Pickful: Your Shared Crypto Community!

· 42m

Basecamp 现已通过全新的、功能全面的命令行界面（CLI）向客服代表全面开放，该界面由一款出色的技能所封装，并依托经过全面升级且大幅扩充的 API。这是一种绝佳的方式，既能让客服代表访问 #Basecamp 中的所有内容，又能将其集成到任何地方。 #cli #ai https://x.com/dhh/status/2036860598785356219

中文 English

0:00 有一整批消费者已经把人工智能等同于产品

0:07 攻击，跟扯淡一样。

0:09 人们并不是讨厌人工智能。

0:11 他们讨厌的是因为你想留住你的

0:16 产品而搞出的糟糕人工智能。

0:17 欢迎收听 Rework，这是 37signals 关于更好工作方式和企业管理的播客。

0:21 你的

0:21 业务。

0:21 我是 Kimberly Rhodes，来自 37signals 团队，与我一同的是 Jason Frieden 和 David

0:26 Heinemayer Hanson。

0:28 我们将谈论科技和人工智能相关的内容。

0:31 Basecamp 最近已支持智能代理。

0:35 这样你的 AI 代理就可以在 Basecamp 内工作了。

0:38 我不是专家。

0:39 接下来由你们来说。

0:40 David，你想先说吗？

0:41 当然。

0:43 我先挑剔一下这个词，就像我们刚开录之前讨论的那样。

0:47 “代理友好”这个词不错，不过对我来说更

0:53 关乎无障碍性。

0:54 在传统网页设计中，这就是我们对无障碍性的争论，意味着

0:58 让视障或

1:05 视力有限、色盲等人群更容易使用网站

1:05 那些特殊因素可能让他们难以使用未考虑这些的应用程序。

1:11 所以我们做无障碍工作。

1:15 进行对比测试。

1:20 如果某两种颜色放在一起，效果不佳。

1:22 色弱者看低对比度内容会很困难。

1:24 所以需要调整。

1:28 引入不同的颜色，增加对比度，

1:31 针对完全失明的用户做这些优化。

1:32 确保键盘访问性很高，能完全用键盘操作。

1:36 在不同元素间切换时，有合适的流程和节奏，

1:39 不会跳来跳去让人摸不着头脑。

1:43 我就是这么看待我们为代理无障碍做的工作。

1:44 这些 AI 代理在许多事情上非常聪明，

1:50 但在使用网站时却还会笨拙。

1:51 它们勉强能用，但速度很慢。

1:55 一个月前我们开始做这项工作时，我做了个大测试，

1:55 想着我们是不是根本不用改进。

1:59 目前的代理、现成模型，云端代码和开源代码等，

2:00 它们能不能直接做到？

2:07 它们到底需不需要帮助？

2:10 我惊讶于这些代理仅用浏览器操作竟然如此成功。

2:15 我能让其中一个代理注册 Basecamp，

2:21 在 Campfire 自我介绍，

2:23 Vizzy 和 Hey 代理也能做到，尽管速度非常慢。

2:29 那种感觉是，我能看到未来，但不知道未来何时到来。

2:33 看起来至少还有一年。

2:33 可能是两年，五年也说不准。

2:35 如果我们今天想让 Basecamp 成为使用代理的好地方，

2:42 得让它速度快起来。

2:46 没人会有耐心等代理在聊天界面慢吞吞地处理任务。

2:47 它们只是机械地输出信息，对吧？

2:52 然而这个速度必须适合在 Basecamp 中使用它们。

2:58 这就是所有代理无障碍工作的意义。

2:59 就像是铺了条小斜坡一样，

3:02 轮椅无障碍，你本可以用别的方法上去，

3:05 但很麻烦甚至有点危险，

3:09 有了斜坡就能轻松滑上去。

3:12 这就是我们给 Basecamp 做的事，

3:15 创建了命令行界面（CLI），

3:21 AI 代理可以用这个工具。

3:24 过去六到九个月中，我们发现给 AI 这些终端工具时，

3:26 它们变得超级强大。

3:29 这就是最近三四个月代理加速开发热潮的原因，

3:31 程序员和设计师突然能做更多事了，

3:34 不仅仅是向训练有素的 LLM 提问，

3:37 然后得到答案。

3:40 而是让代理去执行操作，代理会尝试完成任务。

3:44 就像人一样，有时不会做，

3:48 调用错误工具，犯各种错，

3:49 但如果能持续反馈，反复尝试调整，

3:52 不断试错，最终成功，

3:58 就能快速前进。

4:04 但这需要工具使用快速。

4:09 虽然我很佩服代理能用浏览器，且用得不错，

4:14 但了解它们为何这么慢后就明白了，

4:18 现在它们实际上是在截图，

4:22 通过图像分析分解界面元素和字段，尝试理解，

4:23 非常厉害，但非常繁琐且慢，

4:30 相比之下如果用命令行界面。

4:30 只需要文本。

4:35 这正是这些大语言模型被训练的内容，数万亿的文本数据。

4:39 它们读文本，预测下一个命令，

4:43 下一段文本。

4:43 and then succeed, it can really quickly move on.

4:47 But that loop requires that the tools are fast to use.

4:51 And as impressed as I was that the agents could use a browser and could use a

4:56 browser really well, you also realize that when you know how they do it, why it

5:02 's so

5:02 slow, right now the agents are literally taking like a screenshot of the screen

5:08 and then running it through this image analyze to break down what are all the

5:13 elements, what are all the fields and then trying to reason about it.

5:16 Very impressive, as I said, but also very cumbersome and very slow versus if

5:22 they

5:22 do all this work through a CLI, a command line interface, that's just text.

5:26 Literally what these LLMs have been trained, like trillions of tokens.

5:32 They've gone through text, text, text, what's the next command?

5:35 What's the next token?

5:36 他们在这方面又快又好。

5:39 所以如果它们能保持在那个小循环里，你就能得到每秒的标记数，

5:43 对吧？

5:44 这很接近就像是在写你的傻故事时的速度。

5:47 而这些之间的差别就是一切，对吧？

5:52 速度，正如我们常说，是一个特性。

5:56 当代理能为你做事，而你又有一些

5:59 耐心去等待时，速度就是一切。

6:02 这是“我懒得做”和

6:04 “让我让代理去做更容易”之间的差别。

6:07 我一直在试验我们新代理在 REMC 或 CLI 上的可访问性。

6:14 尤其是最近用的最快的代理之一，

6:20 Kimmy K25，一个超级快的开放式等待模型，真的让我震惊，

6:26 如果你让它写故事或解决逻辑问题，

6:30 可以达到每秒200个标记的速度。

6:32 也就是每秒100多个词。

6:36 超级快，对吧？

6:38 然后我出于好玩，让它帮我设计一个基地营项目，

6:43 规划基地营五号的发布活动。

6:47 当然，我并没有给它所有背景资料。

6:49 所以它不会给我完整的发布计划。

6:53 但它列出待办事项的速度，拆解得很细致，

6:59 加了备注，加了信息，还添加了日程项目。

7:04 比如说如果你四周后要发布，你得提前一周开始做这个。

7:07 你得在前一周做那个。

7:09 我当时简直惊呆了，这太不可思议了。

7:13 很有趣的是，当你想为什么这感觉特别时，

7:19 如果我18个月前问 ChatGPT 3 给我出一个基地营发布计划的文本，

7:25 它也能做出来。

7:29 我当时可能也会觉得挺不错的。

7:32 但我拿到它后，接下来该怎么办呢？

7:34 这就是代理可访问性的魔力。

7:36 你把所有的智慧、洞察、总结，还有我们

7:42 喜欢的大型语言模型的特点，

7:46 变成可用的东西。

7:49 它不仅仅是一堆文本块。

7:51 不，它是一个完整的项目。

7:54 这里有一些待办事项，有些分配给你，

7:56 有些分配给代理，

8:00 有些分配给 Jason 或者其他参与这项工作的团队成员。

8:05 突然间，所有的智能都变得可执行了。

8:08 说这些话的时候真的得小心，别听起来像是 Salesforce 的机场广告。

8:13 智能变得可执行，砰！但有时候这些

8:18 惊奇的发现确实让你感叹，“哇，真是太棒了”。

8:24 就像我用这些代理进行编程一样。

8:30 我曾经很长时间非常好奇，

8:33 也非常着迷。

8:35 我喜欢问它各种问题。

8:36 然后我又自己动手写了点代码。

8:39 就是用我这两只小手，突然一切都变了，因为

8:44 代理开始能做事了。

8:45 它们能用终端，

8:46 能运行代码，

8:47 能运行测试，

8:49 能重写测试失败的代码，

8:51 能进入那个循环。

8:53 我从“哦，这挺不错”

8:56 变成了“这真是太疯狂了”。

9:00 你去做，我需要时会介入。

9:03 如果我们能把这种能力推广到其他一切上，比如项目设置，

9:09 分工，进度检查

9:13 和回访，

9:14 当这些超能力触手可及时，

9:19 因为所有内容都在基地营共享，

9:23 这不仅仅是我和代理的私人对话。

9:26 整个公司，整个项目，每个人都能围绕这些工作

9:30 协作，且不止用一个代理，

9:34 而是多个代理。我们已经在基地营大量实践了，

9:40 它们驱动着一些内部流程。

9:42 我们有很多不同的代理。

9:44 有些是个人人工代理，代表他们行动，

9:48 还有一些自动代理也做这些事情，并进行跟进和提醒。

9:52 我知道这种爆炸式发展在许多先行公司的

9:53 领域内正在发生，

9:57 但还有剩下98%的世界，

10:01 他们或许用过 ChatGPT，但未用过代理。

10:07 他们没用过命令行界面（CLI）。

10:09 如果你很活跃在 X、推特等平台，看到大家高速发现、分享这些，

10:15 你可能会觉得整个世界已经这样在运作了，

10:20 其实不是，完全不是。

10:22 我们现在所做的工作，比如代理可访问版的

10:25 基地营，使用 CLI、技能等，

10:29 甚至还

10:29 没能覆盖

10:33 广大群众。

10:34 我们还差一步，让它像用 ChatGPT 那样简单，

10:35 而 ChatGPT 几亿人都使用了，

10:42 并很快变得主流。

10:44 我们得跑向冰球会去的方向。

10:46 我觉得这就是这项工作令人兴奋的地方，

10:49 近两年来我们内部做了不少 AI 混合特性实验，

10:55 就是那些把 AI 融入产品

11:02 本身的尝试，比如说，它可能帮你总结，

11:06 给你建议，或以其他方式辅助你。

11:06 大多数公司的这类工作

11:10 成绩可以说是不太稳定。

11:11 如果不那么宽容的话，可以说是完全失败了。

11:18 事实上，情况糟糕到一部分消费者

11:22 已经把 AI 等同于产品退化和胡扯。

11:28 这是事实。

11:34 微软上周刚不得不公开道歉，说抱歉，

11:39 我们把AI乱塞进了画图、记事本，还有Windows的各个角落。

11:46 我们听到了，你们不想要那个。

11:50 当然你们不想要，因为这只是附加上去的，实际上

11:54 并不

11:54 够有用。你把它和消费者对明显有用的东西的兴奋感对比一下，

11:59 比如ChatGPT和其他基于聊天的界面，

12:04 几亿人在用它。

12:07 他们认为它是日常生活不可或缺的一部分，并且为此付费，

12:12 这些东西都说明人们不是讨厌AI。

12:15 他们讨厌的是被乱塞进来的糟糕AI，只因为你想把它

12:20 套进你的产品里，对吧？当然有好的方法。

12:24 有技术可以把AI嵌入产品，让它

12:29 完全

12:30 原生且不可或缺，人们不会拒绝，但这很难。

12:35 正如我说的，我们试了几年，看能不能找到很多

12:40 切入点，尝试了很多方法，但几乎没发布任何成果。

12:45 因为没达到最终标准，我们不想像记事本、画图那样发出烂产品遭到反弹。

12:49 所以我们继续努力，我相信我们会成功。

12:53 我们会找到它真正合适的位置，且非常合理的方式。但在那之前，

12:56 我们可以给大家提供一个简约且易用的Basecamp版本，

12:59 一个他们可以用现有喜爱工具操作的友好版Basecamp。

13:06 比如没有人在每天用Plot代码或Open代码时会说我不想要那个。

13:10 没人会说我不想能用指尖控制Basecamp，

13:14 不想把它连接到GitHub或Sentry或者其他任何东西。

13:17 他们会说想，因为这是显而易见的胜利。

13:21 这也是一种建议。

13:24 我想说的是，任何想搞清楚AI在产品中作用的人，

13:28 但你知道吗？

13:31 在你弄明白之前，如果还没弄明白，就先让它易用。

13:36 让代理更容易使用它。

13:38 这就是解锁我们向客户承诺了二十年的API所有潜能的粘合剂。

13:42 但99.9%的人从未使用我们的API，因为那需要程序员，

13:44 还得做各种准备，太贵了。

13:49 让代理易用，CLI，技能，这一整套把所有承诺带了过来，

13:53 虽然不能完全端到桌上，

14:00 但我们差不多了，不那么麻烦

14:03 的一盒子，很多人都能打开。

14:05 我猜年底前

14:11 这实际上会成为主流，因为会以大家都用的界面批量发出。

14:14 好了，现在代理可以用Basecamp了。

14:17 其他产品有什么打算？

14:18 嘿，Fizzy，那些会推出吗，还是只限Basecamp？

14:20 哦，全部都会推出。

14:23 实际上Fizzy的无障碍做得挺好，因为我们的朋友Rob发布了一个开源CLI。

14:27 其实我们还雇了Rob Salkas来帮我们做Basecamp CLI，

14:27 因为他做Fizzy CLI做得很棒。

14:31 我们会应用很多经验，继续打磨Basecamp CLI。

14:35 这是个AI的启发时刻，我们想让Basecamp更易用，

14:37 于是我们做了CLI，

14:42 而代理实际上写了大部分CLI代码。

14:42 CLI绝大多数代码都是代理写的，

14:43 他们很快就完成了65%。

14:51 然后我们花了几周时间把完成度提高到97%，虽然不是100%。

14:55 现在我们可以把学到的经验，应用到Rob做的Fizzy上，

15:01 推出官方完整打磨的Fizzy CLI和技能。

15:04 接着，百分之百我要这个，

15:08 我要让它支持我的邮箱。

15:09 当我用Fizzy或Basecamp时，把邮箱和它们连接起来，

15:14 把所有东西串联，配一个可以听我指令的执行代理。

15:17 比如我经常用的旅行，就是让代理直接去查信息，

15:20 我只想拿邮件里的一个信息，

15:24 不想翻邮件。

15:25 不想看邮件内容，

15:31 只想要邮件里的事实。

15:35 我很期待给邮箱和日历都全面无障碍，

15:36 把它们串联起来。

15:39 我们会全部做到。

15:42 今后这都会是默认配置。

15:47 代理的无障碍不仅适用于直接使用应用，

15:52 也适用于连接更大的生态系统。

15:57 很多成功应用，比如Slack，一个很好的例子。

15:57 Slack之所以成功，是因为它有一个封闭的生态系统和集成。

16:04 代理无障碍正在让这一切变得开放给所有人。

16:07 所有壁垒都被打破了，

16:11 只在一个应用里有什么优势都不存在了。

16:13 实际上，游戏的关键是你的代理能和你用的任何东西对话，

16:17 访问任何数据，

16:18 不管数据在哪里，都能把它调回来或散布到各处，

16:19 把所有东西连接起来。

16:21 I just want the fact that's in the email.

16:23 I'm really excited for having full accessibility for, hey, both on the email

16:29 side

16:29 but also on the calendar side and tying all these things together.

16:32 So we'll do it for all of it.

16:33 And I think anything going forward, this is going to be baked in.

16:36 This is going to be table stakes that your application is, agent accessibility

16:41 accessible,

16:42 both for the direct use with that application, but just as much because it

16:47 kind of plugs in to this broader ecosystem.

16:50 So many successful applications over the year Slack is a good example.

16:54 We're successful because they ended up with this ecosystem that was kind of

16:58 proprietary to Slack and integrations with it.

17:00 What the agent accessibility is doing is basically bringing that to everyone.

17:03 Like all those moats just come tumbling down because there's no specific

17:08 advantage

17:08 to just having something inside of one application.

17:11 In fact, the whole game here is that your agent can talk to anything you use,

17:17 can access anything you use, wherever your data is, it can move it back here or

17:23 everywhere and tie it all together.

17:24 这真令人兴奋。

17:25 好的，你可以在 Basecamp.com/agents 上了解我们正在做的事情。

17:31 这是 37 Signals 的一部作品。

17:33 你可以在我们的网站 37signals.com/podcast 找到单元内容和文字记录。

17:38 视频

17:38 剧集在 YouTube 上。

17:39 如果你有关于更好的工作方式和管理的提问给 Jason 或 David，

17:43 台词是什么？

17:47 正要说那个。

17:48 我想，等等，我需要代理帮我说这句话。

17:52 如果你有关于更好的工作方式和管理的提问给 Jason 或 David，

17:56 业务。

17:57 给我们留言语音，网址是 37signals.com/podcast question。

18:01 或者你可以发邮件到 [email protected]。

18:05 那不是。

18:05 是这样吗？

18:06 你们，这就结束了。

18:08 是邮箱地址，对吧？

18:09 我想是的。

18:10 哦，天哪。

18:11 让代理检查一下。

18:12 真的，我需要代理来做结尾。

18:17 (欢快的音乐)

0:00 An entire contingency of consumers have come to equate AI with product

0:07 aggression, with bullshit.

0:09 It's not that people hate AI.

0:11 What they hate is shitty AI glob done because you want to stick around your

0:16 product.

0:17 Welcome to rework a podcast by 37 signals about the better way to work and run

0:21 your

0:21 business.

0:21 I'm Kimberly Rhodes from 37 Silver Hills team joined by Jason Frieden, David

0:26 Heinemayer Hanson.

0:28 We are talking about tech things and AI things.

0:31 And Basecamp has recently been made agent friendly.

0:35 So your AI agents can work within Basecamp.

0:38 I'm not an expert.

0:39 I will let you guys talk about it.

0:40 David, do you want to kick us off?

0:41 Sure.

0:43 I'll start by quibbling with the term as we did just before hitting record.

0:47 Agent friendly, I think is a fine term, but I think it's actually to me more

0:53 about accessibility.

0:54 And this was the contention we had about what does that mean traditionally in

0:58 web design, accessibility means making it easier for folks who might be blind

1:05 or

1:05 might have limited vision, might be color blind, have all of these factors

1:11 that make it difficult for them to use a website or an application that has not

1:15 been designed within any of those factors in play.

1:20 So then you do accessibility work.

1:22 You do contrast testing.

1:24 Oh, if you have this color next to this color, that doesn't work very well.

1:28 If someone's kind of blinder, they can't see low contrast well.

1:31 So you change it.

1:32 So you get some different colors introduced, you increase the contrast,

1:36 you do all those things for full on blind users.

1:39 You make sure that keyboard accessibility is really high, that you can do

1:43 everything with a keyboard.

1:44 When you tap from one element to the other, it goes in a flow and a rhythm

1:50 that makes sense.

1:51 And it doesn't just jump all over the place in something that doesn't make

1:55 sense.

1:55 So that's actually how I think about the work we've been doing for

1:59 agent accessibility.

2:00 You have these AI agents that are incredibly smart at a ton of things.

2:07 And then they still kind of stumble when they have to use a website.

2:10 They can barely just do it, but they're really slow.

2:15 I did a big test a while back about a month ago when we set off to do this

2:21 work, where it was like, I wonder if we even need to do anything.

2:23 Can the agents we have right now, the models that are out there, the harnesses

2:29 that people run, cloud code and open code and all the other ones, can they

2:33 actually just do it?

2:33 Do they even need any help?

2:35 And I was shocked by how successful the agents were at using just a browser.

2:42 I could tell one of these agents to sign up for Basecamp, to introduce

2:46 itself in campfire.

2:47 The same thing with Vizzy, the same thing with, Hey, but it was really slow.

2:52 It was one of those things where like, I can see the future, but I can't see

2:58 when it's going to arrive.

2:59 This looks like something that might be at least a year out.

3:02 It could be two, it could be five.

3:05 Who knows if we want to make Basecamp a great place to work with

3:09 agents today, we got to make it fast.

3:12 No one's going to have the patience to sit around and wait for agents,

3:15 take minutes and minutes to do work when you ask them in a chatbot interface.

3:21 And they just, they're spitting out facts, right?

3:24 If that kind of pace needs to be captured.

3:26 When you're working with them in Basecamp.

3:29 So that's what all this agent accessibility work is for.

3:31 It's sort of like a little, little ramps, right?

3:34 Like the wheelchair accessibility, like maybe you could get up with your

3:37 wheelchair some other way, but be very cumbersome and maybe even a little

3:40 dangerous, but if we just make it a little ramp, you can just roll right in.

3:44 So that's what we've done for Basecamp here by creating a command line

3:48 interface,

3:49 a CLI tool that the agents can use.

3:52 Because what we found over just the last six to nine months is that AI

3:58 becomes super powerful when you give them these tools inside of a terminal.

4:04 That's what all this explosion of productivity has been about, especially

4:09 the last three or four months on agent accelerated development work,

4:14 like programmers and designers suddenly being able to do way more stuff

4:18 because they're not just asking an LLM that's been trained on the whole

4:22 internet and getting an answer back.

4:23 No, they're asking the agent to do stuff and the agent will try to do something

4:30 .

4:30 And just like a human will realize that often they don't know how to do it.

4:35 They call the tool wrong, they do all these things, but if it can stay in a

4:39 feedback loop where it can try something, adjust its approach and then try

4:43 again,

4:43 and then succeed, it can really quickly move on.

4:47 But that loop requires that the tools are fast to use.

4:51 And as impressed as I was that the agents could use a browser and could use a

4:56 browser really well, you also realize that when you know how they do it, why it

5:02 's so

5:02 slow, right now the agents are literally taking like a screenshot of the screen

5:08 and then running it through this image analyze to break down what are all the

5:13 elements, what are all the fields and then trying to reason about it.

5:16 Very impressive, as I said, but also very cumbersome and very slow versus if

5:22 they

5:22 do all this work through a CLI, a command line interface, that's just text.

5:26 Literally what these LLMs have been trained, like trillions of tokens.

5:32 They've gone through text, text, text, what's the next command?

5:35 What's the next token?

5:36 They're so good and so fast at that.

5:39 So if they can stay in that little loop, you can get the tokens per second,

5:43 right?

5:44 To be quite close to when it's just writing your silly story.

5:47 And the difference between those things is everything, right?

5:52 The speed, as we often say, is a feature.

5:56 And when it comes to agents being able to do things for you and you having some

5:59 patience to wait for it, speed is everything.

6:02 It's the difference between I can't be bothered to.

6:04 It's easier to ask my agent to do it.

6:07 And I've been experimenting with our new agent accessibility on REMC or CLI.

6:14 And I've been mind blown how quickly, especially the fastest agents like

6:20 Kimmy K25 that I've been using recently, which is one of these super fast, open

6:26 wait models that if you just ask it to write a story or solve a problem, logic,

6:30 you can do 200 tokens a seconds.

6:32 That's whatever, 100 plus words a second.

6:36 Super fast, right?

6:38 And then I asked it to actually, just for the heck of it, design me a base

6:43 camp project that spells out the launch campaign for base camp five.

6:47 And of course, I didn't even give it all the context.

6:49 So it wasn't going to give me the launch campaign we're going to ship with.

6:53 But the pace by which it filled out the to-do list, it broke it all down.

6:59 It added comments, it added messages, it added items on the schedule.

7:04 Well, if you want to launch in four weeks, you got to start doing this

7:07 the week before you got to do this.

7:09 I was just like, holy, smoly, this is incredible.

7:13 And it's so funny because when you think about it, why does that feel special?

7:19 If I had asked chat GPT three 18 months ago about just creating a text

7:25 document with a launch plan for base camp, it would have done that.

7:29 And I would probably also been surprised like that's neat.

7:32 But then what am I going to do with it?

7:34 That's the magic of this agent accessibility stuff.

7:36 You take all the wisdom, the insight, the summarization, all the things we

7:42 like about LLMs and then you make it usable.

7:46 It's not just a block of text.

7:49 No, it's an entire project.

7:51 And here's some to do is some of them are assigned to you.

7:54 Some of them are assigned to the agent.

7:56 Some of them are assigned to Jason or whoever is working on this stuff.

8:00 And suddenly all that intelligence becomes actionable.

8:05 I mean, it's so funny when you talk about these things, you really got to be

8:08 careful not to sound like an airport advertisement for Salesforce or something.

8:13 Intelligence becomes actionable, boom, but sometimes some of these

8:18 revelations actually you sit with and you go like, Oh, wow.

8:24 Yes, much the same way as I've had it programming with these agents.

8:30 Like for a long time, I was very curious.

8:33 I was very fascinated.

8:35 I loved asking it stuff.

8:36 And then I went off and I wrote my little code by myself.

8:39 By my two little hands here and then suddenly things flipped because the

8:44 agents could do stuff.

8:45 They could use the terminal.

8:46 They could run code.

8:47 They could run tests.

8:49 They could rewrite code that failed the test.

8:51 They could get into that whole loop.

8:53 And suddenly I went from like, Oh, this is neat.

8:56 But like I'll write it myself to this is freaking incredible.

9:00 You go ahead, I'll intervene if I need to.

9:03 And if we can get to that for everything else, for setting up a project,

9:09 for for figuring out how to divide the work, for checking in on things and

9:13 checking up on things.

9:14 And we can have those kinds of superpowers at our fingertips because it's

9:19 all in base camp, because it's all shared in base camp.

9:23 It's not just like I have a personal little conversation here with an agent.

9:26 No, the whole company, the whole project, everyone can suddenly

9:30 collaborate around this stuff and not necessarily even with one agent, but

9:34 with multiple, we've started doing this extensively already at base camp, where

9:40 you're driving some of these internal processes that we have.

9:42 And we have a bunch of different agents.

9:44 Some have a personal agent that they're just using to act on their behalf.

9:48 And then we have some automated agents that do all of this stuff too and chime

9:52 in and follow up.

9:53 And I know this kind of explosion is going on at a fair number of companies

9:57 sort of on the leading edge, but then there's the rest of the 98% of the world

10:01 who who may have used chat GPT, but we haven't done anything with agents.

10:07 We're not running into CLI.

10:09 And if you're very plugged into X and Twitter and wherever else that people

10:15 are discovering and sharing all this stuff at high velocity, you think that

10:20 this is how the entire world is already working.

10:22 No, they're not, absolutely not.

10:25 And all this work that we're doing right now, the agent accessible version of

10:29 base

10:29 camp that uses the CLI and the skills and so on, even that is not going to

10:33 reach

10:34 the broad masses.

10:35 We are missing a step yet where this becomes as easy as using chat GPT, which

10:42 hundreds of millions of people have used.

10:44 Like that has actually become very mainstream very quickly.

10:46 But we got a skate to where the puck is going.

10:49 And I think that that's what's exciting about this work, that over the last

10:55 almost two years, we've had a bunch of experiments internally with AI infused

11:02 features, like these are the things where you try to bake the AI into the

11:06 product

11:06 itself and like, well, maybe it can summarize some things or can suggest

11:10 some things or it can help you in other ways.

11:11 And the track record for most companies on that work is let's say mixed.

11:18 And if you're less charitable, say completely fucked, right?

11:22 Like in fact, it has been so bad that an entire contingency of consumers

11:28 have come to equate AI with product degradation, with bullshit.

11:34 Microsoft literally last week had to come out with a mea culpa saying, sorry,

11:39 we shoved AI crap into paint, into notepad, into all these crevices of windows.

11:46 We hear you, you don't want that.

11:50 And of course, you don't because it was tacked on and it wasn't actually

11:54 helpful

11:54 enough. And you got contrast that with the consumer excitement for something

11:59 that's obviously good, like chat to BT and the other chat based interfaces,

12:04 like hundreds of millions of people are using it.

12:07 They're considering it integral to their daily lives, they're paying for it,

12:12 all these things. So it's not that people hate AI.

12:15 What they hate is shitty AI globbed on because you want to stick around your

12:20 product, right? Now, there are ways to do this well.

12:24 There are ways to put AI into a product and have it be something that feels

12:29 fully

12:30 native and integral to it and people wouldn't reject, but that's quite hard.

12:35 We've tried, as I said, for a couple of years to see if we could find a bunch

12:40 of angles on this and we tried many things and we shipped virtually none of it.

12:45 Because it didn't pass that final muster because we didn't want the notepad

12:49 paint backlash of shipping something crappy.

12:53 So we continue to work on that. And I'm sure we're going to nail it.

12:56 We're going to find some ways with this really fits and it really makes a lot

12:59 of sense. But until then, we can give everyone the age and accessible version

13:06 of Basecamp, the in friendly version of Basecamp that they can use with these

13:10 tools they already love. Like no one who's using plot code on a daily basis or

13:14 open code are going to tell you like, well, I don't want that.

13:17 I don't want to be able to just command Basecamp with my fingertips and tie

13:21 it together with GitHub or Sentry or anything else.

13:24 They're going to say, yes, please, because this is such an obvious win.

13:28 So that's also a bit of a recommendation.

13:31 I'd actually say for anyone else who's trying to figure out how AI plays a

13:36 role in the product, but you know what?

13:38 Until you figure that out, if you haven't already, just make it accessible.

13:42 Just make it easier for agents to use this.

13:44 This is that glue that unlocks all the promise that we told customers for

13:49 literally 20 years was going to be there with APIs.

13:53 And 99.9% of them never touched our APIs because it required a programmer.

14:00 It required all sorts of stuff to do it.

14:03 It was just too expensive to do.

14:05 Aging, accessibility, CLIs, skills, this whole bucket takes all of that promise

14:11 and then delivers it not quite on a silver platter.

14:14 We're not quite there yet, but like, I don't know, in a slightly less

14:17 cumbersome

14:18 box, quite a few people can open that.

14:20 And then by the end of the year would be my guess.

14:23 This is actually going to go mainstream because it's going to be shipped in

14:27 bulk

14:27 through these interfaces that everyone uses.

14:31 OK, so base camp is now accessible to agents.

14:35 What are we thinking for other products?

14:37 Hey, Fizzy, are those on the horizon or it's just a base camp solution right

14:42 now?

14:42 Oh, it's coming for everything.

14:43 In fact, we already have some pretty good accessibility for Fizzy because our

14:51 friend Rob put out an open source CLI for it.

14:55 In fact, we ended up hiring a Rob Salkas to work with us on the base camp CLI

15:01 because he did such a great job at the Fizzy CLI.

15:04 We're going to apply a bunch of those lessons that we've now taken really

15:08 polishing the base camp CLI.

15:09 This was one of those AI inception moments where like we're trying to make

15:14 base camp more accessible to the agents and we're making the CLI.

15:17 And the agent is actually making the bulk of the CLI.

15:20 Like the vast majority of the code that goes into the CLI was written by an

15:24 agent.

15:25 And it got like 65% of their of that done in no time at all.

15:31 Then we spend weeks and weeks and weeks and weeks getting it to, if not 100%,

15:35 then 97%.

15:36 Now we can take all of those lessons.

15:39 We can apply it to the work that Rob already did on Fizzy.

15:42 And then we can put out an official fully polished Fizzy CLI and skill.

15:47 And then after that, 100% I want this for, hey, I want this for my email.

15:52 I want my email to be tied together with my Fizzy when I'm using that or my

15:57 base camp

15:57 when I'm using that, bring all these things together and having this executive

16:04 agent that I can just tell to do stuff.

16:07 I mean, one of the things I use, hey, for a lot, for example, is travel and

16:11 having an agent

16:13 to be able to just go in and look these things up when I just want a piece of

16:17 information out of my email.

16:18 I don't want to draw through it.

16:19 I don't want the email.

16:21 I just want the fact that's in the email.

16:23 I'm really excited for having full accessibility for, hey, both on the email

16:29 side

16:29 but also on the calendar side and tying all these things together.

16:32 So we'll do it for all of it.

16:33 And I think anything going forward, this is going to be baked in.

16:36 This is going to be table stakes that your application is, agent accessibility

16:41 accessible,

16:42 both for the direct use with that application, but just as much because it

16:47 kind of plugs in to this broader ecosystem.

16:50 So many successful applications over the year Slack is a good example.

16:54 We're successful because they ended up with this ecosystem that was kind of

16:58 proprietary to Slack and integrations with it.

17:00 What the agent accessibility is doing is basically bringing that to everyone.

17:03 Like all those moats just come tumbling down because there's no specific

17:08 advantage

17:08 to just having something inside of one application.

17:11 In fact, the whole game here is that your agent can talk to anything you use,

17:17 can access anything you use, wherever your data is, it can move it back here or

17:23 everywhere and tie it all together.

17:24 That's super exciting.

17:25 OK, well, you can read more about what we're doing at Basecamp.com/agents.

17:31 This has been a production of 37 Signals.

17:33 You can find units and transcripts on our website at 37signals.com/podcast full

17:38 video

17:38 episodes on YouTube.

17:39 And if you question for Jason or David about a better way to work and run your.

17:43 What is the line?

17:47 Literally about to say that.

17:48 I was like, wait, I need an agent to say this for me.

17:52 You have a question for Jason or David about a better way to work and run your

17:56 business.

17:57 Leave us a voicemail at 37signals.com/podcast question.

18:01 Or you can send us an email to [email protected].

18:05 That is not.

18:05 Is that right?

18:06 You guys, that's it.

18:08 Email address, right?

18:09 I think it is.

18:10 Oh, my gosh.

18:11 Have an agent check it.

18:12 Literally, I need an agent to do the exit.

18:17 (upbeat music)

Moon

· 1h

TAO联合创始人：AI圈的人是“普通人”，加密圈的人是“自由斗士”

“你不能一边从萨姆·阿尔特曼那里拿钱——那个试图奴役我们的人——一边又假装自己是‘自由技术’。这到底想干什么？这看起来太奇怪了”

“有时候人们他妈的讨厌加密货币圈的人，但骨子里我们就是自由斗士。这就是为什么Worldcoin行不通，因为我们身边都是真正相信自由的人。这就是我们的本质”

“那些自以为高高在上、了不起的 #AI 圈人士，其实只是普通人。他们全都是‘蓝教堂’里的人。我超级喜欢他们，他们都很聪明，但他们不是追求自由的人。他们不真正理解地缘政治，也不真正关心这些。”

“他们只会去工作，赚点钱，然后去玩他们的桌游。但加密货币圈的人是叛逆者。你别想对我们搞什么Worldcoin。那行不通。”

https://x.com/CryptoMikli/status/2036585881184719127?s=20

中文 English

0:00 你不能从默想奴役我们的山姆·奥特曼那里获得资金，

0:04 然后假装自己是自由技术派，对吧，像你知道的那样，

0:08 你

0:08 想干嘛？我不知道，就是觉得奇怪。Orblep太

0:12 反乌托邦了。是啊，残酷。你为什么要那样做呢？然后

0:15 技术实际上很好，那帮人真的很聪明。

0:18 它有效，对吧？太棒了。我总是在

0:21 一些奇怪的事发生时发推，说有一个平行宇宙里，World Coin

0:24 是比特币的对立面，懂吗？是啊，对吧？这不是这个

0:29 行业的见证吗？人们有时候恨加密货币人，

0:33 但是内心深处，我们是自由斗士。

0:40 这就是World Coin失败的原因，因为我们中间充满了真正

0:44 信仰自由的人，你明白我的意思吗？这才是

0:47 我们真正的样子。顺便说一句，那些AI人以为

0:51 自己很酷、高高在上，实际上他们

0:54 都是普通人，都是朝九晚五的人，我非常喜欢他们，

0:58 他们很聪明，但不是自由派。他们不懂地缘政治。

1:02 他们不理解，也不能理解。

1:05 他们只会为博格工作，赚钱，玩他们的棋盘游戏。

1:08 而加密货币人则像叛徒一样，

1:12 你不能用World Coin那套来做，这行不通。

0:00 You can't get funding from Sam Altman, the guy who's trying to enslave us and

0:04 then pretend to be like, you know, like, free to techno, yeah, like what are

0:08 you

0:08 trying to do? I don't know, like, I just seemed odd. And like the Orblep was so

0:12 dystopian. Yeah, brutal. Well, why would you do it like that? And then the

0:15 technology is actually really good, though, like the guys were really smart.

0:18 It works, right? It's a great. There's, I always, I always tweet this whenever

0:21 some weird shit happens, there's like a parallel universe where world coin is

0:24 like flip, flip Bitcoin, you know? Yeah, right. Isn't that a testament to this

0:29 industry? Like the we actually, you know, people fucking hate crypto people so

0:33 that reason sometimes, but also like deep down, we're freedom fighters. And

0:40 that's why world coin didn't work is because we're permeated by people that

0:44 actually believe in freedom. You know what I mean? Like that that's that is

0:47 actually who we are. And that's not an AI, by the way, the AI people who think

0:51 they're so hot and high and high and mighty, they're all just they're just

0:54 normies. They're all blue church people. I love them to death. They're all very

0:58 smart, but they're not freedom people. They don't really grasp geopolitics.

1:02 You know, I mean, they don't really understand. And they don't really can't.

1:05 They're just going to go work for the Borg and make their money and live in

1:08 the play the board games. But the crypto people are like, we're renegades,

1:12 man. And so you can't try to do world coin on this. It's not going to work.

Moon

· 2h

罗布·格里尔表示，对于新晋的 #AI 开发者而言，最佳的创业平台是Bittensor，而非Y Combinator。

“我认为这是加密货币史上最精妙的代币经济设计。”

https://x.com/therollupco/status/2036204282144309637?s=20

中文 English

0:00 所以，如果你是一个，你知道的，波兰的天才开发者，你会去

0:04 硅谷

0:05 然后去Y Combinator接受稀释股权融资，还是会

0:09 在BitTensor

0:09 开一个子网，矿工就是你的AI开发团队，你甚至

0:13 不用付钱

0:13 ？

0:14 我想显然你会选择BitTensor，然后所有

0:17 这些子网的市值都会累积

0:18 到tau代币上，因为获得

0:21 alpha代币的唯一方式是

0:22 先购买tau。

0:25 我觉得这就是对以太坊的升级，对吧，因为

0:28 关于以太坊的争议在于

0:29 所有这些二层到底带来什么价值累积

0:33 给

0:33 ETH，对吧？

0:34 而tau则必须先买tau，而且这是一个非常有趣的

0:38 正和

0:38 游戏，因为Templar爆发会推高tau价格。

0:45 然后所有其他子网都会觉得太棒了，因为这有利于

0:48 他们的子网

0:49 和他们的发行。

0:50 所以实际上子网之间会互相支持，这是一个正和

0:54 游戏。

0:54 我认为这是加密史上最聪明的代币经济设计，

0:57 我想

0:57 很多人会同意我的看法。

0:00 So, if you're a, you know, a brilliant developer in Poland, are you going to go

0:04 to Silicon Valley

0:05 and go to Y Combinator and get diluted equity funding, or are you going to

0:09 start a subnet

0:09 on BitTensor, where the miners are your AI dev team, which you don't even have

0:13 to pay

0:13 for.

0:14 I think obviously you choose BitTensor, and then all of the market cap built

0:17 from all

0:18 of those subnets accrues the token tau, because the only way you could get

0:21 access to the alpha

0:22 tokens is by first purchasing tau.

0:25 And I think that's like the upgrade over Ethereum, right, because like the

0:28 whole controversy

0:29 over Ethereum is that like all these layer two is like, what is the value accru

0:33 al to

0:33 ETH, right?

0:34 Whereas with tau, you have to purchase tau first, and it's really interesting

0:38 positive

0:38 sum game to because Templar pops off, and that increases the tau price.

0:45 And then all the other subnets are like, that's awesome, because that benefits

0:48 their subnet

0:49 and their emissions.

0:50 So you actually have subnets rooting for other subnets, it's a positive sum

0:54 game.

0:54 I think this is the most brilliant tokenomic design in the history of crypto,

0:57 and I think

0:57 I think a lot of people would agree with me on that.

Love

· 1d

“ #Anthropic 和 #OpenAI 的日志里肯定保存着所有人的访问密钥、API 密钥和持有者令牌。” 如今的人工智能技术栈漏洞远比人们意识到的要多得多。 #IronClaw 解决了这个问题，因为“密钥永远不会接触 LLM”。 @ilblackdragon 联合创始人 @NEARProtocol 正在开发私有 AI 基础设施，其中“模型提供商和硬件提供商实际上都无法访问您用于 #AI 推理的内容”。换句话说，“没有任何一方能够单独解密你的数据”。https://x.com/Bankless/status/2036518654909821158?s=20

中文 English

0:00 人们在使用时没有意识到的一件事是

0:03 Entropic OpenAI，或者更糟的是，你使用其他东西

0:07 进行推理时，Open Cloud 实际上也会发送你所有的秘密

0:11 到那些服务。

0:12 是的。

0:13 所以在 Entropic 和 OpenAI 的日志中某处，

0:15 他们拥有每个人的访问密钥、API 密钥，

0:20 以及用于访问你的 Gmail、Notion 的承载令牌

0:25 和你的——

0:25 我们竟然在做这种事情，真的很疯狂。

0:29 是的。

0:30 首先，我不知道如何解决这个问题。

0:32 密钥绝不会进入 LMM。

0:33 所以即使你用这些集中式提供商，

0:37 虽然不应该，但至少密钥不会

0:39 进入任何 LMM。

0:41 所以这是我们希望做到的，

0:43 这是唯一必须首先做到的事情。

0:45 是的，是的。

0:46 但 Neerai 过去一年一直在努力开发的，

0:50 实际上是如何实现私密 AI？

0:53 也就是说，我们如何提供一种 AI，

0:58 让模型提供商、硬件提供商

1:00 都无法访问你用 AI 推理的内容？

1:05 所以我们有 Neerai Cloud，这是推理云。

1:08 你可以使用开源权重模型。

1:10 它运行在安全环境中。

1:12 它实际上使用——这也就是我

1:14 一开始提到的——

1:15 你会使用我们的多方计算网络，

1:18 这是 Neer 的一部分，用于加密

1:21 解密、备份以及所有内部机制。

1:25 这让你确信，

1:29 没有任何单一方可以解密你的数据。

1:32 没有人可以访问它。

1:35 你需要把所有多方计算网络

1:37 有效地全部集结起来才能访问。

1:40 所以——

1:40 好，那么你的意思是说，

1:43 你提供一种服务，配合

1:46 Iron Claw，类似于保密云

1:50 环境来运行 LMM 实例？

1:53 当然，你得运行开源权重模型。

1:55 但也许一些中国模型

1:57 是最好的，比如 Kimmy 或者类似的，

1:59 或者一些 DeepSeek——

2:00 是的，我们有 Quimmy、Quen、DeepSeek。

2:04 最新的热门我们也会加进去。

2:07 我们还有 OpenAI RSS。

2:10 所以，你可以在它们之间选择。

2:12 现代媒体一团糟。

2:13 每个平台都有偏见，但并不是所有新闻消费者

2:16 都意识到他们所接收平台的偏见。

2:18 这就是为什么当我想知道真相时，

2:20 我会去 Polymarket，世界上最公正的新闻平台。

2:23 从 polymarket.com 获取你的信号。

0:00 One thing that people don't realize when they use

0:03 Entropic OpenAI, or even worse, you use something else

0:07 for inference, Open Cloud actually sends all your secrets

0:11 to those services as well.

0:12 Yeah.

0:13 So somewhere in Entropic and OpenAI logs,

0:15 they have everybody's access keys, API keys,

0:20 and bearer tokens to access your Gmail's and your notions

0:25 and your--

0:25 It's actually insane that we're doing that.

0:29 Yeah.

0:30 And so first of all, I don't know what fixes that.

0:32 The keys never touch a LMM.

0:33 So even if you're using it with those centralized providers,

0:37 which shouldn't, but at least the keys are not

0:39 going ever into a LMM boop.

0:41 So that's something we'd like just like,

0:43 that's just the only same thing to do first.

0:45 Yes, yes.

0:46 But what Neerai has been working on actually for the past year

0:50 is actually developing, how do we do private AI?

0:53 So how do we actually offer AI where neither we, model

0:58 provider, hardware provider, is actually

1:00 able to access what you are using the AI inference with?

1:05 And so we have Neerai Cloud, which is inference cloud.

1:08 You can use open-weight models.

1:10 And so it runs in secure enclaves.

1:12 It actually uses-- and this is kind of what

1:14 I was referring in the beginning--

1:15 you'd use our multi-party computation network, which

1:18 is part of Neer, that is used for encryption

1:21 decryption for backups for all the internal machinery.

1:25 And that's what gives you this kind of knowledge that like,

1:29 hey, there's no single party who can go and decrypt your data.

1:32 There's nobody who can actually access it.

1:35 You would need to collect all the effectively

1:37 multi-party computation network together.

1:40 So there's--

1:40 OK, so is this-- are you saying, then,

1:43 that you offer a service in conjunction

1:46 with Iron Claw, which is almost like a confidential cloud

1:50 type of environment for running LMM instances?

1:53 And of course, you'd have to run the open-weight models.

1:55 But maybe some of the Chinese models

1:57 are kind of the best year, like a Kimmy or something

1:59 like this, or some deep-seak--

2:00 Yeah, we do quimmy, quen, deep-seak.

2:04 Whatever's new hotness will add it as well.

2:07 We have open-air RSS as well.

2:10 So yeah, you can choose between all of them.

2:12 Modern media is a mess.

2:13 Every platform has a bias, but not every news consumer

2:16 is aware of the platform's bias that they're consuming.

2:18 That's why when I want to learn the truth,

2:20 I go to Polymarket, the world's most unbiased news platform.

2:23 Get your signal from polymarket.com.

Love

· 1d

#Groq 提供快速、低成本的推理能力，即使在关键时刻也能稳定运行。

2025年12月24日，NVIDIA宣布以约200亿美元现金收购 #AI 芯片初创公司 Groq 的资产，这是 #NVIDIA 有史以来规模最大的一笔交易。

中文 English

0:00 “推理对于我们在一级方程式所做的一切都至关重要。我们投入了

0:05 巨大的资金

0:06 在技术上。我们运行了数百万次模拟，最终这转化为

0:11 赛道上的表现。

0:12 对与Grock的合作超级兴奋，如果你看一级方程式，

0:16 有一些问题我们真的在努力解决。

0:16 那就是速度，能够

0:19 在严格的成本控制下运营，

0:20 并且在一年参加24场比赛时能够快速扩展。

0:25 每年24场比赛。

0:26 试图同时解决这两个问题非常令人兴奋。

0:31 这正是Grock将

0:33 赋予我们优势的地方。”

Love

· 1d

@davidsenra称，Shopify首席执行官@tobi告诉他，将来我们回望2026年时，会将其视为“全球每一家企业都面临被颠覆的一年”。

“人工智能将席卷一切。”

“届时你会回过头来意识到，这一年本应让人看清：无论现存什么业务，都完全可以重建为原生 #AI 版本。” #Shopify

https://x.com/tbpn/status/2036241506428977519?s=20

中文 English

0:00 托比·卢卡对我来说非常有趣，因为他是你最喜欢的创始人，

0:03 最喜欢的创始人。

0:04 所以你欣赏的每个人，当你问他们欣赏谁时，他们都会提到托比

0:08 。

0:08 他们喜欢他的原因是他对每件事都有独立思考，

0:11 独立思考。

0:11 因此他拥有一系列彼此不相关的信念，

0:15 这意味着当你和他对话时，你无法预测

0:18 他对你的问题或陈述的回应会是什么。

0:21 我们谈了四个小时。

0:24 录音前我们先聊了半小时。

0:26 接着录了三个小时。

0:27 然后我们去了某个地方，录完后又聊了三十分钟。

0:32 我不停地说话。

0:32 我说，伙计，我得赶飞机了。

0:33 那晚我要和查理·罗斯吃晚饭。

0:36 我得飞过去。

0:37 我说我们本来可以继续聊。

0:38 他真让我大开眼界。

0:40 他说的一件事，我觉得你们观众会感兴趣，就是他

0:43 相信。

0:43 我记得他在节目里说过，我们将回顾

0:46 2026年作为这样一年，

0:48 那一年全球所有生意都处于抢夺之中。

0:51 人工智能正全面来袭。

0:54 每一个行业都必须被重新构想，你会回头看，

0:57 那一年你本应该明白，

1:00 你可以

1:01 用人工智能原生版重建所有现有业务。

1:04 很有趣。

1:05 Finn，如果你想让 AI 处理你的客户支持，访问 Finn.ai，

1:10 第一

1:10 客户服务 AI 助理。

0:00 Toby Luca was very fascinating to me because he's your favorite founder,

0:03 favorite founder.

0:04 So everybody that you admire, you ask who they admire and they'll bring up Toby

0:08 .

0:08 And what they like about him is that he thinks to every single thing

0:11 independently.

0:11 So therefore he has all these beliefs that are not correlated with one another,

0:15 which means when you have a conversation with them, you are unable to predict

0:18 what his response to your question or your statement is going to be.

0:21 And that we talked for four hours.

0:24 So we talked for a half hour before recorded.

0:26 Then we recorded for three hours.

0:27 We headed down to whatever was, then we recorded for another 30 minutes after.

0:32 And I kept talking.

0:32 I was like, dude, I got to catch a flight.

0:33 I'm going to have dinner with Charlie Rose that night.

0:36 Like I had to like fly down there.

0:37 I was like, we would have kept talking.

0:38 Like he just blew my mind.

0:40 And one thing he said, I think it'd be interesting to your audience is he

0:43 believes.

0:43 And I think he said this on the episode that we're going to look back on the

0:46 year 2026 as the year

0:48 that every single business in the world was up for grabs.

0:51 That AI is coming for everything.

0:54 Every single other business will have to be reimagined and you're going to

0:57 look back at, this is the year where it should have been obvious to you that

1:00 you can

1:01 rebuild the AI native version of whatever exists out there.

1:04 Interesting.

1:05 Finn, if you want AI to handle your customer support, go to Finn.ai, the number

1:10 one

1:10 AI agent for customer service.

Moon

· 2d

Ceteris 阐述为何加密货币领域的 AI 依然最具前景。

“虽然加密货币领域的 #AI 因代理商操纵市场而饱受诟病，但这仍是最具前景的领域。在这个单人创始人主导的时代，这些创始人并不需要风险投资的支持。他们无需筹集巨额资金。资本形成似乎一直是 #加密货币最具有杀伤力的应用场景。”

https://x.com/Delphi_Digital/status/2035811192128827879?s=20

中文 English

0:00 AI和加密过去一年里确实让大家都很失望

0:03 因为

0:04 代理人们炒高又抛售，但这仍然显然是

0:09 最有意义的领域。

0:09 真的第一次感觉你可以让一个SOAR创始人

0:12 创造一个

0:13 价值一千万到一亿美元ARR的产品。是的。你没见过这种情况。

0:18 我们没有，我们没有，

0:20 但这也正好。真的，完美适合

0:23 这些，

0:24 比如MediDao，SOAR。我觉得还有一种代币化股权的平台，

0:27 这些平台上的

0:29 创始人不适合风险投资支持。也许，有可能，但大多数时候，

0:33 他们不适合风险投资支持。

0:33 就是一个人，变化太快。所有这些东西。

0:36 他们不需要筹集大量资金。

0:37 这真的看起来是，

0:41 资金形成一直是加密的杀手级应用。

0:44 我的意思是，

0:44 我知道我们一直在说这样的话，但有时候事情

0:48 需要时间，

0:49 这真的感觉是加密的巨大机会。

0:00 AI and crypto is definitely, like everyone's soured on it over the past year or

0:03 so because

0:04 agents like pumped and dumped, but it's still obviously kind of like the area

0:09 that makes the most

0:09 sense. For the first time, really, it feels like you can have a solar founder

0:12 create a product that's

0:13 like 10 mil, 100 mil ARR. Yeah. And that's like, you haven't seen that though.

0:18 We haven't, we haven't,

0:20 but that's also perfect. Like that's really, literally perfect for some of

0:23 these like,

0:24 like MediDao, SOAR. I think there's like a tokenized equity one, like these

0:27 platforms where those

0:29 kind of founders are not venture backable. Maybe, maybe, but like for the most

0:33 part,

0:33 they're not venture backable. They're like, it's one guy that it changes too

0:36 quick. All this stuff.

0:37 They don't need to raise a ton of money. It really does seem like this is,

0:41 capital formation has always been crypto's killer use case. And it's like, I

0:44 mean,

0:44 I know we've been saying stuff like this forever, but it sometimes stuff just

0:48 takes time and it

0:49 really feels like this is such a big opportunity for crypto.

Love

· 2d

人工智能正变得更加全球化、更加多模态，并日益深入地融入各行各业。 #ai

中文 English

0:00 [音乐]

0:04 奥利维亚，欢迎。

0:05 >> 谢谢邀请我。

0:06 >> 这一年中最令人激动的时刻，

0:08 就是顶级100报告今天发布了，我想。

0:11 是这样吗？>> 是的。

0:12 >> 这是三年来的第六期。

0:15 跟我们说说什么没变，什么变了，

0:18 你有多兴奋，报告有什么新动向？

0:20 >> 是的，很多方面变化很大，

0:23 自从我们2023年首次发布这份名单以来，增长惊人。

0:25 另一方面，从宏观层面看，我们仍然处于非常早期阶段。

0:28 ChatGPT是迄今为止最大的全球AI产品，

0:32 但全球只有10%的人口每周活跃使用它。

0:36 所以未来还有很大发展空间。

0:41 我认为过去六个月可能是我最喜欢、最激动的时刻，

0:41 因为见证了许多变化。

0:43 其中之一是消费者竞争真正加剧了。

0:47 当然有ChatGPT，还有Gemini和Claude都在加倍下注，

0:51 针对它们各自的消费级和准专业用户。

0:54 我觉得我们开始看到这些平台可能随着时间累积优势。

0:59 这让谁获取最多用户成为一个特别重要且有趣的问题。

1:03 相关的是，这次我们首次纳入了非AI原生但已主导AI的产品，

1:06 例如Canva、Notion和Free Pick。

1:08 Notion甚至宣布他们现在认为一半的新ARR

1:12 来自于AI优先的功能，非常酷。

1:14 最后，我们看到AI在网站或应用提示框之外大幅扩展了，

1:17 例如新出现的各种浏览器，

1:20 像DIA、Comet、Atlas。

1:22 Claude也进入了Excel、PowerPoint和Chrome。

1:25 还有桌面应用如Cursor、Whisper Flow、Granola。

1:29 AI的使用方式呈现出令人兴奋的爆发式增长。

1:32 真是太令人兴奋了。

1:36 这里内容很多要说。

1:39 先从大型基础模型说起吧。

1:43 你能谈谈你认为Gemini和Claude各自的专长领域吗？

1:45 当然，还有ChatGPT，

1:48 因为这更像是一个整体上扬的故事，

1:53 而不是这些模型互相替代。

1:56 是的，我同意。

1:58 尽管上周发生了一些戏剧性事件，

1:59 例如Katy Perry在Twitter上站队品牌之间的争论，

2:00 这事我之前从没料到，

2:02 但从根本上看，ChatGPT仍是明显的赢家。

2:04 在网页端，它的用户是Gemini的2.7倍。

2:07 移动端则是Gemini的2.5倍。

2:08 尽管Twitter上的技术争论频繁，

2:11 ChatGPT网页端用户几乎是Claude的30倍，

2:13 移动端更是Claude的80倍。

2:14 我们看到Sam Altman当时的推文，

2:17 是在超级碗广告大战时期。

2:21 就是那个德州推文。

2:24 是的，他说，德州使用免费版本ChatGPT的人比Claude的全球用户还多，

2:27 这是真的。

2:29 不过，我认为我们看到的趋势是，

2:33 “分歧”也许不是准确词汇，

2:37 但用户使用的产品种类和用途都在扩展，

2:41 市场份额略有变化。

2:45 Claude特别加倍押注准专业用户，

2:47 推出了协同办公、Claude代码，

2:49 以及与Excel、PowerPoint集成。

2:51 如果你看看Claude和ChatGPT的应用商店，

2:52 它们各自都有200多个应用，

2:56 但交集只有11%。

2:58 Claude专注于高端数据源、研究工具、科学和金融数据，

3:00 而ChatGPT更专注消费市场，旅行、营养、消费金融等。

3:04 Gemini则在自己的小领域耕耘，

3:06 主要通过创意工具获得用户和收入增长。

3:10 活跃用户和付费用户几乎与YO3、Nano Banana One、Nano Banana Pro、Nano Banana Two的发布节奏完全同步。

3:13 它们也在准专业用户领域稍作尝试，

3:15 把AI整合到Gmail、Sheets、Calendar中，

3:19 但这些都是现有产品的升级，

3:21 而非全新体验。

3:23 - 我们不如深入聊聊应用商店动态，

3:25 这真是太有趣了。

3:28 你能说说ChatGPT的应用目录的利好吗？

3:30 - 当然可以。

3:32 我看到ChatGPT采取的策略，Sam本人也在Twitter说过，

3:34 就是他们要做“人人皆可用的AI”，

3:37 意味着他们想吸引所有消费者，

3:39 并通过不同方式赚钱。

3:41 Claude则明确了只通过订阅模式盈利，

3:44 这对能付费的用户和企业很好，

3:46 但不是所有人都适合。

3:50 你也能从它们倾向于有偿的高价值订阅插件看出来，

3:52 这些插件多是专业的数据和服务工具。

3:53 - 类似实验室工具。

3:55 - 是的，比如PitchBook，投资人、科学家、数学家会用的工具。

3:58 ChatGPT则更像Google的方式，

4:02 构建普通人愿意用的产品。

4:06 可能转化成订阅用户比例较低，

4:11 但他们能通过广告赚钱，

4:13 也可能通过交易抽成。

4:15 比如如果他们成为预订旅行或各种长尾消费的入口，

4:17 按理说能从中分成，至少是其引流部分。

4:20 所以这就是ChatGPT应用商店的牛市逻辑，

4:22 数据里还没显现，

4:25 但未来一两年会更明显。

4:26 - 真有趣，你刚才谈到报告里“复合优势”的观点，

4:28 也就是上下文效应的累积，

4:30 能详细说说这个概念吗？

4:32 你用什么指标作衡量，比如会话时长、会话次数，

4:34 还是输入数据量，或者其他？

4:37 - 这是个很激动人心的问题，

4:39 因为到目前为止，像ChatGPT、Claude、Gemini、Perplexity这些横向大模型，

4:40 我们一直活在一个上下文和记忆能被轻易导出的世界。

4:43 Claude最近还针对这个做了广告活动。

4:45 但我认为锁定效应会越来越强，

4:48 而且其实这很可能有利于那些更横向、更广泛的工具，比如ChatGPT，

4:50 原因有几方面。

4:54 第一，ChatGPT已经开始开发能让用户通过平台与他人互动的产品，

4:56 比如群聊，想象如果有个更成功的ChatGPT群聊版本，

4:57 你的所有朋友都在上面，

4:58 那当你想离开ChatGPT时，

5:00 你得说服他们都换到另一个产品去。

5:04 - 正是如此。

5:08 第二点，我觉得也类似苹果和谷歌的比较，

5:10 当这些应用商店成熟后，开发者很可能会将时间和精力集中在那些用户最多、

5:12 或者在某些情况下最愿意付费的平台上，

5:15 对于很多消费工具来说，决定因素往往是用户数量。

5:18 这也进一步利好ChatGPT。

5:20 今年我最期待的一点是，Sam Altman暗示的ChatGPT身份验证层，

5:23 也就是说你可以用ChatGPT账号登录，

5:26 携带你的记忆和令牌，

5:30 其他产品就能借此变得更强大、更贴心。

5:35 如果成真，你会希望自己的核心身份存留ChatGPT，

5:37 这样它可以帮助提升你使用其他工具的体验。

5:40 - 这是超级聪明的策略，

5:41 依托它们拥有9亿用户注册的优势，

5:44 第三方开发者理想状况下不用自己支付推理费用，

5:47 用户把推理能力带着走，

5:50 开发者受益，

5:52 ChatGPT获得锁定，

5:53 用户获得个性化，

5:56 一切完美配合。

5:58 - 是的，我完全同意。

6:00 唯一让我存疑的是，

6:02 这点可能对锁定效应正反两面都有影响，

6:04 就是你的工作场景，

6:06 你签的企业合同。

6:07 比如，从某种角度讲，

6:10 如果我公司工作用的是ChatGPT，

6:13 那我学会用它就有优势。

6:17 we've kind of lived in a world where the context

6:20 and the memory is somewhat easily exportable.

6:22 Like Claude ran a campaign around this recently.

6:25 But I think there's gonna be increasing lock-in

6:28 and I do think that probably actually benefits

6:30 the broader, more horizontal tools like Chatsheet BT

6:33 for a few reasons.

6:35 So I think one, we've already seen Chatsheet BT focus on

6:39 or start to build out products where you interact

6:42 with other people on them through the platform.

6:44 So the group chats, like imagine if you,

6:46 if there's an even more successful version

6:48 of Chatsheet BT group chats

6:50 and all of your friends are on there,

6:51 then if you wanted to turn from Chatsheet BT,

6:53 you'd also have to convince them all to go through another product.

6:56 - Exactly.

6:57 I would say the second one is kind of also

7:00 like an Apple Google comparison in that

7:02 as these app stores emerge, it is likely that developers

7:05 might start to concentrate their time and effort

7:08 in who they build for in the most sophisticated way,

7:12 who they ship to first, depending on who has the most users

7:16 or maybe in some cases who's the most willing to pay,

7:19 but for a lot of these consumer tools,

7:20 it'll be who has the most users.

7:22 So I think that also benefits Chatsheet BT.

7:25 And then the other thing probably that I'm most excited

7:27 for this year that Sam Altman had kind of hinted at

7:31 is this like authentication with Chatsheet BT layer.

7:34 So essentially you'd be able to log in

7:36 with your Chatsheet BT account

7:37 and take like your memory and your tokens with you.

7:42 And then that other product would be able to kind of borrow

7:45 those things to be even more powerful and helpful for you.

7:48 And if that's the case, then you're wanting

7:50 to have more of your core identity live on Chatsheet BT

7:54 because then it can lend it to these other tools

7:57 that are even better for you.

7:58 - It's so smart and it really plays to their advantages

8:01 in that they have signups for 900 million people.

8:05 And then the third party developer ideally

8:07 would not want to pay for the inference.

8:09 So the user can bring their inference capacity with them.

8:12 There's an advantage for the developer.

8:13 Chatsheet BT gets the lock in.

8:15 The user gets the benefit of personalization

8:17 and it all kind of works.

8:18 - Yes, I totally agree.

8:19 The one question mark I still have on this

8:21 that I think could play both positive and negative

8:24 in terms of increasing lock in for the consumer product

8:27 is what your work goes with,

8:30 like what your enterprise contract is.

8:33 So for example, in some ways, it's good for me

8:36 if we, if my company uses Chatsheet BT for work

8:40 because then I know how to use the product.

8:42 作为普通消费者，他们可能尝试过

8:44 一两款人工智能产品。

8:45 所以他们更可能觉得习惯

8:47 并继续使用他们已经用过的东西。

8:49 另一方面，有些人可能不想将身份

8:53 和记忆混合在个人和工作使用场景中。

8:56 - 说得对。

8:57 - 所以我很感兴趣。

8:58 我觉得OpenAI最近有暗示，

9:00 但我很想知道如何划分记忆

9:04 在你自己内部的不同角色之间，

9:08 使用这些产品。

9:09 - 不要混淆角色。

9:11 - 对，就是这样。

9:12 - 好吧，换个话题，

9:14 先聊聊Gemini。

9:15 你知道，我在想Google早期AI产品的整体感觉，

9:18 比如Bard，他们永远甩不掉的包袱。

9:21 - 那些时刻很出名。

9:24 - 现在我们看到像Nana Banana这样的产品，

9:25 甚至这个名字“Nana Banana”

9:27 完美体现了Google的进步。

9:30 - 是的。

9:32 - 看起来他们对多模态非常重视。

9:33 - 嗯。

9:34 - 你怎么看他们的策略？

9:35 - 我挺佩服的。

9:36 我觉得他们在某些方面确实犹豫了，

9:38 这种犹豫恰恰符合预期。

9:40 将AI融入核心功能有风险，

9:44 可能会损害自家产品，或者很多人

9:47 已经用了这些工具十多年，

9:51 转换成本比较高。

9:53 他们不想在AI突然出现时吓到用户，

9:56 这我能理解。

10:00 但他们做得非常好的，是这些新创意产品，

10:04 基本上是DeepMind团队驱动的模型，

10:06 我觉得他们整体很棒。

10:09 - 我觉得Notebook LM实际上是首次展现，

10:11 是消费者AI音频领域的真正创新。

10:15 现在又有了图像和视频模型。

10:18 所以这样的大公司，

10:20 他们必须克服自身桎梏，

10:23 实际开始创新。

10:27 看起来他们正在这样做，但你曾在Google工作，

10:31 所以我很好奇你的看法。

10:32 - 很开心你提到Notebook，

10:35 因为Notebook在公司算是个无人争夺的空白地带，

10:38 这避免了十位副总裁争抢资源。

10:40 因此，Notebook的进展非常快。

10:41 他们刚推出了视频生成功能，

10:44 可以很直观地展示你的工作区内容，挺酷的。

10:47 相比之下，像Sheets和Docs这些现有产品，

10:48 有太多过去积累的惯性和势头，

10:50 还有管理层的各种压力，

10:53 他们很难做超出最显而易见的增量改进。

10:54 - 是的，我同意。

10:56 接下来几年会怎样，我们拭目以待。

10:59 我感觉他们会竭力保住这些产品，

11:01 因为不想失去用户群，但正如你说，

11:04 他们已经锁定了众多企业用户，

11:08 短期内可能不必做太多变化，

11:10 也能跟上步伐。

11:13 - 这次谈话隐含的一个点是，

11:14 我们西方体验并讨论很多AI，

11:16 说说全球AI发展趋势吧。

11:18 我看到里面有几个令人惊讶的点。

11:21 - 我们扩大了报告的研究范围，

11:23 结果非常有趣和好玩。

11:25 两个显而易见的与世界其他地方不同的是，

11:26 俄罗斯和中国。

11:28 大家都知道，中国许多AI产品被审查或禁用。

11:31 因此几乎所有的使用，

11:33 他们在ChatGPT、Bard和Gemini的综合使用量上最低，

11:35 只有15%。

11:39 他们主要用的是字节跳动出品的Dall·E，DeepSeek，

11:41 Quinn，Kimi这些模型。

11:43 让我有些惊讶的是，俄罗斯的情况其实很相似，

11:46 他们也有自己的平行AI生态系统，

11:47 这是出于必须，因为制裁等因素，

11:50 阻止他们使用所有美国工具。

11:54 我们看到像GigaChat和Yandex，

11:57 这些俄罗斯本土产品，往往由国家相关公司打造，

11:59 使用量很大，还有DeepSeek。

12:02 DeepSeek在俄罗斯市场仅次于中国，

12:05 排名第二。

12:07 如果你看各国采纳数据，

12:10 是的，某些国家对Claude用得更多，

12:12 有的国家对Gemini用得更多，

12:14 但两个最大异常是俄罗斯和中国，

12:15 他们是巨大市场，

12:18 值得持续关注。

12:21 - 很有趣的是，俄罗斯和中国

12:22 都因使用限制和文化偏好成为例外，

12:25 还有其他国家有地理特定趋势吗？

12:29 还是说这是一种全球AI行为模式？

12:33 - 说到模型开发，

12:34 专有模型开发可让你部署专属AI产品，

12:38 大部分研究来自美国和中国，

12:40 俄罗斯可能也有些贡献。

12:43 其他地方也在见证本土生态出现。

12:46 韩国有几款本土产品，

12:48 比如Neighbor和Cookal，

12:51 做出了不错的语言模型界面。

12:52 印度是我特别关注的另一市场，

12:56 因为那里人口众多，

12:59 可以支持独立大型企业专注印度市场。

13:01 印度的另一特色是语言多样，

13:03 各种语言不全被语言模型和语音产品支持，

13:06 导致主语言用户体验相对较差，

13:08 比如使用类似ChatGPT的产品时。

13:10 目前尚未见大量变种，

13:12 但我不惊讶未来可能有更多创业者，

13:14 甚至美国创业者，针对印度市场做AI。

13:15 另外一点，我们首次制作了个热力图，

13:18 显示哪些国家人均采纳AI最多和最少。

13:21 我们看了十大大型语言模型产品，

13:24 包括网页版和移动端，看看数据如何。

13:25 新加坡排名第一。

13:28 - 太疯狂了。 - 是的。

13:30 紧随其后的是香港、阿联酋、韩国。

13:34 美国排名第20，

13:37 不算特别低，也不算特别高。

13:39 俄罗斯和中国排得很后，排名50之后。

13:43 数据中蕴含许多有趣故事。

13:45 先说这前五名，

13:48 新加坡、韩国、香港，

13:50 劳动力结构高度技术化，白领比例高，技能丰富。

13:52 美国有大量岗位AI还未涉足，

13:55 比如零售、运输等领域。

13:57 另外，关于AI的文化认知多样性极大。

14:01 如果你在美国，

14:02 可能对AI的持续焦虑和质疑感受深刻——

14:04 - 是的，我就打算问你这个问题。

14:07 - 及AI对艺术家的负面影响等，

14:11 这些因素使得人们是否接受AI犹豫不决。

14:14 - 是的。

14:15 去年Edelman全球媒体公司做了项大调查，

14:18 美国的AI信任度较低，约32%，

14:22 而排名靠前的多数国家信任度达50%、60%、70%。

14:25 这也拖累了美国的发展速度，

14:28 尽管最大产品都来自美国，

14:31 人均使用率却不及其他投入更多的市场。

14:35 - 完全正确。

14:39 我读到中国对AI的好感度高达80%。

14:41 - 是的。 - 80%的人持正面看法。

14:44 我知道阿联酋和新加坡，

14:48 他们文化上更倾向技术乐观主义。

14:49 - 是的，确实如此。

14:50 看到一些小国的人均采纳率很有趣。

14:54 在美国，大约有三分之一的人每月活跃使用ChatGPT。

14:56 欧洲一些国家，甚至东欧，

14:59 虽样本较小，人均使用率可达45%到60%。

15:04 他们接受得较快。

15:06 - 是的，非常有意思。

15:09 我关注的是AI应用的全谱系，

15:11 从最实用的，

15:14 几乎是谷歌搜索替代品，

15:17 are very like tech first white collar high skill.

15:20 And the US has a giant chunk of jobs

15:23 where AI hasn't really touched them yet,

15:25 like retail and transportation and some of these other things.

15:29 I think also the cultural norms around AI

15:33 are shockingly diverse.

15:36 If you're in the US, you have probably internalized this

15:39 ongoing angst and questioning around--

15:41 - Yeah, I was going to ask you about this 100%.

15:44 - Or AI is terrible for artists

15:47 or all of these other things that make people pick up

15:50 or not pick up AI. - Yes.

15:51 - There was actually a big survey last year

15:54 from Edelman, the global media company.

15:57 And the US had a fairly low rate of trust in AI.

15:59 It was like 32% and most of these other countries

16:02 that are high on the list are like 50, 60, 70%.

16:06 So that I think has also held the US back

16:08 despite the fact that we are where the biggest products come from

16:12 are per capita usage is lower than a lot of these other markets

16:14 that have maybe smaller populations

16:16 but have embraced it more.

16:19 - I think that's exactly right.

16:20 Now I was reading that in China,

16:22 the sort of favorability views on AI are 80%.

16:25 - Yeah. - 80% hold a favorable view.

16:27 And I know UAE and Singapore,

16:29 I think they've sort of culturally wired to be tech optimistic.

16:34 - Yes. - Which is an advantage.

16:35 - Yes, yes, definitely.

16:37 It's interesting to see some of these smaller countries

16:39 like the per capita adoption rate.

16:42 Like in the US, it's around probably a third of people

16:45 are monthly active users of something like a Chachi BT.

16:49 In some, even some of like the European countries

16:52 or Eastern Europe, it's like 50, 45, 60% on smaller bases.

16:58 But they've kind of embraced it more quickly than we have here.

17:02 - Yeah, really interesting.

17:03 You know, one thing that I'm sort of watching

17:05 and I'm interested in is as you look at the spectrum of AI

17:07 from the most functional,

17:09 almost like a Google search replacement

17:11 对于最具文化性、创意性和个性化的，

17:14 我们应该看到各国之间更多的差异化。

17:17 因为显然文化方面，

17:18 印度拍摄的电影

17:19 与中国或美国拍摄的电影截然不同。

17:21 - 是的。

17:23 - 那他们使用创意工具的方式为何不不同呢？

17:25 - 是的，老实说这也是

17:27 我们开始在这份报告中考虑地理细分的部分原因，

17:30 因为在生成式AI的前两年半，

17:33 三年中，

17:35 绝大多数消费者可能只在使用

17:38 一个产品，现在这种情况已有很大扩展。

17:41 我认为我们将看到更多

17:43 针对特定市场的工具。

17:45 如果它们占领了足够的市场份额，

17:48 比如一些俄罗斯或中国公司，

17:51 它们实际上可以进入全球榜单，

17:53 前提是市场规模足够大。

17:55 说说创意工具的发展，

17:58 你认为这有多少反映了文化、

18:00 是文化驱动的，

18:02 我们何时会跨越那个门槛？

18:04 - 创意工具的趋势令人着迷。

18:07 显然，首个大型生成式AI产品

18:10 实际上是mid-journey，早于Chachi PC发布。

18:13 - 确实，是这样的。

18:14 - 在我们榜单的最初几期，

18:16 创意工具占据主导地位。

18:19 我以前说过，

18:21 创意工具受益于

18:24 早期模型的幻觉效应，

18:27 因为它们生成的内容往往更令人惊喜、美丽或原创。

18:30 因此，一度只有这些创意工具

18:32 在消费者AI中真正奏效。

18:35 现在情况变化很大。

18:36 创意工具仍是榜单的重要部分，

18:39 但作为独立业务的大型创意工具类型已经改变。

18:41 最大的变化是

18:45 我们看到独立的图像生成器越来越少。

18:47 许多这类工作，

18:51 如果制作基本的图像，

18:52 比如表情包、基础营销图像或信息图，

18:55 像Chachi BT和Gemini的核心模型

18:59 现在对此类任务都很擅长。

19:02 因此榜单上仍出现的产品，

19:05 如I-U gram或mid-journey，

19:07 要么是美学主张很强烈，

19:10 要么具有更复杂的工作流程，

19:13 这是Chachi BT类工具无法提供的。

19:16 相比之下，音乐、语音、视频

19:21 似乎是模型最大的公司投入较少的领域。

19:26 所以我们看到像Suno做音乐，

19:28 和11 labs做语音的玩家，

19:31 完全脱颖而出，

19:33 进入榜单前二十、十五名，

19:37 并且长时间保持位置。

19:40 同时还有社区和大量企业客户的复合锁定效应。

19:43 视频领域我有最多疑问。

19:45 OpenAI通过Sora在投资，

19:46 当然谷歌也在用Vio，

19:50 但中国模型非常出色，

19:54 因为它们可以训练任何数据。

19:57 Sea Dance 2可能是最佳例证，

19:59 在某些方面远超美国公司目前的水平。

20:01 我认为这将有利于像Korea这样的平台，

20:03 可以在一个地方使用所有模型，

20:08 因为我姐姐Justine写过相关文章。

20:10 视频的发展趋势显示，

20:14 不太可能出现一个统治所有的单一模型。

20:18 因此你需要能够在模型间切换。

20:21 - 这对大多数模型领域都适用，

20:25 聊天模型、创意模型，甚至代码模型，

20:26 它们都有自己的专业方向。

20:29 你知道，人们会讨论opus的易用性

20:32 和codecs的准确率。

20:35 这是权衡取舍，

20:38 你得选择针对不同问题使用不同工具。

20:40 - 绝对如此。

20:43 - Sora对我来说非常有趣，

20:45 因为它既代表了模型的一大进步，

20:47 也是一个关于社交的雄心勃勃的实验。

20:50 在Sora早期的数据里，

20:51 创作人数百分比显著，

20:54 比之前高出10倍。

20:57 你如何评价

20:59 Sora的社交努力与模型努力？

21:00 你怎么看它的未来？

21:03 - Sora非常吸引人。

21:05 这是一个早期非常有趣的实验，

21:08 它教会了我们许多东西，

21:09 不仅是关于创意工具，

21:12 更重要的是关于AI时代的消费者社交可能是什么样。

21:14 从数据来看，它发布时反响巨大，

21:15 曾连续20天登顶美国App Store，

21:19 这很难做到。

21:22 这意味着你每天大约需要

21:24 15万次下载才能排名第一。

21:27 这下载量很高。

21:29 它实际上比Chachi BT更快达到一百万用户。

21:31 这是个巨大发布。

21:33 事实上，我认为很多人低估了

21:34 它依然有大量活跃用户。

21:36 Sensor Tower数据显示日活有三百万，

21:37 这相当不错。

21:38 Sora的新下载量有所下降，

21:40 可能是因为，

21:42 它11月下载峰值达六百万次每月，

21:45 现在大约一百五十万。

21:46 我认为Sora的成功关键是

21:48 它是个非常好的视频模型。

21:52 它创新并引入了“戏份”（cameos）概念，

21:54 即真实人物

21:57 可以授权自己的形象给Sora，

21:59 让自己和别人生成自己的视频。

22:02 很多人早期制作朋友的搞笑视频。

22:03 杰克·保罗因率先投入Sora

22:05 一炮而红。

22:07 你能看到大量疯狂的杰克·保罗视频，

22:11 我们写过报道。

22:14 是的。

22:15 老实说，他做得很好。

22:17 是的，是的。

22:19 我认为Sora的问题是

22:21 内容可以导出，

22:24 用户会分享到TikTok，

22:25 或Instagram Reels，

22:27 或YouTube，

22:29 在那里它要和人类最佳内容竞争。

22:30 所以整体内容流体验更优，

22:31 因为你看到的是人类和Sora的双重最佳，

22:34 而非仅仅是Sora的最佳。

22:35 我认为我们还没见过纯AI内容的社交产品成功。

22:36 情感投入在某些方面降低了。

22:37 我想我们会看到类似的例子，

22:38 Sora仍有显著使用量，

22:42 作为创意工具带来收入，

22:45 但作为社交应用没那么成功。

22:47 对。

22:49 我不知道是否会有大型AI原生社交网络，

22:50 但目前还看不到它的模样。

22:54 会很有趣。

22:57 我们经常讨论，

22:59 每个社交产品都有自己的身份玩法。

23:00 是的。

23:05 Instagram上的是最火的，

23:07 在X上是最有趣的，

23:11 而在Sora上，涌现的身份玩法是成为最有趣的人。

23:14 是的。

23:18 我觉得这也是内容难以跨界的原因之一，

23:20 因为两者评判“有趣”和“优秀”的标准完全不同。

23:21 同意。

23:23 如果让我想象他们可能找到的细分领域，

23:24 他们现在与迪士尼等大媒体公司达成了多项协议。

23:28 如果Sora是制作

23:30 授权的粉丝视频的唯一平台，

23:31 呈现受欢迎的角色和娱乐人物，

23:32 那将非常有趣。

23:35 完全正确。

23:35 但这还很早，

23:38 我们还在观察它的发展。

23:39 还很早。

23:41 我知道。

23:42 我们可以这么说。

23:43 是的。

23:44 我们谈话时必须提到代理，

23:46 OpenClaw, Manis, Genspar, Moldbook。

23:49 给我们概述一下

23:50 过去60天代理领域发生了什么

23:51 报告告诉了我们什么？

23:53 我说过去，

23:55 甚至过去六个月，

23:57 更准确说是报告中最近两个月，

23:59 是我见过最有趣的时期。

24:02 OpenClaw实际上，正如你将看到的，

24:06 没有上榜，因为它在二月爆发。

24:08 而我们的数据截止到一月。

24:09 但我们提取了二月的数据，

24:10 如果符合条件，

24:12 它会排名我们网络榜单第30名，

24:13 这是一个相当大的首次亮相。

24:14 We can say that.

24:15 Yeah.

24:16 We can't have this conversation without talking about agents,

24:18 open claw, manis, genspar, moldbook.

24:22 Give us an overview of what has happened

24:24 in the last 60 days in the world of agents

24:26 and what does a report tell us?

24:27 I think this is mostly why I say the last,

24:29 you know, even six months,

24:31 but actually even two months of this report

24:33 have been like the most interesting

24:35 that I think we've seen.

24:36 So open claw actually, as you'll see,

24:39 is not on our rankings because it blew up in February.

24:42 Our data ends in January.

24:43 But we did pull the data for February.

24:46 And if it had been eligible,

24:48 it would have been number 30 on our web list,

24:50 which is a pretty big debut.

24:53 我认为关于Open Claw真正有趣的地方

24:55 是它的使用量在技术社区中持续加速增长

24:58 。

24:59 所以现在，我觉得它是第一名，

25:01 GitHub历史上的明星项目。

25:02 它超过了React，也超过了Linux。

25:04 哇，真是非常非常有趣。

25:05 是的。

25:06 非常令人印象深刻。

25:08 但就整体新用户而言，

25:11 增长有点停滞。

25:12 我们查看了一下“开始使用”

25:15 或注册页面的访问情况。

25:16 从二月初开始，这个数据几乎周周持平，

25:20 我认为这表明，

25:21 它是一个了不起的技术产品，

25:24 但还没有完全扩散到非技术人群，

25:27 也就是更广泛的人群。

25:28 他们被OpenAI收购了。

25:32 所以如果要我猜测，

25:33 或者说我希望看到OpenAI做的是

25:35 将Open Claw产品化，

25:37 变成主流消费者可用的东西。

25:39 我也觉得Open Claw架构背后的理念

25:43 激发了许多其他创业者。

25:46 我们每天会听到多少创业者pitch，

25:49 他们都说，

25:51 “我想成为这个领域的Open Claw。”

25:53 完全正确。Open Claw让我意识到这是可能的。

25:54 是的。

25:55 所以我认为Open Claw本身

25:58 会继续成功，成为巨大产品。

25:59 我猜我们会看到更多

26:01 针对不同用例的Open Claw垂直版本。

26:03 是的，这太有趣了，

26:05 因为Open Claw之所以这么有效，

26:07 是因为它可以跨所有模型，

26:09 向各个方向运作。

26:10 我有点好奇，如果它只提供单一模型，

26:11 是否会削弱Open Claw的价值，

26:13 这样它就成了Labs的对立面。

26:16 完全正确。

26:19 他们目前依然保持多模型，

26:21 至少我使用时是这样。

26:24 后续趋势如何我们拭目以待。

26:25 我认为保持多模型使用是明智的。

26:27 是的。

26:28 Manus是面向消费者级的Open Claw，

26:29 还是你如何区分这个团队？

26:31 是的，有人会这么说。

26:33 我实际上认为，

26:35 Manus进入了我们的网络榜单，

26:36 并且在榜单期间被Meta以超过20亿美元收购。

26:38 增长非常惊人。

26:39 他们从零达到1亿、2亿ARR，

26:42 用了大约六到九个月左右，

26:46 这在业内属于顶尖水平。

26:48 我认为Manus成功的原因是，

26:49 它是首个能够相当自主操作，

26:52 跨产品和平台的消费级智能代理。

26:55 你可以连接邮箱，它能浏览网页。

26:59 它能做幻灯片，也能做表格。

27:02 我早期花了很多时间试用，

27:04 比如一年前的ChatGPT操作助手

27:07 或谷歌的Project Mariner。

27:10 它们没有一个可靠。

27:13 Manus在代理的可靠性和易用性上是突破性的。

27:17 我觉得它被收购非常有意义，

27:19 显示了未来的发展方向。

27:23 当人人具备这代理能力时，

27:25 你可能会想，如果它基于核心底层模型，

27:27 作为一个通用产品，

27:30 它其实更适合由Meta或谷歌这类大公司分发，

27:33 而非独立创业公司。

27:35 如果你做的是垂直领域产品，情况就不同了，

27:39 但谷歌如今有资源做出Manus这样的产品，

27:43 创业公司很难竞争。

27:45 大公司确实有很多优先事项，

27:48 不会做到每个领域都一流，

27:52 但这也是我对超广泛的消费者AI应用更为谨慎的原因，

27:55 因为大公司覆盖面广，

27:59 他们本身就有IT批准和企业合同的优势。

28:00 - 对，对，有趣的是我们跨过了

28:02 这样一个文化门槛，

28:06 Manus在产品广度上曾被视为不太明显的赌注。

28:10 现在看起来，他们像是生活在未来。

28:12 - 是的，完全正确。

28:13 他们显然是一个了不起的工程团队，

28:16 产品质量领先市场三到六个月，

28:18 这很难做到，尤其面对成千上万的自由研发者竞争。

28:19 - 完全同意。

28:22 让我们通过这个话题引入

28:26 关于其他水平AI产品的讨论，

28:28 以及那些超越网页界面的应用。

28:30 - 是的。

28:32 - 你意思是？

28:34 - 是的，这是个巨大主题。

28:37 我每天接触的AI产品中，

28:39 很多其实是桌面应用，

28:43 比如Granola、语音转写工具、Cloud Co-work之类。

28:45 这对我们报告来说是个方法论问题，

28:47 因为我们能很好地追踪网站访问，

28:48 也能很好追踪首次下载桌面应用，

28:49 和移动端使用情况，

28:52 但不能太精准地追踪桌面端使用量。

28:55 我认为随着AI产品愈发复杂，

28:57 在专门应用里运行，

28:59 许多会基于桌面，

29:00 因为能与文件交互，也更随时可用。

29:01 这趋势会越来越显著。

29:03 所以后续我们得想办法并行追踪，

29:06 依靠网页、移动端使用量，

29:09 以及收入排名，

29:10 这将是明智的做法。

29:10 因为像Cursor这类，

29:12 一些收入最高的消费级准专业AI应用，

29:15 网页端使用很少，

29:17 大部分都在专用应用内。

29:19 - 是的，非常有趣。

29:22 这也说明OpenAI发布Atlas，

29:25 Anthropic发布Co-work，

29:29 反映了他们的优先级所在。

29:31 - 是的，完全同意。

29:33 AI浏览器本身就是一个有趣的话题。

29:34 我觉得我们仍处于早中期阶段，

29:37 还不知道它会如何发展。

29:40 AI原生浏览器的理念正确，

29:43 如果AI能一直在线，随时可用，

29:45 在你花大量时间浏览的地方，

29:48 这是个好机会。

29:50 Perplexity Comment实际上率先做到了这点。

29:51 - 是个很棒的产品。

29:53 - 是的。

29:56 有趣的是，

29:59 Comment和Atlas下载页面访问量高峰比较，

30:02 Comment是Atlas的五倍，

30:05 这很惊人，因为ChatGPT用户量庞大。

30:07 - 是的，我也觉得如此。

30:09 - 我们看到Comment和Atlas都有忠实用户，

30:12 但对于普通用户来说，

30:13 更换浏览器的成本不低，

30:16 因为你已有固定工作流程，

30:19 自然会打开那个浏览器。

30:20 所以不仅要功能相当，

30:22 AI浏览器必须有一两项杀手级功能，

30:26 且容易被普通人设置和使用。

30:27 我觉得这还没完全实现。

30:28 - 真的很有趣，

30:30 六个月前Sam在一个播客中说，

30:35 有人问他，

30:38 最让你惊讶的是什么？

30:40 他说，是世界变化没想象中大。

30:43 如果看人们大规模使用ChatGPT的趋势，

30:47 仍主要用于作业和类似Google的查询，

30:51 以及一些陪伴功能，

30:54 从某种意义上说，像浏览器那样，

30:56 可以引导用户走不同路径。

30:59 你怎么看普通人今天如何使用AI？

31:00 - 我有几点想法。

31:00 首先，我觉得青少女

31:03 是观察消费端动态的最佳来源。

31:06 - 你没教我这一点，是的。

31:09 - 他们是最大消费成果的早期拥护者，

31:11 不久前Pew研究有个报告，

31:15 专门研究青少年如何使用AI。

31:16 - And I think what we've seen is like

31:18 comment and Atlas still have very dedicated,

31:22 excited user bases,

31:24 but for the average consumer,

31:26 the switching cost of a browser is non-trivial,

31:29 just because like you have workflows set up,

31:32 you naturally just open this one app.

31:35 And so it not only has to be like feature parity,

31:37 there has to be one or two features of the AI browser

31:40 that are really killer and that are easy enough

31:43 for the average person to set up and access.

31:45 And I don't think that we've seen that quite yet.

31:48 - You know, it's really interesting

31:49 because Sam said, I think six months ago in a pod,

31:51 you know, somebody was asking him,

31:53 what has surprised you the most?

31:54 And he said, it's that the world hasn't changed more.

31:57 And if you look at the trends around

31:58 how people are using Chachi BT at scale,

32:00 it's still, you know, homework and Google like queries

32:03 and a little bit of companionship,

32:05 in a sense, something like a browser

32:07 that gives you an opportunity

32:08 to point the user in a different direction.

32:10 What's your view on how the average person

32:12 is using AI today?

32:13 - Yeah, I think a couple of things.

32:17 So one, I feel like teenage girls

32:19 are like the best source of what is happening in consumer--

32:22 - You haven't taught me this, yes.

32:23 - What will be happening in consumer?

32:24 If you look at all of the biggest consumer outcomes,

32:26 like they were the early adopters of all of these products.

32:30 And so there was actually a Pew Research Study

32:32 fairly recently on how teenagers are using AI.

32:35 现在终于，我觉得这是第一次，

32:38 超过一半的人承认

32:40 使用它来做作业。

32:41 所以真实数字可能是99.99%，

32:44 但其中一些人不想惹父母生气。

32:46 - 嗯哼。

32:48 - 38%现在将它用作创作工具。

32:51 比如编辑图片、编辑视频，

32:53 生成图片和视频。

32:55 然后这有一个稍微增长的长尾趋势，

32:58 但我认为最终会成为最大行为之一。

33:02 16%用它进行轻松聊天，

33:05 不是强烈的陪伴产品，

33:08 而是找个人聊聊天。

33:10 然后12%用它做情感支持

33:13 和提供建议。

33:14 我认为所有这些用例最终都会趋近于

33:17 大约100%。

33:21 这些行为可能是目前

33:23 产品服务较少，但未来会被满足的，

33:26 无论是在聊天机器人上，还是

33:27 独立产品上。

33:30 另一个我关注的大事是

33:34 代理。

33:35 我觉得基本上-- - 我们的少女会用代理。

33:39 来吧。 - 情况是这样的。

33:40 我认为代理类似于1990年，

33:44 互联网公司被称为点com公司，

33:47 或者点com成了科技公司的标识。

33:49 - 对。

33:51 - 我觉得代理也会是这样。

33:53 最终每个科技公司

33:55 都会是点com公司。

33:56 我觉得每个AI公司

33:58 以及每个科技公司都会成为代理公司，

34:02 因为模型的发展趋势就是如此。

34:05 如果你能给用户

34:07 提供结果，

34:10 而不仅仅是输入，这作为软件产品更有吸引力。

34:12 所以是的，我认为13岁女孩会用代理，

34:15 但她们不会把它们当作代理。

34:18 但我觉得这会解锁AI在其他消费领域的许多用例，

34:21 比如财务、

34:24 医疗、旅行规划、复杂的趋势购物，

34:29 在有代理之前，数据太多，

34:32 你必须自己去搜集、可靠执行，

34:35 跨系统操作几乎不可能。

34:38 现在这变得可能了。

34:39 所以我认为接下来几个月

34:40 这些其他用例会爆发式增长。

34:43 - 你觉得这要多久才能完成？

34:45 我是说，12个月内大家都会用自己的OpenAI？

34:47 还是五年后？

34:48 这是什么错误的思维模型？

34:49 当我们六个月后在下一个Top 100大会讨论时，

34:50 世界会是什么样子？

34:52 - 我感觉每次我预测的事情

34:55 都比预期发生得快，

34:57 我认为我们每天都能看到这一点，

34:59 创业公司成长速度前所未有。

35:03 但文化变革

35:04 和文化接受速度会比技术变革慢。

35:08 因此我们会持续看到

35:11 这波早期浪潮，通常是技术人员，

35:13 有时非技术人员做先锋，

35:17 六个月后其他人跟进。

35:19 一个我非常兴奋的例子是

35:23 语音，我们谈了很多。

35:26 - 确实谈了很多。

35:28 - 在我看来，这是信息量最大、

35:30 质量最高的媒体形式之一。

35:34 你每日所做的很多事情，

35:36 实际上是基于你说的话上下游运作。

35:38 我觉得过去六个月，第一次见到工程师，

35:41 现在其他技术人员开始用语音输入。

35:46 现在很多公司会议

35:48 几乎成了录音并用AI转录的常态。

35:51 不管是语音输入，

35:55 还是会回答问题或帮你办事的语音助手，

35:58 我觉得接下来六到九个月

36:01 会普及到普通消费者。

36:03 - 真是非常有趣。

36:05 讲到最后，你能聊聊记忆吗？

36:07 - 好的。

36:10 - 你怎么看记忆的发展？

36:12 - 我们之前提过，现在

36:14 记忆功能有点突兀，

36:15 特别是Claude和ChatGPT这方面做得很好。

36:19 就连谷歌的Gemini也推出了“个人智能”，

36:20 它能从文档、邮箱等获取你的信息，

36:24 用AI服务你所有应用。

36:24 正如我说的，目前可能稍微让人不适应，

36:25 因为很多人跟AI谈论生活工作各方面，

36:29 有时AI可能会无意中

36:31 触及它知道你的信息，

36:35 试图更好帮助你，但在错误场景下。

36:36 所以这方面还有很多基础设施工作要做，

36:38 就是如何在各种场景里区分“你是谁”。

36:42 一旦解决，

36:45 我认为记忆会成为

36:48 AI产品的核心优势之一。

36:50 无论是它们自身的记忆，或ChatGPT借用记忆，

36:54 两年后你开始用的产品，

36:55 如果不感觉它了解你，

36:58 那你会觉得它有缺陷。

37:00 产品的上手体验，

37:03 几年后应该是不再存在的概念。

37:05 我认为记忆功能能实现这一点。

37:07 我个人的体会是，我每天都和多款AI交流，

37:11 它们的互动方式和价值，

37:13 在使用两三个月后提升了很多，

37:17 比刚开始用时好太多。

37:19 - 太不可思议了。

37:24 好吧，我不知道未来如何，

37:27 但它会奇妙又奇怪。

37:30 我很期待。

37:31 - Bolivia，非常感谢你。

37:34 今天能聊聊，

37:38 还有这份报告，真的很开心。

37:39 有什么最后想说的吗？

37:40 - 没有，我只是很期待大家读这报告。

37:44 里面有很多有趣的数据，

37:46 六个月后肯定会完全不同。

37:48 到时候我们再见。

37:50 - 非常激动人心。

37:53 告诉我们你的想法，

37:55 谢谢收看。

37:56 - 谢谢。

37:57 （欢快音乐）

37:58 I'm excited for it, yes.

38:00 - Bolivia, thank you so much.

38:01 It was super fun to actually have this conversation today

38:04 and go through the report.

38:05 Any closing comments?

38:06 - No, I'm just excited for people to read it.

38:08 There's a lot of interesting data in there next time

38:10 and I'm sure it will look wildly different six months from now.

38:13 So we'll be back then.

38:14 - Really exciting.

38:15 Well, tell us what you think

38:16 and thanks for checking us out.

38:17 - Thank you.

38:23 (upbeat music)

0:00 [MUSIC]

0:04 Olivia, welcome.

0:05 >> Thanks for having me.

0:06 >> It's the most exciting time of the year,

0:08 which is the top 100 report is coming out today, I think.

0:11 Is that right? >> Yep.

0:12 >> It's been six editions over three years.

0:15 Talk to us about what's the same, what's changed,

0:18 what's your excitement level, what's up with the report?

0:20 >> Yeah, in many ways, so much has changed,

0:23 and there's been just an incredible amount of growth

0:25 since the first time we put out this list in 2023.

0:28 On the other hand, from a macro level, we're still so early.

0:32 Chachibiti is by far the biggest global AI product,

0:36 and still only 10% of the global population is using it on a weekly active

0:41 basis.

0:41 So there's a lot more to come.

0:43 I do think this past six months has been maybe my favorite time

0:47 and the most exciting time because of the shifts that we've seen.

0:51 One of them has been at the race for the consumer is really heating up.

0:54 So Chachibiti, of course, but also Gemini and Claude are kind of doubling down

0:59 on their own ICP within consumer and prosumer.

1:03 And I think we're starting to see how these platforms might have

1:06 compounding advantages over time.

1:08 And so that makes it especially kind of existential or interesting

1:12 of who is acquiring the most users.

1:14 And then on a related note, this was actually the first issue

1:17 that we included products that were non-AI native,

1:20 but are now majority AI-enabled.

1:22 So things like Canva, Notion, Free Pick.

1:25 Notion actually announced that now they think half of their new ARR

1:29 is driven by AI first features, which is very cool.

1:32 And then lastly, I think we've seen a big expansion of AI

1:36 outside of just like the website or app prompt box.

1:39 So we have all of the browsers that have come out, like, you know,

1:43 DIA, Comet, Atlas.

1:45 We have Claude and Excel, PowerPoint and Chrome.

1:48 And then we have desktop apps like cursor, whisper flow, granola.

1:53 And so there's been just a really kind of exciting explosion

1:56 in the ways that people are using AI.

1:58 So exciting.

1:59 There's a ton to cover here.

2:00 So let's start with the big foundation models.

2:02 Can you talk a bit about what you think

2:04 are the respective areas of specialization for Gemini, Claude?

2:07 And then, of course, Chatsheet BT,

2:08 because it feels like it's been a rising tide story more

2:11 than these models trading off with each other.

2:13 Yes, I agree.

2:14 Despite the drama maybe of the past week

2:17 where we have Katy Perry taking sides on Twitter in the LML

2:21 or which is something that I didn't ever see coming,

2:24 I think at a base level still, if you look at AI usage,

2:27 like Chatsheet BT is a very, very clear winner.

2:29 So on web, they're 2.7 times bigger than Gemini.

2:33 On mobile, they're 2.5 times bigger than Gemini.

2:37 And then despite, again, like the kind of tech Twitter discourse,

2:41 Claude, they're almost 30 times bigger than Claude on web

2:45 and almost 80 times bigger than Claude on mobile.

2:47 So we had seen that Sam Altman tweet

2:49 back in the Super Bowl ad wars era.

2:51 The Texas tweet.

2:52 Yes, he was like, we have more people using Chatsheet BT,

2:56 free version in Texas than Claude has,

2:58 like all users globally, which is true.

3:00 That being said, I think we are seeing,

3:04 I don't think bifurcation is the right word,

3:06 but maybe expansion in the number of products people are using

3:10 and what they're using different products for,

3:13 which has kind of changed the market share a little bit.

3:15 Claude in particular is really doubled down on prosumer

3:19 with things like co-work, Claude code,

3:21 Claude and Excel and PowerPoint.

3:23 If you actually look at the app stores

3:25 that are emerging on Claude and Chatsheet BT,

3:28 they both have 200 plus apps,

3:30 but there's only 11% overlap.

3:32 Like Claude is very much doubling down

3:34 on like premium data sources, research tools,

3:37 science tools, financial data.

3:39 And Chatsheet BT is really doubling down

3:41 on like consumer marketplaces, travel, nutrition,

3:44 consumer finance, things like that.

3:46 And then Gemini is kind of in its own little corner as well.

3:50 And the traction has largely been driven

3:52 by creative tools there.

3:53 So if you look at their kind of active users

3:55 and paying users, it's nearly perfectly correlated

3:58 to releases of like VO3, Nano Banana One, Nano Banana Pro,

4:02 Nano Banana Two, they're doing a little more in prosumer,

4:06 they're adding AI to Gmail, Sheets, Calendar,

4:11 but that's all being captured

4:13 by like their existing products

4:15 versus like a net new experience.

4:17 - Maybe let's dig into the app store dynamic a little bit

4:20 because that's so fascinating.

4:22 Can you talk about the bull case for Chatsheet BT

4:25 with I think what they call the apps directory?

4:26 - Yeah, yeah.

4:28 I think the approach we're seeing with Chatsheet BT

4:30 and Sam said this himself on Twitter is like,

4:32 we wanna be the AI for everyone.

4:34 And that means that they're trying to acquire every consumer

4:37 and they'll monetize them in different ways.

4:39 So like I think Claude has been very clear

4:40 that they're just gonna monetize via subscriptions,

4:43 which is great for people and companies

4:45 who can pay for subscriptions, but it won't be everyone.

4:48 I think you see that with the plugins

4:50 that they're leaning into, which are like paid, high ACV,

4:54 like work data tools and data services.

4:56 - Similar labs, things like that.

4:57 - Yeah, things like that.

4:58 Like pitch book, like things that you would use

5:00 if you're an investor or a scientist, a mathematician.

5:04 And Chatsheet BT I think is going more of the

5:08 somewhat of a Google type approach

5:10 in that they're building things

5:12 that like the average person will want to use.

5:15 And maybe a smaller percentage of those convert

5:18 to subscriptions right now, but they will be able

5:20 to monetize those people through ads

5:23 and probably also I would guess through transactions.

5:26 Like if they're building the gateway to like book a trip

5:30 or do all of these other kind of long tail consumer purchases,

5:35 hypothetically they should eventually be able to take

5:37 some kind of cut of that, at least for the traffic

5:40 that they're driving.

5:41 And so I think that that is the bull case

5:44 for the Chatsheet BT App Store that isn't yet showing up

5:47 in the data that will probably become like even more evident

5:50 in the next year or two.

5:52 - Yeah, it's really interesting 'cause it touches

5:53 on your point in the report about compounding advantages

5:56 and how context compounds.

5:58 Can you talk a little bit about that concept

6:00 and then what's your proxy in terms of a metric for it?

6:02 Is it session time, is it number of sessions,

6:04 is it the amount of data you've provided

6:06 or is there something else?

6:07 - Yeah, this is a really exciting question to me

6:10 because I think thus far with these horizontal LLMs

6:13 like Chatsheet BT, Claude, Gemini, perplexity,

6:17 we've kind of lived in a world where the context

6:20 and the memory is somewhat easily exportable.

6:22 Like Claude ran a campaign around this recently.

6:25 But I think there's gonna be increasing lock-in

6:28 and I do think that probably actually benefits

6:30 the broader, more horizontal tools like Chatsheet BT

6:33 for a few reasons.

6:35 So I think one, we've already seen Chatsheet BT focus on

6:39 or start to build out products where you interact

6:42 with other people on them through the platform.

6:44 So the group chats, like imagine if you,

6:46 if there's an even more successful version

6:48 of Chatsheet BT group chats

6:50 and all of your friends are on there,

6:51 then if you wanted to turn from Chatsheet BT,

6:53 you'd also have to convince them all to go through another product.

6:56 - Exactly.

6:57 I would say the second one is kind of also

7:00 like an Apple Google comparison in that

7:02 as these app stores emerge, it is likely that developers

7:05 might start to concentrate their time and effort

7:08 in who they build for in the most sophisticated way,

7:12 who they ship to first, depending on who has the most users

7:16 or maybe in some cases who's the most willing to pay,

7:19 but for a lot of these consumer tools,

7:20 it'll be who has the most users.

7:22 So I think that also benefits Chatsheet BT.

7:25 And then the other thing probably that I'm most excited

7:27 for this year that Sam Altman had kind of hinted at

7:31 is this like authentication with Chatsheet BT layer.

7:34 So essentially you'd be able to log in

7:36 with your Chatsheet BT account

7:37 and take like your memory and your tokens with you.

7:42 And then that other product would be able to kind of borrow

7:45 those things to be even more powerful and helpful for you.

7:48 And if that's the case, then you're wanting

7:50 to have more of your core identity live on Chatsheet BT

7:54 because then it can lend it to these other tools

7:57 that are even better for you.

7:58 - It's so smart and it really plays to their advantages

8:01 in that they have signups for 900 million people.

8:05 And then the third party developer ideally

8:07 would not want to pay for the inference.

8:09 So the user can bring their inference capacity with them.

8:12 There's an advantage for the developer.

8:13 Chatsheet BT gets the lock in.

8:15 The user gets the benefit of personalization

8:17 and it all kind of works.

8:18 - Yes, I totally agree.

8:19 The one question mark I still have on this

8:21 that I think could play both positive and negative

8:24 in terms of increasing lock in for the consumer product

8:27 is what your work goes with,

8:30 like what your enterprise contract is.

8:33 So for example, in some ways, it's good for me

8:36 if we, if my company uses Chatsheet BT for work

8:40 because then I know how to use the product.

8:42 And as a normal consumer, they might have tried

8:44 one or two AI products.

8:45 So they're more likely to be comfortable

8:47 and keep using something that they've already used.

8:49 On the other hand, some people might not want to mix identity

8:53 and mix memory across their personal and work use cases.

8:56 - True.

8:57 - And so I'm really interested.

8:58 I think OpenAI hinted at this recently,

9:00 but I'm really interested in how we kind of segment memory

9:04 across different personas that are within yourself

9:08 that are using these products.

9:09 - Don't cross the streams.

9:11 - Yeah, exactly, exactly.

9:12 - Well, maybe actually switching gears

9:14 to Gemini for a moment.

9:15 You know, I think about the kind of just the vibes

9:18 around Google with their early AI products,

9:21 the Bard, which they'll never live down.

9:24 - Some top times there.

9:25 - So where we are today with products like Nana Banana,

9:27 like even naming it Nana Banana is such a perfect microcosm

9:30 for how far Google has come.

9:32 - Yeah.

9:33 - And it seems like they have a lot of intentions

9:34 around multimodality.

9:35 - Yeah.

9:36 - What's your assessment of their approach?

9:38 - I've been impressed.

9:40 I think they have been hesitant, maybe in some ways,

9:44 more hesitant in same ways is exactly what we would expect.

9:47 So to kind of bake AI into the core features

9:51 because there's a risk of either like cannibalizing

9:53 their own product or like there's so many people

9:56 who have used these tools for 10, 20, 30, 40 years.

10:00 And so the switching costs there is like a little bit high.

10:04 They don't want to scare users when AI is suddenly

10:06 popping up and everything, which I understand.

10:09 But what they've done a really, really good job at

10:11 is these new creative products that are basically

10:15 very model driven from the DeepMind team

10:18 who I think is generally fantastic.

10:20 - I think Notebook LM was actually the first look at this

10:23 and that was something truly new in like consumer AI audio.

10:27 And now we have the image and video models.

10:31 So in some ways with the big company like this,

10:32 they kind of have to get over, get out of their own way

10:35 in terms of being able to actually innovate.

10:38 And it seems like they are, but you also worked at Google.

10:40 So I'd be curious for you.

10:41 - It's interesting to just, I'm glad you brought Notebook

10:44 'cause Notebook is sort of this green field area

10:47 within product area within the company.

10:48 So you don't have 10 VPs fighting over it.

10:50 And as a result, I think just the progress on Notebook

10:53 has been tremendous.

10:54 You know, they just launched a video generation feature

10:56 that helps visually demonstrate all the concepts

10:59 in your sort of workspace, which is cool.

11:01 Conversely, when you look at the existing product surfaces

11:04 like sheets or docs, there's just so much,

11:08 one sort of momentum and inertia from the past,

11:10 but then management overhead around those.

11:13 It's harder for them to do anything other

11:14 than the most obvious incremental thing.

11:16 - Yes, I agree.

11:18 We'll see what happens there in the next few years.

11:21 I feel like they're gonna put up a fight

11:23 on some of those products 'cause they don't wanna lose

11:25 that user base, but to your point,

11:26 they're already locked in with so many enterprises

11:28 that they might not have to do that much,

11:31 at least in the near term to kind of keep up.

11:33 - You know, implicit in this conversation

11:35 is we experience and talk a lot about AI in the West.

11:39 Talk a bit about the sort of global AI trends.

11:41 There was a few surprising things I saw in there.

11:43 - We kind of expanded our scope in terms

11:46 of what we looked at for this report,

11:47 which ended up being like very fun and interesting.

11:50 Two things that are probably obvious

11:54 in terms of how they differ from the rest of the world

11:57 would be Russia and China.

11:59 So Russia, I think China, everyone knows,

12:02 like a ton of AI products are kind of censored or banned.

12:05 And so almost all of the usage,

12:07 they actually have the lowest combined

12:10 Chachi, BT, and Gemini usage of any country.

12:12 It's only 15%.

12:14 So they're mostly using like Dalbau,

12:15 which is made by ByteDance, DeepSeek,

12:18 Quinn, Kimi, those kinds of models.

12:21 The somewhat of a surprise to me

12:22 was that Russia actually is a very, very similar story,

12:25 where they have also their own kind of parallel AI ecosystem

12:29 out of necessity because they have some level of sanctions

12:33 and things like that that prevent them

12:34 from using all the US-based tools.

12:38 So we've seen products like GigaChat and Yandex,

12:40 which are Russia specific built by Russian,

12:43 often state-affiliated companies,

12:46 have big, big usage there, and then DeepSeek.

12:48 So Russia is the number two market for DeepSeek

12:51 after China.

12:52 And so if you look at the kind of like per-country adoption data,

12:56 like, yes, there's some blips where like,

12:59 this country uses Claude a little bit more,

13:01 this country uses Gemini a little bit more,

13:03 but the two huge outliers are Russia and China,

13:06 and those are like big, big markets.

13:08 And so I think it's worth watching what's going on there.

13:10 - It's interesting though, because both Russia and China,

13:12 they're outliers because of restrictions

13:14 around how models can be used,

13:15 and maybe cultural preferences.

13:18 Are there any other countries that have geo-specific trends,

13:21 or is this a sort of global AI behavior set?

13:24 - Yeah.

13:25 I would say in terms of like model development,

13:28 proprietary model development that allows you

13:30 to deploy proprietary AI products,

13:34 most of that research is coming out of the US and China,

13:37 maybe a little bit out of Russia.

13:39 I think we are seeing a few kind of native ecosystems

13:43 in other places.

13:45 I would, Korea has a couple of their own products

13:48 like Neighbor and Cookal that have built out

13:50 nice kind of LLM interfaces.

13:52 India is probably the other one that I watch really closely,

13:55 just because there's so many people

13:57 that you can have standalone big companies focus on India.

14:01 The other interesting thing about India

14:02 is there's so many different languages,

14:04 like such a range that both LLM products

14:07 and even voice products don't necessarily support very well.

14:11 Like it's a worse experience if you're a primary user

14:14 of one of those languages

14:15 and you're trying to use something like a chat UBT.

14:18 So so far we haven't seen a huge amount of variants there yet,

14:22 but I would not be surprised maybe to see more founders,

14:25 even from the US, like targeting the Indian market for AI.

14:28 And then the other thing I wanted to mention,

14:31 we did for the first time also kind of like a heat map

14:35 essentially of which countries are adopting AI the most

14:39 and the least on a per capita basis.

14:41 So we looked across like the 10 biggest LLM products

14:44 to see on web and mobile to see what this might look like.

14:48 So Singapore is number one.

14:49 - Crazy. - Yes.

14:50 Then Hong Kong, then the UAE, then South Korea.

14:54 The US is down at number 20.

14:56 So not super low, not incredibly high.

14:59 Russia and China are like very far down the list, like sub 50.

15:04 And there's a lot of interesting stories,

15:06 I think that live in that data.

15:09 The first one is if you think about those top five,

15:11 like Singapore, South Korea, Hong Kong,

15:14 it's a very like the demographics of the workforce

15:17 are very like tech first white collar high skill.

15:20 And the US has a giant chunk of jobs

15:23 where AI hasn't really touched them yet,

15:25 like retail and transportation and some of these other things.

15:29 I think also the cultural norms around AI

15:33 are shockingly diverse.

15:36 If you're in the US, you have probably internalized this

15:39 ongoing angst and questioning around--

15:41 - Yeah, I was going to ask you about this 100%.

15:44 - Or AI is terrible for artists

15:47 or all of these other things that make people pick up

15:50 or not pick up AI. - Yes.

15:51 - There was actually a big survey last year

15:54 from Edelman, the global media company.

15:57 And the US had a fairly low rate of trust in AI.

15:59 It was like 32% and most of these other countries

16:02 that are high on the list are like 50, 60, 70%.

16:06 So that I think has also held the US back

16:08 despite the fact that we are where the biggest products come from

16:12 are per capita usage is lower than a lot of these other markets

16:14 that have maybe smaller populations

16:16 but have embraced it more.

16:19 - I think that's exactly right.

16:20 Now I was reading that in China,

16:22 the sort of favorability views on AI are 80%.

16:25 - Yeah. - 80% hold a favorable view.

16:27 And I know UAE and Singapore,

16:29 I think they've sort of culturally wired to be tech optimistic.

16:34 - Yes. - Which is an advantage.

16:35 - Yes, yes, definitely.

16:37 It's interesting to see some of these smaller countries

16:39 like the per capita adoption rate.

16:42 Like in the US, it's around probably a third of people

16:45 are monthly active users of something like a Chachi BT.

16:49 In some, even some of like the European countries

16:52 or Eastern Europe, it's like 50, 45, 60% on smaller bases.

16:58 But they've kind of embraced it more quickly than we have here.

17:02 - Yeah, really interesting.

17:03 You know, one thing that I'm sort of watching

17:05 and I'm interested in is as you look at the spectrum of AI

17:07 from the most functional,

17:09 almost like a Google search replacement

17:11 to the most cultural, creative, personal,

17:14 we should see more divergence country by country

17:17 because obviously the culture,

17:18 the movies they make in India

17:19 couldn't be more different than the movies they make

17:21 in China or the US. - Yeah.

17:23 - So why wouldn't their use of creative tools be different?

17:25 - Yeah, and this is honestly part of the reason why

17:27 we started looking at the geographic segmentation

17:30 in this report is because for the first two and a half,

17:33 three years of generative AI,

17:35 the vast majority of consumers were maybe interacting

17:38 with one product and now it's broadening quite a bit.

17:41 And I think that we will see more

17:43 of these market specific tools.

17:45 And if they capture enough of that market

17:48 like some of these Russian companies or Chinese companies,

17:51 they can actually surface up to the global list

17:53 if kind of the market is big enough.

17:55 Talk a bit about the evolution of creative tools

17:58 and how much do you think that is that a reflection

18:00 of culture, is that driving culture,

18:02 when do we cross that threshold?

18:04 - The creative tools trend has been fascinating.

18:07 I mean, obviously the first big generative AI product

18:10 was actually mid-journey, which came out before Chachi PC.

18:13 - True, that's right, yeah.

18:14 - And in our first few editions of the list,

18:16 it was very much dominated by creative tools.

18:19 And I've said this before,

18:21 but the creative tools benefit from kind of hallucination

18:24 of the early models because they produce things

18:27 that are more kind of surprising or beautiful or original.

18:30 And so for a while, those are the only things

18:32 working in consumer AI, really.

18:35 Now it's shifted a lot.

18:36 Creative tools are still a huge chunk of the list,

18:39 but like the type of creative tool

18:41 that is a standalone big business has changed.

18:45 I would say the biggest change

18:47 is we're seeing fewer standalone image generators.

18:51 A lot of this activity,

18:52 if you're making like a basic commodity image,

18:55 like a meme or a basic marketing image or an infographic,

18:59 like the core models in Chachi BT and Gemini

19:02 are quite good at those things now.

19:05 So the products that are still surfacing on the list,

19:07 like an I-U gram or a mid-journey

19:10 are either very aesthetically opinionated

19:13 or they have very more sophisticated workflows

19:16 that you can't get on something like a Chachi BT.

19:21 Contrasting that, I would say like music, voice, video,

19:26 all seem to be things that the model,

19:28 the biggest model companies have maybe invested less in.

19:31 And so we've seen players like Suno and music

19:33 and 11 labs in voice kind of completely break out

19:37 and rise to top 20, top 15 on the list

19:40 and then like hold their spot there over time.

19:43 And then there's like a compounding lock-in

19:45 from like the community and, you know,

19:46 the big base of enterprise customers and all of that.

19:50 Video is where I have the most questions.

19:54 OpenAI has been investing in it with Sora

19:57 and of course Google with Vio,

19:59 but the Chinese models are so good

20:01 because they can train on any data.

20:03 So Sea Dance 2 is probably the best example of this

20:08 where it's just kind of in some ways head and shoulders

20:10 above what the U.S. companies have thus far been able to do.

20:14 So I think we'll see, I think this actually benefits platforms

20:18 like a Korea where you can use all the models in one place

20:21 because my sister Justine wrote an article about this.

20:25 But the way video is shaping out,

20:26 there's it's unlikely to be like one model to roll them all.

20:29 And so you kind of need to be able to switch between them.

20:32 - That seems true of most of the model spaces, you know?

20:35 Chat models, creative models, even code models

20:38 have their areas of specialization.

20:40 You know, people talk about kind of ergonomics of opus

20:43 versus the accuracy of codecs.

20:45 And that's just, that's a trade-off.

20:47 You have to choose what tool you want to use for which problem.

20:50 - Yeah, absolutely.

20:51 - Sora is really interesting to me

20:54 because it represented both a sort of a big step forward

20:57 in the model, but also a really ambitious experiment

20:59 around social.

21:00 And there was data in the early days of Sora

21:03 like the percentage of people that we're creating,

21:05 which is dramatic, these 10x higher than we'd seen before.

21:08 You know, what's your kind of assessment

21:09 of the Sora social effort versus the model effort?

21:12 And where do you see that going?

21:14 - Sora is so fascinating.

21:15 And I think was a very interesting early experiment

21:19 that I think taught us all a lot about kind of

21:22 both creative tools, but also maybe more importantly

21:24 what consumer social in the AI era might look like.

21:27 So by the numbers, they had a massive launch.

21:29 They were number one on the App Store,

21:31 the US App Store for 20 consecutive days,

21:33 which is very hard to do.

21:34 It means you're probably getting,

21:36 to be number one on the App Store,

21:37 you probably have to get these days

21:38 150,000 daily downloads.

21:40 So it's like a high download volume.

21:42 And they actually hit a million users faster

21:45 than Chachi BT itself.

21:46 So like huge launch.

21:48 And actually, I think what a lot of people underestimate

21:52 is it still is very significant usage.

21:54 So three million dows per sensor tower,

21:57 which is not bad at all.

21:59 What has dropped off about Sora is the new download.

22:02 So there may be,

22:03 they peaked like six million a month in November.

22:05 It's looking like a million and a half now.

22:07 I think that what has really worked about Sora

22:11 is that it's a very good video model.

22:14 And they kind of innovated

22:15 and introduced this concept of cameos,

22:17 which is where a real person

22:19 can grant their likeness to Sora

22:21 so that they and others can generate videos of them.

22:24 So like a lot of people in the early days

22:25 were doing like mean videos of their friends.

22:27 Like Jake Paul went viral

22:29 'cause he was the first big celebrity

22:30 to like lean into Sora.

22:31 So you were seeing like insane Jake Paul videos

22:34 that we wrote.

22:35 Yeah.

22:36 I mean, honestly, good for him.

22:37 Yes, yes.

22:38 I think what worked less about Sora

22:42 is that because the content was exportable,

22:45 people would take it to TikTok.

22:47 They would take it to Instagram Reels.

22:49 They would take it to YouTube.

22:50 And there it competed against the best human made content.

22:54 And so the overall feed experience was just better

22:57 because you were seeing the best of both,

22:59 not just like the best of Sora.

23:00 I don't think we've seen a social product yet succeed.

23:05 That's like entirely AI content.

23:07 The emotional stakes are just feel lower in some ways.

23:11 And so I would imagine we'll see more examples like these

23:14 where Sora still has clearly very, very significant usage

23:18 in revenue as a creative tool,

23:20 but not so much as a social app.

23:21 Right.

23:23 And I don't know if there'll be,

23:24 there probably will be a massive AI native social network,

23:28 but we haven't seen what it looks like just yet, I would say.

23:30 It'll be interesting.

23:31 We discussed this frequently,

23:32 but every social product has a status game.

23:35 Yes.

23:35 Instads, it's maybe be the hottest and on X,

23:38 it's be the most interesting.

23:39 And it felt like the emerging status game on Sora

23:41 was be the funniest.

23:42 Yes.

23:43 And I think this is one of the reasons why it's hard

23:44 for the content to cross over.

23:46 Because it's just two different ways of judging

23:49 what is interesting and great.

23:50 I agree.

23:51 What they might do,

23:53 if I had to imagine where they might find more of a niche,

23:55 they have now inked a bunch of deals

23:57 with big media companies like Disney.

23:59 And so if Sora is the only place where you can make

24:02 like licensed like fan videos of like beloved

24:06 kind of characters and entertainment figures,

24:08 then like that's very interesting.

24:09 Totally.

24:10 But we're early, I think in how that rules out.

24:12 It's so early.

24:13 I know.

24:14 We can say that.

24:15 Yeah.

24:16 We can't have this conversation without talking about agents,

24:18 open claw, manis, genspar, moldbook.

24:22 Give us an overview of what has happened

24:24 in the last 60 days in the world of agents

24:26 and what does a report tell us?

24:27 I think this is mostly why I say the last,

24:29 you know, even six months,

24:31 but actually even two months of this report

24:33 have been like the most interesting

24:35 that I think we've seen.

24:36 So open claw actually, as you'll see,

24:39 is not on our rankings because it blew up in February.

24:42 Our data ends in January.

24:43 But we did pull the data for February.

24:46 And if it had been eligible,

24:48 it would have been number 30 on our web list,

24:50 which is a pretty big debut.

24:53 I think the really interesting thing about open claw

24:55 is the usage has just continued to accelerate

24:58 in the technical community.

24:59 So now it's, I think number one,

25:01 GitHub stars of all time.

25:02 It passed React, it passed Linux.

25:04 Wow, it's a really, really interesting.

25:05 Yes.

25:06 Very impressive.

25:08 But in terms of overall new users,

25:11 it's kind of plateaued.

25:12 So we looked at kind of visits to the get started

25:15 or sign up page.

25:16 And that is kind of flat week over week since early February,

25:20 which I think indicates that like,

25:21 it is an amazing product of your technical.

25:24 It has not yet fully escaped containment

25:27 to non-technical people,

25:28 which of course is like a bigger population.

25:32 They were acquired by OpenAI.

25:33 So if I had to guess,

25:35 or what I'd love to see OpenAI do

25:37 is build like productized open claw

25:39 into something that is usable for a mainstream consumer.

25:43 And I think we've also just seen the ideas

25:46 behind the open claw architecture

25:49 inspire so many other founders.

25:51 Like how many pitches do we take a day

25:53 where the founder is like,

25:54 I want to be open claw for this?

25:55 Absolutely. Open claw made me realize this was possible.

25:58 Yes.

25:59 And so I think we're going to see open claw itself

26:01 will continue to succeed and be a massive product.

26:03 And I'm guessing we'll see more

26:05 kind of like verticalized focus versions

26:07 of open claw for different use cases.

26:09 Yeah, it's so interesting

26:10 because it feels like one of the things

26:11 that makes Open Claw work so well

26:13 is it can operate across all models in all directions.

26:16 And I sort of wonder if it dilutes the value

26:19 of Open Claw to have it be sole model provided

26:21 and therefore it's sort of counterpositioned against labs.

26:24 Totally.

26:25 They've kept it, I think, multi-model for now,

26:27 at least in my usage.

26:28 So we'll see how it trends.

26:29 I think it would be smart to keep it that way for usage.

26:31 But yeah.

26:33 Is Manus the consumer grade open claw

26:35 or how do you distinguish the team?

26:36 Yes, some might say that.

26:38 And I do actually think,

26:39 so Manus made our web list.

26:42 And of course they had a $2 billion plus acquisition

26:46 by Meta also in the course of the list.

26:48 Incredible growth.

26:49 Like the ramp that they reported from like zero

26:52 to 100 million, 200 million ARR in the span of like

26:55 honestly six, nine months is really kind of best in class.

26:59 My view on why Manus was so successful was

27:02 it was really the first consumer grade agent

27:04 that could actually operate fairly autonomously

27:07 across products and platforms.

27:10 So you could connect email, you could have it browse the web.

27:13 And it could make slides, it could make spreadsheets.

27:17 I spent a lot of time in the early days

27:19 trying like this was a year ago Chachi BT operator

27:23 or Google's Project Mariner.

27:25 And none of them were reliable.

27:27 And Manus was a breakthrough in kind of agent reliability

27:30 and agent accessibility for the consumer.

27:33 I think the fact that they did the acquisition

27:35 is interesting in terms of where this is going.

27:39 In that once everyone has that agent capability

27:43 and you might imagine they will

27:45 if it's kind of based on the core underlying models,

27:48 then it's actually if you're such a horizontal product,

27:52 you may be better off with the distribution forces

27:55 of a meta or a Google or something like that

27:59 versus a standalone company.

28:00 That's definitely not true if you're building something

28:02 more vertical, but if you imagine that like Google now

28:06 has the resources to create a Manus,

28:10 then that's a really hard thing to keep fighting against

28:12 as a startup, I think.

28:13 And obviously the big companies have a billion different

28:16 priorities so they're not going to do everything

28:18 best in class.

28:19 But it's why I've generally been a little more cautious

28:22 about the very, very horizontal consumer AI apps

28:26 just because it's probably both in scope

28:28 for the bigger companies.

28:30 And they have the advantage of already having IT approval

28:32 and enterprise contracts and all of that.

28:34 - Right, right, it is interesting that we sort of cross

28:37 this cultural threshold though,

28:39 where Manus seemed like a non-obvious bet

28:43 in terms of just the breadth of the offering.

28:45 And now it seems like they're living in the future

28:47 a little bit.

28:48 - Yes, absolutely.

28:49 They were, it's obviously an incredible engineering team.

28:52 Like the quality of the product was like three,

28:55 six months ahead of the rest of the market,

28:57 which is not easy to do when you're competing with teams

28:59 of like thousands of free searchers.

29:00 - Totally.

29:01 Let's use this to kind of leg into a conversation

29:03 about other horizontal AI products,

29:06 things that live beyond the sort of web window.

29:09 - Yes.

29:10 - What are you saying there?

29:10 - Yeah, that has been a massive theme.

29:12 When I think about the products that I interact with

29:15 on a daily basis in the AI world,

29:17 quite a few of them are actually desktop apps,

29:19 like things like granola, voice dictation tools,

29:22 cloud co-work, those kinds of things.

29:25 And it does become a methodology problem for our report,

29:29 because we can track website visits very well.

29:31 And so we can track the first time

29:33 that they download the desktop app.

29:34 We can track mobile app usage very well.

29:37 We cannot track desktop usage that closely.

29:40 And I think that is increasingly,

29:43 as AI products become more sophisticated,

29:45 having them live in their own dedicated application,

29:48 much of which will run on desktop

29:50 because it can interact with your files

29:51 and it can be more ambient.

29:53 I think that's gonna happen more and more.

29:56 And so I think moving forward, finding us finding ways

29:59 to parallel track ranking these products

30:02 by web and mobile usage, but also by revenue,

30:05 is gonna be a pretty good idea.

30:07 'Cause if you think about things like cursor,

30:09 some of the consumer prosumer AI apps

30:12 that are generating the most revenue

30:13 have very few, very little usage on web.

30:16 It's almost all in kind of a dedicated app.

30:19 - Yeah, it's really interesting.

30:20 It also feels like the fact that opening

30:22 I released Atlas and anthropic released co-work

30:26 shows you where their priorities are.

30:27 - Yes, definitely.

30:28 I fully agree.

30:30 The AI browser debate is his own interesting thing.

30:35 I feel like we're still in the early to mid phases

30:38 of how that's gonna play out.

30:40 And I think the instinct behind an AI native browser

30:43 is right in that if you can have AI be

30:47 kind of always on, always available ambient

30:51 in where you're spending a lot of your time online,

30:54 like that's a good opportunity.

30:56 Perplexity comment, I think actually led the way there.

30:59 - It's a great product.

31:00 - It's a great product.

31:00 And the interesting thing is if you look at

31:03 kind of the highest spike for comment and Atlas

31:06 in terms of visits to the download page,

31:09 comment is five times ahead of Atlas,

31:11 which is wild because Chachi BT's audience is like so massive.

31:15 - No, I miss, yeah.

31:16 - And I think what we've seen is like

31:18 comment and Atlas still have very dedicated,

31:22 excited user bases,

31:24 but for the average consumer,

31:26 the switching cost of a browser is non-trivial,

31:29 just because like you have workflows set up,

31:32 you naturally just open this one app.

31:35 And so it not only has to be like feature parity,

31:37 there has to be one or two features of the AI browser

31:40 that are really killer and that are easy enough

31:43 for the average person to set up and access.

31:45 And I don't think that we've seen that quite yet.

31:48 - You know, it's really interesting

31:49 because Sam said, I think six months ago in a pod,

31:51 you know, somebody was asking him,

31:53 what has surprised you the most?

31:54 And he said, it's that the world hasn't changed more.

31:57 And if you look at the trends around

31:58 how people are using Chachi BT at scale,

32:00 it's still, you know, homework and Google like queries

32:03 and a little bit of companionship,

32:05 in a sense, something like a browser

32:07 that gives you an opportunity

32:08 to point the user in a different direction.

32:10 What's your view on how the average person

32:12 is using AI today?

32:13 - Yeah, I think a couple of things.

32:17 So one, I feel like teenage girls

32:19 are like the best source of what is happening in consumer--

32:22 - You haven't taught me this, yes.

32:23 - What will be happening in consumer?

32:24 If you look at all of the biggest consumer outcomes,

32:26 like they were the early adopters of all of these products.

32:30 And so there was actually a Pew Research Study

32:32 fairly recently on how teenagers are using AI.

32:35 Now finally, I think for the first time,

32:38 over half of them are admitting

32:40 to using it for their homework.

32:41 So the real number is probably like 99.99%

32:44 but some of them didn't wanna get in trouble

32:46 with their parents. - Uh-huh.

32:48 - 38% are now using it for a creative tool.

32:51 So editing images, editing video,

32:53 generating images and video.

32:55 And then this like emerging slightly longer tail,

32:58 but I think will ultimately be amongst the biggest behaviors.

33:02 16% are using it for just like casual conversation,

33:05 like not the intense companion products,

33:08 but like just having somebody to talk to.

33:10 And then 12% are using it for like emotional support

33:13 and advice.

33:14 I think all of these use cases will like asymptote

33:17 around probably 100% ultimately.

33:21 And so those are behaviors that maybe have been less well

33:23 served by products so far and will be going forward,

33:26 whether it's on a chat should be tea

33:27 or whether it's on like a standalone product.

33:30 And then the other big thing that I'm looking out for

33:34 is agents.

33:35 Like I think basically-- - Our teenage girl's gonna use agents.

33:39 Come on. - So here's the thing.

33:40 I think that agents similar to like how in 1990,

33:44 an internet company was like a dot com company, right?

33:47 Or tech company like dot com was its own,

33:49 its own designator. - Right.

33:51 - I think that this is what's gonna happen with agents.

33:53 Where ultimately like every tech company

33:55 was a dot com company.

33:56 Like I think ultimately every AI company

33:58 and then every tech company is going to be an agenda company

34:02 because that's just where the models are headed.

34:05 And if you can deliver outcomes

34:07 and not just kind of inputs to your users

34:10 as a software product, that's so much more compelling.

34:12 So yes, I think 13 year old girls will be using agents

34:15 but they will not think of them as agents.

34:18 But I think it does unlock a lot of these other consumer

34:21 use cases of AI like finance,

34:24 healthcare, travel planning, complex trend shopping even

34:29 where pre-agents, there was just so much data

34:32 you had to go out and grab and do it reliably

34:35 and do it across systems that like wasn't really possible.

34:38 And now it is.

34:39 And so I think we're gonna see an explosion

34:40 of those other use cases in the next few months.

34:43 - How long do you think it takes to play out?

34:45 I mean, is everybody using their own open claw

34:47 in 12 months?

34:48 Is that five years away?

34:49 Is that the wrong mental model?

34:50 Like where when we have this conversation,

34:52 perhaps in six months at the next top 100,

34:55 what does the world look like?

34:57 - I feel like every time I predict something

34:59 it happens much more quickly than I would have thought

35:03 which I think is what we're seeing every day

35:04 and that startups are growing faster than they ever have.

35:08 I think people, they're still the cultural change

35:11 and the cultural adoption will be slower

35:13 than the technology change and what's actually possible.

35:17 And so I think what we'll continue to see

35:19 is this early wave of often technical,

35:23 sometimes not technical AI adopters

35:26 like lead the charge on a behavior

35:28 that then six months later everyone else is doing.

35:30 One good example of this that I'm very, very excited about

35:34 is voice, which we've talked about a lot.

35:36 - We have talked about voice.

35:38 - To me, it's like the most information dense,

35:41 high quality source of basically media that we have.

35:46 Like so much of what you do every day

35:48 is actually downstream or upstream of like what you say.

35:51 And we're, I think for the first time in the past six months

35:55 I've seen first engineers and now other people

35:58 within tech companies adopt things like voice dictation.

36:01 Now it's kind of almost a norm at many companies

36:03 that your meetings are gonna be kind of recorded

36:05 and transcribed by an AI.

36:07 Whether that's voice dictation,

36:10 whether that's like a voice pin that answers questions

36:12 or does tasks for you,

36:14 I think that is going to spread

36:15 to the mainstream consumer in the next six to nine months.

36:19 - Really, really interesting.

36:20 Maybe to close, can you talk a little bit about memory?

36:24 - Yes.

36:24 - Do you see that going?

36:25 - Yes, memories as we mentioned earlier right now

36:29 and it can be a little bit jarring

36:31 in that Claude and Chachi PT in particular

36:35 are very good at this.

36:36 Even Gemini, Google has launched something called

36:38 Personal Intelligence where it now can pull information

36:42 it knows about you from your docs, email, et cetera

36:45 to like serve you better with AI across all of the apps.

36:48 And like I said, it can be a little bit jarring now

36:50 because many people are talking to AI about everything,

36:54 personal and professional.

36:55 And so it can sometimes kind of inadvertently cross the line

36:58 of like what it knows about you

37:00 to try to help you better, but in the wrong context.

37:03 So I think there's a lot of work to do,

37:05 kind of on like the infrastructure side almost

37:07 of like how we sort out who someone is in every context.

37:11 Once that is settled,

37:13 I think that memory will be one of the core advantages

37:17 for AI products,

37:19 whether it's their own memory or like Chachi PT lending memory,

37:24 any product that you start to use two years from now,

37:27 if it doesn't immediately feel like it knows you,

37:30 it will feel broken.

37:31 Like the concept of like onboarding to a product

37:34 should not be something that exists in a couple of years.

37:38 And I think that that is something

37:39 that memory is really gonna enable.

37:40 I see it personally for myself where I talked to AI all day,

37:44 talked to several AIs all day.

37:46 And the way that they interact with me

37:48 and the kind of value that they're able to provide

37:50 has been so much higher two or three months in

37:53 and a days when you start using it.

37:55 - Incredible.

37:56 Well, I don't know what the future holds,

37:57 but it's gonna be weird and wonderful.

37:58 I'm excited for it, yes.

38:00 - Bolivia, thank you so much.

38:01 It was super fun to actually have this conversation today

38:04 and go through the report.

38:05 Any closing comments?

38:06 - No, I'm just excited for people to read it.

38:08 There's a lot of interesting data in there next time

38:10 and I'm sure it will look wildly different six months from now.

38:13 So we'll be back then.

38:14 - Really exciting.

38:15 Well, tell us what you think

38:16 and thanks for checking us out.

38:17 - Thank you.

38:23 (upbeat music)

Love

· 3d

Karpathy 的 AI 精神病：App 该消失了，设备只需暴露 API

Karpathy 自述已患"AI精神病"——每天对 Agent 说话16小时、同时跑十几个并行任务、手写代码归零。他认为 App 时代该结束了，智能设备只需暴露 API，由 Agent 统一调度，无需再为每个设备单独开发应用。

Moon

· 3d

我们正在打造 TERAFAB，以弥合当今芯片生产与未来需求之间的差距——一个星际未来。 #TERAFAB #ai #芯片
加入我们 → http://terafab.ai

https://x.com/SpaceX/status/2035542868790558771?s=20