Ben Thompson Matt Garman 和 Sam Altman——Matt,欢迎来到 Stratechery——还有 Sam,欢迎回来[我此前曾于 2025 年 10 月、2025 年 3 月和 2023 年 2 月采访过 Altman]。 Matt Garman and Sam Altman — well Matt, welcome to Stratechery — and Sam, welcome back [I previously interviewed Altman in , , and ]. ¶
Sam Altman 谢谢。 Thank you. ¶
Matt Garman 谢谢,感谢邀请我来。 Thank you, thanks for having me. ¶
Ben Thompson Matt,这是你第一次上 Stratechery。 So Matt, this is your first time on Stratechery. ¶
Ben Thompson 可惜,我觉得有 Sam 在场,我们恐怕没法进行往常那个「了解你」的环节了。 Alas, I think that Sam’s presence is going to preclude the usual getting to know you section. ¶
Ben Thompson 再说了,他可不想听我们回忆在 Kellogg 商学院的时光,不过请一位校友来播客做客真是件好事。 Besides, he doesn’t want to hear us reminisce about our times at Kellogg Business School, but it is good to have a fellow alumnus on the podcast. ¶
Matt Garman 是啊,我很高兴来到这里。 Yeah, I’m happy to be here. ¶
Matt Garman 我下次再来,到时候我们可以聊得更深入一点。 I’ll come back another time and we can do a little deeper dive. ¶
Ben Thompson 那太好了。 That’d be great. ¶
Ben Thompson 你从实习生时期起就在 AWS 工作,如今在这波 AI 浪潮中执掌整个组织。 You’ve been working on AWS since you were an intern, and you’re now in charge of the entire organization during this AI wave. ¶
Ben Thompson 暂且没有更好的说法的话,构建 AI 业务和最初构建商品化算力(commodity compute)业务,哪些方面是一样的,哪些方面是真正不同的? What aspects of building the AI business are the same as building the original commodity compute business, for lack of a better term, and what aspects are really different? ¶
Matt Garman 我觉得相同的部分在于,我看到的还是同样的热情,以及外面的建设者们能做到他们以前根本做不到的事。我们刚创办 AWS 时很棒的一点就是:开发者突然就能用上那些只有坐拥数百万美元、能自建数据中心的巨头公司才用得起的基础设施。 I think that the parts that are the same are that I see that same excitement and builders out there being able to do things that they were never able to do before, and one of the cool things is when we first started AWS, is developers all of a sudden could get their hands on infrastructure that was only available to the largest companies who had millions of dollars to go build data centers. ¶
Matt Garman 靠一张信用卡和几块钱,他们就能启动应用,这真正引爆了那些在互联网上搞建设的人所能做到的一切。 With a credit card and a couple of dollars, they could spin up applications and it really exploded what was possible for people building out there on the Internet. ¶
Matt Garman 我们基本秉持这样一个理念:人们想建什么就建什么,我们不去预设他们该做什么;外面世界的创造力是——只要我们把强大的工具摆在他们面前,他们就会造出有趣而惊人的东西。 We kind of took the idea that people could build whatever they want and we weren’t going to presuppose what they should do and that the creativity of the world out there was, if we could put powerful tools in front of them, they’d build interesting and amazing things. ¶
Matt Garman 我认为这场变革对外面建设者们所能做到的事的改变程度,即便不比当初更大,也至少不相上下。 I think this is as much, if not more, transformational to what it’s enabling builders out there to do. ¶
Matt Garman 想想什么是可能的:你不必非得上十年学、学会编程才能去做一个应用;也不必非得有几百人的大团队、花上一个月又一个月又一个月才能把东西做出来。 As you think about what’s possible, you don’t have to have gone to school and learned for 10 years to code in order to go build an application, you don’t have to have huge teams of hundreds of people and months and months and months of time to go build things. ¶
Matt Garman 你可以用小团队做东西,可以做得快,可以快速迭代,而 AI 正在世界各个层面释放各种各样的创新。 You can build things with small teams, you can build it fast and you can iterate quickly, and AI is unlocking all sorts of innovation across every different aspect of the world. ¶
Matt Garman 我觉得很多方面都非常相似,看到它让客户群体那边迸发出什么可能性,真的超级令人兴奋。 I think in many ways that’s very similar, and it’s super exciting to see what it’s enabling from the customer base out there. ¶
Ben Thompson 不过,AWS 问世那会儿,有段时间只有你们一家,所以种种好处、坏处和一切,差不多都自动找上门来。 There was a bit, though, when AWS came along, , so you get all the upsides and downsides and everything sort of for free. ¶
Ben Thompson 有没有这样一种感觉:AWS 时代讲了很多商品化算力,让它可互换、弹性、便宜——而在 AI、尤其是训练上,胜出的抽象似乎是那些高度垂直整合的超级集群、非常先进的网络,以及软硬件之间非常紧密的联动。 Is there a bit where it felt like in the AWS era, there’s a lot about commodity compute, making it fungible, elastic, cheap — in AI, particularly in training, it feels like the winning abstraction was more about these really vertically integrated super clusters, really advanced networking, and really tight linkages between software and hardware. ¶
Ben Thompson 对现在的你们来说这算是一种意外吗——你们不是白纸一张地进来说「我们是这里唯一的玩家,我们对大规模算力有一套自己的看法」,而在 AI 的头几年里,这两套思路也许并不完全吻合? Was that sort of a surprise for you, where you’re coming at it now — instead of fresh, “We’re the only ones here, we had a particular way of looking at large-scale compute”, and at least for the first few years of AI, it maybe didn’t perfectly align? ¶
Matt Garman 我不觉得对我们来说有什么不同。 I don’t know that it was different for us. ¶
Matt Garman 不过我觉得真正不同的,是采用规模扩张得快得惊人,我想这一点恐怕让所有人都吃了一惊。 I think for what was different though, is just the incredible rapid scale of adoption, and I think that that’s probably surprised everybody. ¶
Matt Garman Sam,如果你不同意可以插话补充,但光是采用的速度、人们抓住那些能力的速度,我觉得就让所有人意外。 Sam, you can weigh in different if you disagree, but just the speed of adoption and how fast people have grabbed onto the capabilities there, I think has surprised everyone. ¶
Matt Garman 这不一样。回到我们当初做云计算的时候——我们花了很长时间去解释为什么一家卖书的人会给你提供算力;要解释什么是云计算,得费好多口舌。 It’s different if you go to the, when we started cloud computing, it took us a really long time to explain why a bookseller would provide your compute power, that was a lot of explanation to explain what cloud computing was. ¶
Matt Garman 有很多苦功夫人们都忘了,但在 2006 年,「世界上的计算就这么迁移过去」并不是理所当然的事,所以那里真的下了很多苦功。 There was a lot of hard work that people forget, but back in 2006, it wasn’t a given that that’s just how the world’s computing would move to and so there was a lot of kind of hard work there. ¶
Ben Thompson 不过你觉得现在是不是还得再费点口舌?因为很多人还锚定在训练时代,而你们说的是「我们在思考的是」,那会是不同的东西——也许你还得把当初那套解释能力重新拿出来? Do you think you had to do a bit of explaining now though, because lots of people were anchoring on the training era and you’re like, “We’re thinking about “, and that’s going to be something different, maybe you still had to get those explanatory powers going again? ¶
Matt Garman 确实还得解释,但人们听懂你在说什么的速度,已经完全不一样了。 You do, but it’s just how quickly people understand what you’re talking about is just totally different. ¶
Matt Garman 所以我认为是的。从大家说「这看起来挺酷,能有一个可以聊天的智能聊天机器人真有意思」,转变到「我可以在你们的企业里真正干活了」,这中间确实需要一点教育引导,但放在技术演进速度的尺度上看,也已经算相当快了。 So I think yes, I think if you move from where people are saying, “That does seem kind of cool, and it’s really neat that I have this intelligent chatbot that I can talk to”, going to, “I can actually do work in your enterprise”, has been a little bit of an education, but it’s also been relatively quick in the scope of how fast technology moves. ¶
Ben Thompson 我保证,我们很快就会聊到今天真正要讲的那个产品。但 Sam——从创业生态的角度回头看,AWS 显然是变革性的——它彻底改变了门槛所在的位置,现在任何人都能起步了。 We’re going to get to the product that we’re here for very quickly, I promise, but Sam — , completely changed where the barrier was, now anyone can get started. ¶
Ben Thompson 你有种子轮,有天使投资人,门槛和分界点因此往后移了:你不必先凭一份 PPT 去筹钱买服务器,你可以先把应用做出来,再去融 A 轮或者别的什么。 You have seeds, you have angel investors, and it sort of moves back the barrier where the cutoff point, you don’t have to get servers on a PowerPoint, you can build an app and then go to your Series A or whatever it might be. ¶
Ben Thompson 不过在你看来,那所释放出的可能性和今天的世界相比,有什么地方不同,什么地方相同? What, though, is different or the same compared to what that enabled versus the world today from your perspective? ¶
Sam Altman 我认为让初创公司得以大规模崛起的平台赋能时刻有四次:先是互联网,然后是云,再然后是移动,然后就是 AI。 I think there have been four great moments for platform enablement of startups at mass scale: there was the Internet, there was cloud, there was mobile, and then there was AI. ¶
Sam Altman 我成年之后亲历的第一次是云,正好在 YC[Combinator]的早期——怎么强调它对初创公司意味着多大的改变都不为过。 The first one of those that I was kind of like an adult for was the cloud and in the early days of YC [Combinator] — it’s like hard to overstate what a change this meant for startups. ¶
Sam Altman 在那之前,初创公司得去租服务器、自己把服务器攒起来、往里面部署东西,那是极其复杂的一件事,你得为此融好多钱。 Before, you had these startups that were like renting and putting together servers and putting stuff in there and it was this like massively complex thing, and you had to like raise all this money. ¶
Sam Altman 然后突然之间——虽说云几乎是在 YC 创办之后紧接着出现的,我猜是第二年。 Then all of a sudden, even though the cloud happened like right after YC got started, I guess it was the year after. ¶
Ben Thompson 我正要问这个——到头来,云和 YC 是不是真的比你当时意识到的更形影不离? I was just going to ask that — is it really at the end of the day, they’re really hand-in-hand more than you realized at the time? ¶
Sam Altman 当时它们感觉就密不可分,YC 从一开始就像是在踩着云的这波浪前进,因为在 AWS 之前就已经有一些早期的例子。 They felt incredibly hand-in-hand at the time, it felt like YC was, you know, surfing this wave of the cloud from the very beginning because there were some early pre-AWS examples. ¶
Ben Thompson 比起从前可能的情形,只要有 AWS 存在,让一家初创公司起步就不需要投入那么多钱了。 You don’t need to put that much money into a startup to get something off the ground if AWS exists compared to what it might’ve been before. ¶
Sam Altman 那是一次巨大的赋能式改变,这也是为什么 YC 当时听起来那么疯狂的原因之一。 It was this huge enabling change and it was part of why YC sounded so crazy at the time. ¶
Sam Altman 人们会说「你不可能用几万美元资助一家初创公司,那是不可能的,一台服务器的钱都不止这个」,所以这彻底改变了初创公司拿少量资本能做到的事。 People were like, “Well, there’s no way you can fund a startup with a few tens of thousands of dollars, it’s impossible, the server costs more than that”, so it was this complete change to what startups could do with small amounts of capital. ¶
Sam Altman 初创公司通常在大平台转换时获胜,因为你可以用比从前更短的周期、少得多的资本去做事——这是初创公司击败大公司的经典路径。我职业生涯早期亲眼见证了云带来这一切;现在看着企业在 AI 上构建东西,方向上其实相当相似,但就像 Matt 说的,那个速度快得疯狂。 Startups generally win when there is a big platform shift and you can do things with a faster cycle time and much less capital than before, that’s a classic way startups can beat big companies, and at the beginning of my career, I really witnessed that happen with the cloud, it actually feels quite directionally similar now watching what companies are doing building on AI, but as Matt was saying, the speed of it is crazy. ¶
Ben Thompson 有没有这样一种情况:在位者、那些大公司,采用 AI 的速度比当初采用云计算快得多? Is there a bit where the incumbents, the large companies, are adopting this way faster than they than they were the cloud? ¶
Sam Altman 那肯定更多了,不过我还指的是初创公司营收增长的速度本身——我最近在 YC 做了次演讲,最后我问「一家好公司在 YC 结束时对营收的预期是多少?」,他们说「嗯,这个差不多每个月都在变,也许这一批开始时的答案和结束时就不一样」,这种事以前从来没有过。 There’s definitely more of that, but I also mean just the the rate that revenue is scaling in at startups — I spoke at YC recently and I kind of asked at the end, “What are the expectations for revenue for a good company at the end of YC?”, and they’re like, “Well it’s kind of changing every month, maybe we’d have a different answer at the beginning of the batch versus the end of the batch”, and this never used to happen before. ¶
Sam Altman 人们在这个新平台上做出规模化业务的速度,是我以前从未见过的。 Just the rate at which people are able to build scaled business on this new platform is unlike anything I’ve seen before. ¶
Ben Thompson Matt,你们几乎是所有初创公司的首选云,那是整个时代的一大优势。 You were the cloud of choice for basically all startups, a huge advantage to that whole era, Matt. ¶
Ben Thompson 今天是什么让你们依然是首选云? What makes you the cloud of choice today? ¶
Ben Thompson 因为你想想,很多人在基于 OpenAI API 做开发;还是说你们的感受其实是「我们是从一个非常不同的角度切入这个市场的:我们拥有庞大的存量客户基础,他们正求着我们要 AI 能力,而对 Sam 说的这一整个创业群体,我们的洞察反而没那么多」? Because you think about a lot of people building on the OpenAI API, or is that something you felt, “Actually we’re coming at this market from a very different perspective, we have a huge installed base who’s begging us to get AI things, and we have less visibility into this whole cohort that Sam’s talking about”? ¶
Matt Garman 我觉得有几方面的原因。 I think there’s a couple of things. ¶
Matt Garman 其中之一是,我们对我们的合作伙伴关系相当兴奋,我觉得这对外面的一大批初创公司会意义重大。 One is, is we’re quite excited about our partnership, and I think it’s going to be really meaningful to a bunch of startups out there. ¶
Matt Garman 但今天,哪怕你去和初创公司聊,绝大多数处于扩张期的初创公司仍然是在 AWS 上扩张的,这背后有一大堆原因。 But today, even if you go and you talk to startups, the vast majority of scaling startups are still scaling on AWS today, and there’s a whole bunch of reasons for that. ¶
Matt Garman 规模在那里,可用性在那里,安全在那里,可靠性在那里,其他 ISV 组成的那种合作伙伴生态在 AWS 上,客户也在 AWS 上。 The scale is there, the availability is there, the security is there, the reliability is there, that kind of partner ecosystem of other ISVs are in AWS, the customers are in AWS. ¶
Ben Thompson (笑)不管愿不愿意,人人都用过 AWS 的控制台,所以他们都习惯了。 (laughing) Everyone’s used the AWS panel whether they wanted to or not, so they’re used to it. ¶
Matt Garman 而且我们会帮助他们。 And we help them. ¶
Matt Garman 我们花大量时间为初创公司赋能,不只是给抵扣额度(credits),还有怎么搭建系统的建议、怎么思考市场进入(go-to-market),这些东西我认为是很多初创公司真心感激的。我们投入大量时间和精力,因为我们真切地觉得初创公司是 AWS 的血脉。 We spend a ton of time enabling startups, whether it’s with credits, but it’s not just with credits, it’s advice on how to set up your systems, how to think about go-to-market, a bunch of those things that are, I think, are really appreciated by a bunch of the startups, we invest a lot of time and effort to make sure because we really feel like the startups are the lifeblood of AWS. ¶
Matt Garman 从 Sam 讲的那个年代开始就是如此,今天依然如此。我仍然每个季度去一次硅谷或其他地方,直接见初创公司,听他们在做什么,确保我们在做的东西能被他们接得住。 They were from the beginning, like when Sam was talking about it, but they remain today, and I still go once a quarter out to Silicon Valley or other places to meet directly with startups to hear what they’re doing, to make sure that what we’re building is landing with them. ¶
Matt Garman 所以对初创公司这份关注的争夺,今天比 20 年前更激烈了,而它对我们的重要性和以往任何时期一样,我们花大量时间确保满足这些初创公司的需求。 So there is more competition today than there was 20 years ago for that startup attention, and it’s just as important for us as it’s ever been and and we spend a ton of time to make sure that we’re meeting the needs of those startups. ¶
Ben Thompson 能不能这么说:直接在 OpenAI API 上做开发(而不是比如 Azure 版本的 OpenAI API)的人,更常见的架构是常规算力用 AWS 这套栈,AI 用 OpenAI? Is it fair to say people building directly on the OpenAI API, as opposed to say the Azure version of it, are more likely to have a stack of AWS for for regular compute and then OpenAI for for their AI? ¶
Matt Garman 我觉得这是今天很多初创公司非常常见的模式,没错。 I think that’s a very common pattern that a lot of startups have today, absolutely. ¶
Ben Thompson 那就说到我们今天的主题:由 OpenAI 驱动的 Bedrock Managed Agents,我想我没说错吧。 Well that brings us to today’s announcement: Bedrock Managed Agents, powered by OpenAI, I think I got that right. ¶
Ben Thompson 据我理解,它的卖点不只是「OpenAI 的模型可以在 AWS 上用」——我觉得那是不被允许的——而是 OpenAI 的前沿模型被打包进了一个 AWS 原生的 agent 运行时:身份、权限状态、日志、治理和部署。 The pitch, as I understand it, is not simply OpenAI models are available in AWS — I don’t think that’s allowed — it’s that OpenAI’s frontier models are being packaged inside an AWS-native agent runtime, identity, permission state, logging, governance, and deployment. ¶
Ben Thompson Sam,这样表述对吗? Sam, is that the right way to articulate it? ¶
Sam Altman 对,说得挺好的。 Yeah, that was pretty good. ¶
Sam Altman 谢谢。 Thank you. ¶
Ben Thompson 这到底是什么? What is this? ¶
Ben Thompson 现在用大白话解释一下。 Now explain it in English. ¶
Sam Altman 我认为,AI 的下一个阶段,将从你给 agent 一些文本、它再返回更多文本,或者你给它一堆代码、它再返回更多代码,转向让这些 agent 在公司内部运行,承担各种各样的工作。 I think the next phase of AI is going from you supply some text to an agent and get more text back, or even you supply a bunch of code and get more code back, to we are going to have these agents running inside of a company doing all different kinds of work. ¶
Sam Altman 「虚拟同事」算是我听过的描述中不那么糟糕的一种,但还没有人真正找到准确的说法来描述这件事。我们正在共同推出一款新产品,帮助那些希望构建这类 stateful agent 并将其提供给用户使用的公司。 Virtual co-workers is kind of my least bad of the ways I’ve heard this described, but no one has quite figured out the right language for this, and we are packaging a new product that we’re working on together to help enable companies that want to build these sorts of stateful agents and make them available. ¶
Sam Altman 再说一遍,我们并不确切知道世界会怎么谈论和使用这些东西,但如果你看看[Codex 上]正在发生的事,我觉得那是个绝佳的例子,让我们能看出这一切将去向何方。 Again, I think we don’t know exactly how the world’s going to talk about these, use these, but if you look at what’s happening [with Codex], I think there’s a great example of where we can see this all going. ¶
Ben Thompson 围绕模型的 harness——那套运行时、工具、状态——照你的说法,这是对你很重要的一个词——记忆、权限、评估(evals),对让 agent 真正能干活到底有多重要? How important is , the runtime around the model, the tools, state — to your point, a very important word to you — memory, permissions, evals, to making agents actually work? ¶
Sam Altman 它有多关键,怎么说都不为过。 Hard to overstate how critical it is. ¶
Sam Altman 我不再把 harness 和模型看成完全可分的两样东西了。以我的使用体验来说,我很清楚一件事:当我在 Codex 里发起一个任务、它给我做出惊人的成果时,我并不总能知道—— I no longer think of the harness and the model as these entirely separable things, like my experience of using these, I am very aware of the fact that I don’t always know when I fire something off in Codex and it does an amazing thing for me. ¶
Sam Altman 该把多少功劳归给—— I don’t know how much credit — ¶
Ben Thompson 是模型惊艳,还是 harness 惊艳? Was it that the model is amazing or the harness was amazing? ¶
Sam Altman 对,就是这样。 Yeah, exactly. ¶
Ben Thompson harness 在多大程度上是和模型协同开发的? To what extent is the harness developed in conjunction with the model? ¶
Ben Thompson 这种整合发生在哪个环节? Where does that integration happen? ¶
Ben Thompson 是在后训练(post-training)阶段吗? Is it in post-training? ¶
Ben Thompson 是在提示词(prompt)里吗? Is it in the prompt? ¶
Ben Thompson 是什么让这种整合真正发挥作用? What makes this integration work? ¶
Sam Altman 两者都有。 Both of those. ¶
Sam Altman 这其实不是预训练流程的一部分,但我想说你可以看到——更有意思的一点是:过去我们见过很多这样的例子,我们以为非常可分的东西,会被越来越多、越来越多地烤进(baked in)内部。 It’s not really part of the pre-training process but I would say you can look at it — there’s a more interesting thing here which is the fact that we’ve seen examples of this many times in the past of where things that we thought were very separable get baked in more and more and more. ¶
Sam Altman 比如我们最初思考工具调用(tool-calling)的方式——它现在是我们使用这些模型的关键一环——当初我们并没有想过要把它深度整合进训练过程,随着时间推移我们做得越来越多。 Like the way we initially thought about tool-calling, which is now a critical part of how we use these models, was not something that we thought about deeply integrating into the training process and over time we’ve done more and more of that. ¶
Sam Altman 我还会怀疑,随着时间推移,模型和 harness 会越来越融为一体;同理,我预计预训练和后训练(post-training)最终也会越来越合流。 I would also suspect that model and harness come together more over time and I would for that matter, I would expect that pre-training and post-training eventually come together more over time as well. ¶
Sam Altman 这话特别老生常谈,但我还是要说,因为我觉得它非常、非常对——我们对这一切的这个范式还处于极早期;就这个行业的成熟度而言,现在还很像 Homebrew Computer Club(家庭手工作坊计算机俱乐部)的年代。 It’s such a cliché to say, but I’ll do it anyway, because I think it’s very, very true — we’re so early in the paradigm of all of this, this is still like the Homebrew Computer Club days of how much this is like really matured as an industry. ¶
Ben Thompson 这正是我觉得它有趣的原因——我几周前也写过这个话题——在任何价值链上,最终都会浮现出一个整合点,那才是真正要紧的地方,这两块必须合在一起才能把事情做成。 This is why I think so interesting, , in any value chain, ultimately a point of integration emerges that that’s where it’s really important, these two pieces have to go together to make it work. ¶
Ben Thompson 随着时间推移,价值显然会大量汇聚在那里。所以我的论点是:这种 harness 与模型的整合,就是那个关键点。 And over time, that’s obviously where a lot of value collects — my thesis then is that this harness-model integration is the key point. ¶
Ben Thompson 这符合你的利益,不过听起来你也是同意的。 It’s to your interest, but it sounds like you agree. ¶
Sam Altman 是符合我的利益,我同意;但我还要说得更宽一点:你真正在意的,是你在 Codex 里输入你想要什么发生,然后它就发生了。 It is to my interest, I do agree, but I also would say even more broadly, what you care about is that you go type into Codex what you want to happen and that it happens. ¶
Ben Thompson 你并不关心实现细节。 You don’t care about the implementation details. ¶
Sam Altman 我觉得你们不会关心。 I don’t think you do. ¶
Sam Altman 在我们一路摸索这一切的过程中,出现过太多次:有些我们必须在系统提示词(system prompt)层面做的事,后来就不需要了。 There have been so many examples as we’ve been figuring all of this out where we had to do something at the level of the system prompt, that later we didn’t. ¶
Sam Altman 这里的一般规律是:随着模型变得更聪明,你让它们按你想要的方式行事就有余地得多——这听起来像句废话,但它确实—— The general observation here is as the models get smarter, you have more flexibility to get them to behave in the ways you want which sounds like an obvious statement, but it is— ¶
Ben Thompson 告诉一个十岁孩子该做什么,比告诉一个五岁孩子更容易。 It’s easier to tell a 10-year-old what to do than a 5-year-old. ¶
Sam Altman 回想在 GPT-3 时代,我们要费多大力气才能从这些模型里挤出一点点效用,而现在这些你根本再也不用做了,因为模型当然开箱就能理解、能做好——这个趋势恐怕还会继续走下去很远。 When I think back to what we had to do to get any drop of utility squeezed out of these models back in the GPT-3 days that now you never would have to, because of course the model just understands and does it well out of the box, that trend may keep going much further. ¶
Matt Garman 我刚想补充一句——我完全同意。而且我想,当你和有想法的客户聊他们到底希望这些系统做什么时,在我们携手做这种联合协作之前,客户基本上是被迫自己把这些东西拼起来,对吧? I was just going to add to that — I completely agree with that and I think when you talk to customers who have ideas exactly what they want these systems to do, previous to this kind of joint collaboration that we worked on together, is that customers were kind of forced to pull that together themselves, right? ¶
Matt Garman 他们希望这些模型和 agent 记得如何与他们良好配合,希望接入他们的现有系统;而且不只是第三方工具,还有他们自己的工具。 They wanted these models and agents to remember that they work together well and they wanted to integrate into their existing systems, and it’s not just third-party tools, it’s their own tools. ¶
Matt Garman 他们希望 agent 了解他们自己的数据、自己的应用和自己的运行环境;而今天至少所有这些整合,都留给每一个客户自己去做。 They want them to learn about their own data, their own applications, and their own operating environment and all of that kind of integration today, at least, is left to every single customer to do on their own. ¶
Matt Garman 所以我们共同推进的这项联合协作,一部分就是共建一种新型产品,真正把这些东西拉到一起,让客户能更容易地做成他们想做的事。身份能力已经内置在那个产品里,连接你的数据库做认证这件事,全部发生在你自己的 AWS VPC[虚拟私有云,Virtual Private Cloud]内部。 So part of this joint collaboration that we were leaning into together is co-building a new type of product that actually brings those things much closer together so that customers can much more easily go accomplish these things that they want to do, where identity is already kind of built into that product, where the ability to go authenticate to your database all happens inside of your AWS VPC [Virtual Private Cloud]. ¶
Matt Garman 如果 OpenAI API 在一头、AWS 在另一头,也有很多事情原则上做得到;但通过一起构建这个东西,我们让客户更容易、更快地获得价值,在他们自己的企业环境里做成他们想做的事。 You can do a bunch of these things that would be possible to do if we were kind of at the OpenAI APIs and AWS over here, but by building this thing together, we make it much easier for customers to much more rapidly get to value and go accomplish the thing they want to do inside of their enterprise environment. ¶
Ben Thompson 所以你的意思是,用通用的 harness 也能搭出有功能的 agent,只是难得多? So you think that you can build a functional agent in a generic harness, it’s just way more difficult? ¶
Ben Thompson 你们是在降低难度? You’re making it easier? ¶
Ben Thompson 还是说,有那么一部分事情,如果不把它们绑在一起,你根本做不到? Or is there a bit where actually there might not even be stuff you can do if you don’t have them tied together? ¶
Sam Altman 回到你之前那个类比:在 AWS 之前的年代,如果你愿意站在机柜间里、买一堆服务器、自己搞定互联、再雇自己的网络工程师,你能做很多事情,也能让很多事情发生。可一旦你能登录 AWS 控制台、点一下「我再来一个 S3 实例」或诸如此类的操作,你能做成的事情就多太多了,因为激活能量——为基本盘付出的工作量——大幅改善了。如今用这些模型,你也能做很多事情。 To go back to your earlier analogy, pre-AWS days, you could do a lot if you were willing to go stand in a cage and buy a bunch of servers and figure out how to connect them and hire your own network engineer, and you could make a lot of things happen and then all of a sudden as soon as you could just like log into an AWS control panel and click, “I need another S3 instance”, or whatever, you could make a lot more things happen because the activation energy, the amount of work that required for the basics, got way better so you can do a lot with the models today. ¶
Sam Altman 但每次我看别人使用我们的模型、或者试着搭建 Matt 刚说的那类工作时,我都很纠结:一边很高兴他们印象深刻、觉得这是神奇的技术,一边又为他们为了把事情跑通所受的痛苦折磨抓狂。这不只是开发这些产品的开发者如此——就算是用 ChatGPT,看着人们把东西从这里复制粘贴到那里、 试图搞出一整套复杂的提示词——我知道这一切会消失,我也为此兴奋。 Yet every time I watch someone use our models or try to set up some of this work Matt was saying, I am torn between being happy they’re so impressed and feel like this is a magical technology and pulling my hair out at how much pain and suffering they’re going through to get anything to work at all, and that’s not just true of developers building these products, even using ChatGPT and watching people copy and paste things from here to there and try to have this complicated set of prompts — I know that’s going to go away, and I’m thrilled. ¶
Sam Altman 现在还处于非常早期,也非常糟糕。 It’s still so early, and so bad. ¶
Ben Thompson 只求别把你们和 BBEdit 的集成拿走——这就是我唯一的要求,它是 ChatGPT 应用里我最喜欢的头号功能。 Just don’t take away your integration , that’s all I ask, my number one favorite feature of the ChatGPT app. ¶
Sam Altman 行。 Fine. ¶
Ben Thompson (笑)谢谢。 (laughing) Thank you. ¶
Sam Altman 第一,这类事情做起来实在太难,我们认为如果能把它变得容易得多,就能给开发者和企业带来多得多的价值;第二,也有很多事情现在无论如何都靠不住、根本做不通。通过我们的联合协作,故事不只是「好用、不用自己去建自己的托管机房(colo)或诸如此类」,我们还要一起想清楚很多要新建的东西,让人们能做出即便吃尽苦头也根本做不到的产品和服务。 A) This stuff is just way too hard to do, and we think if we can make it way easier it’ll bring way more value to developers and businesses, but B) there are a lot of things that you just can’t reliably get to work at all and I think through our joint collaboration not only will it be a story of ease of use and not having to go build out your own colo or whatever, we are going to jointly figure out a lot of new things to build where people will be able to build products and services that just can’t be done even with a lot of pain and suffering. ¶
Ben Thompson 关于「要新建的东西」这一点,我一会儿还想回来。 I actually want to come back to that point about things to be built. ¶
Ben Thompson 不过先快速回到 Codex——Codex 是 harness 加模型,它在本地运行。 But just to go back to Codex real quick — Codex is a harness and model, it runs locally. ¶
Ben Thompson 为什么现在让 agent 在本地跑起来反而更容易? Why is it easier to get agents to work locally right now? ¶
Sam Altman 其实我们一开始是让它跑在云端的,而且我认为最终你是希望它跑在云端的。 Actually, we started with it running in the cloud, and I think eventually you do want it to run in the cloud. ¶
Ben Thompson 那是当然。 For sure. ¶
Ben Thompson 我是在梳理向这个云端产品的过渡。 I’m walking through the transition to this offering, which is in the cloud. ¶
Ben Thompson 但你们为什么退回到本地? But why did you go back to local? ¶
Sam Altman 你的整个环境都在那里:你的电脑已经配好,你的数据在那里,你不必再去操心——就是更容易跑通,尽管那不是终局形态。 You have your whole environment there, your computer’s set up, your data is there, you don’t have to like think about — it was just easier to get to work, even though it’s not the end state. ¶
Sam Altman 但通向一个 agent 真正跑在云端的世界——如果你有非常重型的任务,或者你要合上电脑之类的时候,可以把活儿交给云端去干——我认为那显然会很棒。 But getting to a world where agents do run in the cloud and when you — if you have a very intensive thing, or you need to close your computer or whatever, you can hand stuff off to working on the cloud, I think is clearly going to be great. ¶
Sam Altman 但我们在短期内确实能交付的那种易用性占了上风,所以让它先使用你的本地环境。 But the ease of use that we were able to deliver clearly in the short term, it won out to have it using your local environment. ¶
Ben Thompson 我的一种理解方式是:老派的安全模型是城堡加护城河(castle-and-moat)那一套,而现在你们转向的新安全模型,是每样东西都有恰当的权限结构、身份认证等等这些组成部分。在我看来,本地运行可以理解为你自己给自己修的城堡加护城河—— 所有东西都在里面,我就默认一切都好、做起来容易。 There’s one way that I think about it, is like you have the old school security model, which is like the castle-and-moat sort of thing, and you’re moving to a new security model of and everything having the appropriate permission structure and authenticating and all those bits and pieces, and it feels like to me one way to frame running locally, it’s like your self-imposed castle-and-moat, everything’s on there, I just assume it’s all fine and easy to do. ¶
Ben Thompson 我思考这件事的另一种方式——Matt,你看有没有道理——是要让这些环节在生产环境里真正运转,你根本不能全靠本地;从一开始你就得在这样的环境里运作。这样理解对吗? And a way I’m thinking about this, and Matt, let me know if that resonates with you, is to get all those pieces to actually function in a production environment you just can’t even have that all locally, you have to be operating this environment from the get-go, is that a right way to think about it? ¶
Matt Garman 我不认为有任何计算环境真正甩掉了客户端,本地运行就是有自己的好处。 I don’t know that there’s any computing environment that’s gotten rid of a client, there are just benefits of operating locally. ¶
Matt Garman 你的大多数 iPhone 应用也都有本地组件,这是有原因的:不管是连接性、延迟,还是单纯的本地计算,或是对文件和应用(applications)的访问。 There’s a reason that most of your iPhone apps also have a local component, whether it’s connectivity or latency or just local compute or access to files and applications. ¶
Matt Garman 本地客户端确实有它独到的一面——正如 Sam 所说,它容易、它很好用——但它也是受限的,它有它的上限。 The local client does have a particular — as Sam said, it’s easy, it works really well, it’s constrained, though, there’s limits to it. ¶
Matt Garman 你没法把本地笔记本横向扩展,你手里就这么多资源;而一旦进入企业协作,两个人之间共享就有点难了——权限问题、安全边界问题都会变得有点难。 You can’t scale out your local laptop, you have what you have and once you start getting in an enterprise contract, sharing between two people gets to be a little bit harder — thinking about permissions, thinking about security boundaries gets to be a little bit harder. ¶
Matt Garman 所以有若干这样的环节,我认为——我不会说有本地环境是坏事,它只是不一样——我觉得最终你会想把两边打通(bridge)。 So there’s a number of those pieces where I think that, I wouldn’t say that having the local environment is a bad thing, it’s just a different thing, and I think that you’re eventually going to want to have that bride across both. ¶
Ben Thompson 这正是我的疑问。因为在云时代,有容器帮你把本地环境和生产环境收敛到一致;但这件事给你的感觉是,如果你要处理的是 agent——照你说的,它们是虚拟同事之类——如果它们有自己的身份、自己的权限这类东西,那么你要构建它们,就得身处你将要部署进的那个环境;在我看来似乎应该这样。 That’s my question, because you have in the cloud era, you had containers that helped you converge local and production environments, but it kind of feels like in this case if you have to deal with agents, to your point, say I was like a virtual co-worker and or whatever it might be, if they have their own identity and they have their own permissions and all those sorts of things, to even build them you need to be in the right environment as you’re going to deploy it, it would seem that way to me. ¶
Sam Altman 我觉得这里有太多东西要摸索。 I think there is so much to figure out here. ¶
Sam Altman 举个例子:如果你是一家公司的员工,你使用某个服务时是不是有一个账号?那你的 agent 应该就用你的账号,还是应该用一个不同的账号,好让服务端能分清哪个是哪个? Just to give one example, if you’re an employee at a company, do you want to have one account for when you use some service, and then should your agent just use your account, or should your agent use a different account so that the server can tell which is which? ¶
Ben Thompson 或者,如果你想要很多 agent 呢? Or what if you want lots of agents? ¶
Sam Altman 正是如此。 Exactly. ¶
Sam Altman 我猜我们真正想要的,是我们还没想明白的东西。也许是这样:当 Ben 的 agent 以 Ben 的身份登录时,它用的是 Ben 的账号,但会注明自己是一个 agent、不是真的 Ben。 I suspect that what we actually want is something we haven’t figured out yet, and maybe it’s that when Ben’s agent is logging in as Ben, it uses Ben’s account but it notes that it’s an agent and not the real Ben. ¶
Sam Altman 我们甚至还没有能思考这件事的基本机制,但我们可能很快就得把它搞明白。我的感觉是,还会有其他 50 个类似的问题:随着 agent 加入劳动力、以不断提高的自主性和任务复杂度行事,我们关于软件如何运作、访问控制和权限在公司内部或更广阔的互联网上如何运作的心智模型,全都得演进。 We don’t even have a primitive to think about that, but we may quickly need to figure that out and and my sense is there there are going to be 50 other things like that where as we have agents join the workforce and act with increasing levels of autonomy and complexity of tasks, a lot of the mental models that we have for how software works and how access control and permissions work inside of a company or on the broader Internet, those are all just going to have to evolve. ¶
Ben Thompson Matt,从安全和访问策略等方面,你们是怎么思考 agent 问题的? How do you think about, Matt, in terms of security and access policies and whatnot for agents? ¶
Matt Garman 是的,我确实认为,当你把这些工作负载更多地搬进云端,才能作为一个中心化组织,对其中一些安全环节拥有更多管控。 Yeah, I do think that that’s where when you move more of these workloads into the cloud that you can have as a central organization, more controls over some of the security pieces of it. ¶
Matt Garman 我确实觉得,我们无时无刻不在和客户交流,他们担心的正是这个:「我很期待这些强大的模型和 agent 能带来的可能性,但我怎么确保自己不会摊上一场把公司搞垮的灾难性事件?」外面就是有这样的担忧。 And I do think, when we talk to customers all of the time, it is what they worry about, which is, “I love the promise of what I can do with some of these really powerful models and agents, how do I make sure that I don’t have a company-ending event where I screw it up?”, and there’s the worry out there. ¶
Matt Garman 我觉得我们能帮上忙,因为这些是可解的问题,确实可解。我想给客户信心的一种方式就是:「它在这个 VPC 内运行」,你至少能控制这个边界、知道它能访问什么;或者它经过这个网关,你可以给它授权,就像你在环境的其余部分里给它分配一个角色(role)那样。 I think we can help with that because it these are solvable problems, they are, and I think, giving some customers confidence, “Well, it operates inside of this VPC”, and you can at least then control that boundary and know what it has access to, or it goes through this gateway, and you can give it permissions, much like you give it a role inside of the rest of your environment. ¶
Matt Garman 这些是现成的结构。过去 20 年,我们建立起了非常丰富的一套能力,让使用 AWS 的不只是 Y Combinator 的初创公司,还有全球性银行、医疗机构、全世界的各行各业以及政府机构。围绕安全搭建起了所有这些体系,我认为能帮他们进一步加速利用这项技术,带着这些保障措施跑得快。 These are constructs that over the last 20 years, we’ve built up a really rich set of capabilities, so that it’s not just Y Combinator startups, but it’s global banks and healthcare agencies and everybody in the world and government agencies that can use AWS and having built up all of that security structure around it, I think can help us further accelerate how they take advantage of this technology and kind of have these safeguards to run fast. ¶
Matt Garman 我想很多时候,身处一家公司,尤其是一向规避风险的行业里的公司,有那些安全护栏、让他们可以说「只要它在沙箱里运行,我就放心大胆地快跑」,实际上能帮助我们的很多客户把这些技术用在更广得多的事情上。 I think a lot of times when you’re in a company, particularly companies that are in risk-averse environments, having those safety guardrails where they say, “If it operates inside of the sandbox, I am excited to go fast”, can actually help many of our customers start to use these technologies for a much broader set of things. ¶
Ben Thompson 你说的这些能力——你们 20 年积累下来、现在想配给 agent 的——今天已经通过 AgentCore 暴露出来了。 A lot of these capabilities you’re talking about that you’ve developed over 20 years and you’re trying to put it in place for agents are exposed today . ¶
Ben Thompson 那么由 OpenAI 驱动的 Bedrock Managed Agents 和 Bedrock AgentCore 之间是什么关系? So what is the relationship between Bedrock Managed Agents powered by OpenAI and Bedrock AgentCore? ¶
Matt Garman 我们一起构建的东西,很大一部分是在 AgentCore 的构件之上,把这些环节拉到一起。 A lot of what we’ve built together is building on the building blocks of AgentCore in order to kind of pull some of these pieces together. ¶
Ben Thompson 所以说,那是坐在其上的一层超集(super set)? So there’s like a super set that sits on top of that? ¶
Matt Garman AWS 团队和 OpenAI 团队一起,用 AgentCore 的组件加上 OpenAI 的模型和一批那些环节,共同打造了这款产品。 The AWS team and the OpenAI team used AgentCore components together with the OpenAI models and a bunch of those pieces to go and co-build this product together. ¶
Matt Garman AgentCore 算是我们提供的一套基础构件——就像在 AWS 上,如果你想搭建自己的 agentic 工作流,你完全可以自己动手。 AgentCore is kind of our set of primitives that just like if with AWS, if you want to go and build our own agentic workflows, you can do that. ¶
Matt Garman 你可以有记忆组件,可以有安全的执行环境,可以有权限能力,把所有这些配置起来;今天就有客户在生产环境里跑着这些,做出非常酷的东西。 You can have a memory component, you can have a safe execution environment, you can have a permissioning capability, and you can go and configure all of those and we have customers running those in production today that are doing really cool things. ¶
Ben Thompson 但用的不是 OpenAI。 But not with OpenAI. ¶
Matt Garman 对,不是 OpenAI,他们今天得用别的模型,这没错。 But not with OpenAI, they have to use different models today, that’s true. ¶
Matt Garman 其实,也不对——有人就在用 OpenAI 做。 Actually, that’s not true, we have people doing it with OpenAI. ¶
Ben Thompson 哦,就是远程调用另一个云或诸如此类。 Oh, just calling to another cloud or whatever. ¶
Matt Garman 他们就是直接调用 OpenAI 的模型。 They just call directly to the OpenAI model. ¶
Matt Garman 所以我们今天确实有人用 OpenAI 这么干,只不过不是原生地在 Bedrock 里面,但他们照样在用。 So we actually absolutely have people doing it with OpenAI today, not natively inside of Bedrock, but they’re still using that. ¶
Matt Garman 这是个开放的生态,你可以调用不同的能力去构建任何你想要的东西;我打赌人们会继续这么干。 And it’s an open ecosystem where you can pull different capabilities to go build whatever you want and my bet is that people will continue to do that. ¶
Matt Garman 我们外面那批建设者,呼应 Sam 的类比:今天明明可以不用了,他们还是乐意在家自己攒电脑。人们就是喜欢造东西,我们认为很长时期内都会有人构建自己的 agent;但其中绝大多数人会想要一条更轻松的路,不想自己去配置所有那些环节——这正是我们这次联合发布的一部分内容。 We have builders out there that love to, to Sam’s analogy, love to continue to build computers at home today, even though you don’t have to do that, and even though people like to build and we think that people for a long time will build their own agents, but the vast majority of them are going to want an easier way to do it where they don’t want to have to go configure all of those pieces themselves and that’s part of what we’ve launched in this collaboration together. ¶
Ben Thompson 为了彻底说清楚:你说的是 Bedrock Managed Agents 这套托管体验;但你也照样可以用 AgentCore,去调用模型,无论在 AWS 上还是在别处。 Just to be super clear, you talk about this managed experience with Bedrock Managed Agents, you can also use AgentCore and pull from a model, whether on AWS or somewhere else. ¶
Ben Thompson 再澄清一下——Sam,这个问题问您:比如在 Azure 上的 OpenAI,那只是你直接访问 API;这和亚马逊上这个托管服务是两回事。 And just to make clear, Sam, this is a question for you, this is the distinction between OpenAI on say, Azure, where that’s just you have direct access to the API, and that is distinct from this managed service on Amazon. ¶
Ben Thompson 是这样吗? Is that correct? ¶
Sam Altman 正确,是的。 Correct, yep. ¶
Ben Thompson 而你对这点感觉很好——各方面的边界都划得恰当,今后不会成为问题,对吗? And you feel very good about that, that’s scoped correctly in all terms, it’s not going to be an issue going forward? ¶
Sam Altman 是的,我认为情况会随时间演进,但作为一个起点,我对这件事感觉很好。 Yeah, I think things will evolve over time, but I feel very good about this as a way to start. ¶
Ben Thompson 这会是 AWS 的独家产品吗? Is this going to be an exclusive offering for AWS? ¶
Ben Thompson 还是说你们预计会在其他云上提供这种托管 agent 服务(managed agent)? Or do you anticipate having this sort of managed agent service on other clouds? ¶
Sam Altman 是的,这件事我们只与亚马逊独家合作,我们对此很兴奋。 Yeah, we’re doing this exclusively with Amazon, we’re excited about it. ¶
Ben Thompson 这个「独家」在多大程度上是因为「我们用了亚马逊所有的 API,当然只能在亚马逊上」;还是说它源于托管体验这个整体构想——不只是「我们用了亚马逊的 API」,而是「这东西现在就会长在亚马逊上」? How much of the exclusive is, “Look, we’re using all Amazon’s APIs, of course it’s only on Amazon”, or is this the overall idea of a managed experience, it’s not just a “We’re using Amazon APIs”, it’s, “Right now this is going to be on Amazon”? ¶
Sam Altman 在精神上,我们想把它做成两家公司之间的联合事业。 Spiritually, we want to do this as a joint effort between our companies. ¶
Ben Thompson 明白。 Got it. ¶
Ben Thompson 新闻稿确实提到了这一点。Matt,这又回到你早先说的:你可以去调用别的 API,自己把所有东西粘起来。 The PR does say something, and this goes back to the point you mentioned, Matt, earlier about you could call out to other APIs and glue this all together yourself. ¶
Ben Thompson 在这种情况下,客户数据留在 AWS 内部,那么 OpenAI 究竟能看到什么?这意味着什么? In this case, the customer data stays within AWS, so what exactly does OpenAI see, what does that mean? ¶
Matt Garman 没错。 That’s right. ¶
Matt Garman 整个东西基本都留在你的 VPC 里,所以数据在 Bedrock 环境内部受到保护。 So the whole thing kind of stays within your VPC and so data is protected inside of the Bedrock environment. ¶
Ben Thompson 明白。 Got it. ¶
Ben Thompson 而且这将是通过 Bedrock 来运行 OpenAI 模型的服务——这些会跑在 Trainium 上吗? And this is going to be running on OpenAI models through Bedrock, and these are going to be ? ¶
Matt Garman 它们会是不同方案的混合——一部分跑在 Trainium 上,一部分跑在 GPU 上。 They’ll be through a mix of different – some of it will be on Trainium, some of it will be on GPUs. ¶
Ben Thompson 这只是时机早晚的问题吗? Is that just a function of timing? ¶
Ben Thompson 因为我想几个月前你们宣布的时候—— Because I think as part of your announcement a couple of months ago — ¶
Matt Garman 一部分是时机,一部分是能力。我想我们在共建这个系统的不同组件时会混合搭配,为合适的部分选用合适的基础设施。 Some of it’s timing and capabilities, I think we’ll kind of be mixing in the different components of building the system together, using the right infrastructure for the right parts of it. ¶
Matt Garman 但随着时间推移,越来越多的部分会跑在 Trainium 上。 But over time, more and more of it will be on Trainium. ¶
Sam Altman 我们非常兴奋能让这些模型跑在 Trainium 上。 We are quite excited to get these models running on Trainium. ¶
Ben Thompson 可以想象。 I can imagine. ¶
Ben Thompson 快速问一个问题,Matt,就是关于 Trainium 的一个一般性问题。 One quick question, just a general question about Trainium, Matt. ¶
Ben Thompson Trainium——我是这么想的,所以想确认自己理解得对不对。 Trainium, is it fair to think, and this is the way I’m thinking about it, so I want to make sure I have it right. ¶
Ben Thompson Trainium——这个名字起得实在太不走运,因为往后它其实主要是做推理(inference)的——它最主要的呈现方式将是通过 Bedrock 这样的托管服务,客户甚至未必知道自己在用什么算力。这样理解公平吗? Trainium — very unfortunately named, because it’s really going to be about inference going forward — the number one manifestation will be through managed services like a Bedrock, where the customer doesn’t even necessarily know what compute they’re using, is that a fair way to think about it? ¶
Matt Garman 首先,AWS 所有服务取名糟糕这件事,我来承担责任。 Number one, I take responsibility for bad naming across all AWS services. ¶
Ben Thompson 你看,我运营着一个靠口碑传播、名叫 Stratechery 的网站,所以对取名难我是完全感同身受的。 Look, I have a word-of-mouth site named Stratechery, so I have all sympathy for bad naming. ¶
Sam Altman 我觉得 Trainium 是个很酷的词。 I think Trainium is a cool word. ¶
Matt Garman 它确实是个很酷的词。 It is a cool word. ¶
Ben Thompson 词是很酷,只是感觉它是枚推理芯片,不是训练芯片。 It is a cool word, it just feels like it’s an inference chip, not a training chip. ¶
Matt Garman 没错。 It is. ¶
Matt Garman 不过,抛开名字不谈,训练和推理它都有用。 But, yeah, naming aside, it is useful for both training and inference. ¶
Matt Garman 你看,这是一枚让我们无比兴奋的芯片。无论是当前这几代还是后续迭代,我们认为它将是一盘大生意,也是我们共同做成的许多事的真正赋能者。 And look, it’s a chip that we’re incredibly excited about, and both in the current generations as well as ongoing, we think that’s going to be a huge business and a real enabler for a lot of the things that we do together. ¶
Matt Garman 顺便说一句,即便是 GPU:你们与很多这类加速芯片的交互,也将是通过抽象层进行的。 I think just with GPUs, by the way, you’re going to interact with a lot of these accelerator chips through abstractions. ¶
Matt Garman 所以绝大多数客户其实也不直接和 GPU 打交道,除了也许在笔记本电脑之类设备上用来做图形。 So the vast majority of customers don’t interact with GPUs either, except through maybe like in their laptop or something like that, for graphics. ¶
Matt Garman 当你调用 OpenAI 时,即使它们跑在 GPU 上,你也不是在跟 GPU 说话;你调用 Claude,背后是 GPU、Trainium 或 TPU,你同样不是跟那些芯片说话,你面对的是接口。 But when you’re talking to OpenAI, even if they’re running on GPUs, you’re not talking to the GPUs, if you’re talking to Claude, you’re through GPUs or Trainium or TPUs, you’re not talking to any of those chips, you’re talking to the interface. ¶
Matt Garman 而外部世界绝大多数的推理,都是在屈指可数的几个模型之一上完成的。 And the vast majority of inference out there is being done on one of a handful of models. ¶
Matt Garman 所以无论是 5 个、10 个、20 个还是 100 个,直接针对这些东西编程的不会是千百万人。今后也会如此,就因为这类系统太复杂、规模太大。 And so whether it’s 5, 10, 20, 100, it’s not millions of people that are programming to those things directly, and that’s gonna be true going forward just because these systems are so complex, they’re very large. ¶
Matt Garman 要训练一个模型的话,没多少人有那么多钱去训练模型,也没多少人有真正驾驭它的专业能力。 If you’re going to go train a model, not that many people have enough money to go train a model, not that many people have the expertise to actually manage it. ¶
Matt Garman 它们是非常复杂的系统,而 OpenAI 团队从超大规模算力集群里榨取价值的能力,厉害得不可思议。 They’re very complicated systems, and the OpenAI team is incredible in their ability to squeeze value out of a very large compute cluster. ¶
Matt Garman 但无论芯片是什么,没多少人养得起能做这件事的团队。所以老实说,我认为对所有加速芯片而言这一点都成立。 But not that many people have the team that can do that, independent of what the chip happens to be, and so I think that that’s going to be true for all accelerator chips, honestly. ¶
Sam Altman Ben,我越来越觉得,我们这家公司要做的事,就是当一座 token 工厂。 Ben, I increasingly think of what we have to do as a company is to be a token factory. ¶
Sam Altman 但客户在意的是:我们能以最低的价格交付最好的智能单元(unit of intelligence),要多少有多少,容量管够。 But what the customer cares about is that we can deliver the best unit of intelligence at the lowest price and as much of it as they want, with as much capacity as they want. ¶
Ben Thompson 你觉得定价会一直这样——按 token 计价,长期看合理吗? Do you think we stick with pricing as far as — pricing is based on tokens, does that make sense in the long run? ¶
Sam Altman 不合理。事实上有个挺有意思的例子:就在我们刚发布的那个模型——5.5——上:它的单 token 成本比 5.4 高得多,但拿到同样答案所需的 token 数量少得多。其实你根本不关心答案花了多少 token,你只想把这活儿干完,你要的还是一个价格,以及为此能用到的容量。 No. And in fact, like there was an interesting example of this with , 5.5. where the per-token cost is much higher than 5.4, but it requires a hugely fewer number of tokens to get the same answer, and you actually don’t care about how many tokens the answer takes, you just want the piece of work done, and you want again a price and an amount of capacity you can have for that. ¶
Sam Altman 所以也许我说「token 工厂」说错了,我们更像一座智能工厂之类的东西。 So maybe I was wrong to say “token factory”, but we’re like an intelligence factory or something. ¶
Sam Altman 我们只想以最低的价格获得尽可能多的智能单元:是大模型跑更少的 token,还是小模型跑大量的 token,是 GPU、Trainium 还是别的什么,或者我们在这方面创造性地采用其他种种做法——我认为客户都不在乎。 We just want as many units of intelligence for the lowest price and whether that is a bigger model running fewer tokens, a smaller model running lots of tokens, whether a GPU or Trainium or something else, whether we do any of the other kind of number of things we could do about that creatively, I don’t think customers care. ¶
Sam Altman 事实上,他们根本不会接触到这些。 In fact, they don’t really interact with that. ¶
Sam Altman 当你往 Codex 里布置一个任务,或者在 SRE[Stateful Runtime Environment,有状态运行时环境]里构建一种新的 agent 时,你永远不必操心这些,只会惊叹于花这么少的钱竟能拿到这么多。 When you go put something into Codex or when you go build a new kind of agent in the SRE [Stateful Runtime Environment], you should never have to think about that and you should just be astonished at how much you get for how little cost. ¶
Ben Thompson token 用量的下降,是模型的功劳还是 harness 的功劳? Is the reduced token usage is that model, or is that harness? ¶
Sam Altman 主要是模型,一小部分是 harness。 That’s mostly model, it’s a little bit harness. ¶
Ben Thompson 明白。 Got it. ¶
Ben Thompson 顺便问 Matt——我问过 Sam 独家性的问题了——你们预计会针对其他模型提供类似的托管服务吗? Do you anticipate Matt, by the way, I asked Sam the exclusive question, do you anticipate offering a similar managed service for other models? ¶
Matt Garman 我们目前专注的就是和 OpenAI 做这件事。 We’re focused on doing this with OpenAI right now. ¶
Matt Garman 我们对一起做的事非常兴奋,而所谓「静待时间给出答案」,那个时间可是很长的。 We’re very excited about what we’re doing together, and the fullness of time is a long time. ¶
Ben Thompson 「那个时间很长」——行,这个说法我就让你留着了。 The fullness of time is a long time, I’ll let you stick with that one. ¶
Ben Thompson 没关系,反正这个问题我必须得问。 It’s fine, I had to ask the question. ¶
Ben Thompson 关于客户我确实有个问题。Sam,回到你的点子上,我想听听你们俩的看法,很好奇——当客户真正上生产环境后,OpenAI 的责任在哪里结束,AWS 的责任从哪里开始? I do have a question as far as customers, Sam, to your point, both your input on this, I’m curious — when people are actually in production, where does OpenAI’s responsibility end and AWS’s begin? ¶
Ben Thompson 在我看来,如果所有数据都在 AWS 上、也留在 AWS 里,而运营是在更高的层面,那说到底这就是 AWS 的责任? It sounds to me, if all the data is on AWS and it’s staying there, and they’re operating at a higher level, this is ultimately AWS’s responsibility? ¶
Ben Thompson 这样理解对吗——从用户的角度看,我是这么想的,对吗? Is that the right way — am I thinking about that correctly from a consumer perspective? ¶
Matt Garman 是的,我觉得没错。 Yeah, I think that’s right. ¶
Matt Garman 当你要找人时,你会打给 AWS 支持来帮你解决。它是你 AWS 环境的一部分,你们一起把它建起来,你的 AWS 客户经理会在那里帮你。 When you’re going to call somebody, you’ll call AWS support to help you out, and it’s part of your AWS environment and you build it together and your AWS account reps are going to help you there. ¶
Matt Garman 在构建过程中,我们会把 OpenAI 的同事拉进来,帮你们弄清楚如何最好地利用这套东西,诸如此类。 And we’ll bring in, when we’re building it, we’ll bring in our OpenAI colleagues to help you figure out how to best take advantage of this or whatever. ¶
Matt Garman 到了某个节点,如果我们碰到需要他们协助的 bug,会升级转给他们;但 AWS 会是你打交道的第一线支持。 At some point, if we run into a bug that we need their help with, we’ll escalate over to them, but AWS will be that frontline support that you kind of interact with. ¶
Ben Thompson Sam,你觉得这块业务的规模,和你们核心 API 业务相比会到什么程度? Where do you see the scale of this business, Sam, relative to your core API business? ¶
Sam Altman 我希望它会非常大。我们为此投入了大量精力,我们承诺采购大量算力,我相信那里会有大量营收来支撑这一切。 I hope it’s going to be huge, we’re putting a lot of effort into this, we’re committing to buy a lot of compute, I believe there will be a lot of revenue there to support this. ¶
Sam Altman 我越来越倾向于这样一个框架:只要价格足够低,对智能的需求基本上是上不封顶的。 The increasing framework that I’ve had is that at a low enough price, demand for intelligence is essentially uncapped. ¶
Ben Thompson 所以在这方面它非常有弹性? So is it very elastic in that regard? ¶
Ben Thompson 价格降下来,需求就涨上去? You decrease price, demand goes up? ¶
Sam Altman 肯定有这个成分。但话说回来,水降价,你也许多喝一点,也许从一天洗一次澡变成洗两次;弹性是有的,但到某个点你就会说:「好了,我的水够多了。」 It’s certainly that, but again, you can decrease the price of water and maybe you’ll drink a little more water, maybe you’ll shower twice a day instead of once a day, there’s some elasticity there but at some point you’re like, “You know what, I have enough water”. ¶
Ben Thompson 而且你要是非得用水,多少钱你都买。 Also you will buy water no matter how much it costs if you have to. ¶
Sam Altman 其他公用事业(utility)类的资源,比如电便宜了你肯定用得更多;但如果你把智能看作一种公用事业,据我所知没有别的公用事业是「我就是想要更多——只要价格够低,我就是要用更多、更多」的。 Other utilities, if electricity is cheaper you’ll certainly use more of it, but if you think about intelligence as a utility, there’s no other utility I know of that I’m just like, “I just want more, I’ll just use more as long as the price is low enough, I’ll just use more”. ¶
Matt Garman 我还真要说,有意思的是,算力在很大程度上一直如此:想想今天一次计算周期的成本与 30 年前相比——便宜了多少个数量级我都不知道了——而今天卖出的算力比历史上任何时候都多。 I will say actually and interestingly it’s largely been true of compute power where if you think about the cost of a compute cycle today versus what it was 30 years ago, like I don’t even know how many orders of magnitude cheaper, and there’s more compute being sold today than ever. ¶
Ben Thompson 对。 Right. ¶
Ben Thompson 人们其实不太考虑算力的成本——至少在它高到构成实质影响的水平之前——总体来讲,战略上大家默认你就是有算力的。 People don’t really think about the cost of compute at least until they’re at extremely high levels it’s a material level, but by and large strategically speaking it’s just assumed you have compute. ¶
Ben Thompson AI 要走到那一步——让「我在这儿花了多少钱」不再是第一思考过程——还需要多长的跑道? What’s the runway to getting there with with AI where it’s not the number one thought process, “How much am I spending here?”. ¶
Sam Altman 我不认为那现在是排在第一位的思考。 I don’t think that is the number one thought process. ¶
Sam Altman 现在问我们『不管价格是多少,你们能给我更多吗? Right now we have way more customers asking us, “No matter what the price is, can you give me more? ¶
Sam Altman 『我就是需要更多容量,我多付钱』的客户,远远多于为价格跟我们争执的。 I just need more capacity, I’ll pay you extra”, than we have arguing with us about the price. ¶
Sam Altman 但我确实认为我们会继续把价格疯狂地大幅打下来;而且说不定我们越是这么做,想要流动的财富总量就越是越来越多、越来越多、越来越多。 But I do think we are going to continue to bring the price down crazily dramatically, now maybe the more we do that the amount of wealth that wants to flow and just goes up more and more and more. ¶
Sam Altman 但我有信心,我们会继续相当大幅地降低当前这一档智能水平的成本——有一点多少让我惊讶,我不知道这种情况会不会持续,但至少在今天,整个市场需求中竟然有那么大比例是在绝对前沿(frontier)上的。 But I am confident we will continue to be able to reduce the cost of today’s level of intelligence quite dramatically — one thing that has somewhat surprised me is how much, and I don’t know if this is going to stay the case or not, but at least today how much of the total market demand is at the absolute frontier. ¶
Ben Thompson 对,关于这一点有很多问题。 Right, there’s a lot of questions about that. ¶
Ben Thompson 服务前沿(front end,指 frontier)模型成本非常高,人们大可以用上一代就好,但你的意思是:无论如何人们就是想待在前沿上? It’s very expensive to serve the front end, people can just get the previous one, but you’re saying people just want to be on the front end no matter what? ¶
Sam Altman 到目前为止确实如此。 So far they do. ¶
Matt Garman 我觉得这是个很好的信号:说明我们离想到达的地方还差得远,还有多得多的需求。我真觉得就像看 40 年前的算力需求:那会儿一台电脑贵得离谱,而今天它的算力被人人手里的手机远远甩开,而这些设备的销量还要多卖出几十亿台。 And I think that’s a good signal that you’re not anywhere close to where we want to be and that there’s so much more demand, and I really do think it’s like if you go 40 years ago to compute demand, a computer was crazy expensive, and now it’s dwarfed by the the power that’s in everybody’s cell phone and we sell billions more of those things. ¶
Matt Garman 我确实认为 AI 世界会重演这一幕。在今天这样往前推进的阶段,人人都想用前沿模型,因为要把大量有用的活干成,你需要的就是前沿;而且每个人都被眼下这些能力兴奋着。 I do think that that’s what’s going to happen to the AI world where today you’re pushing, everybody wants to use the frontier because that’s what you need in order to get a lot of useful work, and everyone’s so excited about the capabilities out there. ¶
Matt Garman 顺便说一句,我认为随着时间推移,你会拥有一个混合的模型格局:会有一些更小的模型,能做到连 OpenAI 最新模型都还做不到的事;而且它们会越来越小、越来越便宜、越来越快。也会有超大模型,去尝试攻克癌症之类的难题。 I think over time, you will have a mix of models, by the way, where you will have some smaller models that are able to do stuff that even the latest OpenAI models aren’t able to do yet, but they will be smaller and cheaper and faster over time, and you’ll have the super big ones that are going to go try to cure cancer and other things like that. ¶
Matt Garman 但我认为我们还处在「可能性」的极早期。在可能性还很早期的阶段就能看到这么多需求、这么快增长,未来蕴含的东西令人兴奋。 But I think we’re still at just the early stages of what’s possible and when you see this much demand and this much growth when you’re at the early stages of what’s possible, it’s exciting for what the future holds. ¶
Ben Thompson 这里有没有一种犬儒式的解读:Sam,你有一堆客户说「我们很想用 OpenAI 模型,但我们的东西全在 AWS 上,我们不动」。 Is there a bit of a cynical view here where, Sam, you had a bunch of customers that are like, “We’d love to use OpenAI models, but all our stuff’s in AWS, we’re not moving”. ¶
Ben Thompson 而 Matt,你们这边说「我们东西全在 AWS 上,能不能去把 OpenAI 模型请来?」——这件事正好满足那个需求。结果因为 AWS 是最大的,那个需求量是个天文数字。 And Matt, you’re like, “Look, all our stuff’s in AWS, can you please go get OpenAI models?”, and this is just satisfying that need — and it turns out, because AWS is the biggest, that was an astronomical amount of need. ¶
Ben Thompson 这是不是就是最省事的答案? Is that just the easiest answer? ¶
Ben Thompson 还是说这里面也有一点:你们真心认为能交付某种高度差异化的东西,而且它还会为你们彼此各自吸引来新客户? Or is there a bit here, too, where you actually think you can deliver something highly differentiated that will also draw new customers for each of you? ¶
Sam Altman 我们显然非常高兴能接触到 AWS 的客户,喜欢 AWS 的人实在太多了。 We’re clearly thrilled to get access to AWS customers, and so many people love AWS. ¶
Sam Altman 对,那是句大实话。 Yeah, that is a true statement. ¶
Matt Garman 那部分当然是真的。 That part is definitely true. ¶
Ben Thompson (笑)行吧。 (laughing) Right. ¶
Matt Garman 反过来也一样,我们的客户对能用上 OpenAI 的技术非常兴奋。 And vice-versa, our customers are very excited to get access to OpenAI technology. ¶
Sam Altman 但我确实认为,我们一起有某种不可思议的、全新的东西要去造。我期待一年后人们回顾今天时,谈论的最重要的事不是「哦,终于可以通过 AWS 用上这些模型了」什么的,而是「哇,我们当初没意识到这款新产品有多重要」。 But I do think there is something incredible and new to build together, and I am hopeful that when people look back on this in a year, the most important thing people will talk about is not like, “Oh, finally, you can get access to these models via AWS”, or whatever, but it’ll be like, “Wow, we didn’t realize how important this new product was”. ¶
Sam Altman 我认为在模型、harness 和能力层面,我们已经接近一种完全不同的全新计算形态;它会和人们现有的「我需要一个通往某个模型的 API」之类的想法非常不一样。 I think we are close at a model and harness and capability level to just a completely new kind of computing and that will feel very different than the existing ways people have thought about, “I need an API to this model”, or whatever. ¶
Matt Garman 再同意不过了,正是如此。 I couldn’t agree more, that’s exactly it. ¶
Matt Garman 第一部分很棒也很好,但我觉得第二部分才是我们都会超级兴奋的东西。 The first part is great and is nice and the second part is, I think, what we all get super excited about. ¶
Ben Thompson 说到这个,我前面提过想绕回来的地方:我有一个理论,可能对也可能不对,很好奇你们的看法——关于还有东西要造这一点。 To that point, I mentioned I want to come back to this earlier, but I have a theory, which may or may not be correct, I’m curious your guys’ point about this, about stuff to be built. ¶
Ben Thompson 具体来说,最终可能会真有这么一个中间件或中间层:组织里有各种数据库、SaaS 应用、散落在各处的数据片段,它们可以横跨多个系统;然后这个 agent 层/harness——或者说是配合 harness——坐在最上面;而中间有东西要造,Managed Agents 多少触及了这一点。 Specifically, there may end up being this real middleware or middle layer of where you have all these different databases and SaaS apps and all these bits and pieces of data in an organization that can stretch across things, you have this agent layer/harness or with the harness, I guess, sitting on top, and there’s something to be built in the middle and gets at this a little bit. ¶
Ben Thompson 这算那个中间层的一部分吗? Is this part of this? ¶
Ben Thompson 还是说那是另外有待去造的东西? Or is this something to be built? ¶
Ben Thompson 还是说我完全想歪了,我们根本不需要那层东西? Or am I totally off base and we don’t need that at all? ¶
Sam Altman 那里确实需要有个什么东西,你说的完全对。 You are totally right that we need something there. ¶
Sam Altman 我最近和客户——比如大型企业——聊,他们会说:「我想要某种 agent 运行时环境;我想要一个管理层,能把我的数据接到 agent 上,还能让我搞清楚自己把 token 花在了哪里、没花在哪里,对它保持某种监督;我还想要某种工作空间(workspace)」——希望就是 Codex——「给我员工用的类似东西」。大家开口要的这套东西已经惊人地一致,但要真正把这一整套交付建起来,还有活儿要干。 When I’ve been talking to customers recently, like large enterprises, they’re like, “I want some sort of agent runtime environment, I want a management layer where I can connect my data to agents and also make sure that I understand where I’m spending on tokens and not and have some sort of oversight there, and I want some sort of workspace” — hopefully it’ll be Codex — “something like that for my employees”, and that package of what people are asking for is getting remarkably consistent, but there is work to go off and now go build all that offering. ¶
Ben Thompson 感觉上,几乎需要两层 agent 层。 It feels like there’s like almost a double agent layer that’s necessary. ¶
Ben Thompson 一层 agent 层用来维护那个中间层——不断在各种数据源里做洞穴探险(spelunking);然后还有真正的用户界面层,那才是人们实际交互的地方。 There’s like the agent layer to maintain the middle layer that is constantly spelunking down in all these data sources and then there’s the actual user interface layer that is where people are actually interacting with. ¶
Ben Thompson 这符合我们要去的方向吗,还是跑偏了? Does that sort of fit with where we’re going or is that off base? ¶
Sam Altman 这两点我都同意:这确实是世界今天的样子。 On both of those, I agree that that’s a picture of how the world looks today. ¶
Sam Altman 而当模型真的聪明起来,我并不认为我们确切知道未来的架构会长什么样。 As the models get really smart, I don’t think we know exactly what the architecture of the future is going to look like. ¶
Sam Altman 眼下,在这个姑且称为用户 agent 层的层面,人们确实想与多个 agent 交互;我们让你能为这件事那件事构建 agent,让它们互相对话,如此等等。然后在公司管理层,人们有各种各样的控制手段,管着怎么让 AI 去文件系统里翻找文件做探索。 Right now people do, at this sort of call it user agent layer, want to interact with multiple agents and we make it so that you can build agents for this thing and that thing and they can talk together and whatever else and then at the company management layer, people have all these controls about how you help the AI go spelunk and files in file systems. ¶
Ben Thompson 到某个节点你会意识到,自己纯属无缘无故地抱着过去不放:这些就该直接放进模型里。 And at some point you realize that you’re just holding on to the past for no reason at all, this should just be in the model. ¶
Sam Altman 我正要说这个。 That’s what I was going to say. ¶
Sam Altman 到某个节点,你可能会说:「其实我们已经拥有这么不可思议的能力了,把整个东西重新架构一遍吧。」 At some point, you may say, “Actually, we have such incredible capabilities, let’s re-architect the whole thing”. ¶
Matt Garman 是的,我同意。 Yeah, I agree. ¶
Matt Garman 而且我认为那会是某种不同的东西,我们还没准都知道它具体是什么——但那也正是它美妙的一部分:客户在用、在建,你能从他们身上学习,琢磨怎么为他们把它做得更简单、更快、更好。 And I think there’s something different, and I’m not sure we all know what it is yet, but that’s part of the beauty also, is you get customers using and building and you can learn from them and figure out how you can make that easier, faster, better for them. ¶
Ben Thompson Sam,这是我们第二次做这种产品发布访谈了。上一次是 Kevin Scott 和新版 Bing——那时你对自己给 Google 构成的威胁相当自信。你觉得后来效果如何? Sam, this is the second time we’ve done one of these product launch interviews, it was with Kevin Scott and New Bing — you were pretty confident about the threat you posed to Google then, how well do you think that worked out? ¶
Sam Altman 我觉得我们做得比自己预想的更好。 I think we have done better than I expected. ¶
Sam Altman 我认为 ChatGPT 是 Facebook 之后第一个真正大规模的全新消费产品。 ChatGPT is, I think, the first really large-scale new consumer product since Facebook. ¶
Ben Thompson 这就是答案吗——你们做得比预期好,但主要体现在 ChatGPT 上,而不是其他领域? Is that actually the answer, you’ve done better than you expected, but it manifested mostly through ChatGPT as opposed to other other areas? ¶
Sam Altman 不,我觉得我们在 API 上也做得相当好,尤其是在 Codex 上。但我当时想的不是那个。 No, I think we’ve also done quite well on the API, particularly on Codex, but that was not what I was thinking at the time. ¶
Sam Altman 当时我想的是,这类新的语言接口也许会改变人们在互联网上查找信息的方式。你知道——Google 也绝对是一家了不起的公司;我觉得在很多方面,就其业务的广度和深度而言,Google 至今仍被低估。但相对来看,ChatGPT 的表现我很满意。 At the time, I was thinking maybe these new kinds of language interfaces are going to change the way people find information on the the Internet and you know — Google, also just absolutely phenomenal company, I think in many ways Google is still underrated just in terms of the breadth and depth of what they do, but I am happy with how ChatGPT has performed relatively. ¶
Ben Thompson Matt,类似地,我其实有个关于 Google 的问题想问你。 I actually have a Google question for you Matt, in a similar way. ¶
Ben Thompson Google 这周刚高调亮相:他们那套完全整合的栈,从模型到芯片再到 agent 层一路打通,诸如此类。 Google was just up there this week, their fully integrated stack, all the way up and down from model to chip to to agent layer, all that sort of thing. ¶
Ben Thompson 而你现在和另一家公司的高管同台,按定义在 Amazon 内部并不完全整合。但有没有这样一种情形:此前大家都批评你没有前沿(frontier)模型——而现在进入这种推理(inference)领域,你们本来就服务大量公司、驾轻就熟。 You’re here with another company executive, definitionally not fully integrated within Amazon, but is there a bit where everyone was critical of you not having a frontier edge model — now that we’re in this sort of inference area, you’re used to serving a lot of companies. ¶
Ben Thompson 会不会因为某种程度上的中立,你们反而落在了一个更好的位置? Did you maybe end up in a better spot by being neutral in a way? ¶
Ben Thompson 那是有意为之,还是无意间撞进了一个你们当初都没意识到的好地方? Was that on purpose or did you accidentally end up in a great place that you didn’t realize it was going to be? ¶
Matt Garman 有一点是故意的。 A little bit on purpose. ¶
Matt Garman 我们——从创办 AWS 开始——一直都把合作伙伴视为我们服务终端客户的关键一环。 We, since we started AWS, we have always embraced our partners as a key part of us supporting our end customers. ¶
Matt Garman 从一开始,与合作伙伴深度协同就是我们战略中极其重要的一部分。也许和一些同行不同,我们把成功定义为:伙伴成功、他们在我们之上或与我们一起构建,他们成功我们就成功,那就太好了。 Since the very beginning, it’s been an incredibly important part of our strategy is to lean in with partners and maybe different than some others, we view our success is if the partners are successful and they’re building on top of us or together with us, and if they’re successful, then we’re successful, that’s awesome. ¶
Matt Garman 我们把这看作一起把蛋糕做大,那就是赢——但这不一定是别人看待世界的方式。 We view it as that’s growing the pie together, then that’s a win, and it’s not necessarily how others view the world. ¶
Matt Garman 有时候他们会说「我必须拥有一切」,那没关系,那也是一种人秉持的观点。 Sometimes they say, “I have to own everything”, and that’s okay, that’s a view that people have. ¶
Matt Garman 但我认为这种选择很重要,因为那样,最好的产品才会胜出。 But I think that choice is important, and that way the best products win. ¶
Matt Garman 顺便说一句,在那个世界里既可以有自家(first-party)产品,也可以有大量第三方产品;但我们的观点是,要让客户能为自己挑到最好的东西。 And by the way, you can have first-party products in that world, you can have lots of third-party products in that world, but our view is we want the customers to be able to pick the best thing for them. ¶
Matt Garman 如果最好的东西就是你自己做的,那太好了。 And if the best thing is your own stuff that you’re building, awesome. ¶
Matt Garman 对我们来说,如果最好的东西是合作伙伴做的、但构建在我们之上,我们也视之为赢,因为那是给我们的客户最好的东西。 For us, if the best thing is what our partners are building, but it’s on top of us, we view that as a win as well, it’s because it’s the best thing for our customers. ¶
Matt Garman 我们早就这么想了,实际上我们构建 AI 世界的 Bedrock 平台用的就是这个思路。 We’ve long thought that, and it’s actually how we built the Bedrock platform in the AI world. ¶
Matt Garman 我们想支持尽可能宽的一组模型,想支持尽可能宽的一组能力;确实如此——从数据库到计算平台再到其他这类东西,一直是这样。 We want to support a broad set of models, we want to support a broad set of capabilities, and it’s true, it’s been true across from databases to compute platforms to other things like that. ¶
Matt Garman 所以我认为这是一项有意为之的战略;我认为这是一项客户领情的战略,因为他们喜欢这样;我们很高兴继续全力投入它。 So I think it’s been an intentional strategy, I think it’s a strategy that customers appreciate because they like that, and we’re excited to continue to lean into it. ¶
Ben Thompson 是啊,很有意思。 Yeah, it’s interesting. ¶
Ben Thompson 软件、平台、基础设施之间存在一种平衡,而每家都说自己谁都服务。 There’s the balance between software, platform, infrastructure, and everyone says they’ll serve everyone. ¶
Ben Thompson 不过感觉上,回到 AWS 创办之初:你们是从 I(Infrastructure,基础设施)起步的,而这几乎给了——在我看来——最大的灵活性,让你能和 Sam 在中间地带会合。 But it does feel like you go way back when AWS started, it’s like you start with the I [Infrastructure], and that gives you almost – that gives you the greatest flexibility, it feels like, from my perspective, to meet Sam in the middle. ¶
Ben Thompson Sam 手握很棒的 S(Software,软件),你们俩要一起造出一个 P(Platform,平台)——大概可以这么讲。 Sam’s got a great S [Software], you guys are building a P [Platform] together, I guess is the way to put it. ¶
Matt Garman 没错。 That’s right. ¶
Matt Garman 这确实也有难的地方:你说「我们就一个 S3」,没有别家 S3 可选——这一点属实。 It does make it hard where you say, “We have one S3”, there’s not other S3 offerings, that part is true. ¶
Matt Garman 所以正如你所说,一些核心组件位于基础设施层;在那个层面,我们确实非常倚重我们自己做的东西。 So some of those core components are, like you said, at the infrastructure layer, we do lean in pretty heavily on the stuff that we build. ¶
Matt Garman 但越往栈的上方走,我认为能力谱系越宽。你要是这么看世界:在任何世界里,都不会有一家公司拥有每一个应用。而越往下走,到了模型与服务层,这样的东西就没那么多了;再往下到基础设施层,就更少了。我们的观点是:拥抱整个伙伴生态,对我们、对终端客户都是好事。 But as you move up that stack, I think there’s a broader set of capabilities and if you view the world that — in no world do I think any one company is going to own every application and as you get further down the stack, when you get to kind of the models and services layer, there’s fewer of those and you get down the infrastructure, there’s even fewer of those and our view is kind of embracing that whole set of partners is great for us end customers. ¶
Ben Thompson Sam,最后还有什么要说的吗? Sam, any final words? ¶
Sam Altman 我觉得刚才那番话说得非常到位。 I think that was very well put. ¶
Sam Altman 我真觉得存在一种可能:开发者如今能构建的产品种类会迎来一代新生。考虑到我们预期未来一年模型能力的进步会有多陡,我们要一起走这段旅程、真正去建一个能承载它的平台,这来得正是时候。我觉得人们会爱上它的。 I really do think there’s a potential at a new generation of the kinds of products that developers can now build and given how steep we expect model capability progress to be over the next year, the fact that we’re going to go on this journey together and try to really build a platform to enable it, is coming at a good time, and I think people are going to love it. ¶
Ben Thompson 非常好。 Very good. ¶
Ben Thompson Matt、Sam,感谢做客 Stratechery。 Matt, Sam, thanks for coming on Stratechery. ¶
Matt Garman 太好了。 Awesome. ¶
Matt Garman 谢谢邀请我们来。 Thanks for having us. ¶
Sam Altman 谢谢。 Thank you. ¶