BuildSpeak每日 builder 文摘
今日归档生词本关于
🎙 播客The MAD Podcast with Matt Turck· 2026 年 7 月 16 日· 7,059 词 · 约 35 分钟

OpenAI’s Compute Chief: We Can’t Build Fast Enough | Sachin Katti

SPACE 播放 / 暂停·←→ 上一句 / 下一句
Speaker 100:00 - 00:19
Anytime you have thought you have enough compute, we can slow down. Always negatively surprises like, oh, we should not have slowed down. Demand far outstrips compute supply today. So anything we can bring online, we consume immediately. Our biggest worry is that still at the scale at which we are trying to get compute and build compute.
Speaker 100:00 - 00:19
每当你觉得算力已经够了、我们可以放慢脚步时,结果总会带来负面惊讶——比如,哦,我们本不该放慢。如今需求远远超过算力供给。所以任何我们能上线的资源,都会立刻被消耗掉。我们最大的担忧仍然是:在我们试图获取算力、建设算力的这个规模上,情况依然如此。
Speaker 100:19 - 00:30
The physical world does not move that fast. We do believe that the world of recursion is not that far where AI will design the systems it needs to train and run the next generation of AI. Encourages.
Speaker 100:19 - 00:30
物理世界的推进速度没那么快。我们的确相信,递归(recursion)的世界已经不太遥远了——到那时,AI 将会设计出它为训练和运行下一代 AI 所需要的系统。令人振奋。
Speaker 200:30 - 00:42
Hi. I'm Matt Turk. Welcome to the MAD podcast. My guest today is Sachin Khalti, who holds what might be the most fascinating and relevant title in tech right now, head of industrial compute at OpenAI. Sachin has an incredible background.
Speaker 200:30 - 00:42
嗨,我是 Matt Turk。欢迎来到 MAD podcast。今天的嘉宾是 Sachin Khalti,他拥有一个也许是当下科技界最引人入胜、也最切题的头衔:OpenAI 的 industrial compute 负责人。Sachin 的背景非常惊人。
Speaker 200:42 - 01:13
He was a professor at Stanford, a multi time founder, and most recently the CTO at Intel. Now he's leading what many are calling the largest infrastructure build out in human history. In this episode, we step away from the model layer and dive deep into compute and the physical reality of the AI boom. We talk about the staggering scale of the data centers being built. We get into the weeds on liquid cooled supercomputers, power grid constraints, the potential of nuclear MG, and OpenAI's move into custom silicon with Jalapeno.
Speaker 200:42 - 01:13
他曾是 Stanford 的教授,多次创业者,最近还担任过 Intel 的 CTO。现在,他正领导着许多人称之为人类历史上最大规模的基础设施建设。在这一期节目中,我们暂时离开 model 层,深入探讨算力以及 AI 热潮背后的物理现实。我们会谈到正在建设中的 data center 的惊人规模,也会深入聊 liquid cooled supercomputer、电网约束、nuclear MG 的潜力,以及 OpenAI 通过 Jalapeno 进军定制 silicon。
Speaker 201:13 - 01:28
We also discussed the broadest target strategy and the slightly sapphire reality that AI is now beginning to help design its own chips. It is a fantastic look behind the curtain at what it actually takes to power the future of intelligence. Please enjoy my conversation with Sachin Cotti.
Speaker 201:13 - 01:28
我们还讨论了最广泛的 target strategy,以及一种略显残酷的现实:AI 现在已经开始帮助设计它自己的 chip。这是一场非常精彩的幕后观察,让人看到为智能的未来提供动力,实际到底需要付出什么。请欣赏我与 Sachin Cotti 的对话。
Speaker 301:30 - 01:42
Right, Sachin, welcome. Excited to do this. We are recording this on the sidelines of the RACE conference in Paris. So thank you for braving the heat. It's another heat wave here.
Speaker 301:30 - 01:42
好,Sachin,欢迎。很高兴来做这次访谈。我们这次录制是在 Paris 的 RACE conference 场边进行的。所以感谢你顶着高温前来。这里又迎来了一波热浪。
Speaker 101:42 - 01:44
Thank you. Great to be here. Great to be here.
Speaker 101:42 - 01:44
谢谢。很高兴来到这里。很高兴来到这里。
Speaker 301:44 - 02:12
To start, some people describe what's currently happening in the world of compute data centers as the largest infrastructure build out in history, bigger than the highway, bigger than railroads. And I'm I'm curious, one, if you agree, and and two, what it feels like on from the inside. Like, what's how do you view what you currently building at OpenAI?
Speaker 301:44 - 02:12
先从这里开始。有人把当前算力 data center 世界里正在发生的事,描述为历史上最大规模的基础设施建设,比 highway 更大,比 railroad 更大。我很好奇,第一,你是否同意;第二,从内部看这件事是什么感觉。比如说,你如何看待你目前在 OpenAI 正在建设的东西?
Speaker 102:12 - 02:49
It definitely feels like one of the largest things humanity has ever built effectively. Definitely bigger than many of the things that I've heard of. I'm not old enough to have experienced the highway build out, but no, it's feels exactly like what it sounds like. I'm in the belly of the beast, so to speak. Every day is we are making decisions beyond compute that historically from my previous role, for example, at Intel, we probably take months to make, given the magnitude of those decisions.
Speaker 102:12 - 02:49
这绝对让人感觉像是人类有史以来真正建成的最大规模事物之一。肯定比我听说过的很多项目都更大。我年纪还不够大,没有亲历过高速公路大规模建设的时代,但不,这种感觉和它听起来完全一样。可以说,我现在就在这头巨兽的腹中。每天我们都在做一些不只是 compute(算力)层面的决策;以我之前的经历来说,比如在 Intel,这种量级的决策,我们过去大概需要几个月才会做出。
Speaker 102:49 - 03:03
But the demand is so insatiable that, and it is growing so rapidly that we have to move very quickly. So, it's an intense time, but it's probably the most exciting thing an engineer would want to be part of.
Speaker 102:49 - 03:03
但需求实在是无止境,而且增长速度又非常快,所以我们必须行动得非常迅速。所以,这是一个高强度的时期,但对工程师来说,这大概也是最令人兴奋、最值得参与的事情之一。
Speaker 303:04 - 03:16
Yeah. And I read somewhere that OpenAI was planning on spending about 50,000,000,000 in compute this year. Is that is that still the rough number directionally? Directionally, that sounds about right. Yeah.
Speaker 303:04 - 03:16
对。我之前在某处看到,OpenAI 今年计划在 compute 上花大约 50,000,000,000。这个数量级现在大致还是对的吗?从大方向上看,听起来差不多。对。
Speaker 303:16 - 03:29
And the whole industry itself was gonna be 700,000,000,000 in compute spend this year as well. So like insane numbers, seems.
Speaker 303:16 - 03:29
而且整个行业今年在 compute 上的支出据说也会达到 700,000,000,000。所以这数字看起来真是夸张得惊人。
Speaker 103:29 - 03:41
Yeah. And it's probably continuing to grow, right? And a lot of build happening. So a lot of that is also going to translate to compute usage from people like us in a year or two.
Speaker 103:29 - 03:41
对。而且它大概还会继续增长,对吧?还有很多基础建设正在推进。所以其中相当一部分,在一两年后,也会转化成像我们这样的人实际使用的 compute 用量。
Speaker 303:41 - 04:07
Is the right way to think about this, that for OpenAI, it's a bit of a new world, right? There's obviously not quite a pivot because, obviously, the whole AI research is going, you know, full speed ahead. But, like, building a whole new business within the the the company, is that is that fair? Is that how people think about it? Because, obviously, building models is one thing, building data centers, it's a whole different world.
Speaker 303:41 - 04:07
对这件事,正确的理解方式是不是这样:对 OpenAI 来说,这有点像进入了一个新世界,对吧?显然这不完全算是一次 pivot(业务转向),因为很明显,整个 AI 研究仍然在全速推进。但像是在公司内部打造一整套全新的业务,这样理解公平吗?大家是这么看的吗?因为很显然,构建模型是一回事,建设 data center(数据中心)则完全是另一个世界。
Speaker 104:07 - 04:29
Yeah, mean, I think OpenAI always has had a fundamental belief that computers are the foundation of everything. Compute is the foundation for intelligence. And the way we keep continuing to scale intelligence and distribute intelligence is by having compute. And so that has never been different. I've never always been the belief.
Speaker 104:07 - 04:29
对。我是说,我认为 OpenAI 一直都有一个根本性的信念,那就是计算机是一切的基础。Compute 是 intelligence(智能)的基础。而我们持续扩展智能、分发智能的方式,就是拥有 compute。所以这一点从来没有变过。这一直都是我们的信念。
Speaker 104:29 - 04:54
I think what's becoming clear is to build the kind of compute we need. And at this scale, we have to not just rely on getting compute from our partners. We increasingly have to take a much more active in building and getting that compute that we need. So, it does absolutely feel like a new muscle that we're building in the company.
Speaker 104:29 - 04:54
我认为现在越来越清楚的一点是,要建成我们所需要的那种 compute,而且是在这样的规模之下,我们不能只依赖从合作伙伴那里获得 compute。我们越来越需要以更主动的方式,去建设并获取我们所需要的 compute。所以,这确实让人感觉像是公司正在锻炼出一种全新的 muscle(能力)。
Speaker 304:54 - 05:26
And maybe to anchor the conversation from the beginning, it would actually be very helpful to talk about what a data center is in reality. So I think like everybody knows that data centers are being built. But, you know, gun to one's head, like, I'm not sure that everybody could say, well, what is actually being built? Because we've been building data centers as an industry for cloud for decades at this point. So what is fundamentally different and new about the data centers that we're building for AI today?
Speaker 304:54 - 05:26
也许为了从一开始就给这场讨论建立一个锚点,先谈谈 data center(数据中心)在现实中到底是什么,其实会很有帮助。所以我觉得,大家都知道 data center 正在被建设。但你知道,如果真被逼着回答,我不确定每个人都能说清楚,到底在建的是什么。因为作为一个行业,我们为了 cloud(云计算)建设 data center 已经有几十年了。那么,如今我们为 AI 建设的 data center,究竟有哪些根本性的不同和新的地方?
Speaker 105:27 - 06:00
I think the biggest probably is the scale, right? So we are essentially building large supercomputers as we think about AI. And we build intelligence and deliver intelligence and models become more capable, we use it for more and more complex tasks. We need more and more bigger computers effectively. And so I think the best way to visualize data centers is giant factories that are turning electrons into tokens.
Speaker 105:27 - 06:00
我觉得最大的差别大概是规模,对吧?本质上,当我们思考 AI 时,我们实际上是在建造大型 supercomputer(超级计算机)。随着我们构建 intelligence(智能)并交付 intelligence,而且 model(模型)变得越来越强大,我们会把它用于越来越复杂的任务。这样一来,我们实际上就需要越来越大的计算机。所以我认为,理解 data center 最好的方式,就是把它想象成巨型工厂,把 electrons(电子)转化为 token。
Speaker 106:01 - 06:25
It's a popular phrase nowadays, but it actually has a lot of a ring of truth to it. So how do we take power? How do we take those electrons and actually use it to power chips that effectively are delivering intelligence? And the way I visualize it is large football fields, liquid gold because these chips run really hot. The temperatures on these chips are very, very high.
Speaker 106:01 - 06:25
这句话现在很流行,但它确实非常贴切。那么,我们怎样获取电力?怎样把这些 electrons 真正用于驱动 chips(芯片),而这些芯片实际上是在交付 intelligence?我脑海里的画面是:大到像一整片足球场,里面还有“liquid gold”,因为这些芯片运行时非常热。它们的温度真的非常、非常高。
Speaker 106:25 - 06:35
And so you'll cool them with liquids, you can't cool them with air. So a lot of liquid cooled, basically refrigerators effectively that are sitting alongside the building.
Speaker 106:25 - 06:35
所以你要用液体来给它们降温,不能靠空气冷却。也就是说,会有大量液冷设备,基本上就是成套的“冰箱”系统,实际上就摆在建筑旁边。
Speaker 306:35 - 06:44
And then on that topic while we're at it, the the cooling happens at the data center level or does it happen at the chip level or both?
Speaker 306:35 - 06:44
既然说到这里,冷却是在 data center 层面进行,还是在 chip 层面进行,还是两者都有?
Speaker 106:44 - 06:55
Both. Both. Right. So you need to cool the data halls, but you also need to need to cool the chips individually because it's not going to be enough to do one or the other. And you also have to cool the things that connect chips.
Speaker 106:44 - 06:55
两者都有。两者都有,对吧?你既需要给 data hall(数据机房)降温,也需要给每一块 chip 单独降温,因为只做其中一项是不够的。而且你还得给连接 chips 的那些部件降温。
Speaker 106:55 - 07:10
And so that's why you need cooling pretty much everywhere nowadays. Even the cables that are the transformers that distribute the power become too hot. So they also need to be cold. So everything that processes energy produces heat.
Speaker 106:55 - 07:10
所以这就是为什么现在几乎到处都需要冷却。甚至连那些 cables(线缆)、以及负责分配电力的 transformers(变压器)都会变得太热,所以它们也需要被冷却。凡是处理能量的东西,都会产生热量。
Speaker 307:10 - 07:21
And isn't cooling technology that is being used as something that's well understood and is just getting deployed? Or is there like fundamental new things happening in cooling right now?
Speaker 307:10 - 07:21
那么,正在使用的 cooling(冷却)技术,是那种已经被充分理解、现在只是持续部署的东西吗?还是说,当前 cooling 领域其实正在发生一些根本性的创新?
Speaker 107:21 - 07:49
I think liquid cooling has been around for some time, but has never been deployed at this scale. Innovation is more around how to make it reliable, how to make it cheaper, more scalable. Right. So there's a lot of innovation around that. There's also a lot of new innovation, new kinds of liquids, new kinds of materials that can absorb heat better because anything that can improve the efficiency of heat transfer is very important for data centers.
Speaker 107:21 - 07:49
我认为 liquid cooling(液冷)已经存在一段时间了,但从未以这种规模部署过。创新更多集中在如何让它更可靠、如何让它更便宜、更具可扩展性。对。所以围绕这一点有很多创新。还有很多新的创新,比如新型液体、新型材料,它们能更好地吸收热量,因为任何能提升热传递效率的东西,对 data center(数据中心)都非常重要。
Speaker 107:49 - 08:10
So we can then run the chips hotter, right? And there is a direct correlation between running a chip hotter and how powerful the computers. So the hotter the chip, the more memory bandwidth you get, the more flops you get. And so there's more there's a strong payoff. If you can cool well, that also means you can produce more intelligence.
Speaker 107:49 - 08:10
这样我们就能让 chip(芯片)在更高温度下运行,对吧?而芯片运行得更热,与计算机的性能之间有直接相关性。也就是说,芯片越热,你获得的 memory bandwidth(内存带宽)就越高,得到的 flops(浮点运算性能)也越多。所以这里的回报非常明显。如果你能把冷却做好,也就意味着你能产出更多 intelligence(智能)。
Speaker 308:10 - 08:27
All right. So gigantic factories, lots of cooling. The other part that seems to me very critical to any discussion is power and energy. So how do how do that work starting at a high level? You do you connect to the grid?
Speaker 308:10 - 08:27
好。那么,超大型工厂、大量冷却。另一个在我看来对任何讨论都非常关键的部分,是 power(电力)和 energy(能源)。所以从高层来看,这是怎么运作的?你们是接入 grid(电网)吗?
Speaker 308:27 - 08:29
Do you have your own power generation?
Speaker 308:27 - 08:29
你们有自己的发电能力吗?
Speaker 108:29 - 08:59
I think the early days, we all connected to the grid, and we still all would want to connect to the grid. At this point, we are beginning to hit and we are investing in generation infrastructure for the grid, transmission infrastructure for the grid. So whenever we build a data center anywhere, we make it a hard commitment that we are not taking power away from the grid. In fact, we are investing in the grid to generate new power so that we can consume it for data centers.
Speaker 108:29 - 08:59
我认为在早期阶段,我们都是接入电网的,而且现在我们也仍然都希望接入电网。到了这个阶段,我们开始面临电网容量的限制,因此也在投资电网的发电基础设施和输电基础设施。所以无论我们在哪里建设 data center,我们都会作出一个硬性承诺:我们不会从电网中夺走原有的电力。事实上,我们是在投资电网,新增发电能力,这样我们才能将这些电力用于 data center。
Speaker 308:59 - 09:01
What does that mean practically?
Speaker 308:59 - 09:01
这在实际中意味着什么?
Speaker 109:01 - 09:24
You have a grid somewhere. Has a certain generation and distribution capability, certain number of megawatts. Obviously, data center shows up. If there was spare capacity, then of course, the data center can use it. But if there isn't spare capacity, then we have to new gas or solar or hydro generation infrastructure to the grid.
Speaker 109:01 - 09:24
你在某个地方有一张电网。它有一定的发电和配电能力,有一定数量的 megawatt(兆瓦)容量。显然,data center 一旦出现,如果有富余容量,那 data center 当然可以使用它。但如果没有富余容量,那么我们就必须为电网新增 gas(天然气)、solar(太阳能)或 hydro(水电)发电基础设施。
Speaker 109:24 - 10:02
So we are investing and funding that build out. And then you have to build transmission lines, invest in transformer substations to distribute that power. So wherever we are building data centers, we are funding the development of all of that infrastructure. And so that's one of the things that we do want to emphasize, which is this is infrastructure that would otherwise not have been funded, if not for these data centers. And one of the side benefits of this big data center build out is the grid infrastructure of America and the world for that matter is getting upgraded very quickly.
Speaker 109:24 - 10:02
所以我们会投资并出资推动这些建设。然后你还得建设输电线路,投资 transformer substations(变电站)来分配这些电力。因此,无论我们在哪里建设 data center,我们都在为所有这些基础设施的发展提供资金。这也是我们确实想强调的一点:如果没有这些 data center,这些基础设施原本是不会得到资金支持的。而这轮大规模 data center 建设带来的一个附带好处是,America 乃至全球的电网基础设施,正在非常迅速地得到升级。
Speaker 110:02 - 10:28
And so that's the power piece. So whenever we can do that, we do that and consume power from the grid. But it's also good citizens of the because we are improving the infrastructure for everyone, not just for data centers, but also for households. In some places, are beginning to hit the limits of how much grid power we can build and consume. And so there everyone's looking at behind the meter.
Speaker 110:02 - 10:28
所以这就是电力这一块的核心。只要我们能这么做,我们就会这么做,并从电网消耗电力。但这也意味着我们是在做负责任的公民,因为我们是在为所有人改善基础设施,不只是为 data center,也包括家庭用户。在一些地方,我们开始触及电网电力可建设和可消耗规模的上限。所以现在大家都在关注 behind the meter(电表后)方案。
Speaker 110:28 - 10:43
We are also doing some beyond meter generation where we would have on-site power generation and distribution capability that does not come from the grid. But in fact, the data center becomes effectively self self sufficient Yep. In terms of power.
Speaker 110:28 - 10:43
我们也在做一些 beyond meter 的发电方案,也就是在现场具备发电和配电能力,而这些能力并不来自电网。实际上,这会让 data center 在电力方面变得基本上自给自足。对。
Speaker 310:43 - 10:46
Does gas turbines what what is it?
Speaker 310:43 - 10:46
是 gas turbines 吗?具体是什么?
Speaker 110:46 - 11:03
Today, it's gas turbines, especially in The US, because that's the most dense transportable form of energy and also the one that is quite widely available in The US. But there, are bottlenecked by the supply chain.
Speaker 110:46 - 11:03
目前是 gas turbines,尤其是在 The US,因为那是能量密度最高、又可运输的一种能源形式,而且在 The US 也相当普遍可得。但这方面实际上受制于供应链瓶颈。
Speaker 311:03 - 11:20
Do you think the the nuclear conversation is interesting? Mary recording this in France, which has a bunch of, you know, nuclear nuclear power generation. Nuclear systems have come back to the discussion in The US as well. Is it, like, something that you think about, anything is interesting? Absolutely.
Speaker 311:03 - 11:20
你觉得关于 nuclear 的讨论有意思吗?Mary,我们这期是在 France 录的,而 France 有很多 nuclear 发电。nuclear 系统在 The US 也重新进入了讨论范围。这是你会关注、会觉得有意思的事情吗?当然。
Speaker 111:20 - 11:44
It can't come soon enough. Okay. I think it is the densest form of energy we can all produce and consume, and it also clean. So I think definitely would be a good source of massive, scalable energy for all data centers. Obviously, outside of France, the rest of the world has a lot of catching up to do and building this infrastructure.
Speaker 111:20 - 11:44
它来得越快越好。好。我认为这是我们所有人能够生产和消耗的能量密度最高的一种能源形式,而且它也是清洁的。所以我认为,它显然会成为所有 data center 所需大规模、可扩展能源的良好来源。很明显,除了 France 之外,世界其他地区在建设这类基础设施方面还有很多要补的课。
Speaker 111:44 - 11:49
But I think it's gonna be play a very important role and data center.
Speaker 111:44 - 11:49
但我认为,它将在 data center 中发挥非常重要的作用。
Speaker 311:49 - 12:17
Okay. So that's a great introduction on on on data centers. The the other interesting bit of news that you guys recently had is Jalapeno. Mhmm. So now in for OpenAI, in addition to being the application business, consumer and enterprise, and then being in the model AI research business, and then the compute and data center business, it seems that OpenAI is in the chip business, if that's fair.
Speaker 311:49 - 12:17
好的。关于 data center,这算是一个很好的介绍。你们最近另一条很有意思的消息是 Jalapeno。嗯哼。所以现在对 OpenAI 来说,除了应用业务——面向 consumer 和 enterprise——以及 model AI 研究业务,再加上 compute 和 data center 业务之外,看起来如果这么说算公平的话,OpenAI 现在也进入 chip 业务了。
Speaker 312:17 - 12:26
So, like, completely full stack, but I'm I'm I'm curious, and we'll we'll go into some details about a little later later. But, like, in terms of, like, overall strategy, where where does that fit?
Speaker 312:17 - 12:26
所以,就是说,完全是 full stack(全栈)的。但我我我很好奇,我们我们稍后会再深入谈一些细节。不过,从整体战略的角度来看,这件事处在什么位置?
Speaker 112:26 - 13:13
As we begin to as we serve a pretty big fraction of the world's population, AI usage is exploding, inference is obviously becoming a big fraction of our workload. Right? It's consuming a lot of compute. And one of the other realizations is because we know what is the workload exactly, what is the model we want to run, we can co design the hardware to be super efficient in delivering those models. And so the strategic piece of the Jalapeno is how do we take advantage of knowing what the end workload is, what the model itself is, and design chips that are very efficient in serving those models.
Speaker 112:26 - 13:13
随着我们开始为全球相当大比例的人口提供服务,AI 的使用量正在爆发式增长,inference(推理)显然正在成为我们工作负载中很大的一部分。对吧?它消耗了大量 compute(算力)。而另一个认识是,因为我们确切知道工作负载是什么,知道我们想运行的 model(模型)是什么,所以我们可以对 hardware(硬件)进行协同设计,让它在承载这些模型时具备极高效率。因此,Jalapeno 的战略意义就在于:我们如何利用自己知道最终工作负载是什么、模型本身是什么这一点,来设计出在服务这些模型时非常高效的 chip(芯片)。
Speaker 113:13 - 13:38
So it really allows us to drive efficiency advantage, drive more tokens per watt. So the key metric that Jalapeno is optimizing is maximizing the number of tokens you can produce per watt. And because the world is constrained by power today, so the more tokens you can produce for the same number on on power, it's better for everyone. So we look at it as a very critical ingredient in scaling how we deliver intelligence to the world.
Speaker 113:13 - 13:38
所以这确实让我们能够建立效率优势,提升每瓦可产出的 token 数。Jalapeno 正在优化的关键指标,就是最大化你每消耗一瓦电能生成的 token 数量。由于当今世界受制于电力约束,所以在相同电力下,你能产出越多 token,对所有人就越有利。因此,我们把它看作是将 intelligence(智能)规模化交付给全世界时一个非常关键的组成部分。
Speaker 313:38 - 14:15
Great. So I'll I'll come back to help you in in in a second. But you mentioned you just mentioned inference and, you know, it's such an interesting evolution as well. Without commenting on necessarily what's going on at OpenAI specifically, is inference equally big or much bigger than training these days in terms of, like, usage of compute as as as have we shifted from, you know, being those type of very heavy pre training runs as the major use case for compute to now just inference being the majority?
Speaker 313:38 - 14:15
很好。那么我我我一会儿再回到这个话题帮你展开一下。不过你刚刚提到了 inference,而且,这本身也是一个非常有意思的演变。不一定非要评论 OpenAI 具体在发生什么,但如今就 compute 的使用而言,inference 是不是已经和 training(训练)一样大,甚至大得多?也就是说,我们是否已经从那种非常重的 pre-training(预训练)任务曾经是算力主要使用场景,转向了现在主要是 inference 占多数?
Speaker 114:16 - 14:38
No. Inference is a big, perhaps even the majority on compute. And I think one of the we don't like to make a distinction between training and inference because a lot of training is now inference. So when we train a new model, we are generating synthetic data, for example. That's inference.
Speaker 114:16 - 14:38
是的。Inference 很大,甚至可能已经占了 compute 的大多数。而且我认为,其中一个情况是,我们不太喜欢把 training 和 inference 明确区分开,因为现在很多 training 本身就是 inference。比如说,当我们训练一个新模型时,我们会生成 synthetic data(合成数据),这就是 inference。
Speaker 114:38 - 14:58
When we train a new model, we are doing post training, and that's inference. When you train a model, you're doing test and compute. That's all inference. So when we say training, a lot of the compute actually is inference even in that phase of the work. So inference is a fundamental building.
Speaker 114:38 - 14:58
当我们训练一个新模型时,我们会做 post-training(后训练),而那也是 inference。训练一个模型时,你还会做 test-time compute(测试时计算)。这些全都是 inference。所以当我们说 training 时,实际上在那个工作阶段里,很多 compute 其实也是 inference。因此,inference 是一个基础性的构件。
Speaker 314:58 - 15:32
Yep. Obviously, I cannot resist asking the inevitable question around the potential risk of overbuilding, given the lag between demand and usage and how long it takes to build a data center. I think you mentioned somewhere that you were deliberately very paranoid about the problems ahead in the next three years. Very paranoid about the surprises ahead, which sounds like a very healthy approach. So how do you how do you think about that?
Speaker 314:58 - 15:32
对。显然,我忍不住还是要问那个不可避免的问题:考虑到 demand(需求)和实际 usage(使用)之间存在时滞,再加上建设一个 data center(数据中心)需要很长时间,会不会存在过度建设的潜在风险?我记得你在某个地方提到过,你是刻意对未来三年将要出现的问题保持高度 paranoid(警惕),对未来的意外情况也非常 paranoid,这听起来是一种很健康的做法。那么你是怎么思考这件事的?
Speaker 315:33 - 15:41
Is there any way to mitigate that, or is just, like, how do we play do, like, a deep belief that this is the future and, you know, we should all just go, go, go?
Speaker 315:33 - 15:41
有什么办法可以缓解这种风险吗?还是说,这更像是一种深层信念——坚信这就是未来,而我们都应该一路 go, go, go?
Speaker 115:41 - 15:55
We have deep conviction in scaling. Right? And history has borne us out. So effectively, Redmi, for example, has tracked compute. We tripled compute and we tripled revenue.
Speaker 115:41 - 15:55
我们对扩展规模这件事有很深的信念。对吧?而且历史已经证明我们是对的。所以实际上,比如说,Redmi 一直是跟着 compute(算力)走的。我们把 compute 提高了三倍,收入也提高了三倍。
Speaker 115:55 - 16:13
And we believe that's I mean, that continues to be true. Demand far outstrips compute supply today. So anything we can bring online, we consume immediately. So there's no compute that is going waste as for us at least. So I think that conviction has not changed whatsoever.
Speaker 115:55 - 16:13
而且我们相信——我是说,这一点现在依然成立。如今的需求远远超过 compute 供给。所以任何我们能上线的算力,都会立刻被消耗掉。至少对我们来说,没有任何 compute 会被浪费。所以我认为,这种信念完全没有改变。
Speaker 116:13 - 16:47
And if anything, we are seeing that scaling laws on research and training continue to hold. And, potentially, the pace at which we are doing research is accelerating, right, because of AI itself. So AI is doing a lot of AI research now. And so one of the subtle implications of that is his previously, our researchers used to run experiments, and they needed computer run experiments. But the number of experiments they could run was limited by the number of human researchers they had, which is a scarce resource on the world.
Speaker 116:13 - 16:47
而且如果要说有什么变化,那就是我们看到,research(研究)和 training(训练)方面的 scaling laws(规模定律)仍然成立。并且,由于 AI 本身的作用,我们做研究的速度可能还在加快,对吧,因为 AI 现在已经在做很多 AI research(AI 研究)了。因此,这里面有一个微妙的含义:以前,我们的研究人员会运行实验,而他们需要用计算机来跑这些实验。但他们能做的实验数量,受限于他们拥有多少人类研究人员,而这在世界上是一种稀缺资源。
Speaker 116:47 - 17:08
Right? There's not a lot of people who can do AI research. Now if AI itself can do AI research, the number of experiments we can run explodes, and therefore, the amount of compute you need from research also explodes. So we don't see a world where we will have immune to best compute for the foreseeable future. Right?
Speaker 116:47 - 17:08
对吧?真正能做 AI research 的人并不多。现在如果 AI 本身也能做 AI research,那我们能运行的实验数量就会爆炸式增长,因此,research 对 compute 的需求也会随之爆炸式增长。所以在可预见的未来,我们并不认为会出现一个我们对最好的 compute 需求已经免疫的世界。对吧?
Speaker 117:08 - 17:17
When I was referring to surprises, my worry is more on the downside of we are not able to actually build all the compute we want.
Speaker 117:08 - 17:17
当我提到“意外”时,我更担心的是下行风险:我们实际上没法建成我们想要的全部 compute。
Speaker 317:17 - 17:19
Yeah. And and But this raises
Speaker 317:17 - 17:19
对。然后——但这又引出了
Speaker 117:19 - 17:31
the other way. It's the other way for us. Right? Because that was consistently in the case. We anytime we have thought we have enough compute, we can slow down, always negatively surprises like, oh, shit.
Speaker 117:19 - 17:31
另一面。对我们来说,情况其实是反过来的。对吧?因为一直以来情况都是这样。每当我们觉得自己已经有足够的 compute、可以放慢一点时,结果总会迎来负面的意外,比如,“哦,糟了。”
Speaker 117:31 - 17:45
We should not have slowed down. Right? And so our biggest worry is that still. And at the scale at which we are trying to get compute and break compute, the physical world does not move that fast. Right?
Speaker 117:31 - 17:45
我们当时不该放慢的。对吧?所以这仍然是我们最大的担忧。而且,按我们试图获取 compute 和部署 compute 的这个规模来看,物理世界的运转速度并没有那么快。对吧?
Speaker 117:45 - 17:54
Physical supply chains, factories don't move that fast. Door cannot add capacity that fast. So for us, the surprise is more on that direction than the other direction.
Speaker 117:45 - 17:54
现实中的供应链、工厂,没法移动得那么快。Door 也不可能那么快增加产能。所以对我们来说,出乎意料的更多是这个方向,而不是另一个方向。
Speaker 317:54 - 18:28
Very fascinating. You alluded to communities a minute ago, and obviously, that's a that's a key debate. So curious about your perspective on a on a on a spectrum where, you know, on the one hand, one extreme, you'd say, well, the AI industry and computer industry has a PR problem, and there's no problem. It's just the problems that we cannot explain it well enough to the other extreme, actually, those communities have a point. Where do you think the reality is?
Speaker 317:54 - 18:28
很有意思。你刚才提到了 communities,这显然是一个关键争论点。所以我很好奇你怎么看这样一个光谱:一端的极端观点会说,AI 行业和计算机行业只是有个 PR 问题,实际上并没有什么问题,只是我们没把它解释清楚;而另一端则会说,其实那些 communities 的担忧是有道理的。你觉得现实更接近哪里?
Speaker 118:28 - 19:01
I think anytime there's new technology, which is as as revolutionary as this technology is, there is always disruption that's gonna happen. But, inevitably, we have learned this over history that this always leads to better outcomes for society. Right? And so how do we draw a line from where we are today to that outcome, right, and explain to the world why this is the trajectory we all need to be on? I mean, we it's our responsibility to do that.
Speaker 118:28 - 19:01
我认为,任何时候出现一种像这项技术这样具有革命性的新技术,都一定会带来某种 disruption(扰动、颠覆)。但是,历史已经反复告诉我们,这最终总会为社会带来更好的结果。对吧?那么,我们该如何从今天所处的位置,画出一条通向那个结果的路径,并向世界解释,为什么这就是我们都需要走上的轨迹?我的意思是,这是我们的责任。
Speaker 119:03 - 19:23
On the community side, there's a little bit of a local versus a global issue. Right? On the communities, I think data centers are, even today, a net positive to every community. Because we are building these data centers in rural areas of America, example, right, where there's nothing else that is being built. I'm this clear.
Speaker 119:03 - 19:23
从 community 的角度看,这里面有一点 local 和 global 之间的问题。对吧?就 communities 而言,我认为 data center(数据中心)即使在今天,对每个社区总体上也是净正面的。因为我们建设这些数据中心的地方,比如在 America 的农村地区,对吧,那些地方原本根本没有别的建设项目。我想把这一点说清楚。
Speaker 119:23 - 19:47
So we show up in rural Texas. We build a data center that produces new property taxes sinks for the community, that funds schools, that funds hospitals. We show up, and we invest in new grid infrastructure, which otherwise would never happen because there's no demand. So there's a modernized grid that that AI can enjoy. Basically, we produce jobs.
Speaker 119:23 - 19:47
比如我们来到 Texas 的农村地区。我们建一个 data center,它会为社区带来新的 property tax 收入,用来资助学校,资助医院。我们到来之后,还会投资新的电网基础设施,而这些原本根本不会发生,因为那里没有需求。于是就有了一个现代化的电网,AI 也可以从中受益。归根结底,我们创造了就业。
Speaker 119:47 - 20:15
Nice. And and and so I think one of the things we are investing a lot in is explaining the local benefits every time we build a data center somewhere and making sure that it is well understood the kind of upside that decides. And data centers, once they are built, are essentially very clean citizens. Right? They don't produce any gases or toxic chemicals or anything.
Speaker 119:47 - 20:15
很好。然后,所以我认为,我们投入很多的一件事,就是每次在某个地方建设 data center 时,都去解释它在当地带来的好处,并确保人们充分理解这类项目所带来的 upside(上行收益、好处)。而且 data center 一旦建成,本质上就是非常“干净”的社区成员。对吧?它们不会排放气体,也不会产生有毒化学物质之类的东西。
Speaker 120:15 - 20:19
Right? They're self contained. They just produce intelligence. Yeah.
Speaker 120:15 - 20:19
对吧?它们是自成一体的。它们只是产出 intelligence(智能)。对。
Speaker 320:19 - 20:32
The typical question that comes up is water, and I think that's been debulked quite a bit by by research. But, what would maybe give us just color on how you all think about the the water machine?
Speaker 320:19 - 20:32
通常大家会提到的问题是水,而我认为这一点已经被研究相当程度地 debunk(证伪、澄清)了。不过,也许你可以给我们一些背景,讲讲你们是如何看待 water 这套机制的?
Speaker 120:32 - 20:57
These are liquid cooled, and the liquid is recycled. So we do actually, the water consumption of a data center is shockingly small relative to household water consumption. So I think as you put it, it's been debugged. It's misperception that data centers consume a lot of water. It's anything they consume so little water for what they do, and all of that water is recycled.
Speaker 120:32 - 20:57
这些是液冷(liquid cooled)的,而且冷却液会被循环利用。所以实际上,相比家庭用水消耗,data center 的耗水量小得惊人。所以我觉得正如你所说,这个问题已经被澄清了。认为 data center 会消耗大量水,是一种误解。事实上,就它们所完成的工作而言,它们的用水非常少,而且这些水都会被回收循环。
Speaker 120:57 - 21:03
So we don't net consume new water. Once we get to a particular point, the water just gets recycled as they use the cold.
Speaker 120:57 - 21:03
所以我们并不会净消耗新的水。一旦达到某个运行状态,这些水就只是随着冷却过程被循环再利用。
Speaker 321:03 - 21:13
Yeah. So all those stories of, like, brown water is they just don't make sense because the water at a data center happens in, like, a contained circuit.
Speaker 321:03 - 21:13
对。所以那些所谓“brown water”之类的说法,其实说不通,因为 data center 里的水是在一个封闭回路中运行的。
Speaker 121:13 - 21:16
Right? It's a closed loop. Yeah. It's a closed loop. It's a closed loop.
Speaker 121:13 - 21:16
对吧?它是一个闭环。对,是一个闭环。就是一个闭环。
Speaker 321:16 - 21:30
And by the way, you mentioned Texas and rural areas since we talked about data centers at the beginning of this conversation. Why do OpenAI and other companies pick rural areas? Like, how do you select a site for a data center?
Speaker 321:16 - 21:30
顺便说一句,你提到了 Texas 和农村地区,因为我们在这次对话一开始谈到了 data center。为什么 OpenAI 和其他公司会选择农村地区?比如,你们是怎么为 data center 选址的?
Speaker 121:30 - 21:48
So many factors. So one is, of course, land, like plentiful land. Number two would be permitting. Like, can we build these things? And we wanna build these things such that they are not affecting any neighborhoods.
Speaker 121:30 - 21:48
因素很多。第一当然是土地,比如说土地要充足。第二是审批许可(permitting)。也就是,我们能不能把这些设施建起来?而且我们希望以一种不会影响任何社区居民区的方式来建设这些设施。
Speaker 121:48 - 21:59
Right? So land that is somewhat removed is their ideal candidate. Course, access to power. Right? So a strong grid, strong gas, availability.
Speaker 121:48 - 21:59
对吧?所以,距离居民区有一定距离的土地就是理想选择。当然,还有电力接入。对吧?要有强大的电网(grid)、充足的天然气供应能力。
Speaker 121:59 - 22:11
All of those are important factors. And then four is labor. Right? So how quickly can you build these things? So availability of labor, construction labor, qualified electricians, bloggers, all this can be above.
Speaker 121:59 - 22:11
这些都是重要因素。第四个因素是劳动力。对吧?也就是你能多快把这些设施建起来?所以劳动力的可获得性、建筑工人、合格的电工,以及其他所有这些,都会包括在内。
Speaker 122:11 - 22:27
So all of those factors go into every single site selection decision. I know, obviously, Texas isn't popular because it fits a lot of these criteria, but it's not the only state. I mean, we have data centers all around the country in LA plus. Okay. Great.
Speaker 122:11 - 22:27
所有这些因素都会进入每一个选址决策。我知道,很明显,Texas 之所以受欢迎,是因为它符合其中很多标准,但它并不是唯一的州。我的意思是,我们在全国各地都有数据中心,包括 LA 等地。好。很好。
Speaker 122:27 - 22:28
We're gonna go into all of
Speaker 122:27 - 22:28
我们会进一步展开讲所有这些内容,
Speaker 322:28 - 23:00
this in in in more detail, but let let's talk about you a little bit and and and your journey. So you're the the head of industrial compute at OpenAI, which, by the way, to the beginning of this convention, the title industrial compute is is so I mean, such a perfect title, right, for the moment we're in. But what what does that mean? What what what is the role and how is this whole effort organized within OpenAI to the extent you can talk about it?
Speaker 322:28 - 23:00
但先来多聊聊你和你的经历吧。所以你是 OpenAI 的 industrial compute 负责人,顺便说一句,在这场大会开头提到时,industrial compute 这个头衔,我的意思是,这真是一个非常贴切的头衔,对吧,很符合我们当下所处的这个时代。但这到底是什么意思?这个职位具体是做什么的?以及在 OpenAI 内部,整个工作是如何组织的?在你方便谈的范围内。
Speaker 123:00 - 23:14
Yeah. I I think think of it as my role and our team's role, rather, as how do we bring compute online at IndustryFK. Right? That's effectively what we tell me. That's the entire life cycle.
Speaker 123:00 - 23:14
对。我会把它理解为我的角色,更准确地说,是我们团队的角色:我们如何在 IndustryFK 上线 compute(算力)。对吧?基本上这就是我们对自己的定义。这涵盖了整个生命周期。
Speaker 123:14 - 23:27
So how do we find the ingredients that are going to compete? Land, power, shells, chips. How do we finance them? Right? So because these are massive dollars.
Speaker 123:14 - 23:27
所以,我们如何找到能够构成 compute 的那些要素?土地、电力、shells、chips。我们又如何为它们融资?对吧?因为这涉及的是巨额资金。
Speaker 123:27 - 23:43
And so how do we make sure that we finance the grid infrastructure? How do you finance the construction of the pump shells? How do we finance the chips? Then it's about how do you operationalize all this? So how do you actually make sure these things happen on time?
Speaker 123:27 - 23:43
所以,我们如何确保为电网基础设施融资?如何为 pump shells 的建设融资?如何为 chips 融资?接下来就是,如何把这一切 operationalize(运营落地)?也就是,如何真正确保这些事情能按时发生?
Speaker 123:44 - 23:57
They stay up? How do you operationalize all of this infrastructure? So it's that entire life cycle. And then, of course, how do you actually use the compute? So a big part of my role is capacity allocation inside OpenAir.
Speaker 123:44 - 23:57
它们能持续运行?我们如何把所有这些基础设施 operationalize(运营落地)?所以,这就是完整的生命周期。然后当然,还有你究竟如何实际使用这些 compute(算力)。我职责里很大的一部分,是在 OpenAir 内部分配 capacity(产能)。
Speaker 123:57 - 24:00
So it is always a scarce resource.
Speaker 123:57 - 24:00
所以它始终是一种稀缺资源。
Speaker 324:00 - 24:03
Makes sure you're a very popular guy. I am not very popular.
Speaker 324:00 - 24:03
这能确保你是个非常受欢迎的人。但我并不是很受欢迎。
Speaker 124:04 - 24:39
There is always someone who is unhappy with whatever decisions he make. But, yeah, me, actually, our team provides the input to make the capacity allocation decisions. So we surface what are the different choice points, what are the what if questions on different allocation choices that we have. So capacity planning, and then, of course, using that to forecast how much capacity we need where. Because it's not just more compute, it's also where, what kind, what shape, what chip, what workload you wanna run there.
Speaker 124:04 - 24:39
不管他做出什么决定,总会有人不满意。不过,确实,我们团队实际上会提供输入,来帮助做出 capacity(算力容量)分配决策。我们会梳理出有哪些不同的选择节点,以及针对不同分配方案,有哪些“what if”问题需要考虑。所以一方面是做 capacity planning(容量规划),另一方面当然还要据此预测我们在什么地方需要多少 capacity。因为问题不只是需要更多 compute(算力),还包括在哪里、什么类型、什么形态、什么 chip(芯片),以及你想在那上面运行什么 workload(工作负载)。
Speaker 124:39 - 24:53
So all of those are this team figures out kind of what should be the forecasting and planning, and that informs that closes the loop. And so that informs of where do I go find the next chunk of land and power and chips to put in there.
Speaker 124:39 - 24:53
所有这些,基本上都是这个团队来搞清楚预测和规划应该怎么做,而这些会形成反馈闭环。然后,这又会进一步决定我下一块要去找哪里的土地、电力和芯片,把它们部署进去。
Speaker 324:53 - 25:10
And, again, without getting into anything confidential, although, I guess, when you guys go public, all of this will be soon all of this will be public. But, like, is that thousands of people at this stage? I mean, is that a is is that a, like, multiple different teams, or do you guys gotta, like, outsource a bunch of things and works with a bunch of contractors?
Speaker 324:53 - 25:10
再说一次,不涉及任何机密,虽然我想,等你们上市之后,这些很快也都会公开了。不过,比如说,在现阶段这已经是一个几千人的规模了吗?我的意思是,这是由多个不同团队组成的吗,还是说你们需要把很多事情外包出去,并和一大批 contractors(承包商)合作?
Speaker 125:11 - 25:26
It's a portfolio approach. Right? So we are never going to be in a world where we outsource everything or build anything up source. Right? It's always gonna be a mix because that's the that's the reason that's the reasonable thing to do.
Speaker 125:11 - 25:26
这是一个 portfolio approach(组合式策略)。对吧?我们永远不会处在一个把所有事情都外包出去,或者把所有东西都完全自己建设起来的世界里。对吧?它一定会是混合的,因为那才是合理的做法。
Speaker 125:26 - 25:46
Right? So you don't want to put your rights in all all in one basket. So we will have hyperscalers probably providing a big chunk of a compute a majority of a compute. We will have new clouds, parts of our portfolio. We will be partnering with design build firms that can build the compute that we need.
Speaker 125:26 - 25:46
对吧?你不会想把所有权利都放在同一个篮子里。所以我们会让 hyperscalers(超大规模云服务商)提供很大一部分 compute,可能是大多数 compute。我们也会把 new clouds(新兴云)纳入我们 portfolio(组合)的一部分。我们还会和 design-build firms(设计建造公司)合作,来建设我们所需要的 compute。
Speaker 125:46 - 25:59
And, of course, we will build some of it ourselves. And so we're always gonna have a portfolio approach because at the scale which we need, we will need to tap into all sources of compute. We can't just rely on one particular mechanism.
Speaker 125:46 - 25:59
当然,我们也会自己建设其中的一部分。所以我们始终都会采用 portfolio approach(组合式策略),因为按照我们所需的规模,我们必须动用所有来源的 compute。我们不能只依赖某一种特定机制。
Speaker 325:59 - 26:12
And your background, the default of this, so you are both a professor at Stanford and an entrepreneur or a founder or mostly an academic? Like, just walk us through your journey.
Speaker 325:59 - 26:12
那你的背景呢,作为这一切的起点——所以你既是 Stanford 的教授,也是 entrepreneur(创业者)或者 founder(创始人),还是说主要是学术人士?能不能带我们简单回顾一下你的经历。
Speaker 126:12 - 26:28
Bit of all of the above. So but, yes, I'm at my heart, I'm an academic. So I've been a professor at Stanford since 2010. Recently What do you focus on there? I was a faculty in computer science and electrical engineering.
Speaker 126:12 - 26:28
上面那些都有一点。不过,是的,从内心来说,我是个学者。我从 2010 年起就在 Stanford 任教授。最近——你在那里主要专注什么?——我当时是 computer science 和 electrical engineering 的 faculty(教员)。
Speaker 326:29 - 26:31
With like, particular interest? Yeah.
Speaker 326:29 - 26:31
比如说,有特别关注的方向吗?对。
Speaker 126:31 - 26:53
My area of research was networking. Headworking. So I already built networks, mobile, wireless networks other than the seven days until networks actually. But three, four years ago, I while I was at Stanford, I did a couple of startups. The last startup got acquired by VMware, and that's how I got to know Pat.
Speaker 126:31 - 26:53
我的研究领域是 networking。networking。也就是我一直在构建网络,mobile、wireless networks,而不是过去七天里大家实际讨论的那种网络。不过在三四年前,我在 Stanford 期间做过几家 startup。最后一家 startup 被 VMware 收购了,我也因此认识了 Pat。
Speaker 126:53 - 27:19
Pat Dickens, Intel's CEO. I mean, that's how I ended up at Intel. Most before coming to OpenAI, I was at Intel's SOC team. And so I've kind of seen all the different things, academia, startups, corporate in Intel, and then, of course, a mix of all of the above in Oprior. Because they have a research lab, a startup, and a fast growing company all mixed into one at OpenAI.
Speaker 126:53 - 27:19
Pat Dickens,Intel 的 CEO。我的意思是,我后来就是这样去了 Intel。在来 OpenAI 之前,我在 Intel 的 SOC 团队。所以我算是见过各种不同的环境:academia、startup、Intel 这样的 corporate,以及当然,在 OpenAI 里是上述所有东西的混合体。因为 OpenAI 把 research lab、startup 和一家快速增长的公司都融合在了一起。
Speaker 327:19 - 27:25
Yes. Is it what you what you said what what did you say yes to the job when the job came came up?
Speaker 327:19 - 27:25
对。所以当这个工作机会出现时,这是不是你答应这份工作的原因?
Speaker 127:25 - 27:41
It's actually what I just said. Yeah. That makes this so unique. And it's hard to find anywhere, right, because you always have to choose. But having a world class research environment coupled with the hardest technical problems.
Speaker 127:25 - 27:41
其实就是我刚才说的那个。对。这正是它如此独特的原因。而且这种地方很难找,对吧,因为你通常总得做选择。但这里既有 world-class 的研究环境,又有最难的技术问题。
Speaker 127:41 - 28:04
We are building the largest compute in the world. And so there are a lot of new problems that we need to solve, but also a fast growing business. So this biggest problem I I'd like problems that sit at the intersection of business, technology, and strategies. And so this is, like, very unique time in history and a unique role to some of which are very attractive, obviously.
Speaker 127:41 - 28:04
我们正在建设全球最大规模的 compute(算力)。所以我们需要解决很多新的问题,同时这又是一项快速增长的业务。因此,这就是最大的吸引力。我喜欢那种处在 business、technology 和 strategy 交叉点上的问题。所以这是历史上一个非常独特的时刻,也是一个非常独特的角色,其中有些方面显然非常有吸引力。
Speaker 328:04 - 28:21
Very cool. Alright. So going into a bit more specifics about OpenAI's compute strategy. So maybe let's summarize what you guys currently have. So I think there's some Microsoft.
Speaker 328:04 - 28:21
很酷。好,那我们更具体一点,聊聊 OpenAI 的 compute strategy(算力战略)。也许我们可以先总结一下你们目前拥有什么。我想这里面有一些 Microsoft 的部分。
Speaker 328:21 - 28:36
You did this big $20,000,000,000 deal with Cerberus. There's a bunch of things around this Stargate. Maybe just give us the lay of the land of what you currently have, and then we'll talk about what your question building next. We have compute
Speaker 328:21 - 28:36
你们和 Cerberus 做了这笔规模高达 $20,000,000,000 的大交易。围绕这个 Stargate 有很多事情。也许先请你给我们大致介绍一下你们目前手上有什么,然后我们再谈你们接下来要建设什么。我们有 compute(算力)
Speaker 128:36 - 28:55
from effectively many sources. So Microsoft, obviously, the big partner, important partner. We also have compute, as we have announced, from AWS, ION, Google. So we have compute from all of the hyperscalers effectively. We also have compute from Corby, for example, so a Neo Cloud.
Speaker 128:36 - 28:55
来自很多来源,实际上是如此。所以 Microsoft 显然是重要的大合作伙伴。我们也有算力,正如我们已经宣布的,来自 AWS、ION、Google。所以基本上,我们从所有 hyperscaler(超大规模云服务商)那里都获得算力。比如我们也有来自 Corby 的算力,也就是一种 Neo Cloud。
Speaker 128:55 - 29:37
And then, of course, that chip partners are supplying now, like, Xeripas, and it's up directly producing compute for us. So I think that's the mix roughly today. As we go forward, obviously there'll be building on all of these relationships, but also looking at more options where we design the compute, the data center itself ourselves, or potentially even build the data center ourselves. So all of those are ways of scaling the amount of compute that we have. So I think coming back to my earlier answer, it's the answer always will probably be trite, but it's all of the above.
Speaker 128:55 - 29:37
当然,还有 chip(芯片)合作伙伴现在也在供货,比如 Xeripas,而且它现在直接为我们生产算力。所以我认为,这大致就是我们今天的组合。展望未来,显然我们会继续深化所有这些合作关系,但也会寻找更多选择,比如由我们自己来设计算力、设计 data center(数据中心)本身,甚至可能自己建设 data center。所以这些都是扩大我们所拥有算力量的方式。所以回到我之前的回答,我知道这个答案可能总是有点老套,但答案就是:以上全部。
Speaker 129:37 - 29:38
It's all you know?
Speaker 129:37 - 29:38
也就是说,全部都要,是吗?
Speaker 329:38 - 29:41
Yeah. Yeah. No. No. That that's that's helpful.
Speaker 329:38 - 29:41
对,对。不,不。这个很有帮助。
Speaker 329:41 - 30:06
And, you know, obviously, driver's education makes the sense in the world given the scarcity. And then it seems that Target has evolved. So from the what was gonna be a joint venture with Oracle and SoftBank to what now seems like it seems like it's more like an umbrella term for the big strategy these days. Is a fair way to describe it?
Speaker 329:41 - 30:06
而且,你知道,考虑到资源稀缺,driver's education 显然在各种意义上都说得通。然后看起来 Stargate 也已经演变了。所以它原本像是 Oracle 和 SoftBank 的一个 joint venture(合资项目),而现在看起来更像是如今这套大战略的一个总括性名称。这样描述公平吗?
Speaker 130:07 - 30:21
Yeah. Yeah. I think we look at StarVirt as our compute strategy. Yep. And it is varying degrees of us designing or building the compute ourselves.
Speaker 130:07 - 30:21
对,对。我认为我们把 StarVirt 看作是我们的 compute strategy(算力战略)。没错。它体现的是不同程度上由我们自己来设计或建设算力。
Speaker 130:22 - 30:54
For example, with Oracle, close partnership, We help them design. We help them with how we need to operate AI compute, which is, in fact, a very new kind of compute for us. We work with SoftBank Energy, which is public. We basically have co designed the Warm Shell with them, and they're executing on that Warm Shells. And we will be kind of figuring out how to operate our chips in these data centers ourselves, even the new chips.
Speaker 130:22 - 30:54
例如,和 Oracle 的合作就非常紧密,我们帮助他们做设计。我们帮助他们理解我们需要如何运营 AI compute(AI 算力),因为这实际上对我们来说是一种非常新的算力类型。我们也和 SoftBank Energy 合作,这是公开信息。基本上,我们已经和他们共同设计了 Warm Shell,他们正在执行这些 Warm Shell。至于如何在这些 data center 里运行我们的芯片,甚至包括新芯片,我们会自己逐步摸索出来。
Speaker 130:54 - 31:21
So Stargate to us is that umbrella strategy for across all of these different things. And think of it as an evolution that we continuously be on because it's never gonna be tomorrow when we wake up wake up and do only one kind of way of building compute. Yep. I think Stargate to us is a continuous way of learning how to scale compute, and they'll be adding on more and more capability later. Right?
Speaker 130:54 - 31:21
所以,对我们来说,Stargate 就是覆盖所有这些不同事项的那个总括性战略。你可以把它理解为一种我们会持续推进的演进过程,因为不可能明天一觉醒来,我们就只用一种单一方式来构建 compute(算力)。对。对我们来说,Stargate 是一种持续学习如何扩展 compute 的方式,而且之后还会不断加入越来越多的能力。对吧?
Speaker 331:21 - 31:35
And as part of that, there are data centers being built, right, like, like, Abilene, Texas. Yes. And so maybe walk us through that. What what is currently being built for for people to have some situational awareness?
Speaker 331:21 - 31:35
而且作为其中的一部分,正在建设 data center(数据中心),对吧,比如说 Texas 的 Abilene。是的。所以也许你可以带我们过一遍这个情况。为了让大家对现状有一些整体认知,目前到底在建设什么?
Speaker 131:36 - 31:53
We obviously have a big partnership with Oracle. That's the Abilene Data Center. That's where, for example, we are training our newest models. So very excited about that. That's a very big GB Black Belt cluster for our needs.
Speaker 131:36 - 31:53
很明显,我们和 Oracle 有一个重要的合作。那就是 Abilene Data Center。比如说,我们最新的模型就是在那里训练的。所以我们对此非常兴奋。就我们的需求而言,那是一个规模非常大的 GB Black Belt cluster。
Speaker 331:54 - 31:56
So that's up and running. It's up and running.
Speaker 331:54 - 31:56
所以它已经上线运行了。它已经上线运行了。
Speaker 131:56 - 32:12
It's been used for training the last two models more, actually. So it's a super threat. And you are seeing the results, You're seeing how quickly the models are becoming more capable. Yep. It's because of these kinds of compute.
Speaker 131:56 - 32:12
实际上,它已经被用于最近两个模型的训练了,甚至不止。所以这是一个非常强的助力。你也已经看到了结果,你正在看到这些模型的能力提升得有多快。对。正是因为有这类 compute。
Speaker 332:12 - 32:19
Yep. I think data centers are currently being built that are Yes. Good out there. Yeah.
Speaker 332:12 - 32:19
对。我认为目前确实有一些 data center 正在建设中。是的。已经在推进了。对。
Speaker 132:19 - 32:49
Yes. So Oracle is building a number of data centers, all of which are public, so across Michigan and Texas and other places. So these are coming online in the next couple of years as they get built and put whatever chips on that time training are the latest. But these are really meant to be very big clusters that allow us to do both our training, but also product inference kind of compete.
Speaker 132:19 - 32:49
是的。所以 Oracle 正在建设多个 data center,这些都是公开信息,分布在 Michigan、Texas 和其他一些地方。因此,随着这些设施建成,并安装上到时用于训练的最新 chips(芯片),它们会在未来几年内陆续上线。但这些项目真正的目标,是建设非常大的 cluster,让我们既能进行训练,也能支撑产品 inference(推理)这类计算需求。
Speaker 332:49 - 32:56
Right. But the way those deals are structured, the Oracle is the prime building those Oracle is the cloud.
Speaker 332:49 - 32:56
对。但这些交易的结构方式是,Oracle 是主导建设的一方,Oracle 也是 cloud(云)提供方。
Speaker 132:56 - 32:57
Is the cloud. Company.
Speaker 132:56 - 32:57
是云。公司。
Speaker 332:58 - 33:09
Of the tenant. Yes. Core tenant from the interesting. How does the for the stuff that you're building, how does the financing strategy work? I mean, obviously, you you you all raised what was it?
Speaker 332:58 - 33:09
作为租户。对。核心租户,这点挺有意思。对于你们正在建设的这些东西,你们的融资策略是怎么运作的?我的意思是,很明显,你们都融资了,是多少来着?
Speaker 333:09 - 33:27
122,000,000,000. Was it the number recently? So there's no shortage of cash. Although, you know, given all those expenses, I don't know. But, you know, the COREs of the world are famous for being very strategic users of debts.
Speaker 333:09 - 33:27
1220 亿美元。最近是这个数字吗?所以资金显然并不短缺。虽然,你知道,考虑到那些支出,我也不确定。但是,像 CORE 这样的公司,一向以非常善于战略性使用债务而闻名。
Speaker 333:27 - 33:31
Is that part of their financing strategy as well? How do you how do you all think about it?
Speaker 333:27 - 33:31
这也是他们融资策略的一部分吗?你们是怎么思考这件事的?
Speaker 133:31 - 33:55
I mean, with all of these compute that we have today, we have amazing partners who actually are handling that for us. So Microsoft, Google, Amazon, Oracle are we are off the off take. So we are the tenants, as you put it. So we commit to consuming that compute, to buying that compute, whether it's online.
Speaker 133:31 - 33:55
我的意思是,面对我们今天拥有的所有这些 compute(算力),我们有很棒的合作伙伴,实际上是他们在替我们处理这些事情。所以 Microsoft、Google、Amazon、Oracle——我们是 off-take 方。所以按你的说法,我们是租户。也就是说,我们承诺会消耗这些算力、购买这些算力,不管它是否已经上线。
Speaker 333:55 - 34:12
Oh, so across the board, like everything, your because you're you're building stuff as well. So building the partners are building? Partners are building. And you're always the, across the board, the the tenants, not the owner at the Correct. Okay.
Speaker 333:55 - 34:12
哦,所以是全面如此,像所有这些都是,因为你们自己也在建设东西。那么是合作伙伴在建设?合作伙伴在建设。而且无论在哪一方面,你们始终都是租户,而不是拥有者,对吗?对。好的。
Speaker 334:12 - 34:31
So therefore, financing is outsourced to apartments. Okay. We talked about the jalapeno. Let's let's into a bit more detail there because in particular, it seems that you guys went incredibly quickly and designing it. I think I read somewhere in nine months from design to tip out.
Speaker 334:12 - 34:31
所以也就是说,融资被外包给了这些伙伴。好的。我们刚才谈到了 jalapeno。我们再稍微展开讲讲,因为尤其是看起来,你们在它的设计上推进得快得惊人。我记得我在哪儿读到过,从设计到 tip out 只用了九个月。
Speaker 334:31 - 34:35
So maybe maybe walk us through that, and what was the reason it went so quick?
Speaker 334:31 - 34:35
所以,也许你可以带我们过一遍这个过程,以及它为什么会推进得这么快?
Speaker 134:35 - 34:49
Yes. It was incredibly quick. Nine months is very, very fast, probably the fastest I've seen in my career. I think several reasons. One is it it's a team.
Speaker 134:35 - 34:49
是的。速度快得惊人。九个月非常非常快,可能是我职业生涯里见过最快的一次。我觉得有几个原因。一个原因是,这是一个团队的成果。
Speaker 134:49 - 35:13
It's a strong team. They have many many the team have designed TPU chips at Google in the past, so very well experienced team. We have a great partner in Broadcom that has a very strong track record of delivering XPOS ASICs. So I think a strong partnership with Broadcom and and making this happen. Three, I think, an OpenAI unique point.
Speaker 134:49 - 35:13
这是一个很强的团队。团队里有很多人过去曾在 Google 设计过 TPU 芯片,所以这是一支经验非常丰富的团队。我们还有一个很棒的合作伙伴 Broadcom,它在交付 XPOS ASICs 方面有非常强的履历。所以我认为,和 Broadcom 的强强合作促成了这件事。第三,我觉得这是 OpenAI 的一个独特优势。
Speaker 135:15 - 35:59
In most chip companies or when you design chips, you don't know what you're designing it for because you're a vendor, the customer who eventually runs the workload is someone different. There's a unique advantage here of us knowing what the future models might look like, and therefore being able to short circuit a lot of the decisions you need to make, design decisions you need to make on the chip side. That's super helpful. And finally, increasingly, AI itself helping design and optimize the chip. That is usually the one that takes the longest time because human you're basically limited by how many human human much human time is there to process all this data and run the experiments, and we can do a lot of those iterations much faster With AI.
Speaker 135:15 - 35:59
在大多数芯片公司里,或者说当你设计芯片时,你并不知道自己究竟是在为谁、为怎样的工作负载而设计,因为你是 vendor(供应商),而最终运行这些 workload(工作负载)的客户往往是另一方。而这里一个独特的优势在于,我们知道未来的模型可能会是什么样,因此能够大幅缩短很多你在芯片侧必须做出的决策,也就是芯片设计决策。这一点特别有帮助。最后,AI 本身也越来越多地在帮助设计和优化芯片。通常这部分最耗时,因为人类的限制本质上在于,可用于处理这些数据和运行实验的人力时间是有限的,而借助 AI,我们可以把很多这样的迭代做得快得多。
Speaker 135:59 - 36:00
Using AI.
Speaker 135:59 - 36:00
使用 AI。
Speaker 336:00 - 36:03
Yeah. So AI is building its own chips now?
Speaker 336:00 - 36:03
对。所以现在是 AI 在为自己制造芯片了吗?
Speaker 136:03 - 36:20
Yes. Yeah. I think that world is not very far. I mean, right now, is assisting in chip design, but we do believe that the world of Likertion is not that far where AI will design the systems it needs to train and run the next generation of AI.
Speaker 136:03 - 36:20
是的,没错。我觉得那样的世界离我们并不远。我的意思是,现在 AI 还主要是在辅助芯片设计,但我们确实相信,那样一个阶段已经不远了——AI 将会设计出它训练并运行下一代 AI 所需要的系统。
Speaker 336:20 - 36:32
And including chips? Including chips. Including chips. You also released a few weeks ago MRC, which is a networking protocol. Walk us through through that.
Speaker 336:20 - 36:32
其中也包括芯片?包括芯片。包括芯片。几周前你们还发布了 MRC,这是一个 networking protocol(网络协议)。给我们详细讲讲这个吧。
Speaker 336:32 - 36:34
What is it, and why is it a big deal?
Speaker 336:32 - 36:34
它是什么?为什么这件事很重要?
Speaker 136:34 - 37:15
It's a new networking protocol routing technology, if you will, to scale these really large cluster fabrics. So imagine you have a 100,000 GPUs. They need to be connected together. And when you're doing large training runs, they're constantly communicating with each other because these models are so large that the processing of the models is happening over the entire 100,000 GPU cluster, for example. And you can imagine imagine the number of links and switches and NIC cards that need to be there to connect all of these chips together.
Speaker 136:34 - 37:15
这是一种新的网络协议路由技术,可以这么说,目的是让这些超大规模 cluster fabric(集群网络)实现扩展。你可以想象一下,你有 100,000 个 GPU,它们需要彼此连接在一起。而当你在进行大规模训练任务时,它们会持续不断地相互通信,因为这些模型太大了,以至于模型的处理过程是分布在整个 100,000 个 GPU 组成的集群之上的,比如说。你也可以想象,为了把所有这些芯片连接起来,需要多少链路、交换机和 NIC card(网卡)。
Speaker 137:16 - 37:41
At this scale, failures are right? It happens all the time. You can't really even enumerate all the ways things could fare. So the strategy behind MRC is how do you design algorithms and protocols that can gracefully mask all these failures and make sure that the training workload does not get impacted. Right?
Speaker 137:16 - 37:41
在这个规模下,故障是必然会发生的,对吧?这类情况一直都在发生。你甚至很难把所有可能出问题的方式一一列举出来。所以 MRC 背后的策略是:你该如何设计算法和协议,来优雅地屏蔽所有这些故障,并确保训练 workload(工作负载)不会受到影响。对吧?
Speaker 137:41 - 37:55
They just the network is an abstracted system that the training job does not ever worry. It's always gonna be there. It's always gonna find a path even if a link fails. So it is all about reliability. It is all about availability.
Speaker 137:41 - 37:55
对训练任务来说,网络只是一个被抽象掉的系统,它根本不需要去操心。网络总会在那里。即便某条链路失效,它也总能找到一条路径。所以这件事的核心就是 reliability(可靠性),核心就是 availability(可用性)。
Speaker 137:56 - 38:32
So how do we design protocols that contain the complexity of such a common such a big cluster and make sure that we don't get stopped because of failures which are very common in the system? So MRC is like a multipath spraying protocol where you can spray packets over multiple parts. So between any two chips, there are many, many routes to get there, kind of like between any two points in the city, there are many routes. So instead of picking just one route, we will we will send traffic across all of them, and whichever one succeeds, we take them. Right?
Speaker 137:56 - 38:32
那么,我们该如何设计协议,把这样一个庞大而复杂的集群的复杂性控制住,并确保我们不会因为系统中非常常见的故障而停摆?MRC 有点像一种 multipath spraying protocol(多路径喷洒协议),你可以把数据包喷洒到多条路径上。所以在任意两个芯片之间,通常都有非常多的路由可达,就像城市里任意两点之间都有很多条路一样。因此,我们不会只选择一条路,而是会把流量同时发到所有这些路径上,哪一条成功,我们就采用哪一条。对吧?
Speaker 138:32 - 38:46
And so that way, even if any one of them fails, it's not a showstopper. So that's kind of the basic intuition behind it. It's obviously a lot more sophisticated than that. But doing this at scale and that speed is hard, and so that's why it's quite innovative.
Speaker 138:32 - 38:46
这样一来,即使其中任何一条路径失败了,也不会成为致命问题。这大概就是它背后的基本直觉。当然,实际实现显然要复杂得多。但要在这种规模和这种速度下做到这一点是很难的,所以这也是它相当有创新性的原因。
Speaker 338:46 - 39:01
What are the bottlenecks that you experience these days? So it seems like the nature of the bottleneck keeps changing in the compute industry. Obviously, people talk a lot about memory these days. Is that is that one of them, or what what else?
Speaker 338:46 - 39:01
你们现在遇到的瓶颈是什么?看起来,计算行业里的瓶颈性质一直在变化。显然,这些天大家谈论很多的是 memory(内存)。那是不是其中之一?或者还有什么别的?
Speaker 139:01 - 39:25
I think there are bottlenecks everywhere, to be honest, to the supply chain. So I don't think there's any one. Right? We have bottlenecks in the data center building itself around permitting, around availability of gas turbines, transformers. Those industries have historically have not added much capacity over the last decade or so Yep.
Speaker 139:01 - 39:25
老实说,我觉得到处都有瓶颈,甚至包括 supply chain(供应链)。所以我不认为只有某一个瓶颈,对吧?我们在 data center(数据中心)建设本身上就有瓶颈,比如许可审批、gas turbine(燃气轮机)和 transformer(变压器)的可用性。这些行业在过去大约十年里,历史上一直都没有增加太多产能。没错。
Speaker 139:25 - 39:34
And they've suddenly experienced a demand shock. Yeah. And it takes years before you can add capacity to produce more turbines and transformers. So they're trying to quick catch up.
Speaker 139:25 - 39:34
而它们现在突然遭遇了一次需求冲击。是的。而要增加产能、生产更多涡轮机和变压器,往往需要好几年的时间。所以它们现在是在努力快速追赶。
Speaker 339:34 - 39:45
And to the jump thing that you that you mentioned earlier, is there, like, a shortage of, like, electricians and and my technical people, trade people that know how to build those things
Speaker 339:34 - 39:45
还有,回到你刚才提到的那个增长问题,是不是像 electrician(电工)以及那些懂得如何建这些东西的技术人员、工种人员,也存在短缺?
Speaker 139:45 - 39:48
at once? Or you're Absolutely. Yeah. Absolutely. So we That's
Speaker 139:45 - 39:48
是一下子就都短缺了吗?还是——绝对是。对,绝对是。所以我们——这也是
Speaker 339:48 - 39:55
why we should all do as AI replaces knowledge work, should pick up electricians. No. I think
Speaker 339:48 - 39:55
为什么我们大家都应该在 AI 取代知识型工作时,去学 electrician。不是,我觉得
Speaker 139:55 - 40:29
there is definitely a shortage of electricians, plumbers, all kinds of tricks. You name it. So anything we can do to train more folks to be able to do those, there are very well paying jobs that lot of us, all of the hyperscalers, all of the labs would actively hire you for if you had the, like, conversations. So I think that is definitely a bottleneck. I I think they become an increasing bottleneck because we are all trying to build more and more, and we have a limited number of these these capabilities.
Speaker 139:55 - 40:29
electrician、plumber(水管工)以及各种技术工种,肯定都存在短缺。你能想到的都有。所以,凡是我们能做的、能培训更多人去胜任这些工作的事,都值得去做。这些工作薪酬很高,我们很多公司、所有 hyperscalers 以及所有 labs,都会非常积极地雇用你——前提是你具备这些能力。所以我认为这绝对是一个 bottleneck(瓶颈)。而且我觉得,它会越来越成为 bottleneck,因为我们都在努力建设越来越多的东西,而这类能力的人数是有限的。
Speaker 340:29 - 40:42
Great. Let's talk for a minute about the the business side of things. Another thing you launched recently is in guaranteeing capacity Mhmm. For customers to lock in compute. That which is interesting.
Speaker 340:29 - 40:42
很好。我们来花一点时间聊聊业务层面的事。你们最近推出的另一项东西,是 capacity(产能)保障,嗯哼。让客户能够锁定 compute(算力)。这一点很有意思。
Speaker 340:42 - 40:52
Right? It almost feels like OpenAI is also becoming a utility company, providing computing to others. What what's the story behind that and the the strategy?
Speaker 340:42 - 40:52
对吧?这几乎让人感觉 OpenAI 也正在变成一家 utility company(公用事业公司),向其他人提供 computing(计算能力)。这背后的故事和策略是什么?
Speaker 140:52 - 41:04
Yeah. So guaranteed capacity is guaranteed tokens. Okay. Right? So we are effectively saying we will guarantee you a certain dollars worth of tokens of intelligence.
Speaker 140:52 - 41:04
对。所以 guaranteed capacity(保障产能)其实就是 guaranteed tokens(保障 token)。明白吧?对吧?也就是说,我们实际上是在说,我们会向你保证某个特定美元价值的 intelligence(智能)tokens。
Speaker 141:05 - 41:19
And, I mean, it makes sense. Right? So in a world where compute is a shortage, therefore tokens are always gonna be at a premium, and there's a shortage of tokens that we can produce given the limited compute that we have. And as this becomes a fundamental input to the enterprise. Right?
Speaker 141:05 - 41:19
而且,我是说,这很合理。对吧?在一个 compute(算力)短缺的世界里,tokens 自然总会是稀缺溢价资源,而鉴于我们拥有的 compute 有限,我们所能生产的 tokens 也是短缺的。并且,随着这逐渐成为 enterprise(企业)的基础性投入。对吧?
Speaker 141:19 - 41:40
So enterprises are going to need intelligence, more and more of it, to run. Right? And so this is a way of enterprises gaining assurance that the tokens of intelligence that they need will be there for them, and so they they don't have to take business risk. Right? And so I think it's it's good business hygiene.
Speaker 141:19 - 41:40
所以,企业要运转下去,将会需要 intelligence,而且会越来越多,对吧?因此,这是一种让企业获得保障的方式,确保它们所需要的那些 intelligence token(智能 token)到时会在那里可供使用,这样它们就不必承担业务风险,对吧?所以我认为,这是一种很好的 business hygiene(经营卫生/经营基本功)。
Speaker 141:40 - 41:57
If you have a critical supply resource in every enterprise, intelligence is kind of the most important supply item, if you will. It makes business sense to make sure you secure that supply. And so that that's the demand we're trying to fulfill. You know? I know it's a new concept.
Speaker 141:40 - 41:57
如果说每家企业都有某种关键的供应资源,那么 intelligence 可以说就是其中最重要的供应项。确保你把这项供应锁定下来,从商业角度看是合理的。所以,这就是我们试图满足的需求。你知道的,我也明白这还是个新概念。
Speaker 141:57 - 42:08
Like, what does it mean to have guaranteed capacity for the intelligence thing? But I think that's what intelligence is becoming. It is becoming a supply unit, not every digital enterprise.
Speaker 141:57 - 42:08
比如,给“intelligence 这种东西”提供有保障的 capacity(容量),这到底是什么意思?但我认为,intelligence 正在变成这样的东西。它正在成为每一家数字化企业的一种 supply unit(供给单元)。
Speaker 342:08 - 42:22
Right. So maybe to end on the fun one, and you can answer either with your OpenAI hat on or or not. Data centers in space, is that is that is that exciting? Is that science fiction? Is that needed?
Speaker 342:08 - 42:22
好的,那也许最后说个有意思的话题,你可以用 OpenAI 的身份来答,也可以不这么答。太空中的 data center(数据中心),这件事令人兴奋吗?这是科幻吗?这是有必要的吗?
Speaker 342:23 - 42:28
Is that something that people like to talk about just because it's cool, or where where where do you land?
Speaker 342:23 - 42:28
这是不是人们喜欢谈论、只是因为它很酷的东西,还是说——你对此的判断是什么?
Speaker 142:28 - 43:04
For for the geek in me and the engineer in me, it's definitely exciting. It's one of those things can be super cool to see a constellation of satellites producing compute. I actually think it is it will become feasible, right, as engineering problems to be solved, but they can be solved, with enough time and investment. Whether it is needed, I think there is room for orbital compute. I don't think it's gonna solve all the compute beats, but it's definitely going to be a component in the arsenal.
Speaker 142:28 - 43:04
对我内心那个 geek 和工程师来说,这当然很令人兴奋。看到一整个卫星星座在产生 compute(算力),会是那种超级酷的事情。我确实认为,这件事会变得可行,对吧——它本质上是工程问题,而工程问题是可以被解决的,只要有足够的时间和投入。至于它是否有必要,我认为 orbital compute(轨道算力)是有空间的。我不认为它会解决所有的 compute 需求,但它肯定会成为工具库中的一个组成部分。
Speaker 143:04 - 43:28
I think what we are waiting for is when does the economics of the launching satellite change and when does the economics of the hardware change? Because we need to get to a point where it's cheap to launch the hardware, And if something fails, it's cheap to throw it away. I think we can't go up and fix it, unlike on the ground. And so I think that inflection point hopefully happens soon, and at that time, it becomes wilder.
Speaker 143:04 - 43:28
我认为我们在等待的是两个经济性拐点:发射卫星的经济性什么时候改变,硬件的经济性什么时候改变。因为我们必须走到这样一个点:把硬件发射上去足够便宜,而如果有东西坏了,把它直接扔掉也足够便宜。我觉得我们没法像在地面上一样上去维修它。所以我希望那个拐点很快到来,而一旦到了那个时候,事情就会变得更疯狂。
Speaker 343:28 - 43:32
Correct. Sachiv, it's been wonderful. Thank you so much for spending time with us. We appreciate it.
Speaker 343:28 - 43:32
没错。Sachiv,今天非常精彩。非常感谢你抽时间和我们交流。我们非常感激。
Speaker 143:32 - 43:34
Thank you. It's been great to have this chat.
Speaker 143:32 - 43:34
谢谢。这次聊天非常愉快。
Speaker 243:34 - 43:54
Hi. It's Matt Turk again. Thanks for listening to this episode of the MAD podcast. If you enjoyed it, we'd be very grateful if you would consider subscribing if you haven't already or leaving a positive review or comment on whichever platform you're watching this or listening to this episode from. This really helps us build a podcast and get great guests.
Speaker 243:34 - 43:54
你好,我又是 Matt Turk。感谢你收听这一期 MAD podcast。如果你喜欢这期内容,而你还没有订阅的话,我们会非常感激你考虑订阅,或者在你观看或收听这一期节目的平台上留下积极的评价或评论。这对我们持续打造这档 podcast 并邀请到优秀嘉宾真的很有帮助。
Speaker 243:54 - 43:55
Thanks and see you on the next episode!
Speaker 243:54 - 43:55
谢谢,我们下期节目见!
原文 ↗https://www.youtube.com/watch?v=wEZBlmvxx4o
BuildSpeak — 关于本项目BUILT IN PUBLIC · 跟随 builders 而非 influencers