24 financials yeah now they

52:25 might do it in they might say oh we're going to just do it here in this

52:28 particular application

52:30 space or here in this particular usage space but this this idea that there are

52:35 that nobody is going

52:36 to go in and because they're so terrified that the CFO is going to get fired or

52:39 something no no

52:40 it's just crazy yeah and but then there are going to be CFOs who make a mistake

52:44 and like okay everybody

52:45 gets a little yes well if they do that that's a complete fail yes because but

52:49 but also or like you

52:50 get you like you there is a really interesting like um on you know finesse here

52:55 which is like

52:56 you don't really want your engineers right now having to think about compute

52:59 budget because we're

53:00 still developing the oh okay so that set you over there but I was feeling like

53:05 we've been having

53:06 the discussion for 15 years when it comes to clag this is totally new like only

53:11 like 10 only like

53:13 10% of your engineering had to think about it in two in two thousand sixteen in

53:17 two thousand

53:18 eighteen time frame there's a whole set of companies that was basically like

53:22 like the dashboard for

53:24 what was it called phenoms where the developers is very cool right now because

53:28 the developers would

53:28 have would developers have access because cloud spends are getting out of

53:32 control yeah P.I.

53:33 spend were getting out of control yeah and so it was like you know here's your

53:35 Twilio spin here's

53:36 but but you know it's it's pretty different and I'm gonna wait for all the

53:40 comments to come in on

53:41 YouTube to call you out of this like it's it's like you can get into a

53:45 conference room and just be

53:46 like hey can you make that one you know kind of algorithm a little bit more

53:49 efficient so you don't

53:50 use as much you know of our cluster at this time of night or whatever and then

53:53 you get out of the

53:54 meeting somebody goes improves it and you're good this is like every single

53:58 prompt that every

53:59 engineer is doing like do you like you have to decide like if you want it to be

54:03 a long running

54:04 prompt do you want that you want to be a long running agent do you want to

54:07 parallelize that

54:07 like like do you want like what is your comfort level of wasted tokens like for

54:12 me right now I'm

54:13 like yeah we should probably waste a lot of tokens because that means that we

54:16 're like trying new

54:17 things yeah and like like should your head of engineering be happy if if you

54:21 run 10 experiments

54:23 in parallel and thus you're obviously going to be wasting 90 percent of the

54:26 tokens um but you're

54:28 going to choose one of the successful past or do you want to tell the team know

54:31 before you go do

54:32 that make sure to like like really go and design the perfect system like we

54:36 actually have a whole

54:36 bunch of open questions that are going to start to happen like literally as of

54:40 this recording time

54:41 people are freaking out right now on the new cloud code max plan you know whole

54:45 like like

54:45 because they're getting blocked after like three prompts well this is this is

54:49 going to be like a

54:50 very like real topic until we can actually find a way to build data center

54:53 capacity oh that's a

54:54 different problem okay no because no one is well assuming well wait no you can

54:58 assume that if we

54:59 build more capacity the price will drop because there is more capacity and we

55:02 're priced now based

55:03 on limited capacity whatever but like this is just going to get worked out and

55:08 I feel bad for the

55:09 those that have to make a decision immediately yeah about which 17 people get

55:14 no more tokens this

55:15 week or whatever and that the whole company is walking around with like a token

55:18 card and

55:19 and the person in the lunch line yeah yeah like it's the person in the lunch

55:25 line is punching their

55:26 card every time they do set but but like I you know I I don't know like

55:30 somebody we were talking

55:31 about today about performance and how like you know we used to write command

55:35 line tools that spit

55:36 out the time it took after you ran a command line just so you knew and if you

55:41 knew you were getting

55:42 better or worse and you know but you the thing is is this is all going to go

55:48 away there's absolutely

55:49 no doubt yeah that this just goes away I think I'm 10 or 10 and the biggest

55:53 reason it does is

55:55 is because you you have to do the penny off kind of math which is if you're

55:59 paying an enterprise

56:00 salesperson you know a million dollars a year you have to ask how much is their

56:04 tool worth yeah

56:06 and if you're paying an engineer X dollars a year well at some point their

56:11 their tooling is worth

56:12 it's absolutely worth it and it's not going to even be an issue and and yeah

56:17 yeah I don't think it's

56:18 um I think so if there's a capacity thing in the short term yeah that's a

56:22 different that is a

56:23 different problem driving the price than this just we're gonna forever have to

56:27 be in some budgeting

56:27 exercise I think I think a lot of large numbers solves this because because

56:31 eventually you have

56:32 enough engineers using this much compute but like we're in a transition phase

56:35 we're like

56:36 most people thought you know the two year ago level of spend on AI which is a

56:40 guy it's a chat

56:41 bot and yeah yeah but they were wrong yeah right okay but they were wrong but

56:44 they were wrong we

56:45 tried to warn them no but they were wrong because they saw it as this this

56:50 particular use case yes

56:52 and then but again like you you know like the vacuum tube thing you made fun of

56:58 yeah but like

57:00 there was a time when they thought that the that like whole like all of the Dak

57:04 otas would be covered

57:05 in vacuum tube warehouses and people on roller skates would be running up and

57:09 down the isles

57:10 replacing vacuum tubes just so we could fight World War II I mean like that was

57:15 how that was the

57:16 and they thought that and then someone said hey how about a transistor right

57:20 and like we are going

57:21 to have a transistor moment with all of this it might just be more supply the

57:24 way we think of it

57:26 but it also might be an actual algorithmic fundamental change it could be a

57:30 change in the

57:31 hardware there's a lot of stuff that can happen yeah that changes this

57:34 particular moment in time

57:36 it's just this I think is particularly weird that everybody has just gotten to

57:40 token yeah which is

57:42 the same thing that happened with IBM and mainframes people were on MIPS and

57:45 then one day the reality

57:47 was IBM was selling more MIPS for fewer dollars every year yeah and didn't even

57:52 realize it and they

57:53 were still pricing their mainframes by MIPS until it got pointed out to them

57:58 that they were on a

57:59 decreasing curve because they were making MIPS faster than they can charge for

58:02 and that's what's

58:04 going to happen guaranteed I just said that in a hardcore way I thought that

58:07 sounds great yeah

58:09 like sounds really great to sound like I know what I'm making guaranteed I

58:12 actually probably

加载中...