尊敬的用户您好，这是来自FT中文网的温馨提示：如您对更多FT中文网的内容感兴趣，请在苹果应用商店或谷歌应用市场搜索“FT中文网”，下载FT中文网的官方应用。
{"text":[[{"start":7.62,"text":"Martin Casado is a technologist and investor who has backed some of Silicon Valley’s best known start-ups. He leads an AI investment team at Andreessen Horowitz, also known as a16z, one of the world’s most prominent and outspoken venture capital firms."}],[{"start":22.92,"text":"Earlier in his career, Casado worked on simulations of nuclear weapons at Lawrence Livermore National Laboratory and, after the attacks on September 11 2001, moved into conducting research for the intelligence community on networking and cyber security. During a subsequent PhD at Stanford University, he helped pioneer software-defined networking, enabling the movement of data between computers via software rather than hardware. He then launched his own start-up, Nicira Networks, which sold for more than $1bn."}],[{"start":55.84,"text":"In this conversation with FT’s venture capital correspondent George Hammond, he argues that recent progress in AI is an industrial-revolution scale event but warns that the ability of the bigger players in the sector to raise “cheap money” is time-limited and the value will quickly move downstream."}],[{"start":null,"text":"
"}],[{"start":72.6,"text":"George Hammond: Is the AI cycle that we’re living through at the moment like anything you’ve seen before?"}],[{"start":79.42,"text":"Martin Casado: Sometimes I think it’s just like the internet, which changed the social order, because you could talk to anybody, independent [of] geographic distance. AI reminds me of that. It has cultural impact, it has social impact, it changes what it means for a human to communicate and interact with something."}],[{"start":96.04,"text":"Sometimes I feel it’s more like electricity or even the industrial revolution, just because it also feels like it totally changes the economic dynamics of innovation. Things like the job market, things like the funding market. With the internet, there was a lot of demand early on but we didn’t know how to monetise it. And here, it feels like there’s infinite demand and we’re monetising immediately."}],[{"start":117.84,"text":"The industrial revolution lifted the world out of poverty and was a great equaliser. It solved many serious problems humanity faced. AI could do quite the same thing except . . . in a much more dramatic fashion."}],[{"start":129.42,"text":"GH: And over a much more compressed time period? The industrial revolution played out over decades. This is happening in the space of years. What are the consequences of that speed?"}],[{"start":140.72,"text":"MC: Well, this is what we don’t know, right? Is this technology primarily about improving the productivity of a human? To what extent can you use AI to replace a human? To what extent does AI not need humans at all? Could these become independent entities that can create businesses? I think we have zero evidence of that but we’re in a disruptive technology. Who knows what happens?"}],[{"start":162.64,"text":"GH: We are seeing some evidence of productivity gains now. What will keep that going and what could that look like over the next few years?"}],[{"start":170.32,"text":"MC: The one that’s very obviously working is code. Assume there’s 30mn professional software developers or something like that, and they make an average of $100,000 a year. This is a $3tn market. I mean, we’re just in the very, very early innings of this massive market."}],[{"start":187.12,"text":"A single lawyer can now do more cases, the accuracy of healthcare will increase. We know that the call centre stuff is working. We know customer support is working. But then there’s a long tail of stuff that we’re not quite sure about."}],[{"start":199.58,"text":"It may also be the case that AI makes us feel more productive than we actually are. That’s a kind of phantom productivity. How often would you talk to ChatGPT, and it’s like, ‘that was such a great idea.’ Then you correct it and it’s like . . ."}],[{"start":200.08,"text":"GH: ‘That’s a good catch!’"}],[{"start":216.54,"text":"MC: That’s right. There’s some disconnect between how productive [AI] makes us feel versus how productive they are, but there is no question that [chatbots] are actually creating value."}],[{"start":225.12,"text":"But I think when this is all said and done the big innovation here is we can actually deploy large amounts of money to solving certain classes of problems and we’ve not been able to do that in the past."}],[{"start":232.72,"text":"In my opinion, whether or not AI can solve the problem is now a capital question not a technology question. If we look back and ask ‘what was the big shift here?’ I think it is that we changed things from an engineering problem to a capital problem. To me, that’s like industrial revolution level change. Because it changes the nature of innovation."}],[{"start":252.82,"text":"GH: AI is clearly very generally effective, but it’s not high fidelity in all the ways that we would need it to be to take on knowledge work and go beyond the narrow use cases. How do we get to that next?"}],[{"start":264.76,"text":"MC: Pre-training, where you pour a whole bunch of money in and you train a base model, that’s basically done. The next phase is RL [reinforcement learning], where you start tackling specific problems that you have a reward signal for. We’re at that point right now."}],[{"start":279.68,"text":"And so I think the question is less: can it do a particular vertical? I think the answer is yes in most cases. The question is: is it cost effective to do that?"}],[{"start":289.28,"text":"Here’s where I feel the industry is: computers hit this point where you needed a computer to create a computer, this point of having an autocatalytic effect. Right now we’re at the point where people are starting to use AI to create AI. That’s where the feedback cycle will start happening. That tends to dramatically drop the price of whatever you’re doing."}],[{"start":null,"text":"
You’ve got the fastest depreciating assets we’ve ever seen, and the fastest growth rate companies you’ve ever seen. And nobody can answer the question of where this converges.
"}],[{"start":308.04,"text":"GH: On the capital question, the big AI labs have raised more and grown revenue faster than any previous generation of start-up. But they are still losing money. How do you resolve that?"}],[{"start":319.24,"text":"MC: Yeah, this is honestly the big question. If you look at the economics of a frontier lab relative to the training run of the last model, they look great. But if you look at the economics of the current training run, they look terrible."}],[{"start":331.48,"text":"And how long do you think a model is relevant? Three to six months. Then you have to do that next training run, these models have no permanence at all. So you’ve got the fastest depreciating assets we’ve ever seen, and the fastest growth rate companies you’ve ever seen. And nobody can answer the question of where this converges."}],[{"start":345.88,"text":"GH: So as an investor in AI companies, how do you square the circle?"}],[{"start":350.6,"text":"MC: We have to have exposure to every potential outcome. One is that the large labs get 80 per cent of the value. So you have to have exposure to large labs. The other is, at some point in time, the market will rationalise this burning of money. The labs won’t be able to raise this much money so cheaply, and then the value will quickly move downstream to the apps and the smaller models. So we have to have exposure to that."}],[{"start":375.12,"text":"Nobody can predict the future, so you have to have enough exposure [to] any given path."}],[{"start":380.88,"text":"What’s most important is can [AI labs] raise more capital than any company that depends on them? If Anthropic raises $30bn, the entire downstream ecosystem can’t raise that much money altogether. The amount of money we’re talking about is so insane. But on the other hand, the model permanence is so low that I don’t see how they can maintain that. So I think it’s more a question of ‘how long can they continue to soak up this capital?’ To me that is the question. After that happens, the world will be a very different place."}],[{"start":411.98,"text":"GH: Well, what would the world look like when we get to that point?"}],[{"start":415.52,"text":"MC: At some point in time, and I don’t know if that’s two years or it’s six years, the market will rationalise this and say there are other opportunities here, you’re subsidising all of these things. So, the cheap money will start to slow down."}],[{"start":428.62,"text":"At that point, the big model labs will be brand monopolies, they’ll have massive account control in the enterprise, they will be entrenched incumbents. But I think that’s when their growth will slow dramatically and the value will accrue to one level downstream, which is the people that will go after specific verticals, applications, smaller models, et cetera."}],[{"start":447.36,"text":"GH: What are the next pieces of model development or entirely new architectures that might come along to sustain the boom?"}],[{"start":453.76,"text":"MC: I just think we have enough general tools to build the next thing. And we’re still learning to use those tools. I think we’re past the event horizon at which the world is going to change."}],[{"start":463.8,"text":"If I have one hot take, it's this: the more I do this, the more I don’t think it’s that hard to build these models. I think the reason that we’re like, ‘oh, these are amazing,’ is because it was kind of a niche field, and there weren’t that many people doing it."}],[{"start":null,"text":"
Innovation is no longer about being ingenious, it’s about amassing resources
"}],[{"start":477.6,"text":"I consistently see non-standard teams — in the sense they don’t have traditional backgrounds, don’t have the PhD, didn’t go to Stanford — with money build great models. It literally is like, can you get the data? Can you get the computer? That changes the nature of innovation; innovation is no longer about being ingenious, it’s simply about amassing resources."}],[{"start":498.18,"text":"GH: You’ve invested in companies founded by Ilya Sutskever [Safe Superintelligence] and Fei-Fei Li [World Labs] — some of the most celebrated researchers in the field. Clearly at some level you believe that innovation comes from very high-end research teams. What is it that they’re doing that differs from what you’ve just described, which is kind of brute-forcing it?"}],[{"start":520.1,"text":"MC: Well, so there’s a pragmatic reality: The most famous people will raise the most capital. So even in the case that they’re not the most ingenious, they will be able to raise most capital. There’s also the case that the most famous people will be able to hire the top 20 people."}],[{"start":531.96,"text":"I do feel that right now it’s early enough that you need people that actually understand it. You actually have to have the people that are the experts now. [That] doesn’t mean that you’re smart or not. You were literally just in the room."}],[{"start":543.64,"text":"Like the number of people that would just happen to be close to OpenAI was just this random smattering of people, right? It could be the case that the access to the knowledge is the hard thing right now because it’s so early."}],[{"start":553,"text":"These people actually have the access: Mira [Murati] was in the room. Ilya was in the room. Fei-Fei created deep learning in many ways. And so they have the networks and the access to the knowledge. But over time, we all will, and it very well could be that these things aren’t that hard and it really just becomes a capital view."}],[{"start":571.24,"text":"I say this a tiny bit facetiously and I don’t mean to be categorical, but it’s almost contrary to how everybody views it. Everybody’s like, ‘it’s these brainiac geniuses, and only they can do it’. It could actually be the case that like, no, this stuff is pretty straightforward. It hasn’t evolved a lot. And if you can aggregate the money, then a lot of people can do it."}],[{"start":592.8,"text":"GH: You’ve written about the US’s AI race with China. If it is the case that it’s a capital game, requiring commitment from private enterprise but also government, then who wins?"}],[{"start":603.4,"text":"MC: China has total government investment, very little free market, and the ability to get regulation out of the way very efficiently. In the US, we have a free market, free capital, tons of regulation and no obvious way to get the regulation out of the way. I think that’s the race."}],[{"start":619.12,"text":"Clearly, if we as a government would get the regulations out of the way, the US would just dominate."}],[{"start":624.32,"text":"GH: Why is it so important that the US beats China, particularly if model advantage lasts for weeks rather than months?"}],[{"start":null,"text":"
I think that there’s three potential outcomes. The US wins, China wins or both grow and are at détente. And I actually prefer the third: innovation wins, humanity wins, the world wins.
"}],[{"start":634.76,"text":"MC: Inasmuch as there are competing ideals, I believe in western ideals. So if you’re asking at that level, that’s kind of my core belief."}],[{"start":640.36,"text":"I think that there’s three potential outcomes. The US wins, China wins or both grow and are at détente. And I actually prefer the third: innovation wins, humanity wins, the world wins."}],[{"start":652.48,"text":"I have no need for the United States to be totally dominant locally in all aspects, but we cannot be in a position where we rely on China."}],[{"start":659.76,"text":"Remember, we went through this with the internet. We created the internet. We actually had the leading companies in the industry, Cisco [Systems], Juniper [Networks]. And then China copied them, and they’ve had huge implications in the emerging world. The span of [Chinese] influence increases. You can imagine this being much more the case in AI. So minimally, we cannot be relying on China’s AI and we want to provide our allies access to US technology."}],[{"start":686.68,"text":"GH: Related to that, I want to talk about the role of AI companies and the government in the US. Anthropic vs the Pentagon has been the big story there. You started your career at Lawrence Livermore [National Laboratory, an R&D centre focusing on nuclear weapons development, which is part of the US Department of Energy]. I’m curious in general what you think the relationship between private tech companies and the government should be when it comes to marshalling AI?"}],[{"start":713.38,"text":"MC: There are great analogies between what happened just post-9/11 to what is happening today. So I was actually in DoE [Department of Energy] during 9/11, and I was in the weapons programme, which was like an anachronism from the cold war."}],[{"start":727.32,"text":"Everything had changed: the nature of the threat changed with the internet and then terrorism. And many of the same questions that we were asking then are being asked today: we know that terrorists are using the internet to communicate. How can we find them, and how can you do that without surveilling US citizens? They’re exactly the same questions."}],[{"start":746.84,"text":"The way it shook out then is companies decided to work with government or decided not to work with the government. And both were totally fine. And then we kind of bubbled our way through that. This time it’s the exact same question: how do you use AI when the laws are not up to date? Companies can decide whether they work with the government or not."}],[{"start":766.02,"text":"But the thing that does not work, and the thing that’s different this time, is Anthropic has decided that they want to work, but actually have control of the government’s operational processes. This is entirely unprecedented and new that an independent company wants to actually control this decision process of the government."}],[{"start":783.42,"text":"So my personal view is, it’s totally fine not to work with government. What’s not OK is saying ‘we’re gonna make your decisions for you, government, because you don’t understand the technology.’
I think that’s what’s being litigated, right? And, by the way, I think that Microsoft, and OpenAI, all these companies that are going to Anthropic’s defence, are wrong to do it because this is so different."}],[{"start":807.8,"text":"GH: Why could this not have been resolved by the government just saying, as you were describing, ‘we don’t have to work together, our terms of service differ from your expectations, so we won’t work together’. Why have they gone further than that with the supply-chain risk designation?"}],[{"start":822.44,"text":"MC: Well, I think Anthropic could have done the same thing. They literally could have just decided not to, but that’s not what they did. Anthropic was trying to retain some control over what was happening within the government."}],[{"start":834.52,"text":"It would seem to me because of how pervasive this technology is — and we’re going into Martin’s opinion territory — that if Anthropic demonstrated that they are willing to pull support for critical infrastructure, which is exactly what happened, you could imagine that being billed as a supply-chain risk."}],[{"start":850.6,"text":"To my knowledge, that never happened during the time of the Patriot Act . . . post-9/11, where a company was like, ‘we will shut this down, even though you depend on it.’ And that’s effectively what happened here, right?"}],[{"start":862.6,"text":"So this would be like if Cisco says, if we don’t like what you’re doing, even if it is totally legal, we’re going to shut down the routers. Or like the telephone company saying that."}],[{"start":872.12,"text":"GH: Is another analogy, something like, you know, WhatsApp or Telegram saying, ‘we won’t insert a back door into our services?’"}],[{"start":872.62,"text":"MC: Totally different."}],[{"start":873.12,"text":"GH: Why so?"}],[{"start":879.36,"text":"MC: Because being able to deny availability of something that is used for an operation, it’s entirely different than providing access to something post factum. Literally, imagine war fighters are communicating on the field in the middle of an operation and then the telco decides to shut that down. Which is the perfect analogy to what is happening here."}],[{"start":899.64,"text":"GH: So pulling back from the details of the Anthropic contract, what I’m really curious about is how this impacts the working relationship between the government and Silicon Valley, more generally. Does this change the way they work together?"}],[{"start":912.84,"text":"MC: I don’t think so. There’s always been this tension with the government. I fully, fully respect any company that does not want to work with the government. I fully respect them, I will invest in them."}],[{"start":924.36,"text":"I’m glad we’re having the discussion now of what happens if a company threatens to reduce access to critical infrastructure?"}],[{"start":null,"text":"
What does the government do? I think we need to figure this out because AI is a very powerful technology
"}],[{"start":930.32,"text":"GH: I think it speaks to what has been a theme in the conversation throughout, which is that AI as a technology is powerful in a way that we have never really encountered."}],[{"start":939.64,"text":"MC: Yeah. Or maybe there’s an ideology around AI that is powerful that we haven’t encountered. It could be that this is perfectly analogous to the internet. And then the difference is the people that run these things just, like, want to have more control than the people that built the internet. That really could be the case."}],[{"start":956.72,"text":"GH: You think that’s the motivating ideology, rather than that they are just more fearful about the effects of this than people were of the internet?"}],[{"start":964.1,"text":"MC: I’m just saying it could be. They could be more fearful, but that’s the thing: it’s like it’s the ideology, not the technology, in that case. The people controlling the technology are different, not the technology. That absolutely could be the case. And we don’t know that here."}],[{"start":964.6,"text":"GH: Do those two things not interrelate?"}],[{"start":978.64,"text":"MC: They could, but it could also be the case that this is just like any other computer technology that we’ve been dealing with for 70 years. And, you know, you just have cultures that have evolved around that, that view it as something that it’s not, and they’re using that to actually dictate governmental policy without having any clue what the implications of that are. And I don’t know what it is. Which is why I’m actually very glad that this discussion is happening with the current administration. It’s a very healthy discussion."}],[{"start":1007.76,"text":"Maybe it turns out the government’s totally wrong, which I’m very open to. It may be the case."}],[{"start":1013.52,"text":"This transcript was edited for brevity and clarity."}],[{"start":1021.2,"text":""}]],"url":"https://audio.ftcn.net.cn/album/a_1776479917_1510.mp3"}