Builder

VIP Juror

Web3来时路

Jun 2025

3 Following 18 Followers 598 Post

Post Media

Builder

· 2d

Anthropic 刚刚展示了一个 24 分钟的工作坊视频，内容是关于如何实际为 Claude 生成提示语。

由那些曾经建造它的人来传授知识。

免费。无需注册。没有付费门槛。

我见过一些收费 300 美元的课程，但他们在前 8 分钟里并没有讲解任何内容。

请观看并立即保存这个内容吧。

中文 English

0:00 (欢快的音乐)

0:02 - 你好。

0:05 大家好，我是Boris。

0:08 我是Anthropic的技术团队成员，

0:11 我创建了quad code。

0:13 今天和大家聊聊

0:15 使用quad code的一些实用技巧。

0:18 会很实用的。

0:20 我不会过多讲解其历史

0:22 或者理论之类的内容。

0:24 嗯，开始之前，

0:27 我们先快速举个手好吗？

0:28 谁之前用过quad code？

0:30 好啊，这才对。

0:34 没举手的朋友们，

0:36 我知道讲课时不该这样，

0:38 但如果你能打开笔记本电脑，

0:38 输入这段命令，

0:42 它会帮你安装quad code，

0:48 这样你后面能跟上讲解。

0:55 只要你的工作环境有Node.js就行。

0:58 (含糊不清)

1:04 嗯，你不必跟着敲，

1:06 但如果还没装，

1:08 现在是装它的好机会，可以跟着学。

1:11 那quad code是什么？

1:15 Quad code是一种全新的AI助手，

1:19 有过几代

1:20 为编程设计的AI助手。

1:23 大多数都是一次补全一行代码，

1:26 或几行代码。

1:28 而quad code并非如此，它是完全自主的。

1:31 它用于构建功能，

1:33 编写完整函数、完整文件，

1:37 同时修复整个bug。

1:39 quad code很酷的一点是，

1:42 它能配合你所有的工具使用，

1:44 你不用改变工作流程，

1:46 也不用更换一切设置来开始用它。

1:48 无论你用什么IDE，无论是VS code，

1:51 还是Xcode或JetBrains系列，

1:56 Anthropic内部有些人

1:57 死也不肯放弃他们的IDE，

2:00 但他们用quad code，因为quad code适用

2:03 所有IDE和终端。

2:05 无论本地、远程SSH，

2:08 或者TMUX等任何环境，都可运行它。

2:12 它是通用的，如果

2:17 你过去没用过这种自由形式

2:20 的代码助手，

2:22 可能会觉得难以入门，

2:24 因为打开后只有一个提示栏，

2:26 你可能会想，我该干嘛？

2:28 我该输入什么？

2:29 它是个强大工具，可以做很多事情，

2:32 但功能强大也意味着

2:34 我们不会指引你用固定流程，

2:36 毕竟你作为工程师，

2:38 应该能自由使用它。

2:40 首次打开quad code时，

2:47 我们建议做几件事

2:49 来配置你的环境。

2:51 这些操作都很简单。

2:52 运行terminal set up，可以让你用shift+回车

2:54 换行，不必用反斜杠输入新行。

2:56 这让使用体验更好。

2:58 你懂的，让它更顺手。

3:01 输入/slash theme设置浅色或深色模式，

3:04 或者其他漂亮主题。

3:05 你也可以输入/slash install GitHub app。

3:09 今天我们发布了GitHub应用，

3:12 可以让你在任何GitHub问题或拉取请求中

3:15 添加或提及一个cloud。

3:17 安装时，只需在终端运行该命令。

3:20 你可以自定义允许使用的工具列表，

3:24 这样不必每次都被提示。

3:25 这非常方便。

3:27 对于频繁提示的设置，

3:28 我肯定会这样定制，

3:30 避免每次都要确认。

3:32 还有一点，我很多提示

3:35 不会手动输入到quad code里。

3:37 如果你用macOS，可以进系统设置

3:40 在辅助功能的听写里启用它。

3:44 我常用的方法是双击

3:45 听写快捷键，然后直接口述提示。

3:49 拥有具体的提示真的很帮忙。

3:52 这实在太棒了。

3:53 你可以像跟其他工程师讲话一样

3:55 和quad code交流，不用敲太多代码。

3:58 刚开始用quad code时，

4:04 它非常自由，功能全。

4:06 那从哪里入手呢？

4:07 我最推荐的就是

4:09 从代码库问答开始。

4:11 直接提问，向你的代码库提问。

4:13 这是我们在Anthropic给新员工

4:15 的第一课。

4:18 技术入职第一天，

4:20 你会了解quad code，下载并配置好它。

4:23 然后你立刻开始提问

4:25 关于代码库的问题。

4:26 过去在做技术入职培训时，

4:28 这对团队来说是个很大的负担，对吧？

4:31 你必须问团队里的其他工程师。

4:34 你要查看代码，这需要一些时间。

4:36 你得弄明白怎么使用工具，

4:37 这也需要很长时间。

4:39 使用 Quad Code，你只需问它，

4:42 它就会自动浏览代码库。

4:43 它会回答这类问题。

4:45 所以在 Anthropic 入职时，

4:47 技术岗位一般需要两到三周。

4:49 现在大约只需要两三天。

4:50 Q&A 另一个很酷的地方是，

4:56 我们不做任何索引。

4:57 所以没有远程数据库存储你的代码。

4:59 我们不上传代码到任何地方。

5:00 你的代码保持本地。

5:02 我们不在代码上训练生成模型。

5:05 代码依然存在，

5:06 由你控制。

5:06 没有任何索引之类的东西。

5:09 这也意味着没有设置步骤。

5:11 你下载 Quad，启动它，

5:13 没有索引过程。

5:14 你无需等待。

5:15 可以马上开始使用。

5:16 这是一场技术讲解，

5:20 我会展示一些非常具体的提示语

5:22 和代码示例，

5:24 希望能帮助你提升 Quad Code 的使用体验。

5:28 可以问的问题比如，

5:30 这段代码具体是怎么用的，

5:33 或者我该如何实例化它？

5:34 Quad Code 不只是简单的文本搜索

5:36 来回答这个问题。

5:37 它通常会进一步深入，试图找到

5:41 这个类是如何被实例化的示例，

5:42 它是怎么被使用的？

5:44 它会给你更深入的答案。

5:45 这是你从基础文档中难以获得的，

5:48 而非简单的按 Command+F 查找。

5:52 我还经常用它查询 Git 历史。

5:55 例如，为什么这个函数有 15 个参数？

5:59 为什么参数名称这么奇怪？

6:01 我在所有代码库里都加了这类信息。

6:03 你会有类似这样的函数或类，

6:07 Quad Code 能查看 Git 历史，

6:09 找出这些参数是怎么被引入的，

6:11 是谁引入的，是什么情况，

6:14 这些提交关联了哪些问题？

6:16 它会浏览所有这些并做总结。

6:18 你不需要告诉它所有细节，

6:21 只要提问就行了。

6:22 你说，看一下 Git 历史，

6:23 我不知道该怎么做。

6:26 顺便说一句，它之所以知道，

6:27 不是因为我们特别提示了它。

6:28 系统提示里没有关于查看

6:30 Git 历史的任何信息。

6:32 它知道是因为这个模型非常强大。

6:34 如果你让它用 Git，它就知道怎么用 Git。

6:37 我们能基于这么优秀的模型开发，真的很幸运。

6:40 我也经常问关于 GitHub issue 的事。

6:43 它可以用网络抓取功能，获取 issue 并查找

6:47 相关上下文。

6:48 这非常棒。

6:51 我每周一的例会上都会这样做，

6:54 问“这周我发布了什么？”

6:57 Quad Code 会查看日志，

6:58 它知道我的用户名，

6:59 然后给我一份漂亮的发布内容清单。

7:02 我只需复制粘贴到文档里即可。

7:04 这是第一个技巧。

7:10 对于没用过 Quad Code 的人来说，

7:12 如果你第一次给团队演示，

7:15 我们非常推荐从代码库 Q&A 开始。

7:17 不要一开始就用复杂工具，

7:18 不要先修改代码，

7:19 先从问代码库的问题开始。

7:21 这能教会大家怎样提问。

7:23 也会慢慢让他们了解 Quad Code 的边界，

7:25 什么是它能做的，

7:27 什么是它能胜任的，

7:28 哪些需要你多引导一点。

7:31 哪些可以一次性完成？

7:32 哪些需要两步、三步操作？

7:33 哪些得用交互模式 REPL 来操作？

7:35 当你对 Q&A 比较熟悉后，可以深入做

7:37 代码编辑，

7:40 这是下一步。

7:42 用 LM 代理方式的酷点在于，

7:44 你给它工具，它能神奇地

7:46 自己学会使用工具。

7:51 这非常神奇。

7:53 它能自己找到用法。

7:54 我们在 Quad Code 中给它的工具很少，

7:56 数量不多。

7:58 It's not a lot.

8:00 所以它有一个编辑文件的工具。

8:02 它有一个运行bash命令的工具。

8:03 它有一个搜索文件的工具。

8:05 它会把这些工具串联起来，探索代码

8:08 头脑风暴，然后最终进行编辑。

8:12 你不必特别提示它使用这个

8:14 工具和那个工具，以及其他工具。

8:15 你只需要说，做这件事，它就会想办法去做

8:17 它。

8:17 它会以正确的方式把它们串联起来。

8:19 这对quad代码来说很有意义。

8:20 有很多使用它的方法。

8:26 我有时喜欢做的一件事是，在让

8:30 quad开始写代码之前，我会让它先头脑风暴

8:33 一下或者制定计划。

8:35 这是我们强烈推荐的。

8:37 我有时看到的是，人们会用quad代码

8:39 让它实现一个庞大的3000行功能。

8:43 有时它第一次就能准确做到。

8:45 但有时它做出来的东西

8:47 根本不是你想要的。

8:49 而得到你想要结果的最简单方法是让它

8:51 先思考。

8:55 所以先头脑风暴想法。

8:56 制定计划，告诉我。

8:57 在写代码之前先征求我的同意。

8:58 你不必使用计划模式。

9:01 你也不必用任何特殊工具来做这件事。

9:03 你只要问Claude就行。

9:05 它就知道该怎么做。

9:06 所以只说，写代码之前先制定计划。

9:07 就这样。

9:10 这也是——我想用它想想这个提交

9:13 推送的步骤。

9:15 这是我常用的一种方法。

9:15 没什么特别，但Claude足够聪明能理解它。

9:18 它会做一次提交。

9:20 它会推送到分支，新建分支，然后在GitHub上

9:21 为我创建一个拉取请求。

9:22 你不必做任何解释。

9:23 它会自己查看代码。

9:25 它会查看历史记录。

9:26 它会自己浏览Git日志，搞清楚提交格式等细节。

9:27 然后它会正确地提交并推送。

9:29 我们没有用系统提示让它这么做。

9:30 它就是知道怎么做。

9:32 模型很强大。

9:36 当你逐渐进阶时，你会想要

9:39 开始接入你团队的工具。

9:40 这正是Claude代码真正闪光的地方。

9:41 通常有两类工具。

9:46 一类是bash工具。

9:48 举个例子，我编了个叫burly的CLI工具。

9:51 它并不真实存在。

9:53 但你可以说，使用这个CLI做某事。

9:54 你可以告诉Claude代码这个工具。

9:57 你可以让它用--help参数来了解用法。

9:57 这样很高效。

9:59 如果你常用这个工具，也可以把信息

10:01 存进你的Claude MD，我们稍后会讲。

10:02 这样Claude可以跨会话记住它。

10:06 这是我们在在非营利组织中常用的模式，

10:07 外部客户也用得很多。

10:09 MCP工具同理。

10:11 Claude代码可以用bash工具。

10:13 也可以用MCP工具。

10:15 只要告诉它工具信息。

10:18 然后你可以添加MCP工具。

10:21 告诉它怎么用。

10:23 它就会开始用它。

10:24 这非常强大，因为当你开始用

10:25 新代码库时，你可以把所有

10:27 团队已有的工具给它。

10:29 Claude代码可以代你使用它们。

10:30 有几种常见工作流程。

10:33 之前说过，这个就是一种。

10:36 先做点探索，制定计划，然后开始写代码前征求我确认。

10:39 右边的其他工具非常强大。

10:40 当Claude有办法检查自己的工作——

10:43 比如写单元测试，或用puppeteer截屏，或截iOS模拟器屏幕——

10:45 它就可以迭代改进。

10:53 这很棒，比如你给它一个mock示例，告诉它做这个网页UI，

10:55 它会做得蛮好。

10:57 但迭代两到三次后，通常能做到几乎完美。

11:01 关键是给它一些工具，能用于反馈，检查结果。

11:02 基于此，它能自主迭代，

11:05 你会得到更好的成果。

11:07 无论你的领域是什么，是单元测试，还是

11:11 in puppeteer, or screenshotting the iOS simulator--

11:14 then it can iterate.

11:16 And this is incredible, because if you give it, for

11:18 example, a mock, and you say, build this web UI, it'll

11:20 get it pretty good.

11:22 But if you iterate two or three times, often it gets it

11:25 almost perfect.

11:26 So the trick is give it some sort of tool that it can use

11:29 for feedback to check its work.

11:31 And then based on that, it will iterate by itself, and you're

11:33 going to get a much better result.

11:35 So whatever your domain is, if it's unit test, or

11:37 集成测试，或截图，或应用程序，或网页，或

11:40 任何东西，只要给它一种查看结果的方式，它就会

11:43 迭代并变得更好。

11:44 所以这就是接下来的步骤。

11:50 教Claude如何使用你的工具，找出

11:52 正确的工作流程。

11:53 如果你想让Claude跳进代码，或者想让它

11:55 稍微清理一下，做个计划，如果你想让它

11:57 迭代，基本上要有些这样的意识。

11:59 这样你就知道如何提示Claude去做你想要的事。

12:02 当你深入到工具之外时，你要开始给

12:09 Claude更多的上下文。

12:10 上下文越多，决策越聪明，

12:12 因为作为在代码库中工作的工程师，你脑海中有

12:14 大量关于系统的上下文，所有

12:17 历史以及其他信息。

12:19 所以有不同的方法可以把这些给Claude。

12:22 随着你给Claude更多上下文，它表现会更好。

12:25 实现方式有很多种。

12:27 最简单的是我们称之为Claude MD的方式。

12:30 Claude.md是特殊的文件名。

12:33 最简单的放置位置是项目根目录。

12:36 也就是你启动Claude的那个目录。

12:38 在那放一个Claude MD文件。

12:40 它会在每次会话开始时自动被读入上下文中。

12:42 本质上，第一次用户交互会包含

12:43 Claude MD内容。

12:45 你也可以有一个本地的Claude MD，这个文件

12:48 通常不会提交到版本控制。

12:52 所以Claude MD文件，应提交到版本控制，与团队共享，

12:53 这样写一次可以共享给团队。

12:56 这个本地文件不提交。

12:58 仅供你自己使用。

13:00 你放在Claude MD里的内容，

13:01 比如常用bash命令，常用MCP工具，架构决定，

13:03 重要文件，任何你通常需要知道以便在此代码库工作

13:05 的信息。

13:09 尽量保持简短，因为太长会消耗大量上下文，

13:12 且通常没那么有用。

13:14 所以尽量保持尽可能简洁。

13:15 举例来说，在我们的代码库中，我们有常用bash命令，

13:18 有代码风格指南，有几个核心文件，

13:20 类似这样的东西。

13:21 其他的Claude MD文件可以放在其他

13:23 嵌套的子目录中，Claude会根据需求加载它们。

13:27 这些Claude MD文件会被自动加载。

13:30 你也可以在嵌套目录放置Claude MD文件，

13:32 当Claude在那些目录工作时，也会自动加载。

13:34 当然，如果你是公司，可能需要一个

13:37 跨所有代码库共享的Claude MD，由你代表用户管理，

13:39 可以放在企业根目录，它会自动加载。

13:41 有很多种方式可以加载上下文。

13:42 我为了表达可以用的各种方式，组装这段内容时

13:45 碰到了不少困难。

13:48 但Claude MD会自动加载。

13:51 你也可以使用快速命令（flash commands）。

13:54 这是.Dot Claude的快速命令。

13:56 它可以在你的主目录，也可以提交到项目中。

13:58 这是用来执行快速命令的。

14:00 这里有几个我们在Claude代码里用到的快速命令示例。

14:01 比如说，如果你在Claude代码库里看到问题被自动标记，

14:07 实际上就是这个工作流在运行。

14:09 它会给GitHub Issues打标签。

14:11 我们有个GitHub动作在运行，就是今天早上提到的，

14:12 Claude代码会执行这个快速命令，

14:15 它就是个快速命令。

14:16 执行后会给问题自动打标签，省去人工标记的时间。

14:18 节省了我们很多时间。

14:21 当然，你也可以添加引用文件，把它们纳入上下文。

14:21 正如我之前说的，嵌套目录里的Claude MD，

14:24 当Claude在那目录工作时会被加载。

14:29 所以给Claude更多上下文。

14:31 花时间调优上下文绝对值得。

14:34 你可以用提示改进器来优化。

14:37 考虑这个上下文是给谁的。

14:38 你是想每次都加载，还是按需加载，

14:40 想跟团队共享，还是个人使用，都去调优。

14:42 如果做得好，会显著提升性能。

14:46 随着你更深入，你会想更细致考虑

14:47 不同层级和方式来加载所有内容。

14:50 have to.

14:50 It just saves us a bunch of time.

14:53 And of course, you can add mention files to pull them into

14:55 context.

14:56 And like I said before, Claude MD is in a nested directory,

14:59 get pulled in when Claude works in that directory.

15:01 So give Claude more context.

15:08 And it's definitely worth taking the time to tune context.

15:11 You can run it through a prompt improver.

15:12 Consider who the context is for.

15:15 If you want to pull it in every time, if you want to pull

15:17 it in on demand, if you want to share it with a team, if it's

15:20 a personal preference, definitely take the time to tune

15:22 it.

15:22 This will improve performance dramatically if you do it

15:25 right.

15:25 As you get more advanced, you're going to want to think

15:33 about this a little bit more of this kind of hierarchy of

15:35 different ways to pull in everything.

15:36 不仅仅是Claude MD，也包括配置等。

15:42 关于Claude的一切，你都可以以这种方式引入

15:43 分层的方式。

15:45 所以项目是针对你的Git仓库的，你可以

15:48 提交它，或者只为自己使用。

15:51 你也可以有全局配置，应用于你所有的

15:53 项目，或者你可以有企业策略。

15:56 这本质上是一个你为

15:58 所有员工统一推行的全局配置，团队中的每个人

16:00 都会自动应用。

16:02 这张幻灯片信息量很大，但

16:05 关键是它适用于很多情况。

16:07 你可以用它来管理斜杠命令。

16:08 你也可以用它来管理权限。

16:10 比如说，如果你有一个批量命令，想

16:12 让所有员工都能运行，比如让所有员工使用

16:15 这个测试命令。

16:17 你实际上可以把它直接写进企业

16:19 策略文件里。

16:21 然后任何员工运行这个命令时，会

16:23 自动获得批准，非常方便。

16:26 你还可以用它来阻止命令。

16:27 比如说，有个URL绝对不能

16:29 被抓取。

16:31 只要把它添加到配置里。

16:32 这样员工就不能覆盖它了。

16:35 那个URL就永远无法被抓取。

16:38 这既方便解放员工，也

16:40 有助于保护你的代码库安全。

16:43 MCP服务器也一样。

16:45 有个MCP JSON文件，提交到代码库。

16:47 这样每当有人在代码库中运行quad代码时，

16:50 都会提示安装MCP服务器并与团队

16:53 共享。

16:54 如果你不确定用哪种，这其实是一个

17:00 相当复杂的矩阵。

17:01 因为我们支持很多东西。

17:03 工程师的工作流非常灵活。

17:04 每个公司都不一样。

17:05 所以我们想支持所有情况。

17:07 如果你不知道如何开始，我建议

17:09 先从共享项目上下文开始。

17:12 你写一次，

17:13 然后与你团队所有人分享。

17:14 这样就会产生网络效应，有人做点工作，

17:17 团队所有人都能受益。

17:19 Quad内置了很多管理工具。

17:26 比如说，运行/memory，你可以看到所有

17:29 被引入的不同记忆文件。

17:31 比如我有企业策略。

17:33 我有用户记忆。

17:34 我有项目的quad MD。

17:36 还有可能有嵌套的quad MD。

17:38 只会在某些目录下被引入。

17:42 同理，运行/memory时，你可以编辑

17:45 特定的记忆文件。

17:46 当你输入井号记住某事时，你可以

17:48 选择想要存到哪个记忆里。

17:49 这就是下一步。

17:55 花时间配置Quad MD、MCP服务器，配置

18:00 团队使用的所有东西，这样你只用配置一次，

18:02 然后与所有人共享。

18:05 一个例子是在我们的人类学应用仓库里。

18:11 这里是我们所有网页和应用代码所在的仓库。

18:16 这里有个puppeteer MCP服务器，我们共享给

18:18 团队。

18:19 并且有一个MCP JSON文件被提交。

18:21 所以任何在该仓库工作的工程师都能用puppeteer

18:24 来自动化pilot、测试和截图

18:26 并进行迭代，这样每个工程师

18:28 就不用自己安装了。

18:29 这是关于专业技巧的讲解。

18:35 我想稍作插曲，讲讲一些

18:38 大家可能不知道的常用快捷键。

18:40 为终端开发很难，

18:41 但也很有趣。

18:43 这感觉像是在重新发现一种新的设计语言。

18:46 但终端有个特点，就是极简。

18:49 所以有时很难发现这些快捷键。

18:52 这里是个快速参考表。

18:54 任何时候你都可以按Shift+Tab接受编辑，

18:58 这会切换到自动接受编辑模式。

19:00 bash命令不需要批准，编辑则

19:02 自动接受，你随时可以让Quad

19:04 稍后撤销它们。

19:06 例如，当我知道Quad方向正确，

19:09 或者写单元测试并迭代时，我

19:11 通常会切换到自动接受模式。

19:13 这样我不用每个编辑都确认。

19:17 任何时候你想让Claude记住某事，比如

19:19 它没正确使用某工具，你想让它

19:22 以后正确使用，只需输入井号，告诉它要记住什么，

19:24 它就会记住。

19:26 它会自动将其纳入Quad MD。

19:26 如果你想切换到bash模式，执行一个

19:30 bash命令，可以按感叹号并输入

19:32 命令。

19:35 命令会在本地执行，同时也会进入

19:35 上下文窗口，Quad下一次能看到。

19:38 这对长时间运行的命令很有用。

19:42 如果你

19:44 确切知道你想做什么，或者任何你

19:46 想要进入上下文的命令。

19:47 Quad 会看到命令和输出。

19:51 你可以添加上千个文件夹。

19:53 任何时候你都可以按 Esc 停止 Quad 的操作，

19:57 无论 Quad 在做什么，你总能看到我们是否按了

19:58 Esc 键。

19:59 这不会破坏会话。

20:00 也不会弄乱任何东西。

20:02 比如 Quad 正在编辑文件。

20:04 我按下 Esc。

20:04 告诉它要做不同的操作。

20:06 或者它建议了一个 20 行的修改，我觉得

20:09 19 行完美，但有一行你应该

20:11 改动。

20:12 我按下 Esc。

20:12 告诉它。

20:13 然后让我告诉它重新做一次。

20:17 你可以按两次 Esc 回到历史状态。

20:20 结束会话后，你可以用 resume 来恢复那个会话，

20:22 或者用 --continue。

20:25 任何时候如果想看更多输出，按

20:28 控制键（controller）。

20:30 它会显示完整输出，跟 Quad 在上下文窗口里看到的一样。

20:31 And that'll show you the entire output, the same thing

20:34 that Quad sees in its context window.

20:35 接下来我想讲的是 Quadcode SDK。

20:44 我们之前顶部提过这个。

20:46 紧接着这之后，说它在做一个会话，我觉得，

20:49 就在走廊对面。

20:50 他会深入讲解 SDK。

20:52 如果你没试过这个，或者在 Quad 里用过 -P 参数，

20:55 这就是 SDK。

20:58 我们这几周不断增加功能，让它更强大。

21:00 few weeks to make it even better.

21:03 所以你可以基于它开发。

21:05 能做很多酷炫的事。

21:06 这正是 Quadcode 使用的东西。

21:08 完全相同的 SDK。

21:11 比如，你可以用 Quad -P。

21:13 这是命令行 SDK。

21:15 你可以传入提示（prompt）。

21:18 可以传入允许的工具，甚至可以包括

21:20 特定的 bash 命令。

21:22 可以告诉它你想要什么格式。

21:24 比如 JSON，或者流式 JSON，

21:27 方便你做处理。

21:29 这对构建基于它的应用很棒。

21:31 我们在持续集成里一直用它，也用来应急响应，

21:34 各种流水线里都用它。

21:37 非常方便。

21:38 把它当作一个类 Unix 工具就行。

21:39 你给它一个提示，它给你 JSON。

21:41 你可以随意使用它。

21:42 可以向它传输输入，也可以从它传出输出。

21:44 管道功能也很酷。

21:50 比如用 Git status，然后把结果传给它，

21:53 再用 JQ 选择结果。

21:56 组合无穷无尽。

21:58 这是个新思路。

21:59 像一个超级智能的 Unix 工具。

22:02 我觉得我们才刚刚开始摸索用法。

22:04 我们还在探索中。

22:07 你可以从 GCP 存储桶读取巨量日志，传给它，

22:11 让 Cloud 发现日志里有什么

22:12 有趣的内容。

22:14 你可以获取 Century CLI 的数据。

22:18 也同样可以输入数据，让 Cloud 做些处理。

22:20 最后一个，也是我们见过最高级的用例，

22:29 我算是个 Quad 普通用户。

22:31 通常我只开一个 Quad。

22:34 有时会开多个终端标签页，

22:36 运行几个不同代码库。

22:38 看过 Enthropic 的高手们，

22:41 几乎都开 SSH 会话。

22:44 用 Tmux 隧道连接 Quad 会话。

22:47 会有多个同一仓库的检出，

22:50 这样他们可以并行运行多个 Quad。

22:53 或者用 Git worktree 做隔离。

22:55 我们正在积极优化使用体验。

22:56 但现在这些就是在 Cloud 里并行工作的思路。

22:59 你可以开任意多会话。

23:03 可以同时完成很多事情。

23:05 就是这样。

23:08 我还想留点时间回答问题。

23:10 我想这是我最后一页幻灯片了。

23:14 如果大家有问题，

23:17 两边都有麦克风。

23:19 我们很乐意回答任何问题。

23:21 [掌声]

23:23 我明白了。

23:27 嗨，Boris。

23:31 感谢你开发 Cloud Code。

23:42 我想知道对你来说，开发中最难的部分是什么？

23:50 我觉得有很多棘手的地方。

23:51 我觉得最难的一部分是

23:54 我们为了让 bash 命令安全做的工作。

23:59 I think there's a lot of tricky parts.

24:03 I think one part that is especially tricky

24:06 is the things that we do to make bash commands safe.

24:10 Bash 本质上是相当危险的。

24:11 它可能以意想不到的方式改变系统状态。

24:14 但与此同时，如果你必须手动

24:18 批准每一条 bash 命令，

24:20 作为工程师会非常烦人。

24:21 而且你几乎无法高效工作，

24:23 因为你必须不断地验证每条命令。

24:25 怎样安全地进行操作，

24:28 并能适应不同代码库的规模，

24:30 很有挑战，因为不是每个人

24:32 都在 Docker 容器中运行代码。

24:35 我们最终确定的是，

24:37 某些命令是只读的。

24:39 我们会做一些静态分析，

24:40 来判断哪些命令可以安全组合，

24:43 以安全的方式。

24:44 然后我们有一个相当复杂的分级权限

24:47 系统，可以在不同层级允许或屏蔽命令，

24:49 实现白名单和无阻断命令管理。

24:53 你好，Boris。

24:54 你提到给 Cloud Code 传递一个镜像，

24:57 这让我想知道是否有我不知道的多模态

25:00 功能。

25:02 那是——你只是把它指向

25:03 文件系统上的镜像还是别的？

25:05 是的，Cloud Code 完全支持多模态，

25:08 从一开始就是这样的。

25:09 它在终端里，所以不太容易发现。

25:12 但你可以拖拽一个镜像进来，

25:15 这会奏效。

25:16 你也可以提供一个文件路径，同样有效。

25:18 你还能复制粘贴镜像，这也可以。

25:23 我经常这样用，比如我有个模拟文件，

25:26 我就拖拽进来。

25:28 告诉它去实现，

25:29 给它多级服务器以便迭代测试。

25:32 是的，这完全自动化。

25:33 嘿，你为什么做一个 CLI 工具而不是 IDE？

25:43 这是个好问题。

25:44 我觉得有两个原因。

25:45 一是我们开始时是为 Anthropic 做的，

25:48 Anthropic 的人用各种不同的 IDE。

25:51 有的人用 VS Code，有的人用 said、Xcode、Vim 或 Emacs。

25:58 做一个适合所有人的工具很难。

26:00 终端就是最通用的选择。

26:03 第二， Anthropic 近距离见证了模型的快速进步，

26:10 越来越好。

26:10 我觉得很可能今年年底，

26:13 人们就不再使用 IDE 了。

26:15 所以我们想为这个未来做好准备。

26:17 想避免在 UI 和其他层面过度投入，

26:21 因为模型发展速度快，

26:24 这些工作很快就可能没用了。

26:26 是的。

26:32 你使用 Cloud Code 做机器学习建模有多少？我不知道这段录音有没有开。

26:35 你使用 Cloud Code 做机器学习建模有多少？

26:39 还有那种 AutoML 的体验。

26:40 我很好奇到目前为止的体验如何。

26:44 问题是，我们用 Cloud Code 做机器学习和建模的频率如何？

26:48 其实我们用得挺多的。

26:50 Anthropic 的工程师和研究员每天都用 Cloud Code，

26:51 大约 80% 的技术人员每天都用。

26:56 希望你能从产品中看到，我们投入了很多爱和自用。

27:00 这包括用笔记本工具编辑和运行笔记本的研究员。

27:01 好的，非常酷。

27:04 谢谢。

27:04 好了，我想就这些。

27:06 谢谢。

27:10 [掌声]

27:12 [音乐播放]

27:13 [音乐播放]

27:16 All right, I think that's it.

27:17 Thanks.

27:18 [APPLAUSE]

27:20 [MUSIC PLAYING]

27:48 [MUSIC PLAYING]

0:00 (upbeat music)

0:02 - Hello.

0:05 Everyone, I'm Boris.

0:08 I'm a member of technical staff here at Anthropic,

0:11 and I created quad code.

0:13 And here to talk to you a little bit

0:15 about some practical tips and tricks for using quad code.

0:18 It's gonna be very practical.

0:20 I'm not gonna go too much into the history

0:22 or the theory or anything like this.

0:24 And yeah, before we start, actually,

0:27 can we get a quick show of hands?

0:28 Who has used quad code before?

0:30 Yeah, all right, that's what we like to see.

0:34 For everyone that didn't raise your hands,

0:36 I know you're not supposed to do this

0:38 while people are talking,

0:38 but if you can open your laptop and type this,

0:42 and this will help you install quad code

0:48 just so you can follow along for the rest of the talk.

0:55 All you need is no JS if you have a dish at work.

0:58 (mumbling)

1:04 Yeah, well, you don't have to follow along,

1:06 but if you don't have it yet,

1:08 this is your chance to install it so you can follow along.

1:11 So what is quad code?

1:15 Quad code is a new kind of AI assistant,

1:19 and there's been different generations

1:20 of AI assistants for coding.

1:23 Most of them have been about completing line at a time,

1:26 completing a few lines of code at a time.

1:28 Quad code is not for that, it's fully agentic.

1:31 So it's meant for building features,

1:33 for writing entire functions, entire files,

1:37 fixing entire bugs at the same time.

1:39 And what's kind of cool about quad code

1:42 is it works with all of your tools,

1:44 and you don't have to change out your workflow,

1:46 you don't have to swap everything to start using it.

1:48 So whatever IDE you use if you use VS code

1:51 or if you use Xcode or if you use JetBrains IDEs,

1:56 there's some people out on Tropic

1:57 that you can't pry them from their cold, dead hands,

2:00 but they use quad code because quad code works

2:03 with every single ID, every terminal out there.

2:05 It'll work locally over remote SSH,

2:08 over Tmox, whatever environment you're in, you can run it.

2:12 It's general purpose, and this is something

2:17 where if you haven't used these kind of free form

2:20 coding assistants in the past,

2:22 it can be kind of hard to figure out how to get started

2:24 'cause you open it up and you just see a prompt bar

2:26 and you might wonder like, what do I do with this?

2:28 What do I type in?

2:29 It's a power tool, so you can use it for a lot of things,

2:32 but also because it can do so much,

2:34 we don't try to guide you towards a particular workflow

2:36 because really you should be able to use it

2:38 however you want as an engineer.

2:40 As you open up quad code for the first time,

2:47 there's a few things that we recommend doing

2:49 to get your environment set up.

2:51 And these are pretty straightforward.

2:52 So run terminal set up, this will give you shift enter

2:54 for new lines so you don't have to do

2:56 like back slashes to enter new lines.

2:58 This is, you know, it makes it a little bit nicer to use.

3:01 Do slash theme to set light mode or dark mode

3:04 or delta nice themes.

3:05 You can do slash install GitHub app.

3:09 So today when we announced a GitHub app

3:12 where you can add, mention a cloud

3:15 on any GitHub issue or a pull request.

3:17 So to install it, just run this command in your terminal.

3:20 You can customize the set of allowed tools that you can use

3:24 so you're not prompted for it every time.

3:25 This is pretty convenient.

3:27 For stuff that I'm prompted about a bunch,

3:28 I'll definitely customize it in this way

3:30 so I don't have to accept it every time.

3:32 And something that I actually do is for a lot of my prompts,

3:35 I won't hand type them into a quad code.

3:37 If you're on macOS, you can go into your system settings

3:40 under accessibility as dictation and you can enable it.

3:44 And so something I do is you just hit

3:45 like the dictation key twice and you can just speak your prompt.

3:49 And it helps a lot to have specific prompts.

3:52 So this is actually pretty awesome.

3:53 You can just talk to quad code and like you would

3:55 another engineer and you don't have to type a lot of code.

3:58 So when you're starting out with quad code,

4:04 it's so freeform and it can do everything.

4:06 What do you start with?

4:07 The thing I recommend above everything else

4:09 is starting with code base Q&A.

4:11 So just asking your question, asking questions

4:13 to your code base.

4:15 This is something that we teach new hires at Anthropic.

4:18 So on the first day in technical onboarding,

4:20 you learn about quad code, you download it, you get it set up.

4:23 And then you immediately start asking questions

4:25 about the code base.

4:26 And in the past, when you're doing technical onboarding,

4:28 it's something that taxes the team a lot, right?

4:31 You have to ask other engineers on the team questions.

4:34 You have to look around the code and this takes a while.

4:36 You have to figure out how to use the tools

4:37 that this takes a long time.

4:39 With quad code, you can just ask quad code

4:42 and it'll explore the code base.

4:43 It'll answer these kind of questions.

4:45 And so at Anthropic onboarding, you

4:47 should take about two or three weeks for technical hires.

4:49 It's now about two or three days.

4:50 What's also kind of cool about Q&A is we

4:56 don't do any sort of indexing.

4:57 So there's no remote database with your code.

4:59 We don't upload it anywhere.

5:00 Your code stays vocal.

5:02 We do not train generative models on the code.

5:05 So it's there.

5:06 You control it.

5:06 There's no indices or anything like this.

5:09 And what that means is also there's no setup.

5:11 So you start quad, you download it.

5:13 You start it, there's no indexing.

5:14 You don't have to wait.

5:15 You can just use it right away.

5:16 This is a technical talk.

5:20 So I'm going to show some very specific prompts

5:22 and very specific code samples that you can use

5:24 and hopefully improve and up level your quad code experience.

5:28 So some kind of questions that you can ask is,

5:30 how is this particular piece of code used

5:33 or how do I instantiate this thing?

5:34 And quad code, it won't just do a text search

5:36 and try to answer this.

5:37 It'll often go a level deeper and it'll try to find examples

5:41 of how is this class instantiated?

5:42 How is it used?

5:44 And it'll give you a much deeper answer.

5:45 So something that you would get out of a weekier documentation

5:48 instead of just like Command F.

5:52 Something that I do a lot also is ask it about Git history.

5:55 So for example, why does this function have 15 arguments?

5:59 And why are the arguments named this weird way?

6:01 And this is something I've added in all of our code bases.

6:03 You have some function like this or some class like this.

6:07 And quad code can look through Git history.

6:09 And it'll look to figure out how did these arguments get

6:11 introduced and who introduced them and what was the situation?

6:14 What are the issues that those commits linked to?

6:16 And it'll look through all this and summarize it.

6:18 And you don't have to tell it that in all this detail,

6:21 you just ask it.

6:22 So you say, look through Git history.

6:23 And I don't know who to do this.

6:26 The reason it knows it, by the way,

6:27 is not because we prompted it to.

6:28 There's nothing in the system prompt about looking

6:30 through Git history.

6:32 It knows it because the model is awesome.

6:34 And if you tell it to use Git, it'll know how to use Git.

6:37 So we're lucky to be building on such a good model.

6:40 I often ask about GitHub issues.

6:43 So it can use web fetch and it can fetch issues and look up

6:47 contexts on issues, too.

6:48 And this is pretty awesome.

6:51 And this is something that I do every single Monday in our

6:54 weekly stand-up is I ask, what did I ship this week?

6:57 And quad code looks to the log.

6:58 It knows my username.

6:59 And it'll just give me a nice readout of everything I shipped.

7:02 And I'll just copy and paste that into a docs.

7:04 So yeah, that's tip number one.

7:10 For people that have not used quad code before, if you're

7:12 just showing it to someone for the first time onboarding your

7:15 team, the thing we definitely recommend is start with code

7:17 base Q&A.

7:18 Don't start by using fancy tools.

7:19 Don't start by editing code.

7:21 Just start by asking questions about the code base.

7:23 And that'll teach people how to prompt.

7:25 And it'll start teaching them this boundary of like, what

7:27 can quad code do?

7:28 What is it capable of versus what do you need to hold this

7:31 hand with a little bit more?

7:32 What can be one-shotted?

7:33 What can be two-shotted, three-shotted?

7:35 What do you need to use interactive mode for in a

7:37 REPL?

7:40 Once you're pretty comfortable with Q&A, you can dive into

7:42 editing code.

7:44 This is the next thing.

7:46 And the cool thing about any sort of agentic, like using

7:51 an LM in an agentic way is you give it tools, and it's just

7:53 like magical.

7:54 It figures out how to use the tools.

7:56 And with quad code, we give it a pretty small set of tools.

7:58 It's not a lot.

8:00 And so it has a tool to edit files.

8:02 It has a tool to run bash commands.

8:03 It has a tool to search files.

8:05 And it'll string these together to explore the code

8:08 brainstorm, and then finally make edits.

8:12 And you don't have to prompt it specifically to use this

8:14 tool and this tool and this tool.

8:15 You just say, do this thing, and it'll figure out how to do

8:17 it.

8:17 It'll string it together in the right way.

8:19 That makes sense for quad code.

8:20 There's a lot of ways to use this.

8:26 Something I like to do sometimes is before having

8:30 quad jump in to write code, I'll ask it to brainstorm a

8:33 little bit or make a plan.

8:35 This is something we highly recommend.

8:37 And something I see sometimes is people, they take quad code

8:39 and they ask it, hey, implement this enormous 3,000

8:43 wine feature.

8:45 And sometimes it gets this right on the first shot.

8:47 But sometimes what happens is the thing that it builds is

8:49 not at all the thing that you want it.

8:51 And the easiest way to get the result you want is ask it to

8:55 think first.

8:56 So brainstorm ideas.

8:57 Make a plan, run it by me.

8:58 Ask for approval before you write code.

9:01 And you don't have to use plan mode.

9:03 You don't have to use any special tools to do this.

9:05 All you have to do is ask Claude.

9:06 And it'll know to do this.

9:07 So just say, before you write code, make a plan.

9:10 That's it.

9:13 This is also-- I want to think with this one, this commit

9:15 push figure.

9:15 This is a really common incantation that I use.

9:18 There's nothing special about it, but Claude is kind of

9:20 smart enough to interpret this.

9:21 So it'll make a commit.

9:22 It'll push it to the branch, make a branch, and then make

9:23 a pull request from me on GitHub.

9:25 You don't have to explain anything.

9:26 It'll look through the code.

9:27 It'll look through the history.

9:29 It'll look through the Git log by itself to figure out the

9:30 commit format and all of the stuff.

9:32 And it'll make the commit and push it the right way.

9:36 Again, we're not system prompting to do this.

9:39 It just knows how to do this.

9:40 The model is good.

9:41 As you get a little bit more advanced, you're going to want

9:46 to start to plug in your team's tools.

9:48 And this is where Claude code starts to really shine.

9:51 And there's generally two kinds of tools.

9:53 So one is bash tools.

9:54 And an example of this, I just made up this, like, burly

9:57 CLI.

9:57 This isn't a real thing.

9:59 But you can say, use the CLI to do something.

10:01 And you can tell Claude code about this.

10:02 And you can tell it to use, for example, dash, dash help

10:06 to figure out how to use it.

10:07 And this is efficient.

10:09 If you friend yourself using it a lot, you can also dump

10:11 this into your Claude MD, which we'll talk about in a bit.

10:13 So Claude can remember this across sessions.

10:15 But this is a common pattern we follow at on-profit, and we

10:18 see external customers use too.

10:21 And same thing with MCP.

10:23 Claude code can use bash tools.

10:24 It can use MCP tools.

10:25 So just tell it about the tools.

10:27 And you can add the MCP tool.

10:29 And you can tell it how to use it.

10:30 And it'll just start using it.

10:33 And this is extremely powerful, because when you start to use

10:36 code on a new code base, you can just give it all of your

10:39 tools.

10:40 All the tools your team already uses for this code base.

10:43 And Claude code can use it on your behalf.

10:45 There's a few common workflows.

10:53 And this is the one that I talked about already.

10:55 So kind of do a little bit of exploration, do a little bit

10:57 of planning, and ask me for confirmation before you start

11:01 to write code.

11:02 These other tools on the right are extremely powerful.

11:05 When Claude has some way to check its work--

11:07 so for example, by writing unit tests, or screenshotting

11:11 in puppeteer, or screenshotting the iOS simulator--

11:14 then it can iterate.

11:16 And this is incredible, because if you give it, for

11:18 example, a mock, and you say, build this web UI, it'll

11:20 get it pretty good.

11:22 But if you iterate two or three times, often it gets it

11:25 almost perfect.

11:26 So the trick is give it some sort of tool that it can use

11:29 for feedback to check its work.

11:31 And then based on that, it will iterate by itself, and you're

11:33 going to get a much better result.

11:35 So whatever your domain is, if it's unit test, or

11:37 integration test, or screenshots, or apps, or web, or

11:40 anything, just give it a way to see its result, and it'll

11:43 iterate and get better.

11:44 So these are the next steps.

11:50 Teach Claude how to use your tools, and figure out the

11:52 right work flow.

11:53 If you want Claude to jump in a code, if you want it to

11:55 rinse from a little bit, make a plan, if you want it to

11:57 iterate, kind of have some sense of that.

11:59 So you know how to prompt Claude to do what you want.

12:02 As you go deeper, beyond tools, you want to start to give

12:09 Claude more context.

12:10 And the more context, the smarter the decisions will be,

12:12 because as an engineer working in a code base, you have a

12:14 ton of context in your head about your systems, and all

12:17 the history, and everything else.

12:19 So there's different ways to give this to Claude.

12:22 And as you give Claude more context, it'll do better.

12:25 There's different ways to do this.

12:27 The simplest one is what we call Claude MD.

12:30 And Claude.md is the special file name.

12:33 The simplest place to put it is in the project root.

12:36 So the same director, you start Claude in.

12:38 Put a Claude MD in there.

12:40 And that'll get automatically read into context at the

12:42 start of every session.

12:43 And essentially, the first user turn will include the

12:45 Claude MD.

12:48 You can also have a local Claude MD, and this one, you

12:52 don't usually check into source control.

12:53 So Claude MD, you should check into source control, share

12:56 with your team so that you can write it once and share it

12:58 with your team.

13:00 This one, you don't check in.

13:01 It's just for you.

13:03 The kinds of things you put in Claude MD, it's like

13:05 common bash commands, common MCP tools, architectural

13:09 decisions, important files, anything that you would kind

13:12 of typically need to know in order to work in this code

13:14 base.

13:15 Try to keep it pretty short, because if it gets too long,

13:18 it's just going to use up a bunch of context.

13:20 And it's usually not that useful.

13:21 So just try to keep it as short as you can.

13:23 And for example, in our code base, we have common bash

13:27 commands, we have a style guide, we have a few core files,

13:30 kind of things like that.

13:32 All the other Claude MDs, you can put them in other

13:34 nested child directories, and Claude will pull them in on

13:37 demand.

13:39 So these are the Claude MDs that will get pulled in

13:41 automatically.

13:42 But then also, you can put in Claude MDs in nested

13:45 directories, and those will get automatically pulled when

13:48 Claude works in those directories.

13:51 And of course, if you're a company, maybe you want a

13:54 Claude MD that's shared across all the different code

13:56 bases, and you want to manage it on behalf of your users,

13:58 and you can put it in your enterprise route, and that

14:00 will get pulled in automatically.

14:01 There's a ton of ways to pull in context.

14:07 I actually had a lot of trouble putting this fight

14:09 together, just to communicate the breadth of ways you can do

14:11 this.

14:12 But Claude MD is pulled in automatically.

14:15 You can also use flash commands.

14:16 So this is dot Claude's flash commands.

14:18 And this can be in your home directory, or it can be checked

14:21 into your project.

14:21 And this is for flash commands.

14:24 And over here, we have a few examples of the flash commands

14:29 that we have in Claude code itself.

14:31 And so for example, if you're in the Claude code repo and

14:34 you see issues getting labeled, that's actually this

14:37 workflow running here.

14:38 It's labeled GitHub Issues.

14:40 And we have a GitHub action running, the same one we

14:42 talked about this morning, where Claude code will run this

14:46 command, and it's just a splash command.

14:47 It'll run, and it'll label the issues, so humans don't

14:50 have to.

14:50 It just saves us a bunch of time.

14:53 And of course, you can add mention files to pull them into

14:55 context.

14:56 And like I said before, Claude MD is in a nested directory,

14:59 get pulled in when Claude works in that directory.

15:01 So give Claude more context.

15:08 And it's definitely worth taking the time to tune context.

15:11 You can run it through a prompt improver.

15:12 Consider who the context is for.

15:15 If you want to pull it in every time, if you want to pull

15:17 it in on demand, if you want to share it with a team, if it's

15:20 a personal preference, definitely take the time to tune

15:22 it.

15:22 This will improve performance dramatically if you do it

15:25 right.

15:25 As you get more advanced, you're going to want to think

15:33 about this a little bit more of this kind of hierarchy of

15:35 different ways to pull in everything.

15:36 So like not just Claude MD, but also config and kind of

15:42 everything about Claude, you can pull in in this

15:43 hierarchical way.

15:45 So projects are specific to your Git repo, and this you can

15:48 check in, or you can make it just for you.

15:51 You can also have global configs that are across all your

15:53 projects, or you can have enterprise policies.

15:56 And this is essentially a global config that you roll out

15:58 for all of your employees, everyone on your team

16:00 automatically.

16:02 And this slide is like pretty information dense, but the

16:05 point is this applies to a lot of stuff.

16:07 So you can do this for slash commands.

16:08 You can do it for permissions.

16:10 So for example, if you have a batch command that you would

16:12 run for all your employees, like all your employees use

16:15 this like test command, for example.

16:17 You can actually just check it into this enterprise

16:19 policies file.

16:21 And then any employee, when they run this command, it will

16:23 be auto approved, which is pretty convenient.

16:26 And you can also use this to block commands.

16:27 So for example, let's say there's a URL that should never

16:29 be fetched.

16:31 Just add it to this config.

16:32 And that'll make it so an employee cannot override it.

16:35 And that URL can never be fetched.

16:38 So pretty convenient both to unblock people and also just

16:40 to keep your code base safe.

16:43 And then same thing for MCP servers.

16:45 Have a MCP JSON file, check it into the code base.

16:47 That way, anytime someone runs quad code in your code base,

16:50 they'll be prompted to install the MCP servers and share it

16:53 with the team.

16:54 If you're not sure which of these to use, this is like a

17:00 kind of an insane matrix.

17:01 Because we support a lot of stuff.

17:03 And engineer workflows are very flexible.

17:04 And every company is different.

17:05 So we kind of want to support everything.

17:07 So if you're not sure how to get started, I would recommend

17:09 start with shared project context.

17:12 You write this once.

17:13 And then you share it with everyone on the team.

17:14 And you get this kind of network effect where someone does a

17:17 little bit of work and everyone on the team benefits.

17:19 There's a lot of tools built into Quad to manage this.

17:26 So as an example, if you run /memory, you can see all the

17:29 different memory files that are getting pulled in.

17:31 So maybe I have an enterprise policy.

17:33 I have my user memory.

17:34 I have project quad MD.

17:36 And then maybe there's an nested quad MD.

17:38 That's only pulled in for certain directories.

17:42 And then similarly, when you do /memory, you can edit

17:45 particular memory files.

17:46 When you type pound sign to remember something, you can

17:48 pick which memory you want it to go to.

17:49 So yeah, that's the next step.

17:55 Take the time to configure Quad MD, MCP servers, all the

18:00 stuff that your team uses, so that you can use it once,

18:02 configure it once, and then share it with everyone.

18:05 An example of this is in our apps repo for anthropic.

18:11 This is the repo that we have all of our web and apps code in.

18:16 There's a puppeteer MCP server, and we share this with the

18:18 team.

18:19 And there's an MCP JSON checked in.

18:21 So any engineer working that repo can use puppeteer in

18:24 order to pilot and to when test and to screenshot

18:26 automatically and iterate, so that every engineer doesn't

18:28 have to install it themselves.

18:29 This is a talk about pro tips.

18:35 I just want to take a quick interlude to talk about some

18:38 common key bindings that people may not know.

18:40 It's very hard to build for Terminal.

18:41 It's also very fun.

18:43 It feels like rediscovering this new design language.

18:46 But something about Terminal is it's extremely minimal.

18:49 And so sometimes it's hard to discover these key bindings.

18:52 And here's just a quick reference sheet.

18:54 So anytime you can hit Shift+Tab to accept edits, and this

18:58 switches you into auto-accept edits mode.

19:00 So bash commands don't need approval, but edits are

19:02 auto-accepted, and you can always ask Quad to undo them

19:04 later.

19:06 For example, I'll do this if I know Quad's on the right track,

19:09 or if it's writing unit tests and iterating on tests, I'll

19:11 usually just switch into auto-accept mode.

19:13 So I don't have to OK every single edit.

19:17 Any time you want Cloud to remember something, so for

19:19 example, if it's not using a tool correctly and you want it

19:22 to use it correctly from then on, just type the

19:24 pound sign and then tell it what to remember, and it'll

19:26 remember it.

19:26 It'll incorporate it into Quad MD automatically.

19:30 If you ever want to drop down to bash mode, so just run a

19:32 bash command, you can hit the exclamation mark and type

19:35 in your command.

19:35 That'll run locally, but that also goes into the

19:38 context window, so Quad will see it on the next turn.

19:42 And this is pretty good for long running commands, if you

19:44 know exactly what you want to do, or any command that you

19:46 want to get into context.

19:47 And Quad will see the command and the output.

19:51 You can add mention thousand folders.

19:53 Any time you can hit escape to stop what Quad is doing, no

19:57 matter what Quad is doing, you can always see if we hit

19:58 escape.

19:59 It's not going to corrupt the session.

20:00 It's not going to mess anything up.

20:02 So maybe Quad is doing a file edit.

20:04 I'll hit escape.

20:04 I'll tell it what to do differently.

20:06 Or maybe it suggested a 20-line edit, and I'm like, actually,

20:09 19 of these lines look perfect, but one line, you should

20:11 change.

20:12 I'll hit escape.

20:12 I'll tell it that.

20:13 And then I'll tell it to redo that.

20:17 You can hit escape twice to jump back in history.

20:20 And then after you're done with the session, you can start

20:22 Quad with a resume to resume that session, if you want, or

20:25 dash, dash, continue.

20:28 And then any time if you want to see more output, hit

20:30 controller.

20:31 And that'll show you the entire output, the same thing

20:34 that Quad sees in its context window.

20:35 The next thing I want to talk about is the Quadcode SDK.

20:44 So we talked about this at the top.

20:46 Right after this, say it is doing a session, I think, just

20:49 across the hallway.

20:50 And he's going to go super deep on the SDK.

20:52 If you hadn't played around with this, if you used the

20:55 dash P flag in Quad, this is what the SDK is.

20:58 And we've been winning a bunch of features over the last

21:00 few weeks to make it even better.

21:03 So yeah, you can build on top of this.

21:05 You can do cool stuff.

21:06 This is exactly the thing that Quadcode uses.

21:08 It's exactly the same SDK.

21:11 And so for example, something you can do is Quad dash P.

21:13 So this is the CLI SDK.

21:15 You can pass a prompt.

21:18 You can pass some allowed tools, which could include

21:20 specific bash commands.

21:22 And you can tell it which format you want.

21:24 So you might want JSON, or you might want streaming JSON,

21:27 if you want to process this somehow.

21:29 So this is awesome for building on.

21:31 We use this in CI all the time, we use this for incident

21:34 response, we use this in all sorts of pipelines.

21:37 So really convenient.

21:38 Just think of it as a Unix utility.

21:39 You give it a prompt, it gives you JSON.

21:41 You can use this in any way.

21:42 You can pipe into it, you can pipe out of it.

21:44 The piping is also pretty cool.

21:50 So you can use, for example, Git status and pipe this in,

21:53 and use JQ to select the result.

21:56 The combinations are endless.

21:58 And it's sort of this new idea.

21:59 It's like a super intelligent Unix utility.

22:02 And I think we barely scratched the surface of how to use this.

22:04 We're just figuring this out.

22:07 You can read from a GCP bucket, read a giant log and pipe it

22:11 in and tell Cloud to figure out what's

22:12 interesting about this log.

22:14 You can fetch data from the century CLI.

22:18 You can also pipe it in and have Cloud do something good.

22:20 The final thing, and this is probably the most advanced

22:29 use case as we see, I'm sort of a quad normy.

22:31 So I'll have usually one quad running at a time.

22:34 And maybe I'll have a few terminal tabs

22:36 for a few different repos running at a time.

22:38 When I look at power users in and out of enthropic,

22:41 almost always they're going to have SSH sessions.

22:44 They'll have Tmux tunnels into their quad sessions.

22:47 They're going to have a bunch of checkouts of the same repo

22:50 so that they can run a bunch of quads in parallel in that repo.

22:53 Or they're using Git work trees to have some kind of isolation

22:55 as they do this.

22:56 And we're actively working on making this easier to use.

22:59 But for now, these are some ideas for how

23:03 to do more work in parallel with Cloud.

23:05 You can run as many sessions as you want.

23:08 And there's a lot that you can get done in parallel.

23:10 So yeah, that's it.

23:14 I wanted to also leave some time for Q&A.

23:17 So I think this is the last slide that I have.

23:19 And yeah, if folks have questions,

23:21 there's mics on both sides.

23:23 And yeah, we'd love to answer any questions.

23:27 [APPLAUSE]

23:31 I did.

23:42 Hey, Boris.

23:50 Thanks for building Cloud Code.

23:51 And I was wondering what was the hardest implementation,

23:54 like part of the implementation for you, of building it?

23:59 I think there's a lot of tricky parts.

24:03 I think one part that is especially tricky

24:06 is the things that we do to make bash commands safe.

24:10 Bash is inherently pretty dangerous.

24:11 And it can change system state in unexpected ways.

24:14 But at the same time, if you have to manually

24:18 approve every single bash command,

24:20 it's super annoying as an engineer.

24:21 And you can't really be productive,

24:23 because you're just constantly proving every command.

24:25 And just kind of navigating how to do this safely

24:28 in a way that scales across the different kinds of code

24:30 bases people have, because not everyone

24:32 runs their code in a Docker container was pretty tricky.

24:35 And essentially, the thing we landed on

24:37 is there's some commands that are read-only.

24:39 There's some static analysis that we

24:40 do in order to figure out which commands can be combined

24:43 in safe ways.

24:44 And then we have this pretty complex tiered permission

24:47 system so that you can allow list and blockless commands

24:49 at different levels.

24:53 Hi, Boris.

24:54 You mentioned giving an image to Cloud Code,

24:57 which made me wonder if there's some sort of multimodal

25:00 functionality that I'm not aware of.

25:02 Is that-- are you just pointing it out

25:03 in the image on the file system or something?

25:05 Yeah, so Cloud Code is fully multimodal.

25:08 It has been from the start.

25:09 It's in a terminal, so it's a little hard to discover.

25:12 But yeah, you can take an image and just drag and drop it in.

25:15 That'll work.

25:16 You can give it a file path that'll work.

25:18 You can copy and paste the image in, and that works, too.

25:23 So I'll use this pretty often for if I have a mock of something.

25:26 I'll just drag and drop in the mock.

25:28 I'll tell it to implement it.

25:29 I'll give it up a tier server so it can iterate against it.

25:32 And yeah, it's just fully automated.

25:33 Hey, why did you build a CLI tool instead of an IDE?

25:43 Yeah, it's a good question.

25:44 I think there's probably two reasons.

25:45 One is we started this ad entropic.

25:48 And ad entropic people use a broad range of IDs.

25:51 And some people use VS Code, other people use said, or Xcode, or Vim, or Emacs.

25:58 And it was just hard to build something that works for everyone.

26:00 And so terminal is just a common denominator.

26:03 The second thing is ad entropic, we see up close how fast the model is getting

26:10 better.

26:10 And so I think there's a good chance that by the end of the year,

26:13 people aren't using IDs anymore.

26:15 And so we want to get ready for this future.

26:17 And we want to avoid over-investing in UI and other layers on top,

26:21 given that the way the models are progressing,

26:24 it may not be useful work pretty soon.

26:26 Yeah.

26:32 How much have you-- I don't know if this is on.

26:35 How much have you used Cloud Code for machine learning modeling?

26:39 And almost that AutoML experience.

26:40 I was curious what the experience has been so far with that.

26:44 Yeah, I think the question was, how much are we using Cloud Code for machine

26:48 learning and modeling?

26:50 We actually use it for this a bunch.

26:51 So both engineers and researchers at anthropic use Cloud Code every day.

26:56 I think about 80% of people at anthropic that are technical use Cloud Code

27:00 every day.

27:01 And hopefully you can see that in the product and the kind of the amount of

27:04 love

27:04 and dog fooding we've put into it.

27:06 But this includes researchers who use tools like the notebook tool to edit and

27:10 run notebooks.

27:12 OK, very cool.

27:13 Thank you.

27:16 All right, I think that's it.

27:17 Thanks.

27:18 [APPLAUSE]

27:20 [MUSIC PLAYING]

27:48 [MUSIC PLAYING]

Builder

· 2w

Spacex 的失败与成功！
感谢埃隆·马斯克和 SpaceX，感谢你们从未放弃将意识延伸到地球之外的探索。

中文 English

0:00 [音乐播放]

0:30 [音乐播放]

0:33 好了。

0:33 所有准备就绪，倒计时开始。

0:35 飞行指挥，准备发射。

0:44 [音乐播放]

1:00 [音乐播放]

1:03 保持间隔距离。

1:10 [音乐播放]

1:14 我们可以显示任何点火启动。

1:19 [音乐播放]

1:30 [听不清]

1:32 [音乐播放]

1:35 [听不清]

1:44 [音乐播放]

1:48 [欢呼]

1:51 是的。

1:53 就这样。

1:55 好了。

1:56 通往火星的大门。

1:57 [笑声]

2:00 我们到了。

2:01 这里是新成立的星基地德州。

2:08 [欢呼]

2:13 这是多年后，美国建造的新城，

2:18 至少我是这么听说的。

2:22 名字也很酷。

2:25 它之所以这么命名，是因为这里

2:28 将开发必要技术，

2:30 首次将人类、文明和生命，

2:34 送往另一颗星球，

2:37 这是地球45亿年历史上的第一次。

2:40 [欢呼]

2:43 接下来请看这小段视频。

2:52 一开始几乎什么都没有。

2:56 星基地最初只是一个沙洲，一无所有。

3:02 甚至我们建的那些小设施。

3:07 那是最初的疯狂麦克斯火箭。

3:26 [音乐播放]

3:28 你会发现照明

3:36 对疯狂麦克斯火箭来说非常重要。

3:37 [笑声]

3:39 看，没多久以前。

3:55 这里几乎一片空白。

3:57 短短五六年的时间里，

4:01 多亏了SpaceX团队的卓越工作，

4:05 我们建起了一座小城市。

4:06 还建造了两个巨型发射台，

4:11 和一个为巨大火箭服务的巨大工厂。

4:17 很酷的是，看这视频的朋友们

4:20 其实可以来参观，

4:24 因为我们的生产基地和发射场

4:28 都位于公共公路旁。

4:30 来德州南部的任何人都可以近距离

4:35 观看火箭，参观工厂。

4:38 所以，任何想看地球上最大飞行物的

4:43 朋友都能随时来，

4:45 沿着公路开车过去

4:47 亲眼见证，非常棒。

4:50 然后我们进展到了现在的位置。

4:55 星基地2025。

4:57 现在我们可以大约每两三周

5:07 生产一个宇宙飞船。

5:08 我们不总是每两三周生产一艘，

5:14 因为还在改进设计。

5:18 但最终目标是能

5:20 每年生产约1000艘，平均每天3艘。

5:23 这就是目前的情况。

5:31 我正站在那栋楼里。

5:32 那是我们的气垫船。

5:48 我们正把助推器沿路运往发射场。

6:06 你可以看到超级车间。

6:16 我说这视频的看客们最酷的是

6:19 你们可以真的来这里，

6:22 沿着这条路开车过去亲眼见证，

6:24 这是历史上首次可以这样做。

6:28 所以这一切酷炫设施，左边那条路，

6:33 那条公路是公共的，你随时可以来看，

6:37 强烈建议来看看。

6:38 我觉得非常鼓舞人心。

6:40 那就是我们的超级车间。

6:46 我们正在扩建整合，

6:48 目标是每年生产1000艘宇宙飞船。

6:49 嗯，这还没建好，

6:56 但我们正在建设中。

6:59 那真是个巨大结构。

7:01 将成为世界上最大建筑之一。

7:03 按某些标准可能是全球最大建筑。

7:08 它专为一年产1000艘飞船设计。

7:11 我们也在佛罗里达建超级车间，

7:18 这样德州和佛州各有一个基地。

7:22 其实很难衡量这些建筑的规模，

7:28 你得以人的比例来参照，

7:30 看到人与建筑相比有多渺小，

7:33 才能真正体会它们的巨大。

7:39 比较我们的年度产量，

7:44 波音和空客制造飞机，

7:49 未来星舰可能会

7:53 制造和波音、空客制造商用飞机一样多，

7:58 专为火星打造的星舰。

8:00 所以这真的是一个规模，非常巨大的规模。

8:04 每个星舰都将拥有一定的能力。

8:08 每艘星舰都比747或A380还要大。

8:14 它真的非常巨大。

8:15 至于星链卫星，

8:23 第三版卫星，年产大约5000颗，

8:27 或许某个时候接近一年一万颗。

8:30 这些Sonic V3卫星大小大约相当于一架737。

8:37 它们相当大。

8:40 这相当于二战时期的B-24轰炸机。

8:43 不过相比特斯拉还算小。

8:51 所以，特斯拉未来可能会，

8:56 产量翻倍，甚至更多。

8:58 这让我们看清事实，

9:00 实际上建造大量

9:04 星际飞船是可能的。

9:08 即使从吨位上比较，和其他汽车公司相比，

9:10 其他车企仍在制造比SpaceX更复杂的大吨位产品，

9:13 这其实说明了

9:17 这是非常可行的事情。

9:20 这些数字虽然按照传统航天标准来说

9:22 极其庞大，

9:26 但因为其他行业已经达到过，

9:28 人类完全可以实现。

9:33 进展的衡量标准是

9:35 建立火星自给自足文明的时间线。

9:37 这就是我们在Starbase衡量进展的方式。

9:40 因此，每次发射，尤其是在星舰的早期，

9:44 每次发射都是为了更多了解

9:48 让生命多星球发展的需求，

9:51 并改进星舰，直到

9:54 能最终载送数十万，

9:56 甚至数百万的人到火星。

10:00 理想情况下，我们可以带任何想去火星的人，

10:02 带去火星，并带上所有必要设备，

10:05 让火星实现自给自足，

10:11 让火星能够在最糟糕的情况下自我发展。

10:17 达到那种关键分叉点，

10:19 人类命运的基本转折点

10:22 就是火星能够独立发展，

10:29 即使地球的物资供应因任何原因中断。

10:31 那时，我们实现了文明韧性，

10:33 火星有可能在地球出问题时救援地球，

10:39 或者地球救援火星。

10:43 拥有两个自给自足且强大的星球，

10:45 我认为对文明长期存续至关重要。

10:47 所以，我觉得任何文明

10:56 如果是多星球文明，

11:00 其存续时间可能是单一星球文明的十倍，

11:04 甚至更长。

11:08 因为人类总有可能做出疯狂的事情，

11:11 比如第三次世界大战，虽然希望不会，

11:16 但这有可能发生，或者可能出现

11:18 自然事件，比如陨石、超级火山，

11:22 或其他无法预料的灾难。

11:25 如果我们只有一个星球，

11:28 那将是一场灾难的终结。

11:34 但如果有两个星球，我们可以继续前行。

11:36 最终我们会超越火星，

11:39 到达小行星带、木星的卫星，甚至更远，

11:41 最终去往其他恒星系统。

11:44 我们将置身星辰之间，

11:46 实现科幻不再是虚构。

11:53 为了实现这一目标，

11:56 我们必须制造快速可重复使用的火箭，

11:58 使每次飞行的成本、每吨货物到火星的成本

12:01 尽可能低，快速可重复使用的火箭是关键。

12:03 我说那实际上是四个小时，或者就像海盗那样，

12:06 快速可重复使用且可靠的火箭是关键。

12:12 R-R-R-R。

12:18 我们SpaceX团队在捕获巨型火箭方面

12:21 取得了令人难以置信的进步。

12:26 SpaceX团队能够多次用一种非常新颖的方法

12:29 用巨大的筷子从空中捕获史上最大的飞行物体，

12:35 真是令人震惊。

12:39 你见过这种操作吗？

12:52 [笑声]

12:56 再次恭喜你们。

13:01 这是一次了不起的成就。

13:26 [欢呼声]

13:34 再次恭喜你们。

13:56 这是一项了不起的成就。

13:58 [欢呼声]

14:00 我们以这种方式捕获，这是前所未有的，

14:02 目的是实现火箭的快速可重复使用部分。

14:13 如果超重助推器如此巨大，

14:14 直径约30英尺，

14:22 它如果用着陆腿降落在着陆台上，

14:28 我们必须抬起它，收起着陆腿，然后放回发射台。

14:31 运输这么大的物体非常困难。

14:36 但如果我们用同一个发射塔

14:43 来捕获它，这个发射塔正是最初用于安装它的，

14:48 那是快速重复使用的最佳方式。

14:53 它实际上是被同一双机械臂

14:59 从发射环上放置的地方接住的，

15:03 But if we catch it with the same tower

15:06 that it's used to put it in the launch mount to begin with,

15:12 that is the best case outcome for rapid reuse.

15:15 So it literally gets caught by the same arms

15:18 that placed it in the launch ring,

15:22 然后它会立即被放回发射环中。

15:26 所以原则上，超重助推器

15:28 可以在着陆后一小时内重复飞行。

15:33 所以它大约五六分钟内返回，

15:36 总之不管怎么样。

15:38 然后它会被塔臂抓住，再次固定回

15:43 发射座架中。

15:45 然后你可以用大约三四十分钟

15:50 重新加注推进剂，放上飞船。

15:53 原则上，每小时可以重复飞行整个助推器。

15:58 可能每两小时一次，给它一点额外时间。

16:02 但我们就说这已经接近快速重复使用的极限。

16:13 接下来我们需要做的是回收飞船。

16:17 这我们还没做过，但会做。

16:19 这就是我们希望今年晚些时候展示的。

16:37 也许就在两三个月内。

16:44 然后飞船会被放在助推器上方，

16:48 再用推进剂重新加注，重新飞行。

16:54 带着飞船，需要的时间更长，

16:55 因为它需要绕地球几圈，

16:57 直到轨迹回到发射台上空。

17:01 但飞船也打算每天多次重复飞行。

17:04 这是新款猛禽3发动机，非常棒的引擎。

17:09 我们不能告诉猛禽团队这事。

17:13 这非常令人兴奋。

17:14 猛禽3设计时不需要基本的隔热瓦，

17:15 节省了底部大量质量，

17:28 实际上还提高了可靠性，

17:29 如果比如猛禽发动机出现轻微燃料泄漏，

17:31 它会直接泄漏到已有的火焰等离子体中，

17:35 没什么大影响。

17:38 而发动机包裹在箱体内时燃料泄漏，

17:41 确实非常可怕。

17:44 这就是猛禽3。

17:49 可能需要多次尝试，

17:51 但它将在有效载荷能力、发动机效率和可靠性

17:55 上带来巨大提升。

17:57 这真的是一次革命性的发动机。

18:02 猛禽3其实，

18:05 我认为，是一种外星技术的火箭发动机。

18:08 我说的是，就连业内专家，

18:11 看到猛禽3照片时，

18:12 都说那个发动机还没完成。

18:21 然后我们说，好吧，这个未完成的发动机在测试中，

18:23 效率达到了前所未有的水平。

18:26 这是一台极其干净的发动机。

18:30 为了做出这样的发动机，

18:35 我们简化了设计中的许多零件，

18:44 并将辅助流体回路和电子设备

18:51 内置于发动机结构中。

18:54 所有部件都被封装和保护，

18:58 这是正确工程的典范。

19:00 另一个对火星任务关键的技术

19:04 是轨道推进剂加注。

19:06 可以把它想象成飞机的空中加油，

19:14 但这里是在轨道上为火箭加注，

19:21 这以前从未实现过。

19:24 但技术上是可行的。

19:27 我总觉得这些事有点不太适合公开说。

19:31 你得以某种方式转移流体，

19:33 这是不得不做的。

19:35 两艘星舰会会合，

19:45 一艘星舰会转移燃料和氧气。

19:51 其实大部分质量是氧气，

19:56 大约80%是氧气，

20:03 稍微超过20%是燃料。

20:08 然后你送一艘满载有效载荷的星舰上轨，

20:10 再发射多艘星舰，

20:14 为那艘星舰加注推进剂。

20:16 一旦推进剂罐几乎注满，

20:23 就可以起航前往火星或月球等地了。

20:25 这是一项重要技术，

20:28 希望明年能够演示成功。

20:31 [掌声]

20:36 接下来要解决的难题之一，

20:38 是可重复使用的隔热盾。

20:42 没有人真正开发出过可重复使用的轨道隔热盾，

20:45 所以这极其难实现。

20:54 即使航天飞机的隔热瓦也

20:58 需要几个月的维修，

21:03 修补损坏瓦片，测试每块瓦片。

21:07 因为重返大气层时承受的极端高温高压

21:10 极难抵抗，

21:13 只有高级陶瓷材料才能承受，

21:17 基本就是玻璃、氧化铝，有时是碳-碳复合材料。

21:21 但几乎没有材料能真正实现可重复使用，

21:24 在不侵蚀、脱落或开裂的情况下，

21:31 经受住重返大气层带来的应力。

21:40 这是首次开发这样的可重复使用轨道隔热盾。

21:47 并且显然需要极高的可靠性。

21:52 我认为这将是未来几年努力的重点，

21:57 不断完善隔热盾。

22:01 这完全是可实现的，

22:04 所以我们不是在尝试无法实现的事情。

22:08 这完全在物理领域内可行。

22:11 So this will be something they'll be working on for a few years,

22:16 I think, to keep honing the heat shield.

22:20 It's a very-- it is an achievable thing,

22:22 so we're not trying to do something that isn't achievable.

22:24 It is when the realm of physics to get this done,

22:27 只是一个极其困难的任务。

22:31 而火星——火星的大气是二氧化碳，

22:35 这起初看起来可能更好，但实际上

22:40 会更糟，因为当二氧化碳变成等离子体时，

22:46 你实际上会得到

22:47 更多的自由氧进入火星大气，

22:51 比地球大气更多。

22:52 地球大气中氧气约占20%，

22:58 而火星上基本上是它的两倍，

23:02 甚至可能是三倍，当二氧化碳变成等离子体时，

23:07 会产生碳和氧气。

23:10 这会氧化热防护罩，

23:13 基本上燃烧热防护罩。

23:15 这就是为什么我们在二氧化碳

23:19 大气中进行了非常严谨的测试，

23:22 因为它不仅要适用于地球，

23:23 还要适用于火星。

23:33 我们希望为地球使用和火星相同的热防护罩，

23:36 因为热防护罩还有很多其他因素，

23:38 如确保瓷砖不破裂或掉落等问题。

23:42 所以我们想要在地球上和火星上使用相同的热防护罩结构，

23:45 相同的材料，

23:47 这样我们可以在地球上测试数百次，

23:50 然后有信心送到火星时能够发挥作用。

23:53 所以我们正在开发一些下一代星舰，

24:06 相比当前这代有不少改进。

24:16 例如，它更高，

24:21 并且在飞船和助推器之间有更好的级间结构。

24:22 你可以看到那些支撑杆。

24:31 这使得火焰更容易通过，就在那儿。

24:34 进行热级联时，也就是当飞船发动机点火，

24:38 助推器发动机仍然在工作，

24:44 飞船发动机的火焰可以更容易地通过

24:46 新的相间开放支撑杆排出。

24:49 这次，我们会回收这些支撑杆，

24:54 而不是扔掉它们。

25:00 这里稍微高一点，从大约69米升到72米。

25:03 推进剂容量——我想我们可能会提高，

25:05 大概是3,700吨。

25:19 长期来看，我猜测是大约4,000吨。

25:21 以及稍超过8,000的推力。

25:25 每一点提升，比如8,000——还会继续增长。

25:30 我猜，最终我们会在这里达到4,000吨，接近

25:36 10,000吨推力。

25:39 但这是下一阶段，或者说是超级重型助推器的下一版本。

25:43 因此助推器底部看起来有点单薄，

25:44 因为Raptor 3发动机不需要热防护罩。

25:48 看起来好像少了零件，

25:50 但只是因为Raptor 3不需要热防护罩。

25:56 所以它就那样站着，被炽热的等离子火焰包围。

26:01 但重量大减。

26:05 没错。

26:08 集成热级联。

26:12 我觉得看起来很棒。

26:25 飞船稍长，能力更强。

26:28 推进剂容量达到1,550吨。

26:29 我猜长期来看，这可能会多出20%。

26:33 你可以看到热防护罩更流线型。

26:34 边界更平滑。

26:45 热防护罩朝背风面的边界非常光滑，

26:50 不再是锯齿状的瓷砖。

26:55 我觉得它看起来非常流畅。

27:03 这个版本我们仍然有六台发动机，

27:06 但未来版本会有九台。

27:12 同样得益于Raptor 3，我们减少了质量，

27:20 提高了比冲。

27:21 星舰3号版本是我认为

27:29 实现所有关键要素的版本。

27:32 一般来说，任何新技术

27:39 需要经过三次重大迭代

27:44 才能真正稳定且高效地工作。

27:51 Raptor 3，星舰3号，和助推器3号，

27:55 应该能实现我刚才提到的所有目标，

27:56 那就是快速重复使用、可靠的火箭

28:02 具备轨道加注能力。

28:04 基本上，实现多星球生活所需的一切要素，

28:10 都将在星舰3号版本中实现，

28:12 我们计划今年年底首次发射该版本。

28:16 [掌声]

28:19 你可以看到左边是现在的情况，

28:22 中间是今年年底的状态。

28:26 正如我所说，长期来看会是这样，

28:29 142米。

28:31 中间那个版本将完全具备火星任务能力。

28:40 之后会有大量性能提升。

28:45 就像猎鹰9号一样，

28:48 我们最终都会让火箭更长、有效载荷增加。

28:54 这就是计划，很简单。

28:58 但必须强调，即使是我们今年年底

29:05 将要发射的这枚火箭，

29:09 也有能力实现多星球生活。

29:12 之后，只是继续

29:16 优化火箭效率和性能的问题。

29:17 longer and increasing payload.

29:20 So that's the game plan, pretty straightforward.

29:26 But it's important to emphasize, even with the rocket

29:29 that we'll be launching just at the end of this year,

29:31 it will be capable of making life multi-planetary.

29:34 And thereafter, it's just about continuing

29:37 to hone the efficiency and capability of the rocket

29:40 降低每吨和每人的成本

29:43 去火星。

29:44 就像我说的，最终，做到

29:46 任何想移居火星的人

29:48 并帮助建设新文明都能实现。

29:53 所以任何人在外面，这会多酷啊？

29:57 即使你不想去，

29:58 也许你的儿子或女儿想去

30:02 或者你的朋友想去。

30:05 我觉得这将是一场冒险，

30:08 但最好的冒险，是可能实现的，

30:12 去新星球上帮助建设新文明。

30:16 [掌声]

30:18 所以，是的，最终我们会有42台发动机，

30:27 这是必然的。

30:32 正如伟大的先知道格拉斯·亚当斯在他的书中预言的那样

30:38 《银河系漫游指南》，

30:45 生命意义的答案是42。

30:50 所以星舰堆叠必然会有42台发动机。

30:54 [掌声]

31:02 至于运载能力，值得注意的是

31:05 它是200吨——

31:08 星舰将有200吨的轨道运载能力

31:10 且完全可重复使用。

31:12 这是土星五号月球火箭能力的两倍。

31:16 土星五号月球火箭是一次性的。

31:20 星舰完全可重复使用，但轨道运载能力是

31:23 能到轨道的下一个最大火箭的两倍，

31:28 也就是土星五号。

31:31 实际上，如果没有可重复使用，星舰

31:35 可运载大约400吨到轨道。

31:39 所以我说的是，这是一枚非常大的火箭。

31:45 你需要大火箭去实现多行星生活。

31:51 途中，我们还能做一些很酷的事，

31:53 比如建一个月球基地，像月球基地阿尔法。

31:57 很久以前，有个关于月球基地阿尔法的电视节目。

32:03 你不能太较真它的物理学，

32:06 因为显然基地在慢慢远离地球。

32:09 不过不管怎样，我们应该建一个月球基地阿尔法，

32:12 这是阿波罗计划后的下一步，

32:14 在月球上建一个基地。

32:16 你可以建一个巨大的科学站，

32:23 在月球上研究宇宙的本质。

32:26 那会非常酷。

32:27 那么关于——你什么时候能去火星？

32:33 你每两年或每26个月可以去一次火星。

32:36 下一次火星机会是在明年年底，

32:41 大约18个月后。

32:42 所以11月、12月是下一个火星机会。

32:45 我们会努力赶上那个窗口。

32:47 如果幸运的话，我认为我们现在有

32:48 大约五五开的机会，因为我们

32:52 必须解决轨道加注问题，才能

32:55 具备足够能力去火星。

32:58 如果及时实现轨道加注，

33:01 我们明年年底将发射首个无人星舰

33:05 前往火星。

33:07 [掌声]

33:09 这张图示说明了宇宙飞船

33:19 如何从地球飞往火星？

33:20 蓝色是地球，红色是火星。

33:25 实际上飞行的距离大约

33:37 是去月球的1000倍左右。

33:43 所以你不能直接飞向火星。

33:47 你必须创造一个椭圆轨道，地球是

33:50 一个焦点，火星在另一侧，

33:52 椭圆的远端。

33:54 然后调整你在轨道上的位置

33:59 和椭圆轨道，以便与火星相交。

34:01 这就是轨道转移，

34:05 如何从地球轨道转移到火星轨道。

34:08 如果你看你的星链Wi-Fi路由器，

34:11 你会看到这张图。

34:13 因为星链Wi-Fi或星链互联网

34:17 是资助人类到火星的方式。

34:23 所以我要感谢所有加入星链的人，

34:24 因为你们在确保

34:28 文明的未来，帮助

34:31 实现多星球生活，助力

34:33 人类成为太空文明。

34:35 谢谢大家。

34:36 [掌声]

34:39 这是一个初步的计划，

34:48 我们希望实现的目标是

34:53 每个发射窗口大幅增加飞往火星的航班频率。

34:58 所以大约每两年，飞往火星的飞船数量

34:59 会大幅增加。

35:05 最终，尝试达到每次火星会合时

35:10 1000到2000艘飞船。

35:18 我是说，这只是一个大致估计——只是猜测，

35:19 但我猜，我们需要大约一百万吨

35:26 运到火星表面，

35:29 才能让火星文明实现自给自足。

35:31 同样，要达到关键点，

35:36 如果来自地球的补给船

35:37 因任何原因停止，火星依然能成功。

35:40 火星依然能够发展。

35:43 所以不能遗漏任何东西，

35:45 甚至不能缺少类似维生素C的东西。

35:46 不能缺少任何东西。

35:48 或其他任何东西。

35:49 你必须拥有火星发展的所有必需品。

35:51 这是必不可少的。

35:55 所以我猜大概是一百万吨，

35:56 但也可能是一千万吨。

35:58 希望不是一亿吨。

36:00 那就太多了。

36:02 但我们希望尽快达到那个目标

36:05 保障文明的未来。

36:07 尽可能快。

36:08 所以我们在考察不同的地点。

36:14 目前的首选是阿卡迪亚地区。

36:18 火星有很多地方可选，

36:21 但结合所有因素后，说实话，

36:26 不能离极地太近。

36:28 需要靠近冰层以获取水源。

36:34 且不能太多山地，不利于火箭发射，

36:39 这样就缩小到一个较小的区域。

36:42 阿卡迪亚也是我一个女儿的名字，

36:47 实际上是其中一个选择。

36:51 我们会在火星部署首批星舰，

36:54 收集关键数据，那是那里最初的光亮，

36:57 我们会派出奥普蒂姆斯机器人。

37:00 我们可以派它去勘探，为

37:03 人类做准备。

37:04 如果明年发射成功，

37:07 这将是非常酷的画面。

37:10 技术上我们将于2027年抵达，

37:13 那将是一幅史诗般的画面，

37:17 看到奥普蒂姆斯在火星表面行走。

37:20 两年后发射，

37:29 我们将送人类过去。

37:30 如果首批任务成功，

37:33 成功着陆，

37:34 下一次任务就会派人类登陆。

37:37 我们会真正开始建设

37:40 火星的基础设施。

37:42 不过，为了安全起见，

37:45 可能会用奥普蒂姆斯进行两次着陆，

37:50 第三次才派人类，我们拭目以待。

37:54 就像经典的帝国大厦工人照片那样。

37:58 在火星上的通信，

38:06 我们会使用改版的星链，

38:08 为火星提供互联网。

38:10 是的，光速——

38:25 即使是光速，最佳情况

38:28 也需要大约三分半钟到达火星。

38:32 最坏情况则是22分钟甚至更久，

38:36 因为火星位于太阳和地球的另一侧。

38:41 所以，

38:45 与火星实现高带宽通信非常具有挑战性。

38:48 但星链能够实现这一点。

38:49 然后第一批人类

38:55 将奠定火星表面永久存在的基础。

38:57 是的，正如你所说，

39:02 目标是尽快让火星实现自给自足。

39:06 这只是对火星首个城市可能样貌的

39:15 粗略设想。

39:17 我猜我们会把发射台

39:18 和着陆台稍微分开一点，以防万一。

39:24 在火星，我们将需要大量太阳能。

39:29 目前无法在火星表面自由行走，

39:36 至少目前不能，除非火星地球化成像地球一样。

39:41 你需要穿着火星太空服行走，

39:44 最初还得待在玻璃穹顶里。

39:49 但这是可行的。

39:51 最终，我们可以将火星变成类似地球的星球。

39:54 我们希望达到每个火星转移窗口

40:02 转移超过一百万吨的规模。

40:06 那才是真正的文明规模。

40:10 每个转移窗口一百万吨。

40:12 是的，我们会有很多太空港。

40:22 因为你不能连续飞往火星，

40:25 必须在窗口期转移，

40:28 会汇聚一千艘、两千艘，

40:31 甚至更多的飞船。

40:33 这让我想到《太空堡垒卡拉狄加》般的场景，

40:35 所有飞船都在轨道上，

40:36 等待起飞。

40:38 然后它们一起起飞，

40:44 这将是令人惊叹的画面。

40:47 肯定需要大量火星的着陆台。

40:51 或者你得快速把飞船驶离着陆台。

40:53 如果有几千艘飞船抵达，

40:57 你至少需要几百个

40:59 着陆台。

41:02 反正我们以后会解决这个问题。

41:06 是的，这真是令人难以置信的事迹，

41:08 火星上建立这样一座惊艳的城市。

41:14 火星上第一座城市，

41:22 另一个星球上的第一座城市。

41:26 一个新世界。

41:29 这也是一次机会，

41:35 让我认为

41:40 火星人可以重新思考他们想要什么样的文明。

41:47 你可以重新思考，想要什么样的政府形式，

41:50 想要哪些新规则？

41:51 火星上有很多自由和机会，

41:55 可以重新构建文明，

41:57 这是一次重新整合文明的机会。

42:01 将由火星人决定。

42:02 那么，好吧，开始吧。

42:09 [掌声]

42:13 谢谢大家。

0:00 [MUSIC PLAYING]

0:30 [MUSIC PLAYING]

0:33 All right.

0:33 All up it is on countdown one.

0:35 Flight directors go for launch.

0:44 [MUSIC PLAYING]

1:00 [MUSIC PLAYING]

1:03 Stay in separation.

1:10 [MUSIC PLAYING]

1:14 We can show any burn startup.

1:19 [MUSIC PLAYING]

1:30 [INAUDIBLE]

1:32 [MUSIC PLAYING]

1:35 [INAUDIBLE]

1:44 [MUSIC PLAYING]

1:48 [CHEERING]

1:51 Yeah.

1:53 There you go.

1:55 All right.

1:56 The gateway to Mars.

1:57 [LAUGHTER]

2:00 So here we are.

2:01 Here we are at the newly incorporated Starbase Texas.

2:08 [CHEERING]

2:13 This is the post, New City, made in America in, I think,

2:18 quite a few decades, at least that's what I'm told.

2:22 And a very cool name.

2:25 And it's named that because it is where

2:28 we are going to develop the technology,

2:30 necessary to take humanity and civilization and life,

2:34 as we know it, to another planet for the first time

2:37 in the 4 and 1/2 billion year history of Earth.

2:40 [CHEERING]

2:43 So go with this little video here.

2:52 This is how it started off with basically nothing.

2:56 So Starbase started off as basically a sand bar with nothing.

3:02 Even those little things we built, obviously.

3:07 That's the original sort of Mad Max rocket.

3:26 [MUSIC PLAYING]

3:28 This is where you discover lighting

3:36 is very important for that Mad Max rocket.

3:37 [LAUGHTER]

3:39 See, not that long ago.

3:55 There was basically nothing here.

3:57 And in the space of about five or six years,

4:01 thanks to the incredible work of the SpaceX team,

4:05 we've built a small city.

4:06 And we've added we've built two gigantic launch pads

4:11 and a gigantic rocket factory for a gigantic rocket.

4:17 So-- and the cool thing is for anyone out there

4:20 who's watching this, you can actually come and visit

4:24 because our entire production facility and launch site

4:28 are on a public highway.

4:30 So anyone who comes to South Texas can come and see the rocket

4:35 pretty close up and see the factory.

4:38 And so anyone who's interested in seeing the largest flying

4:43 object on Earth can come here any time they want

4:45 and just drive down the public highway

4:47 and see it, which is pretty cool.

4:50 So then we progress to where we are now.

4:55 Start base 2025.

4:57 So we're now at the point where we can produce a ship

5:07 roughly every two or three weeks.

5:08 Now, we don't always produce the ship every two or three weeks

5:14 because we are making design upgrades.

5:18 But ultimately, we're aiming for the ability

5:20 to produce 1,000 ships a year or so, three ships a day.

5:23 So that's where things are now.

5:31 I'm standing in that building.

5:32 That's our hovercraft.

5:48 We're driving a booster down the road to the launch site.

6:06 You can see the megabays.

6:16 And I said, what's the cool thing for those out there watching

6:19 this video is that you can actually just literally come here,

6:22 drive down the road and see it, which

6:24 is the first time in history that that's been possible.

6:28 So all this cool stuff, that road on the left there,

6:33 that highway is public, and you can just come here and see it,

6:37 which I recommend doing.

6:38 I think it's very inspiring to see.

6:40 So that's our Gigabay.

6:46 So where we're expanding integration

6:48 to produce 1,000 saushes per year.

6:49 Well, yeah, that hasn't been built yet,

6:56 but we're building it.

6:59 That is a truly enormous structure.

7:01 That'll be one of the biggest structures.

7:03 I think by some measures, the biggest structure in the world.

7:08 And it's designed for 1,000 saushes per year.

7:11 We're also building a Gigabay in Florida, bringing--

7:18 so we'll have two facilities, one in Texas and one in Florida.

7:22 It's actually difficult to gauge the size of these buildings,

7:28 because you need a kind of human per scale,

7:30 where you see how tiny a human is next to that building,

7:33 and you realize just how enormous it is.

7:39 So when we look at our build comparison, vehicles per year,

7:44 and so you look at Boeing and Airbus making airplanes,

7:49 Starship will be making, at some point, probably

7:53 as many star ships for Mars as Boeing and Airbus

7:58 make commercial airplanes.

8:00 So this is really at a scale, an enormous scale.

8:04 And each Starship will have a capability.

8:08 Each Starship is bigger than a 747 or an A380.

8:14 Like it's truly enormous.

8:15 And then in terms of Starling satellites,

8:23 the version 3 satellites, making on the order of 5,000 a year,

8:27 maybe at some point closer to 10,000 a year.

8:30 And those Sonic V3 satellites are each the size of roughly a 737.

8:37 They're pretty big.

8:40 That compares to the B-24 bomber in World War II.

8:43 Now it's still small compared to Tesla.

8:51 So, and Tesla will probably be doing, you know,

8:56 double or trouble at volume in the future.

8:58 So it just puts things into perspective

9:00 that it is actually possible to build a vast number

9:04 of interplanetary star ships.

9:08 And even when you can compare things

9:10 on the tonnage standpoint, and other car companies,

9:13 still building far more complex manufacturer tonnage

9:17 than SpaceX, which is really a way of saying

9:20 that it's very achievable.

9:22 Like these numbers, while they are insanely high

9:26 by traditional space standards,

9:28 are achievable by humans because they

9:33 have been achieved in other industries.

9:35 Progress is measured by the timeline

9:37 to establishing a self-sustaining civilization on Mars.

9:40 That's how we're gauging our progress here at Starbase.

9:44 So with each launch, especially in the early days of Starship,

9:48 each launch is about learning more and more about what's

9:51 needed to make life multiplanetary,

9:54 and to improve Starship to the point

9:56 where it can be taking ultimately hundreds of thousands,

10:00 if not millions of people, to Mars.

10:02 Ideally, we can take anyone who wants to go to Mars,

10:05 we can take to Mars, and bring all of the equipment necessary

10:11 to make Mars self-sustaining, so Mars can grow by itself

10:17 in a worst-case scenario.

10:19 Getting to the point where the fundamental fork in the road

10:22 for human destiny is where Mars can continue to grow,

10:29 even if the supply shifts from Earth

10:31 and stuff coming for any reason.

10:33 At that point, we've achieved civilizational resilience

10:39 where Mars can potentially come to the rescue of Earth

10:43 if something goes wrong, or maybe Earth

10:45 can come to the rescue of Mars.

10:47 But having two planets that are both self-sustaining and strong,

10:56 I think it's going to be incredibly important for the long-term

11:00 survival of civilization.

11:04 So just, I think, any given civilization

11:08 is likely to last maybe, I don't know, 10 times longer,

11:11 maybe much longer, if it is a multi-planet civilization,

11:16 then if it is a single-planet civilization.

11:18 Because there's always some chance that us humans could

11:22 do something crazy, like World War III, hopefully not,

11:25 but it's possible, or that there could

11:28 be some natural event, like meteors, or super volcanoes,

11:34 or something that we don't expect.

11:36 And then, if we only have one planet,

11:39 the neck of that could be curtains.

11:41 But if we have two planets, we keep going.

11:44 And then we go beyond Mars, ultimately,

11:46 to maybe the asteroid belt, the moons of Jupiter, and beyond,

11:53 and ultimately to other star systems.

11:56 And we can be out there among the stars,

11:58 making science fiction no longer fiction.

12:01 So in order to achieve this goal,

12:03 we have to make rapidly reusable rockets,

12:06 so that the cost per flight, the cost per ton to Mars,

12:12 is as low as possible, for that rapidly reusable rockets,

12:18 I said, it's actually four hours, or it's like a pirate,

12:21 or it's like rapidly reusable, reliable rockets is the key.

12:26 R-R-R-R.

12:29 Now, we've progressed the SpaceX team

12:35 on making incredible progress on catching a giant rocket.

12:39 So it's really mind blowing that the SpaceX team

12:52 has been able to catch the largest flying object ever

12:56 made multiple times, using a very novel method

13:01 of catching it out of the air with giant chopsticks.

13:26 I mean, have you ever seen that before?

13:34 [LAUGHS]

13:56 Yeah, congrats again.

13:58 That was an incredible achievement.

14:00 [CHEERING]

14:02 Yeah, congrats again.

14:13 That was an incredible achievement.

14:14 [CHEERING]

14:22 So the reason we are catching it in this way, which has never

14:28 been done before, is in order to achieve

14:31 the rapidly reusable portion of the rocket rapidly reusable.

14:36 So if the super heavy booster, which is gigantic,

14:43 it's like 30 feet in diameter, if it were to land with landing

14:48 legs on a landing pad, we would then have to pick it up,

14:53 stow the legs, and put it back in the launch pad.

14:59 And that's quite difficult to transport such a large thing.

15:03 But if we catch it with the same tower

15:06 that it's used to put it in the launch mount to begin with,

15:12 that is the best case outcome for rapid reuse.

15:15 So it literally gets caught by the same arms

15:18 that placed it in the launch ring,

15:22 and then it is placed back in the launch ring immediately.

15:26 So in principle, the super heavy booster

15:28 can be reflown within an hour of landing.

15:33 So it comes back in about five or six minutes,

15:36 one way or another.

15:38 And then it gets caught by the tower arms placed back

15:43 in the launch mount.

15:45 And then you can refold propellant in about 30 to 40 minutes

15:50 and place a ship on top of it.

15:53 And in principle, refly the entire booster every hour.

15:58 Maybe every two hours to give it a little bit of extra time.

16:02 But let's just say it's in the limit of rapid reuse.

16:13 And the next thing we need to do is catch the ship, too.

16:17 So we haven't done this yet, but we will.

16:19 So that's what we hope to demonstrate later this year.

16:37 Maybe as soon as two or three months from now.

16:44 And then the ship would be placed on top of the booster

16:48 and then, again, refold with propellant and flown again.

16:54 With the ship, it takes a bit longer

16:55 because it's got orbit Earth a few times

16:57 until the ground track comes back over the launch pad.

17:01 But the ship is also intended to be reflown multiple times

17:04 per day.

17:09 This is the new Raptor 3, which is an awesome engine.

17:13 We can't tell the Raptor team for this.

17:14 This is very exciting.

17:15 So Raptor 3 is designed to require no basic heat shield,

17:28 saving a lot of mass on the bottom,

17:29 and actually improving reliability

17:31 so that if there is, for example, a small fuel leak

17:35 from the Raptor engine, it will simply

17:38 leak into the existing flaming plasma

17:41 and not really matter.

17:44 Whereas a fuel leak when the engines are contained in a box

17:49 is a very scary thing indeed.

17:51 So this is a Raptor 3.

17:55 It'll take probably a few kicks at the can,

17:57 but it will be as a massive increase in payload capability,

18:02 in engine efficiency, and in reliability.

18:05 So this is really a revolutionary engine.

18:08 Now, Raptor 3 is really, I'd say, kind of alien technology

18:11 rocket engine.

18:12 I mean, even industry experts, when

18:21 we showed a picture of the Raptor 3,

18:23 said that engine is not complete.

18:26 So then we said, well, here's the engine not complete firing

18:30 at a level of efficiency that has never been achieved before.

18:35 So I mean, that is one clean engine.

18:44 So in order to make the engine like that,

18:51 we had to simplify so many parts of the design,

18:54 incorporate secondary fluid circuits and electronics

18:58 in the structure of the engine itself.

19:00 So everything is contained and protected.

19:04 It is a model of engineering correctly.

19:06 Then one of the other technologies

19:14 that's key for Mars is doing orbital propellant transfer.

19:21 So you can think of this similar to aerial refueling

19:24 for airplanes.

19:27 But in this case, it's orbital refilling of rockets,

19:31 which has never been done before.

19:33 But it is technically feasible.

19:35 I always feel like these things are a little NSFW, sort of.

19:45 Listen, you've got to transfer fluid somehow.

19:51 There's no-- this has got to be done.

19:56 So the two starships would get together.

20:03 And one starship would transfer fuel and oxygen.

20:08 And actually, most of the mass is oxygen.

20:10 It's almost 80% oxygen that gets transferred,

20:14 a little over 20% fuel.

20:16 And so you send a starship to orbit that's full of payload.

20:23 And then you send a bunch of other starships up,

20:25 and you would refill the propellant on that starship.

20:28 And once the propellant tanks are mostly full,

20:31 then you can depart for the Mars or the Moon or-- yeah.

20:36 So this is an important technology,

20:38 which we should hopefully demonstrate next year.

20:42 [APPLAUSE]

20:45 So then with-- one of the toughest problems to solve

20:54 is the reusable heat shield.

20:58 So no one has ever developed a truly reusable orbital heat

21:03 shield, so that it's extremely difficult to do so.

21:07 Even the shuttle's heat shield required

21:10 several months of refurbishment, basically

21:13 fixing broken tiles, testing each tile.

21:17 And because it's an extremely hard problem

21:21 to be able to withstand the extreme heat and pressure

21:24 of reentry, and the only things that can really withstand

21:31 this level of heat are advanced sort of ceramics, basically

21:40 glass, alumina, sometimes of carbon carbon.

21:47 But very, very little actually can survive with reusability,

21:52 without getting-- without eroding or falling off or cracking,

21:57 can survive the stresses of reentry.

22:01 So this will be the first time that

22:04 is done, that a reusable orbital heat shield is developed.

22:08 And it needs to be obviously extremely reliable.

22:11 So this will be something they'll be working on for a few years,

22:16 I think, to keep honing the heat shield.

22:20 It's a very-- it is an achievable thing,

22:22 so we're not trying to do something that isn't achievable.

22:24 It is when the realm of physics to get this done,

22:27 just an extraordinarily difficult thing to get done.

22:31 And Mars-- the Mars atmosphere is carbon dioxide,

22:35 which at first may seem better, but actually it ends up

22:40 being worse, because when the CO2 turns into a plasma,

22:46 and you've got-- you actually end up

22:47 with more free oxygen entering on a Mars atmosphere

22:51 than on Earth atmosphere.

22:52 So Earth's atmosphere is only around 20% oxygen,

22:58 and Mars ends up being basically more than double that,

23:02 maybe triple that, when the CO2 becomes a plasma,

23:07 and you get carbon and O2.

23:10 So that wants to oxidize the heat shield,

23:13 basically burn the heat shield.

23:15 So that's why we tested very rigorously in the CO2

23:19 atmosphere, because it's got to work not just for Earth,

23:22 but also for Mars.

23:23 And we want to use the same heat shield for Earth

23:33 that we use for Mars, because there are many other factors

23:36 with the heat shield, such as making sure

23:38 the tiles don't crack or fall off or anything like that.

23:42 So we want to have the same heat shield structure,

23:45 same material on Earth as on Mars,

23:47 so we can test it hundreds of times

23:50 on Earth before going to Mars, and be confident that when

23:53 it goes to Mars, it will work.

24:06 So we're developing some next-generation starships,

24:16 which have a number of improvements versus our current

24:21 gen.

24:22 So it's taller, for example, and has a better kind of the

24:31 interstage between the ship and the booster.

24:34 You can see that sort of the sort of struts there.

24:38 That makes it easier for the flame, like right there.

24:44 When doing hot staging, which is when we light the ship's

24:46 engines while the booster's engines are still firing,

24:49 the flame from the ship engines can more easily exit

24:54 through the open struts of the new interstage.

25:00 And in this case, we'll bring the struts

25:03 interstage back with us instead of throwing it away.

25:05 So a little more height here, 72 meters from around 69.

25:19 Repelent capacity-- I think we'll probably push that up a

25:21 little, maybe 3,700 tons.

25:25 Long term, my guess is we're maybe around 4,000 tons.

25:30 And just sort of over 8,000.

25:36 Every point, like 8,000-- this will keep getting up.

25:39 My guess is, ultimately, we're 4,000 tons here, close to

25:43 10,000 tons of thrust.

25:44 But this is kind of the next level, or the next version

25:48 of the Super Heavy.

25:50 So the booster will look a little naked on the bottom

25:56 because the Raptor 3 engines don't require a heat shield.

26:01 So it looks like there's parts missing, but that's just

26:05 because the Raptor heat shield does not--

26:08 but the Raptor 3 does not need a heat shield.

26:12 So it's just standing there in a bathed in flaming plasma.

26:25 But it's a lot lighter.

26:28 Yep.

26:29 Integrated hot stage.

26:33 I think it looks amazing.

26:34 And then the shift's a little longer, a little more capable.

26:45 Moving to 1,550 tons of propellant capacity.

26:50 My guess is, this probably ends up 20% more than this long term.

26:55 And yeah, you can see the heat shield is sleeker.

27:03 So much smoother boundaries.

27:06 As the boundary of the heat shield going to the leeward side

27:12 is very smooth, no more jagged tiles.

27:20 I think it looks very sleek.

27:21 And this version, we still have six engines,

27:29 but a future version will have nine.

27:32 But with the Raptor 3 again, we have improved reduced mass,

27:39 higher specific impulse.

27:44 The Starship version 3 is really the version that

27:51 is, I think, achieves all of the key elements.

27:55 Generally, with any new technology,

27:56 it takes three major iterations of any new technology

28:02 to have it really, really work well.

28:04 And with Raptor 3 and Starship and Booster version 3,

28:10 it should be able to achieve all of the things that I just

28:12 mentioned, which is a rapidly reusable, reliable rocket

28:16 with orbital refilling.

28:19 Basically, all of the ingredients necessary to make

28:22 life multi-planetary will be achieved with version 3

28:26 of Starship, which we're aiming to launch for the first time

28:29 at the end of this year.

28:31 [APPLAUSE]

28:40 Yeah, so you can see this is where things are on the left,

28:45 where things will be end of this year in the middle.

28:48 And as I was saying, where things will be probably long term,

28:54 yeah, 142 meters.

28:58 So the one in the middle will be fully capable of doing Mars.

29:05 And thereafter, it'll be a lot of performance

29:09 improvements.

29:12 And as has been the case with Falcon 9,

29:16 we always end up making the rocket

29:17 longer and increasing payload.

29:20 So that's the game plan, pretty straightforward.

29:26 But it's important to emphasize, even with the rocket

29:29 that we'll be launching just at the end of this year,

29:31 it will be capable of making life multi-planetary.

29:34 And thereafter, it's just about continuing

29:37 to hone the efficiency and capability of the rocket

29:40 and reduce the cost per ton and reduce the cost per person

29:43 to Mars.

29:44 And like I said, ultimately, make

29:46 it so that anyone who wants to move to Mars

29:48 and help build a new civilization can do so.

29:53 So anyone out there, how cool would that be?

29:57 And even if you don't want to do it,

29:58 maybe that you have a son or daughter who wants to do that

30:02 or a friend who wants to do it.

30:05 And I think it would be the adventure,

30:08 but the best adventure that it, when it could possibly do,

30:12 is to go and help build a new civilization on a new planet.

30:16 [APPLAUSE]

30:18 So yeah, ultimately, we'll have 42 engines,

30:27 which it was inevitable.

30:32 As the prophecy foretold by the great prophet Douglas Adams

30:38 in his book, "The Hitchhiker's Guide to the Galaxy,"

30:45 the answer to the meaning of life is 42.

30:50 And so inevitably, the Starship stack will have 42 engines.

30:54 [APPLAUSE]

31:02 And in terms of payload to orbit, what's remarkable

31:05 is that it's 200 tons--

31:08 Starship will have 200 tons payload to orbit

31:10 with full reusability.

31:12 So this is twice the capability of the Saturn V moon rocket.

31:16 Saturn V moon rocket was fully expendable.

31:20 Starship is fully reusable, but will have twice the payload

31:23 to orbit of the next biggest rocket that made it to orbit,

31:28 which is a Saturn V.

31:31 In fact, with out reusability, Starship

31:35 would have about 400 tons of payload to orbit.

31:39 So this is a very big rocket, is what I'm saying.

31:45 But you need a big rocket to make life multi-planetary.

31:51 And then along the way, we could do very cool things,

31:53 like have a moon base, like moon base alpha.

31:57 Long ago, there was a TV show about moon base alpha.

32:03 You couldn't think about the physics of that too much,

32:06 because apparently the moon base was drifting away from Earth.

32:09 But anyway, we should have a moon base alpha,

32:12 which is the next step after the Apollo program

32:14 would be to have a base on the moon.

32:16 So you could have a gigantic science station

32:23 doing research about the nature of the universe on the moon.

32:26 And it would be very cool.

32:27 So in terms of-- like, when can you go to Mars?

32:33 So you can go to Mars every two years, or every 26 months.

32:36 So the next Mars opportunity is at the end of next year,

32:41 in about 18 months.

32:42 So November, December is the next Mars opportunity.

32:45 So we'll try to make that opportunity.

32:47 If we get lucky, I think we'll probably

32:48 have a 50/50 chance right now, because we've

32:52 got to figure out orbital refilling in order

32:55 to have enough capability to go to Mars.

32:58 But if we achieve orbital refilling in time,

33:01 then we will launch the first uncrewed Starship

33:05 to Mars at the end of next year.

33:07 [APPLAUSE]

33:09 So this gives you an illustration of how does the spaceship

33:19 go from Earth to Mars?

33:20 So you've got blue Earth there and red Mars.

33:25 And I mean, the actual distance traveled on the arc

33:37 is close to it like 1,000 times further than the moon.

33:43 So you can't just go straight to Mars.

33:47 You have to create this elliptical orbit with Earth

33:50 at one point and Mars at the other side,

33:52 at the far end of the ellipse.

33:54 And then time the where you are and that ellipse

33:59 to intersect with Mars.

34:01 And this-- so this is the orbital transfer--

34:05 how you do orbital transfer from Earth to Mars.

34:08 And if you look on your Starlink Wi-Fi router,

34:11 you'll see this image.

34:13 Because the Starlink Wi-Fi or Starlink internet

34:17 is what's being used to pay for humanity getting to Mars.

34:23 So I'd just like to thank everyone out there who

34:24 is board Starlink because you're helping secure

34:28 the future of civilization and helping

34:31 make life multi-planetary and helping

34:33 make humanity a space-braying civilization.

34:35 Thank you.

34:36 [APPLAUSE]

34:39 So this is a tentative game plan here

34:48 where we're hoping to achieve, where

34:53 we increase the cadence of flights to Mars dramatically

34:58 with every launch window.

34:59 So every roughly two years, we are dramatically

35:05 increasing the number of ships that go to Mars.

35:10 And ultimately, try to get to 1,000 or 2,000 ships per Mars

35:18 rendezvous.

35:19 I mean, as a rough order of magnitude-- this just guesses,

35:26 obviously, but we need to get about a million tons,

35:29 as my guess, to the surface of Mars

35:31 to make a civilization on Mars self-sustaining.

35:36 And again, to that critical point,

35:37 where if the resupply ships from Earth

35:40 stop coming for any reason, Mars still succeeds.

35:43 Mars can still grow.

35:45 And so you can't be missing anything.

35:46 You can't be missing even the equivalent of vitamin C

35:48 or anything.

35:49 You've got to have everything you need for Mars to grow.

35:51 That's essential.

35:55 So my guess is that's about a million tons,

35:56 but it might be 10 million tons.

35:58 I hope it's not 100 million tons.

36:00 That'd be a lot.

36:02 But we want to try to get to that point

36:05 and secure the future of civilization as quickly as

36:07 possible.

36:08 So we're looking at different locations.

36:14 The lead candidate right now is the Arcadia region.

36:18 So Mars has a lot of real estate,

36:21 but when you combine all of the factors and say, OK,

36:26 we can't be too close to the poles.

36:28 We need to be near ice to get water.

36:34 And can't be too mountainous for the rockets,

36:39 then it narrows down to a smaller region.

36:42 So Arcadia is one of my daughters, the name is Arcadia,

36:47 actually, and is one of the options.

36:51 So we've got the first starships on Mars,

36:54 gather critical data, which are the first lights there

36:57 we'll send with the Optimus robot.

37:00 So we can go out there and explore and prepare the way

37:03 for humans.

37:04 And that'll be a very cool image if we're

37:07 able to achieve it by launching at a next year.

37:10 We would actually technically arrive in 2027,

37:13 but that would be an epic fixture

37:17 to see Optimus walking around on the surface of Mars.

37:20 And then with launching two years later,

37:29 we would be sending humans.

37:30 Seeing the first missions are successful,

37:33 and they land successfully.

37:34 We would send humans on the next mission.

37:37 And we'd really start building the infrastructure

37:40 for Mars.

37:42 So anyway, maybe just to be safe and we

37:45 might just do two landing episodes with the Optimus

37:50 and do the third one with humans, we'll see.

37:54 So that classic picture of the workers on the Empire state.

37:58 And then for communications on Mars,

38:06 we'll be using a version of Starlink

38:08 to provide internet on Mars.

38:10 Yeah, so the speed of light--

38:25 even at the speed of light, the best case scenario

38:28 is, I guess, running around 3 and 1/2 minutes to Mars.

38:32 And then worst case is 22 minutes or more,

38:36 because Mars is on the other side of the sun from Earth.

38:41 So anyway, it's quite challenging

38:45 to do high bandwidth communications with Mars.

38:48 But Starlink will achieve that.

38:49 Yeah, and then we'll have the first humans

38:55 lay the groundwork for permanent presence on the surface.

38:57 And yeah, the goal, like you said,

39:02 will be to make Mars self-sustaining as quickly as possible.

39:06 So it's just a rough idea of what things will be like

39:15 for the first city on Mars.

39:17 My guess is we'll probably put the launch pads

39:18 a little further away with the landing pads, just in case.

39:24 But in four Mars, we're going to need a lot of solar power.

39:29 You can't really walk around on the surface of Mars,

39:36 at least as yet, until Mars is terraforming to be like Earth.

39:41 You need to walk around with a Mars suit

39:44 and be initially in glass domes.

39:49 But it would work.

39:51 And eventually, we can make Mars into an Earth-like planet.

39:54 And we want to get to the point where we're transferring

40:02 over a million tons at every Mars transfer window.

40:06 And then we-- that's like a serious civilization.

40:10 A megaton per transfer window.

40:12 So yeah, we have a lot of spaceports.

40:22 I mean, because of the fact that you can't fly there

40:25 continuously and you have to transfer in these windows,

40:28 you'd have a gathering of 1,000 ships, or 2,000 ships,

40:31 or more than that.

40:33 So we're looking at this kind of like Battlestar Galactica

40:35 feel, where all these ships are in orbit,

40:36 where they need to depart.

40:38 And then they're all depart, I think

40:44 an amazing image of all these ships departing at once.

40:47 And then you're going to need, obviously, a lot of landing

40:51 pads on Mars.

40:53 Or you'll need to move the ships off the landing pad

40:57 pretty fast.

40:59 So if you've got, I don't know, a few thousand ships

41:02 inbound, probably need at least a few hundred

41:06 pads, landing pads.

41:08 And anyway, we'll solve that problem later.

41:14 So yeah, anyway, this is like an incredible thing

41:22 to have this amazing city on Mars.

41:26 The first city on another planet.

41:29 And a new world.

41:35 And it's also an opportunity to, I think,

41:40 for the Martians to rethink how they want civilization to be.

41:47 So you can maybe rethink, like, what kind of former government

41:50 do you want?

41:51 What new rules do you want to have?

41:55 There's a lot of freedom and opportunity in Mars

41:57 to do a recompile on civilization, which

42:01 will be up to the Martians.

42:02 So all right, let's get it done.

42:09 [APPLAUSE]

42:13 Thank you, everyone.

Builder

· 1M

埃隆·马斯克：“5 年后，数字智能的总和将超过人类智能的总和。”

在五年之内，人类形机器人的数量可能会达到至少 1 亿台，甚至可能达到 10 亿台。

由于人工智能和机器人技术的应用，生产力可能会大幅提升，因此该国的经济规模有可能在 5 到 7 年内实现翻倍。

变革的步伐将会快得惊人，短短几年之内，世界可能会变得面目全非。

中文 English

0:00 五年内，差不多，也就是说，大约在2031年，我认为

0:11 数字智能

0:12 将超过所有人类智能的总和，五年内可能会有，

0:22 大概，

0:24 可能至少有一亿个类人机器人，也许高达十亿。

0:36 我预测经济规模在五年内可能是当前的两倍，

0:49 可能六年

0:51 五到七年内，你将进入一个经济产出快速翻倍的时期，

0:59 经济

1:00 产出增长如此之快，前后几年的误差，你会看到巨大的

1:06 变化。

Builder

· 1M

全球去中心化金融领域的总价值锁仓量较 10 月的峰值下降了 50%

自去年 10 月达到历史最高点以来，全球去中心化金融领域的总价值已正式下降了 50%以上。

最新数据显示，总 TVL 仍面临巨大压力，有数十亿美元的资本流向了那些受监管的、能带来收益的金融产品。

Builder

· 1M

那位亿万富翁 CZ 刚刚在 CNBC 上进行了现场直播：

“我认为，在 2026 年下半年，我们将见证历史上最庞大的加密货币超级周期。”

我们谈论的是加密货币与比特币领域中最重要的人物之一。

问题是……他到底知道些什么呢？

👀

中文 English

0:00 你最近经常交易加密货币吗？

0:03 不，我根本不交易。

0:05 我不交易，我只是持有大币，持有B&B。

0:09 我不做日内交易，我不买卖，

0:10 我不试图抓住市场。

0:12 我就是不擅长那种事。

0:13 多年前，大约20年前，我试过，亏了钱，

0:17 后来想，我是开发者，做系统的，不适合交易。

0:21 那么，能不能简单说说，

0:23 你怎么看比特币的价格和未来走向？

0:26 你觉得我们处于冬天、夏天，还是某个周期？

0:30 你支持凯西·伍德的观点吗，认为比特币会涨

0:32 到一百万美元？

0:34 每当有人问这个问题时，

0:35 我总要问他们，时间范围是多久？

0:38 如果你看的是今天、明天或日常，

0:40 我根本无法预测。

0:42 如果你看的是五年、十年的时间段，

0:44 很容易预测，我们会涨。

0:48 如果看一年时间，

0:49 这也是最常见的视角，

0:54 2026年，我不确定，也许吧，我很有信心

0:59 那会是一个超级周期。

1:01 - 你说超级周期？

1:02 - 是的，2026年。

1:03 - 针对比特币。 - 针对比特币。

1:04 - “超级周期”对你来说是什么意思？

1:06 - 通常比特币历史上遵循四年周期。

1:09 看历史数据，每四年都有新高，

1:11 然后下跌。

1:14 但我认为今年，鉴于美国如此支持加密，

1:18 其他国家也有跟进，

1:20 我觉得我们可能会看到，可能甚至会提前突破

1:23 传统周期的高点。

0:00 Do you trade a lot of crypto these days?

0:03 No, no, I don't trade at all.

0:05 I don't trade, I just, I hold big coins, I hold B&B.

0:09 I don't do day trading, I don't buy and sell,

0:10 I don't try to attend the market.

0:12 I'm just not good at that.

0:13 Years ago, like 20 years ago, I tried that, lost money,

0:17 just figured, I'm a developer, I do systems, I don't trade.

0:21 So, do you have any thought though, real quick,

0:23 on just like the price of Bitcoin and where this is all headed?

0:26 Do you think we're in a winter, a summer, a cycle?

0:30 Are you with Kathy Wood and thinks it's going to turn

0:32 into a million dollars of a coin?

0:34 So, whenever people ask that question,

0:35 I always got to ask them, what's the time frame you're looking at?

0:38 If you're looking at today, tomorrow or a daily basis,

0:40 there's no way I can predict.

0:42 If you look at a 10, five, 10 year horizon,

0:44 it's very easy to predict, we're going to go up.

0:48 If you look at a one year horizon,

0:49 which is like what the most common framework is,

0:54 2026, I'm not sure, maybe, I think I have very strong feelings

0:59 that you will probably be a super cycle.

1:01 - You're super cycle?

1:02 - 2026, yeah.

1:03 - For Bitcoin. - For Bitcoin.

1:04 - And what super cycle to you means what?

1:06 - Well, normally Bitcoin follows four year cycles historically.

1:09 If you look at historic data, like every four years,

1:11 you know, there's an all-time high and then there's a drop.

1:14 But I think this year, given the US being so pro-crypto

1:18 and every other country's kind of following,

1:20 I do think we will see this, we will probably break

1:23 before your cycle.

Builder

· 1M

对 Solana 和 XRP 的持有者来说，这是个坏消息：

🇺🇸

高盛刚刚公开区分了赢家和输家。

已售出：所有 Solana ETF 持仓。
已售出：所有 XRP ETF 持仓。

保持：对比特币的持有。
保留：以太坊的基础设施依然发挥着重要作用。

贝莱德：在以太坊领域投资了 70 亿美元。比特币 ETF 依然占据主导地位。
高盛：退出 SOL 和 XRP 业务。

美国两家最强大的银行。
与比特币和以太坊保持一致。
忽略其他所有内容。

这并非市场信号。
这是一份定罪声明。

Builder

· 1M

“以太坊是实现代币化的必经之路。”

约瑟夫·查洛姆称，拉里·芬克致力于推动代际更替。

稳定币是最早的实践成果。

现在，其他的一切就都顺其自然了吧：

🔹

股票

ETFs

抵押品/担保物

全天候开放的市场

纽约证券交易所、纳斯达克、DTCC 以及 Bullish 的走势都一致。

黑岩集团认为，市场结构的变化即将到来。

中文 English

0:00 我认为拉里专注于代际变革，我觉得他的描述方式是

0:05 稳定的

0:07 稳定币是代币化的第一个证明。

0:11 稳定币与代币化没有区别，它只是代币化的美元

0:15 很快还会有

0:16 港币、韩元和日元。

0:19 这是代币化的首个案例，如果你能编程某样东西，

0:23 即时

0:24 结算，这东西恰好是现金，这可以改变金融服务和

0:29 资金

0:30 的流动方式。

0:31 第二点，在某种程度上，这为眼下即将到来的事做了练习，

0:35 那就是所有资产的代币化。

0:36 他在达沃斯展示了一系列关于金融未来的幻灯片。

0:37 其中一张说代币化将改变世界，以太坊是

0:42 通向代币化的收费公路，这是幻灯片顶部的引用。

0:48 我认为撇开价格差异不谈，

0:49 从未有过一个时期，信号如此明朗，表明许多资产

0:52 的代币化正到来。

0:57 我给你五个过去几周的证明点。

1:00 纽约证券交易所转为24小时代币化，纳斯达克也转为24小时。

1:01 DTCC表示他们能使抵押品无纸化，代币化，并允许

1:02 人们分批移动，

1:06 每次15分钟。

1:11 就在本周早些时候迈阿密的这栋楼和共识大会上，你听到

1:16 有人看好买下了Quinity。

1:17 这里没有人听说过Quinity。

1:20 Quinity是2900只上市股票的枯燥过户代理，包括Chainlink。

1:26 他们会以模拟形式持有账簿，会支付你的股息。

1:27 这不是玩笑，这是一个极其看涨的信号，说明有一家本土加密公司

1:29 拥有上市股票，

1:32 相信他们能更好、更快地将2900只股票代币化，

1:37 并大大加快链上处理，不需一家家公司慢慢来。

1:38 这是一个大胆的声明。

1:44 当这发生时，稳定币正在增长，代币化就在眼前，

1:49 逐步

1:50 实现代币化功能。

1:55 你将开始看到这些资产进入去中心化金融（DeFi）。

1:56 所以你可以借贷的不仅是稳定币和加密货币，还有特斯拉股票

2:02 和ETF。

2:04 我认为拉里更专注于市场结构的变化，因为

2:07 最终

2:08 这对积极参与的机构是极其正面的，

2:10 对投资者而言，则在风险与回报上极其正面。

2:13 所以我认为他少关注迈克尔·塞勒，更关注接下来

2:19 未来十年会发生什么，他想引领这个趋势。

2:20 And I think Larry's more focused on that market structure change, because

2:23 ultimately

2:24 it's going to be incredibly positive for institutions who lean in and

2:27 incredibly positive

2:29 for investors in terms of risk and return.

2:31 So I think he's less focused on Michael Saylor, more focused on what is going

2:36 to happen over

2:36 the next 10 years, and he wants to lead.

0:00 I think Larry is focused on generational change, and I think the way he's

0:05 described it is stable

0:07 coins is the first proof point for tokenization.

0:11 Stable coins are not different than tokenization, it's just tokenized dollars

0:15 and soon to be

0:16 Hong Kong dollars and Korean Wan and Yan.

0:19 It is the first proof case for tokenization that if you can program something,

0:23 settle

0:24 it instantly, it happens to be cash, that can change how financial services and

0:29 money is

0:30 moved.

0:31 The second thing is in some ways that's practice for what's coming right now,

0:35 and that is tokenization

0:36 of all assets.

0:37 He in Davos had a series of slides talking about the future of finance.

0:42 One of the slides was tokenization is going to change the world, and Ethereum

0:48 is the toll

0:49 road to tokenization, that's the quote from the top of the slide.

0:52 And I think you're seeing put aside price divergence.

0:57 There's never been a period of time where the signals are clear that token

1:00 ization of

1:01 many assets are coming.

1:02 I'll give you five proof points from the last several weeks.

1:06 New York Stock Exchange going 24/7 on tokenization, NASDAQ going 24/7.

1:11 The DTCC saying they can dematerialize collateral, tokenize it, and allow

1:16 people to move it in

1:17 while it's 15 minutes at a time.

1:20 And in this building here in Miami and Consensus earlier this week, you heard

1:26 that bullish is

1:27 buying a quinity.

1:29 No one in this building had heard about a quinity.

1:32 Quinity is the boring transfer agent for 2,900 public equities, including sharp

1:37 link.

1:38 They will hold a ledger in analog form, they will pay your dividends.

1:44 This is a no pun intended, a massively bullish signal that you have a crypto

1:49 native company

1:50 with public stock who believes they can do a better job, quickly tokenizing 2,

1:55 900 public

1:56 equities, and bring it on chain much faster than going one company at a time.

2:02 It is a bold statement.

2:04 And when that happens, stable coins are growing, tokenization is right in front

2:07 of us, step

2:08 function tokenization.

2:10 You'll start seeing some of these assets make their way into DeFi.

2:13 So you can borrow and land not just stable coins and crypto, but Tesla stock

2:19 and an ETF.

2:20 And I think Larry's more focused on that market structure change, because

2:23 ultimately

2:24 it's going to be incredibly positive for institutions who lean in and

2:27 incredibly positive

2:29 for investors in terms of risk and return.

2:31 So I think he's less focused on Michael Saylor, more focused on what is going

2:36 to happen over

2:36 the next 10 years, and he wants to lead.

Builder

· 1M

所有机构都选择了同一条链

72.6%的所有代币化 ETF 运行在以太坊上。

全球 ETF 市场：$20,000,000,000,000。
迄今为止已代币化：$441,900,000。

0.002% 链上。
99.998% 仍在等待。

Subtitle processing failed

Builder

· 1M

CZ：我们仍然将 70%–80%的资本投资于区块链

在 5 月 7 日于 Consensus Miami 2026 的演讲中，币安创始人赵长鹏表示，他和 Yzi Labs 仍然将 70%到 80%的资本分配给区块链，因为区块链仍然是他们的核心专业领域。他说，人工智能代理将创造未来支付的重大需求，而区块链是它们更原生和自然的支付工具。

CZ 还表示交易所应该为 AI 代理做准备：用户应该能够使用自然语言让 AI 代理审查投资组合并执行交易，而不是手动阅读图表并输入价格和数量。

中文 English

0:00 所以，对我来说，对于Easy Labs，我们仍然将70-80%的资金投向区块链。

0:04 那里是我们的核心专长所在。

0:05 但从根本上讲，人工智能你的智能体可以与另一个

0:09 世界另一端的智能体互动。

0:09 我是一个来自，不知道，新加坡的人，想去迈阿密参加一个

0:12 会议。

0:13 智能体会花时间帮你找，比如说，这里很贵，而且都订满了。

0:15 全部这些都应该由智能体来完成。

0:17 然后当那个人想在新加坡付款时，信用卡可以用，

0:18 但智能体不知道怎么刷你的信用卡。

0:20 信用卡没有API。

0:22 所以智能体最本地化的工具显然是区块链。

0:24 这是最自然的技术。

0:27 我认为未来对智能体的支付需求会很大。

0:28 智能体会做更多微交易。

0:31 今天如果你打开币安App，有一堆图表，

0:31 一堆数字，然后你得选一个图表，放大，

0:33 看价格，在手机上输入价格，

0:35 输入数量，然后点击红色或绿色按钮。

0:38 这非常不直观。

0:40 所以你应该直接拿起手机说，“给我看我的投资组合

0:43 。”

0:43 过去24小时的情况，以及最新消息，

0:44 例如《明确法案》通过了。

0:47 然后你说，“好的，我想把10%的稳定币换成B&B

0:47 。”

0:49 如果你持仓量大，智能体会在后台帮你操作。

0:51 技术上今天已经有条件实现，只是没人做。

0:54 我其实挺沮丧没人做这个。

0:54 这就是智能体和你的互动方式。

0:56 更自然的人机交互。

0:57 未来，应该就是智能体在交易。

1:00 所以所有这些变化，都需要支持智能体。

1:00 很多不同的东西，我觉得人工智能是改变世界多方面的基础技术，

1:01 但人工智能仍然需要资金。

1:03 所以区块链不会消失。

1:05 我们不会因为有了智能体就突然不需要钱了。

1:06 钱是经济的基本概念，驱动一切。

1:08 所以区块链不会消失。

1:11 我看到很多区块链开发者转向人工智能。

1:13 这没问题，你们可以做人工智能。

1:14 但看看NVIDIA，多年里显卡并不是最热门的东西，

1:15 We're not going to suddenly not need money because we have agents.

1:17 CPU才是，对吧？

1:20 我们花了40年才到现在，现在人工智能来了，

1:22 并且他们成了全球最有价值的公司。

1:24 所以我认为对区块链开发者来说，资金不会消失。

1:26 所以对我来说，对于Easy Labs，我们仍然将70-80%的资金投向区块链。

1:30 那里是我们的核心专长所在。

1:30 然后20%投人工智能，也许10%投生物技术。

1:31 So they took 40 years for us to come to now, and now the AI come around,

1:35 and they're the most valuable company in the world.

1:36 So I think for blockchain builders, money's not going away.

1:38 So for me, for easy labs, we still invest 70, 80% of our phones in blockchain.

1:42 That's where our core expertise are.

1:44 And then the 20% AI, and then maybe 10% in biotech.

0:00 So, for me, for easy labs, we still invest 70-80% of our phones in blockchain.

0:04 And that's where our core expertise are.

0:05 But AI fundamentally, your agent can interact with another agent across a

0:09 different part of the world.

0:09 I'm a guy from, I don't know, Singapore wants to travel to Miami for a

0:12 conference.

0:13 And the agent will spend all the time finding you, like, no, it's pretty

0:15 expensive here and they're fully booked.

0:17 All of that should be done by the agent.

0:18 And then when that guy wants to pay in Singapore, credit card will work,

0:20 but the agent doesn't know how to swipe your credit card.

0:22 Credit cards don't have an API.

0:24 So the most native thing for the agent to use is obviously a blockchain.

0:27 It's the most natural technology.

0:28 I think there's going to be a lot of agent requirement for payments in the

0:31 future.

0:31 And the agent is going to do much more microtransaction.

0:33 Today, if you open the Binance app, there's a bunch of charts,

0:35 there's a bunch of numbers, and then you have to select one chart, zoom in,

0:38 figure out the price, type in the price on the phone,

0:40 type in two numbers of price on the quantity, and then hit a red or green

0:43 button.

0:43 That's very counterintuitive.

0:44 So what you should do is just open up your phone and say, "Show me my portfolio

0:47 ."

0:47 We're having the last 24 hours, and when news are there,

0:49 clarity act gets passed.

0:51 And then you say, "Okay, I want to convert 10% more of my stable coins into B&B

0:54 ."

0:54 And if you have a large position, that's just happening in the background with

0:56 an agent.

0:57 The technology is there to make it work today, just so that nobody has built it

1:00 yet.

1:00 I'm actually quite frustrated that nobody has built it.

1:01 So that's now agents interacting with you.

1:03 It's a much more natural human interaction.

1:05 In the future, it should just be agents trading.

1:06 So all this change, it needs to be agent-ready.

1:08 A lot of different stuff that I think AI is a fundamental technology

1:11 that can change many different parts of our world,

1:13 but AI still needs money.

1:14 So the blockchain is now going away.

1:15 We're not going to suddenly not need money because we have agents.

1:17 Money is a fundamental concept of the economy, which drives everything.

1:20 So the blockchain is now going away.

1:22 So I see a lot of blockchain builders that go to AI.

1:24 That's fine, you know, you can be building AI.

1:26 But if you look at NVIDIA, for years, graphics cards were not the sexiest thing

1:30 .

1:30 CPUs were, right?

1:31 So they took 40 years for us to come to now, and now the AI come around,

1:35 and they're the most valuable company in the world.

1:36 So I think for blockchain builders, money's not going away.

1:38 So for me, for easy labs, we still invest 70, 80% of our phones in blockchain.

1:42 That's where our core expertise are.

1:44 And then the 20% AI, and then maybe 10% in biotech.

Builder

· 1M

Ai在 1、2 或 3 年内会发展到哪里？

中文 English

0:00 让我跟你分享一些想法。

0:03 第一点是，在过去几个月里，

0:07 AI工具的生产力发生了质的飞跃，

0:10 变得大大提升了。

0:13 它比九个月前强大得多。

0:20 对我们Citadel来说，这使得

0:23 我们能释放更多AI的应用场景。

0:29 这非常有趣，

0:31 坦白说，以前我们通常

0:35 用硕士、博士和金融专业人员

0:39 几周甚至几个月完成的工作

0:43 现在由AI代理在几小时或几天内完成。

0:47 这并不是中等水平的白领工作，

0:52 而是极高技能的工作。

0:56 我用“由AI代理自动化”形容这个词。

1:02 说实话，有个星期五我回家时，

1:07 对这一点感到相当沮丧，

1:11 因为你能明显看到这将对

1:15 社会产生多么剧烈的影响。

1:18 当你亲眼见证这点时，

1:20 看到过去几周或几个月的工作

1:23 几天内就被完成，真是震惊，

1:27 那是我第一次看到我们内部真正的影响。

0:00 Let me just share a few thoughts with you on this.

0:03 Number one is in the last few months,

0:07 there has been a step change function

0:10 in the productivity of the AI toolkit.

0:13 It is profoundly more powerful than it was just nine months ago.

0:20 And for us at Citadel, that has allowed

0:23 us to unleash a much broader array of use cases for AI.

0:29 And it has been really interesting

0:31 to watch, to be blunt, work that we would usually

0:35 do with people with masters and PhDs and finance

0:39 over the course of weeks or months being done by AI agents

0:43 over the course of hours or days.

0:47 So these are not, these are not mid-tier white-collar jobs.

0:52 These are like extraordinarily high-skilled jobs.

0:56 I'm going to pick a word being automated by agenic AI.

1:02 And I got to tell you, I went home one Friday, actually

1:07 fairly depressed by this, because you could just

1:11 see how this was going to have such a dramatic impact

1:15 on society.

1:18 And when you win this in your own four walls,

1:20 when you see work that used to be manures of work being done

1:23 in days or weeks, it's like, wow,

1:27 that's the first time I've seen real impact in our four walls.