The New Primitives: Building AI Native Software — Kwindla Kramer, Daily

频道: AI Engineer
视频: https://www.youtube.com/watch?v=LZuWZRze3MU
原文语言: en
统计: 共 65 轮

[0:01]

[music]

[音乐]

[0:12]

Good morning. I know a lot of you in this room. It's great to see you. Welcome to the voice track at AI Engineer World's Fair. For those of you who don't know me, my name is Quinn La Holman Cramer. I work at a company called Daily. We make developer infrastructure for real-time audio, video, and AI. And we're the team behind Pipe Cat, which is the most widely used framework for building voice agents today. Pipe Cat is open source and vendor neutral. It's used by companies like AWS and Nvidia and Anthropic and thousands of startups and scale-ups and enterprises. And today I'm going to talk about what kind of agents we're building today, including voice agents, but not just voice agents, and what I'm interested in building next. And I'm going to try to put all this in the context of the roughly 80-year history of digital computing so far. So, we've got a lot to cover. We're going to go fast. But we're going to start in 1945 with an essay called As We May Think, written by an engineer, an academic, a civil servant named Vannevar Bush. Bush deeply understood technologies ranging from analog computers to photography to radio to radar. As We May Think is a extraordinary piece of writing.

早上好。在座有很多人我都认识，很高兴见到大家。欢迎来到 AI Engineer World's Fair 的语音专场。不认识我的朋友，我叫 Kwindla Hultman Kramer，在一家叫 Daily 的公司工作。我们做的是实时音频、视频和 AI 的开发者基础设施，也是 Pipecat 背后的团队——Pipecat 是今天使用最广泛的 voice agent 开发框架，开源，而且不绑定任何厂商。AWS、Nvidia、Anthropic 这样的公司，还有成千上万的初创公司、成长期公司和大企业都在用它。今天我想聊两件事：我们现在在造什么样的 agent（包括 voice agent，但不只是 voice agent），以及我接下来想造什么。我会尽量把这些放进数字计算大约 80 年的历史里来讲。要讲的内容很多，节奏会比较快。我们从 1945 年开始，从一篇叫《As We May Think》的文章讲起，作者是 Vannevar Bush——一位工程师、学者，也是公职人员。从模拟计算机到摄影、无线电再到雷达，Bush 对这些技术都有极深的理解。《As We May Think》是一篇非同寻常的文字。

[1:31]

The essay predicts the development of, among other things, document display on a screen and document scanning and OCR and speech-to-text and text-to-speech and programming languages and hypertext and search engines and data networks. Something like the GoPro camera, something weirdly like the Amazon Kindle store, and voice interfaces and brain computer interfaces. And I've been thinking a lot about As We May Think lately because Bush wrote this essay right at the very beginning of the computing age. And I think it feels to most of us like we're working right at the beginning of a new age, the intelligence age. So, what will we build? Well, at the moment we're building agents and we're having a lot of fun doing it. And a lot of the AI engineering work we're all talking about this week is focused on building a full coherent software stack for AI agents. Here is Satya Nadella talking a couple weeks ago on a crossover episode of the No Priors and Latent Space Pod about the challenges of building agents in 2026.

这篇文章预言了很多东西：屏幕上的文档显示、文档扫描和 OCR、speech-to-text 和 text-to-speech、编程语言、hypertext、搜索引擎、数据网络，还有类似 GoPro 的相机、一个诡异地接近 Amazon Kindle 商店的构想，以及语音界面和脑机接口。我最近老在想《As We May Think》，因为 Bush 写下这篇文章的时候，正是计算时代刚刚开始的那一刻。而我觉得我们大多数人今天的感受也差不多——我们正站在一个新时代的开端，智能时代的开端。那我们会造出什么？眼下我们在造 agent，而且造得很开心。这周大家聊的很多 AI engineering 工作，重点都是为 AI agent 搭出一整套完整、自洽的软件栈。下面这段是 Satya Nadella 几周前在 No Priors 和 Latent Space 两档播客联合录制的那期节目里，谈 2026 年造 agent 的难点。

[2:38]

That's sort of

差不多就是这样——

[2:39]

That's right. So, so in some sense you kind of want to harness to define the models, the the data, uh and the tools. And so that you have a loop across those three. And so what we are trying to first of all make sure is each of our products that we build, right? Whether it's GitHub Copilot or the security copilot the stuff we showed with M dash or even the discovery for science, it doesn't matter. All of them are multimodal harnesses um with tools access so that you can do this progressive uh disclosure of tools even so that they're token efficient. Uh and then you're feeding it with very rich context.

没错。所以从某种意义上说，你需要一套 harness 来把模型、数据和工具这三样东西定义清楚，让它们之间形成一个循环。我们首先要确保的是，我们做的每一个产品——不管是 GitHub Copilot、Security Copilot，还是我们用 M 系列展示的那些东西，甚至是面向科学发现的产品，都一样——全都是带工具访问能力的多模态 harness，能对工具做渐进式披露，这样才够省 token。然后你再往里面喂非常丰富的 context。

[3:20]

So, if you were here last year at AI Engineer World's Fair, you could draw a through line from the things we were talking about last year to loops and tool calls and context engineering and the stuff we're focused on this year to some emerging ideas. Uh you can hear that in Nadella's clip just there. I think of this is kind of agents plus plus, like multi-model harnesses and software copilot embedded in every single piece of software, organization level harnesses. So, how do we go from agents to agents plus plus to the next thing beyond agents. Well, the last time we had this kind of massive change in how we write software and what we write software for and to do was the early days of the World Wide Web. And I was around for the early days of the World Wide Web. I was a baby programmer in 1995 and the thing we talked about all the time in 1995, the way we talk about agents today, is web pages. I spent a lot of time writing HTML by hand and building web server software in C and indexing and search software in C and authoring tooling and management infrastructure for web pages in Pearl. I was as excited about HTML in 1995 as I am about agents today. And the web page is still with us and it's still important and useful. But today we talk a lot more about web applications and native mobile applications than we talk about web pages.

如果你去年也在 AI Engineer World's Fair，你能画出一条从去年到今年的连线：从 loop、tool call、context engineering，一路连到今年我们关注的这些刚冒头的想法。刚才 Nadella 那段里就能听出来。我把这些叫做 agent++——多模型 harness、嵌进每一个软件里的 copilot、组织层面的 harness。那我们要怎么从 agent 走到 agent++，再走到 agent 之后的下一样东西？上一次「软件怎么写、写来干什么」发生这种量级的剧变，还是万维网早期那段时间。那段日子我是亲历者。1995 年我还是个刚入行的程序员，而 1995 年我们天天挂在嘴边的东西，就跟今天我们谈 agent 一样，是网页。我花了大量时间手写 HTML，用 C 写 web server，用 C 写索引和搜索软件，用 Perl 写网页的创作工具和管理系统。1995 年我对 HTML 的兴奋，跟今天对 agent 一模一样。网页今天还在，依然重要、依然有用。但今天我们谈 web 应用、谈原生移动应用，远比谈网页要多。

[4:44]

So, just like we went from web pages to full-blown web and native mobile, clearly we're going to chart a path to a new fully AI native software that comes after agents and agents plus plus. So, let's keep going back in time to in order to think about this future. Here's a timeline Vannevar Bush lays out and as we may think, he talks about the abacus, which was both an immensely useful device for doing practical everyday mathematical calculations and also an incredibly important theoretical tool that led to ideas like numeric place value and the concept of zero. And Bush talks about the massive jump from the abacus to the state-of-the-art electromechanical keyboard calculating machines that he had in 1945. And then he posits that we're about or he is about to witness and help create an equally large leap to what he calls the arithmetical machine. And then he goes a step even further than that and he invents or designs in that essay a device he calls the memex. And we have a little bit of an advantage over over Bush in 1945. We've seen 80 years of computing play out. So, we can modify his timeline a little bit. We can go from the abacus to the stored program computer to 40 years later the personal computer and 40 years after that this AI agents era that we're all collectively helping to invent and create and bring into being.

所以，就像我们从网页一路走到成熟的 Web 和原生移动应用一样，我们显然也会走出一条路，通向 agent 和 agent++ 之后那种真正 AI 原生的新软件。为了想清楚这个未来，我们继续往回看。这是 Vannevar Bush 在文章里画出的时间线。他从算盘讲起——算盘既是做日常实用计算的极好用的工具，也是一件极重要的理论工具，催生了数位制、零这样的概念。然后 Bush 讲了从算盘到他 1945 年手上那种最先进的机电式键盘计算机器，是多大的一次跃迁。接着他断言，人类——或者说他本人——即将见证并参与创造一次同等量级的飞跃，跃向他所说的「算术机器」（arithmetical machine）。他还更进一步，在那篇文章里发明、或者说设计了一台他称之为 memex 的设备。相比 1945 年的 Bush，我们有个优势：我们已经看了 80 年计算机的发展。所以可以把他的时间线改一改：从算盘，到存储程序计算机，到 40 年后的个人电脑，再到 40 年后的今天——我们所有人正在共同发明、创造、把它带进现实的 AI agent 时代。

[6:18]

So, the question for me is what did we build to go from those very first digital computers in the 1940s to the personal computer in the 1980s? Well, in the 1950s the big job was to figure out more effective ways of transmitting human intent to these new computing machines. We built the first programming languages. We wrote the first compilers. And the the theoretical underpinnings here were figuring out how to combine the elegance of mathematical formalisms with something a little bit more like natural language. And then building on that in the 1960s the challenge was to make these machines interactive. Make these machines capable of a two-way dialogue with humans. The '60s also saw the birth of graphical programming with systems like Ivan Sutherland's Sketchpad. And the '60s were an amazing era for science fiction. Even though almost nobody had access to a computer the computer became a big part of the popular imagination. The idea of a computer really resonated with people and ideas matter. For example, here is the idea of the computer in Star Trek.

所以我关心的问题是：从 1940 年代最早的数字计算机，到 1980 年代的个人电脑，这中间我们究竟造了些什么？1950 年代最大的活儿，是找出更有效的方式，把人的意图传达给这些新机器。我们做出了第一批编程语言，写出了第一批编译器。背后的理论功课，是琢磨怎么把数学形式化的优雅，跟更接近自然语言的东西结合起来。到了 1960 年代，接着往上的挑战是让机器变得可交互，让机器能跟人做双向对话。60 年代也诞生了图形化编程，比如 Ivan Sutherland 的 Sketchpad。60 年代还是科幻的黄金年代。尽管当时几乎没人真正碰得到计算机，计算机却成了大众想象里的重要角色。「计算机」这个想法真的打动了人——而想法是有力量的。举个例子，这是《Star Trek》里的计算机。

[7:29]

Put her on record.

记录下来。

[7:31]

Recording.

记录中。

[7:34]

Come. Captain's log supplemental. Engineering officer Scott informs warp engines Can be made operational and re-energized.

进来。舰长日志补录。轮机长 Scott 报告，曲速引擎可以修复并重新启动。

[7:45]

Computed and recorded, dear.

已计算并记录，亲爱的。

[7:49]

Computer, you will not address me in that manner. Computer.

电脑，不许这么称呼我。电脑。

[7:53]

Computed, dear.

已计算，亲爱的。

[8:00]

I love the background sound of punch cards going through a punch card reader. So, like you know the computer is working even though it's talking to you about what it's actually computing. There were of course a bunch of other talking computers in in science fiction of the '60s and the next year after this, the Kubrick movie that was a interpretation of Arthur C. Clarke's 2001: A Space Odyssey had the HAL 9000 computer. This is a much, much more dystopian view of a talking computer than the Star Trek computers. And by the 1970s, computers had become powerful enough that the next big job was designing abstractions that could scale to much larger amounts of data and much more powerful computing substrates. We got relational databases, which introduced new theoretical underpinnings for data manipulation. And we got declarative languages, which leveraged those new theoretical insights. And programming languages in general continued to evolve in what to me at least are really amazing ways. We got Smalltalk and object-oriented programming in the '70s. And all of this set the stage for the personal computer in the 1980s. The Macintosh shipped in 1984. Windows 1.0 shipped in 1985. And Microsoft's mission statement was a computer on every desk and in every home. And incredibly, Microsoft delivered on that mission statement.

我特别喜欢背景里那个穿孔卡片读卡机的声音。这样你就知道计算机确实在干活——哪怕它嘴上正在跟你聊它到底算了什么。60 年代的科幻里当然还有一堆会说话的计算机。这一集之后的第二年，Kubrick 改编 Arthur C. Clarke 的《2001: A Space Odyssey》，里面就有 HAL 9000。那是对「会说话的计算机」远比 Star Trek 反乌托邦得多的一种想象。到了 1970 年代，计算机已经强到一个程度，接下来最大的活儿变成了设计新的抽象，让它能撑住大得多的数据量和强得多的算力底座。我们有了关系型数据库，它给数据操作带来了新的理论基础；我们有了声明式语言，把这些新的理论洞见用了起来。编程语言整体也在继续演化，至少在我看来是以特别精彩的方式演化。70 年代我们有了 Smalltalk 和面向对象编程。这一切为 1980 年代的个人电脑铺好了路。Macintosh 在 1984 年发售，Windows 1.0 在 1985 年发售。而微软的使命宣言是：让每张办公桌上、每个家庭里都有一台电脑。不可思议的是，微软真的兑现了这句话。

[9:17]

And we got a computer on every desk and in every home because these new personal computers delivered real, amazing, tangible benefits. Take VisiCalc, for example, which was the first spreadsheet program. A truly new abstraction for doing computation, numerical computing, two-dimensional, interactive, so durable and so useful that probably most of us in this room use a direct descendant of VisiCalc regularly, Google Google Sheets or Microsoft Excel or whatever. Or put another way, this was a spreadsheet in 1957. And this was a spreadsheet in 1985. And I think a lot about VisiCalc these days too because I think VisiCalc is an example of how transformative new technologies can be in the way of delivering a capability that used to require a lot of specialized people and specialized knowledge and making it generally accessible. And I think VisiCalc is a potentially a counter-argument to the argument or the fear or the concern that AI is going to lead to mass unemployment. Because VisiCalc didn't put accountants out of business. Instead, it made much much much much more accounting-like work possible. And it made new categories of work possible that we couldn't even really conceive of when a spreadsheet or doing a screen's worth of calculations as we think about it today took a roomful of people.

我们之所以真的做到了每张桌子、每个家庭一台电脑，是因为这些新的个人电脑带来了实实在在、了不起、看得见摸得着的好处。比如 VisiCalc，第一个电子表格程序。那是一种全新的计算抽象——数值计算、二维、可交互——好用又耐久到什么程度呢？在座大多数人今天多半还在经常用它的直系后代，Google Sheets 也好、Microsoft Excel 也好，反正是它那一脉。换个说法：这是 1957 年的「电子表格」，这是 1985 年的电子表格。我最近也常常想到 VisiCalc，因为我觉得它是个绝好的例子，说明变革性的新技术能干什么——把一件原本需要一大批专门的人、专门的知识才能做的事，变成人人都能上手的能力。而且我觉得，VisiCalc 可能正是对「AI 会带来大规模失业」这种担忧的一个反例。因为 VisiCalc 并没有让会计失业，反而让多得多得多的、类似会计的工作变得可行，还催生出一整类全新的工作——在「一屏的计算量」还得靠一屋子人来做的年代，那些工作我们根本想象不出来。

[10:41]

So if we were here in the Moscone Center in 1985 and these two interfaces, the Macintosh System 2 and Windows 1.0 were state of the art, what would we have said the world would look like in 10 or 20 or 30 or 40 years? Well, we actually have a really great example of a prediction from that time. Like as we may think another famous document in the history of human computer interaction, concept video from Apple made in 1987 called Knowledge Navigator. This is very much worth tracking down online and watching all of if you haven't seen it. I'm just going to play about 20 seconds from the middle.

所以，假如 1985 年我们就坐在 Moscone Center，而当时最先进的界面就是 Macintosh System 2 和 Windows 1.0，我们会说 10 年、20 年、30 年、40 年后的世界是什么样？其实那个年代还真留下了一份非常棒的预测。跟《As We May Think》一样，它是人机交互史上另一份著名文献：Apple 在 1987 年做的概念视频，叫 Knowledge Navigator。如果你没看过，非常值得上网找出来完整看一遍。我这里只放中间大概 20 秒。

[11:22]

You have three messages. Your graduate research team in Guatemala, just checking in. Robert Jordan, a second semester junior, requesting a second extension on his term paper and your mother reminding you about your father's

你有三条留言。你在危地马拉的研究生团队报个平安；Robert Jordan，大二下学期的学生，申请把学期论文再延期一次；还有你母亲提醒你，你父亲的——

[11:36]

surprise birthday party next Sunday.

——惊喜生日派对在下周日。

[11:40]

So, the video shows a foldable tablet, a touchscreen interface, [clears throat] a conversational voice assistant with a really strong personality, access to both global and personal information, real-time video generation, real-time computer vision, seamless video call integration, delegation of complex tasks for autonomous execution, and what we might call today continual learning. And it's really, really clearly influenced by AS we may think, but it's also quite different. It really is updated for 40 years of progress, and it really does sort of presage this AI agent era we're in now in a way that Vannevar Bush's Memex didn't and maybe couldn't. The Knowledge Navigator video divides our timeline, I think, quite neatly in half, and hold that thought cuz we're going to come back to it. The 1990s were about the network, first local area networks and dial-up, and then the internet and the web, and with the benefit of hindsight, I now think that the single most important thing about the web was that it was multimodal from the very beginning. More even than the GUIs of the 1980s, the web anticipated that text and audio and video and data were not different things to be used in different programs, they belonged together. And in a real sense, the web was an attempt, and a conscious attempt on the part of a lot of people building the web to make that Knowledge Navigator video real.

这段视频里出现了：可折叠平板、触摸屏界面、一个个性极强的对话式语音助手、对全球信息和个人信息的双重访问、实时视频生成、实时计算机视觉、无缝的视频通话集成、把复杂任务委派出去自动执行，还有我们今天可能会叫做 continual learning 的东西。它明显深受《As We May Think》影响，但又相当不同。它确实是按 40 年的技术进展做了更新，也确实以 Vannevar Bush 的 memex 做不到、甚至可能根本做不到的方式，预示了我们现在所处的这个 AI agent 时代。我觉得 Knowledge Navigator 这段视频把我们的时间线相当漂亮地劈成了两半——这点先记着，等会儿还要回来说。1990 年代的主题是网络：先是局域网和拨号上网，然后是互联网和 Web。现在回头看，我认为 Web 最重要的一点，是它从一开始就是多模态的。比起 80 年代的图形界面，Web 更早地预见到：文本、音频、视频和数据不是要用不同程序分开处理的不同东西，它们本来就该在一起。而且从很实在的意义上讲，Web 是一次尝试——对很多参与建设 Web 的人来说是一次自觉的尝试——要把 Knowledge Navigator 那段视频变成现实。

[13:03]

Then in the first decade of the new millennium, the big job was to make all of this computing stuff mobile and continually connected, to put this new multimodal networked computer in your pocket, literally, to give a supercomputer to everybody in the world that they could carry around in their hand. And as with the 1960s, there was an efflorescence of like futurism on screen in the first few years of the new millennium. And I think it was because computers you could carry around with you and cameras everywhere and a kind of Moore's law for pixels making screens super cheap really gave us a chance to think through what we thought the future would look like in a new way. A lot of stuff we could almost but not quite build was cohering in the minds of people working on these machines. And the best and most famous Hollywood computers from that era were created by John Underkoffler [clears throat] for the films Minority Report and Iron Man. Here's Minority Report from 2002.

进入新千年的头十年，最大的活儿是让这一整套计算的东西移动起来、持续联网，把这台多模态的联网计算机真真切切塞进你口袋里，让世界上每个人手里都能拿着一台超级计算机。跟 60 年代一样，新千年最初几年，银幕上又爆发了一波未来主义想象。我觉得原因是：能随身带的计算机、无处不在的摄像头，再加上像素上的那种「摩尔定律」让屏幕变得极其便宜——这些让我们有机会用一种全新的方式去想象未来长什么样。一堆我们差一点点就能造出来的东西，在做这些机器的人脑子里逐渐成了形。那个年代最好、最有名的好莱坞计算机界面，都出自 John Underkoffler 之手，为《Minority Report》和《Iron Man》两部电影设计。这是 2002 年的《Minority Report》。

[14:05]

It's no longer there.

已经不在那儿了。

[14:07]

Time frame?

还剩多长时间？

[14:08]

13 minutes.

13 分钟。

[14:09]

Hey Chief, investigator from the Feds here.

头儿，联邦那边的调查员到了。

[14:12]

Yeah, I don't need some twink from the Fed poking around right now.

我现在可没工夫让联邦来的哪个愣头青在这儿瞎转悠。

[14:14]

John, I wrote it down on your calendar. I left you a message at your house.

John，我记在你日历上了，还往你家里留了口信。

[14:16]

Check in with the Favors ahead of Ford and see if the neighbors knew where they went. Check all relations.

去问问周围的邻居，看他们知不知道人去哪儿了。所有亲属关系也全查一遍。

[14:19]

Check the neighbors and relations.

查邻居，查亲属关系。

[14:20]

But John

可是 John——

[14:21]

It's actually

其实……

[14:21]

Just get him some coffee. Tell him some stories how I save [music] your ass every day and you can't do without me.

给他倒杯咖啡，再跟他讲讲我怎么天天给你收拾烂摊子、你离了我根本不行。

[14:25]

I got coffee. Thank you.

咖啡我有了，谢谢。

[14:27]

Danny Witwer. Twink from the Fed. Oops. Gone.

Danny Witwer，联邦派来的那个小白脸。哎哟，没了。

[14:32]

So the gestural interface in Minority Report was implemented on screen as special effects, but it was actually based on John's PhD work at the MIT Media Lab. In a real sense, this was real technology. John had brought the UI out of the small screen and into the world with us in a bunch of really interesting and lovely ways. John also consulted on Iron Man, which is a very different view of the future than you Minority Report, which was Spielberg working in like the American Kubrick dystopian tradition. Iron Man is really squarely in that Star Trek goofy futurist tradition. But I think you could see the common elements in the UI depicted on screen. It's still from the same era.

《少数派报告》里那套手势界面，在银幕上是当成特效做出来的，但它其实是基于 John 在 MIT Media Lab 的博士研究。从某种意义上说，那是真实存在的技术。John 用一堆非常有意思也很漂亮的方式，把 UI 从小屏幕里解放出来，带进了我们所处的真实空间。John 还给《钢铁侠》当过顾问，而《钢铁侠》对未来的想象跟《少数派报告》完全不一样——《少数派报告》是 Spielberg 在走「美国版 Kubrick」的反乌托邦路线，《钢铁侠》则是彻彻底底站在《Star Trek》那种憨憨的乐观未来主义传统里。但我觉得，你能在银幕上呈现的 UI 里看到共同的元素，它们仍然属于同一个时代。

[15:12]

Wake up, Daddy Sean.

醒醒，老爸回来了。

[15:13]

Welcome home, sir. [music] Congratulations on the opening ceremonies. They were such a success. As was your summit hearing.

欢迎回家，先生。恭喜开幕式大获成功，参议院听证会也一样。

[15:20]

[music]

[音乐]

[15:20]

And may I say how refreshing it is to finally you in a video with your clothing on, sir. You! I swear to god, I'll dismantle you or short your motherboard. I'll turn you into a wine rack.

另外容我说一句，先生，终于能在视频里看到您穿着衣服，真是让人耳目一新。——你给我等着！我发誓我要把你拆了，或者烧了你的主板，把你改造成一个红酒架。

[15:37]

I co-founded a startup with John in 2006 to make the Minority Report interface into a commercial product. This is our demo reel from 2012, 6 years into that work.

2006 年我和 John 一起创办了一家创业公司，想把《少数派报告》里的那套界面做成真正的商业产品。这是我们 2012 年的 demo 视频，当时这件事已经做了 6 年。

[15:49]

[music] [music] [music] [music] [music] [music] [music] [music] [music] [music]

[音乐]

[17:10]

This long project to build the multi-modal, multi-device, multi-screen, multi-player, ubiquitously connected computer is still what I'm working on 15 years later. In 2010s, we built out the cloud, which laid the groundwork for the infrastructure and data centers and data capacity we would need to scale up AI training and inference, which brings us to now. We're building agents. And we're starting to think about agents plus plus. But I think we can also start to think about the next thing, the AI native software that is to agents what today's internet is to the web pages of 1995. And one way to think about the story is this. We went from the calculator to the computer to the personal computer to the global cloud computer. And now we actually have the ability to build the Memex and Jarvis from Iron Man and knowledge navigator from that 1987 video for real, completely working. And by building those things, we'll figure out what we want to build next. A couple of weeks ago the team at Tavis released a reimagined knowledge navigator video this time entirely built on real and available technology. I'll just play another 20 seconds of this, but like the original knowledge navigator video, this is worth tracking down and watching in full.

这个漫长的项目——造一台多模态、多设备、多屏幕、多人协作、无处不在联网的计算机——15 年后仍然是我在做的事。2010 年代，我们把云建了起来，为后来扩大 AI 训练和 inference 规模所需的基础设施、数据中心和数据容量打下了地基，这就一路走到了今天。现在我们在造 agent，也开始琢磨 agent++。但我觉得我们同样可以开始想下一件事了：AI 原生软件。它之于 agent，就像今天的互联网之于 1995 年的网页。这条脉络也可以这么讲：我们从计算器走到计算机，再到个人电脑，再到全球规模的云计算机。而现在，我们真的有能力把 Memex、《钢铁侠》里的 Jarvis，以及 1987 年那支视频里的 Knowledge Navigator 造出来了，是真能跑起来的完整东西。而在造这些东西的过程中，我们才会搞清楚下一步该造什么。几周前，Tavus 团队发布了一支重新演绎的 Knowledge Navigator 视频，这一次完全建立在真实、现成可用的技术之上。我只放 20 秒，但和 1987 年那支原版一样，这支视频值得你专门找出来完整看一遍。

[18:31]

Good evening, Hassaan. I adore that houndstooth jacket you're wearing today. Anything I can help with or would you like to review tomorrow's schedule?

晚上好，Hassaan。你今天这件千鸟格外套我太喜欢了。有什么需要我帮忙的吗？还是先看看明天的日程？

[18:38]

Thanks so much, Tom. Yeah, let's review tomorrow's schedule and see how busy it is.

太谢谢了，Tom。好啊，我们看看明天的日程，看有多满。

[18:44]

Opening your calendar now. Here is the quick version since it is late. Tomorrow morning is slammed. Investor meeting at 9:00, then back-to-back one-on-ones and internal meetings until 1:00.

这就打开你的日历。时间不早了，我说个简版：明天上午非常满。9 点有投资人会议，然后是一场接一场的一对一和内部会议，一直排到 1 点。

[18:55]

So, the full 4-minute video is one take, completely real. And when you watch it, it really does feel both like the knowledge navigator video from 1987, familiar but built on real technology, and like something brand new. And I'll just close with a massively multiplayer game project I've been working on with some friends as a canvas to really think about what AI native software can be. This game is built from the ground up with LLMs as the core of every interaction. At every moment in the game, there are hundreds of inference calls happening and we couldn't have built anything like this even a year ago. Oh, sorry.

完整的 4 分钟视频是一镜到底，全部真实。你看的时候会有两种感觉同时出现：一方面它很像 1987 年那支 Knowledge Navigator 视频，熟悉，只不过这回是建立在真实技术上；另一方面它又像是某种全新的东西。最后我想用一个大型多人在线游戏项目来收尾，这是我和几个朋友一起在做的，把它当成一块画布，认真去想 AI 原生软件到底能是什么样。这款游戏从底层开始，就把 LLM 当作每一次交互的核心。游戏里的每一刻都有几百个 inference 调用在跑，哪怕就在一年前，这样的东西我们也造不出来。哦，抱歉。

[19:36]

Welcome to Gradient Bang, a multiplayer game [music] that showcases real-time agent orchestration. Gradient Bang demonstrates several patterns for AI sub agents such as asynchronous non-blocking context compression.

欢迎来到 Gradient Bang，这是一款展示实时 agent 编排的多人游戏。Gradient Bang 演示了若干 AI sub agent 的模式，比如异步非阻塞的上下文压缩。

[19:50]

Okay, make a note for later. We are going to eliminate Heliotrope from existence.

好，记一笔备忘：我们要把 Heliotrope 从这世上彻底抹掉。

[19:54]

Noted.

已记录。

[19:55]

Long-running sub agents that share context. [music]

共享上下文的长时运行 sub agent。[音乐]

[19:58]

Eagle is on five trade loops. Hawk and Raptor are on five exploration loops [music] each. Your fleet is busy.

Eagle 正在跑五条贸易循环，Hawk 和 Raptor 各自在跑五条探索循环。你的舰队很忙。

[20:06]

Progressive skills loading.

渐进式技能加载。

[20:08]

How much does your average ship cost?

你们的飞船平均多少钱一艘？

[20:10]

Ships range quite [music] a bit, Captain.

价格差得挺多的，舰长。

[20:12]

Dynamic user interface generation.

动态生成用户界面。

[20:14]

Show my task history.

把我的任务历史调出来。

[20:16]

Certainly.

好的。

[20:17]

Uh hide the map.

呃，把地图收起来。

[20:19]

Okay.

好。

[20:20]

And conversational [music] voice.

以及对话式语音。

[20:21]

No, I don't want to exchange it. I just want to sell it for cold, hard cash, please.

不，我不想换购。我就是想把它卖了换现金，谢谢。

[20:26]

I'm afraid [music] the galaxy doesn't allow you to be shipless and hitchhike.

恐怕不行[音乐]，银河系不允许你没有飞船、只靠搭便车。

[20:31]

So, I went long after wrap-up, but I will say that the first version of this new draft talk was an hour. So, I have a lot more things I'm super excited to talk about with all of you. So, if you are interested in this stuff, come find me. We have a booth on the show floor. I'm online everywhere and I'm excited to build agents cuz agents are awesome, but also to build the next next thing, too. Thank you.

好，我讲超时了，都到收场了还在讲。不过我得说一句：这版新讲稿的第一稿有整整一个小时。所以还有一大堆东西我特别想跟大家聊。如果你对这些感兴趣，来找我，我们在展区有个展位，我在网上各个平台也都能找到。我很期待去做 agent——agent 太棒了——但同样期待去做下一个、再下一个新东西。谢谢大家。

[20:54]

[applause] [music]

[掌声][音乐]

[21:13]

Mhm.

嗯。