LLMs Are Reaching a Stage of Diminishing Returns: What is the Next S Curve

with

So I was I had a very fortunate childhood. I grew up in sort of ex industrial small town in the North Of England. My mom was a teaching assistant. My dad was a social worker. I have three siblings. It was like a big happy sort of family. But I I think, like, probably the formative experience for me, at least in retrospect, was where I grew up. If you want if you were a teenager and you wanted to, like, make money, you basically could work in Greggs.

That was the option. And I really didn’t wanna work in Greggs. I think, like, a really formative thing was I remembered that there was a lightning storm in our village, and one of my parents’ friends, it sort of somehow damaged their computer. And I remember them saying to me, do you think you can fix this? I knew nothing about fixing computers. But I was like, yes. And I basically just bought a new motherboard and pumped everything back in. And and then it started. I never went to work at Greg’s.

I ended up sort of, like, building a little business sort of age 13 fixing and building computers for first my parents’ friends and then their friends and then sort of like branching out into building them websites and taking spyware and malware off their computers, you know, like, was like the early days of the consumer Internet where everything was horrible. I think just realizing you can do stuff. You can just say yes and figure it out. You don’t need anyone’s permission. I think life was really life changing for me.

I think, like, so much of everything I’ve done since comes from believing you can just say yes and you can just do stuff.

Yeah.

We believe one one of our kind of cool ways that we evaluate founders at Entrepreneurs First is that the best predictor of future behavior is past behavior. So we absolutely think that it’s very unlikely that you rock up with a perfect CV, but, like, no sign whatsoever of ever having done anything that you weren’t told to do by a teacher or a boss, and suddenly you’re gonna figure out how to do things that you aren’t told to do. So in that sense, I think I very strongly agree.

Guess I what I would say though is that culture and default paths really matter. And so one thing we’ve learned about evaluating talent is what it looks like to do something that your parents, teachers, bosses didn’t tell you to do in the Bay Area is very different from in Singapore. You know? Like, if you grew up in the Bay Area, go to Stanford, and you graduate, and you never thought about starting a company, and that’s a very negative signal on you as a founder. If you grew up in Polany in, you know, rural India, I don’t think I need to have seen that you started a company before.

But what I’m doing when I’m interviewing you is figuring out what is the behavior, what is your equivalent of the story I just told you, what is your equivalent of that entrepreneurial activity that is a predictor of you being able to succeed in an unstructured environment? That’s what I care about.

I was very lucky to start investing in AI about ten years ago, a bit more than that now, when the first sort of deep learning wave kicked off. What’s really interesting is that really, certainly the last five years, but arguably longer, the story of AI is really a story largely about the deployment of just enormous amounts of compute and enormous amounts of data and not that many new ideas, candidly. You know? I mean, like, there are I’m saying there are none, but, like, broadly, if you had to, like, do a pie chart of where the process has come from, it’s largely come from massive investment in compute and application of that.

I think what’s interesting, if you haven’t gone to my end and said, like, where are we today? I think that we are seeing the flattening off of the value of just adding more compute and more data to language models. I think we’ve seen truly extraordinary progress over the last five years. I think of most technologies of s curves, you know, you sort of get slow progress, then fast progress, then slow. And it strikes me that we’re in this moment where the incremental value of of just continuing on this path of scaling is leveling off.

Mhmm. And what that means is that the value of ideas is about to go up a lot relative to the value of just scale. And what that means to me is that whenever that ideas become valuable again, you have an opportunity for startups. Frankly, I don’t think startups today should be trying to compete on building bigger and bigger LLMs. It doesn’t make sense. It’s mainly a capital game. Maybe you can do a mister Allen raise, you know, like, hundreds of millions in year one, but even they might be under capital at history.

I mean, that’s the kind of crazy thing. And so today, I think the real opportunity for founders is find the next S curve, and I think we’re in a moment where that’s actually paused.

It could be. I mean, I think the things that I’m excited about and I’m looking at pretty hard right now are search. And there’s been a lot of buzz about this recently in the sort of AI research community, like how do you use search? The techniques that actually made things like AlphaGo work, how do you kind of mix those with LLMs to get better results? I think, clearly, there’s this big moment around multimodality that is just getting started where if you can figure out very smart ways to think about different data types than text, that might be another S curve.

Yeah. I mean, look, my point’s not that, like, the big guys won’t be able to do this. I suppose my point is really that the value of ideas over the last five years have been relatively small. Like, the question was, could you get the scale compute? I think the value of ideas over the next few years might be high, and it might be enough to unlock. I’m not saying it’s suddenly gonna become this low capital game. I think if you wanna build an AGI company, you’re gonna need a lot of capital.

But I think until now, the only way to do that was to have trained a large language model at one of the labs, come out and raise a bunch of money, and say we’re gonna carry on doing it. I think there may be a new approach that’s possible today where you actually just have an idea that’s, like, really quite different from what people have been doing, and you can use that a little bit like DeepMind ten years ago to do a demo that makes people say, And that unlocks the capital that allows you to go up against the big guys.

So I think it’s certainly true data is the one of those that is least obvious how you continue to scale. At the moment, you can just translate dollars into compute relatively easily. And so, you know, you can imagine at least a couple of orders of magnitude of of growth of compute quite straightforwardly. It’s not obvious that you can do the same in data. That said, I think people probably underestimate what we haven’t done yet partly because of compute limits. You know, these models are not trained on all video.

You know? In fact, most of that hasn’t happened yet. Some of the things that people like Andreessen Karpathy have been talking about recently is, like, the potential of thinking about video as being a really great way for a model to build a world model and, you know, to understand more about how the world works. So, yeah, data is a bottleneck, but I wouldn’t bet against smart people figuring out ways to either create or ingest new types of data. I guess what I’m really arguing, when I say we’re on an s curve, is I don’t think there’s a lot further to go in just finding more text.

I think we may be at the flattening out of the s curve on text. But I suspect the next s curve could just involve finding ways to use video, to use, like, interactive experience in in very different ways.

Yeah.

Well, I think it’s a race. As in, like, I do think that if we can’t find anything better than just trying to scale LLMs, then, yeah, the flattening out will be brutal. If I had to bet, my bet is that the enormous amount of capital and talent that has been aggregated, you know, around this relatively small number of companies because of the hype will be enough to unlock the next S curve, and so you’ll see, you know, kind of further improvement. The returns to good ideas, I think, will be very high over the next few years.

So actually, another way of thinking about it is one thing you’ve seen over the last couple of years since GPT-four was released is actually a convergence of capabilities. You know, when GPT-four came out, it was very clear that OpenAI were in the lead. And now, like, there’s quite a few companies that have a GPT-four ish level model. You’ve seen convergence. Right? My guess is that you may see more divergence over the next few years because the value of ideas goes up, because people aren’t just scaling.

I think the pure LLM approach absolutely is commoditized. Will be either is or will be very quickly commoditized. I guess what I’m saying is I don’t think GPT-five is just gonna be a bigger language model. I think it’s gonna involve something different. It’s gonna involve a lot of productizations, maybe search as we already, you know, discussed and things. And so, like but these are not commodity ideas yet. It’s not yet the case that everyone will be able to copy that. And so I think broadly, you know, I was talking recently to someone who was on the Lama three team, and they were like, you’re probably underhearing the extent to which Lama three is literally just Lama two with more compute and and more data.

There’s not a lot that’s different other than, you know, that. I guess what I’m saying is LLM a four will not just be a bigger LLM a three. It will have to be different. It

Yeah. Well, particularly because norms around secrecy around ideas have increased over time. I mean, you it’s easy to forget Have they, or do

Yeah. That’s certainly true. But I guess what I mean is the the knowledge around how to build, how to train, do pretraining on, you know, a very large model is now extremely well known. But I think it is now quite appreciated in the labs that if you can find an idea that has this, like, disproportionate impact on capabilities, guarding that is gonna be very important. Now, again, I’m not saying that that will remain secret forever, but I think in the short term, I think we’re likely to see more divergence, you know, over the next couple of years than convergence.

I think I would not be. I mean, I’m a huge fan of OpenAI. I think it’s extraordinary what they’ve done. But, I mean, on the other hand, if they’re really doing whatever it is, 3,500,000,000 here and, you know, they can continue to productize, you know, maybe it’s not crazy, but I do think that we’re gonna get a lot more data in the next a lot more evidence in the next year whether anyone has another good idea, and I think that’s what it’s gonna take.

Yeah.

Yeah. I mean, I I have found myself using Claude three Absolutely. All the time.

Yeah. Yeah. Definitely. Yeah. And there’s a big opportunity in that. One thing I often think about AI is like anyone who claims they know what’s gonna happen, I think, is like, is wrong.

You you you have to maintain a set of quite radically different scenarios and just change the weight that you put on them. I can totally see a world in which everything I just told you about the value of new ideas is wrong. And, actually, it really is just like we can’t find anything that fundamentally find is the next S curve. Everything commoditizes. Everything flattens out. And, you know, it’s really a, you know, race to the bottom on price. In which case, obviously, none of these companies are gonna make any money.

Do you think China is two years behind? This is what Alex said. This is what Alex disputed. Right? So Alex has become a a a good friend. I would I would put this again in the category of, like, it’s very hard to say. You know? As as you know, I I spent quite a lot of time last year working on the the geopolitics of AI in in government, and including, like, spending time in Beijing and, you know, negotiating with the Chinese government on AI. What is that level of sophistication?

I think the Chinese government is very sophisticated on AI. You know, for example, the guy I was meeting to negotiate with was the the minister of science and technology, you know, senior minister on the, you know, standing committee at the CCP. You know, he’s a computer scientist. He knew his staff. Is that a higher level of

the UK

I guess I would say, like, one thing I learned doing doing that job, albeit briefly, was that The UK’s capacity in the space is high relative to most countries. I remember being on the initial kickoff call. You know, I was running this this summit for the prime minister, and there are 28 countries coming. We did this kickoff call on Zoom or Teams. The best way to improve productivity of government is just, like, allow them to not use Teams. And, you know, like, honestly, without naming names, there were countries where they were like, how could we possibly do this?

We don’t have an AI policy division. We don’t have an AI minister. You know? Like so I I think, like, The UK actually has been ahead of the curve on this, but I do think China is very sophisticated.

One of the things that’s interesting and I think underrated is that, actually, China is more paranoid about AI safety than probably any other government in the world. So if you look at Chinese AI regulation, it makes Brussels look like a soft touch. It’s incredibly onerous.

Think you just gotta see the CCP as being, like, fundamentally have an extreme priority on stability. The the whole story of the CCP and its dominance is eliminate threats to stability. And, you know, whatever you think about AI, whether you’re bullish or bearish, whether you care about safety, whether you don’t care about safety, think everyone agrees that it’s like a destabilizing force. And so I think the last thing they want is either companies gaining a lot of power through AI or AI itself being destabilizing. So, you know, like, if you read the regulation that’s already coming through for example, if you train a large model, you have to, like, supply random samples of the training data to the CCP or to the government.

You have to, like, show it’s, like, not gonna undermine any of the, you know, stated positions of the CCP on various issues. So, it’s actually a very, very regulated environment. Now that doesn’t mean that ultimately the Chinese state won’t try and harness powerful AI for its own ends as as you’re sort of talking about, but I think it’s easy to understate the level of paranoia they have about it. So there there are two factors that make me think that maybe I’m not quite in Alex’s camp.

One is that that actually it’s a very highly regulated space there.

Well, I think if you look at, like, why why it’s been so fascinating in the West, you know, basically because, you know, certainly OpenAI and Anthropic and to an extent, especially now, you know, under the sort of unified leadership here in London of of of DeepMind, you basically just have this extraordinary permissionless to go back to that idea explosion of, like, ambition. What, you know, Sam and Demis and Dario are doing is this almost untrammeled, very visionary ambition. As we’ve seen in other parts of the tech industry in China, the CCP is very nervous about entrepreneurs being given that much that much room.

It’s not trivial to create the conditions in which a Chinese Sam Altman or a Chinese Demis Hassabis can just go to the races in the same way because there’s just much more I mean, we all saw what happened to Jack Ma when it looks like he was too powerful. But I think that’s an important ingredient in building, you know, these sort of AGI companies is you need someone at the top that actually is all out to build, you know, like, the final invention as some of them call it.

Yeah.

Yeah.

Well, I I think one thing that’s really interesting is the asymmetry in how we in the West learn about China and how they learn about us. You know, when Leopold Ashenbrenner published his situational awareness thing, which obviously, like, you know, kind of blew up a couple of weeks ago, Within three days, I hear from a friend who follows China very closely, there were five Chinese translations circulating in the Chinese AI ecosystem. If an equivalent thing is written in China, it does not get translated into English.

It does not get circulated.

Yeah. Yeah. So so I think in the world where actually either my thesis that ideas really are gonna matter a lot is wrong or where it’s just a very narrow window and then you find a new idea and then it’s just scale, scale, scale, then, yeah, I think China has a very strong hand there. But, you know, the other fact you talked about algorithms, data, compute, the other thing that is still a factor, and it really comes down to, like, timelines, is, you know, The US export controls on, you know, the semiconductor supply chain have had an impact.

Not a binary impact, but it is certainly true that it is harder if you are a big Chinese company to build a access a 100,000 GPU cluster. You can do it. You know, in February, Kai Fu Lee was, like, going around Europe, including at Davos, saying like, oh, you know, you know, I traded my latest model on these Huawei designed SMIC manufactured GPUs. And, you it wasn’t claiming they were as good as NVIDIA’s, but, know, it was like 70%. And but but, know, I think in a world where scale compute matters, that friction, that discount if you like on the Chinese system, it does compound.

And I think, you know, if you wanna be a China bear on AI, I think you have to believe that we’re gonna get very radically improved progress in the West quickly enough that they don’t build their domestic chip ecosystem fast enough.

You know, although I would say there’s been fascinatingly, like, the one bipartisan issue in Washington is screw China. It’s certainly true that I remember writing about in ’20, you know, ’18, ’19 about, like, Trump export controls, and it was interesting is that the Biden White House has been as aggressive on exactly the same thing. Been a remarkable continuity on this. And so, yeah, you’re right. I think whoever wins the election, they’re gonna continue to see NVIDIA and others as a key national security asset.

I think the bare case for China is that the gap matters, that actually they can’t harness the same level of compute over a relevant time period. Got you. And then, sorry, the bull case for China? The bull case is either, like, it takes longer and, you know, the sort of the domestic ecosystem, because they can’t buy ASML machines, they eventually make their own, and then we have no levers. Right? You know, it doesn’t matter what we ban, they’ve got their own. And at which point, you know, your argument that when they turn on the industrial policy tap, it really turns on, at that point, I think they’re in a they’re in an extraordinary position.

You know, one of my favorite tech essays that I find myself referring to all the time is that, you know, Joe Spolsky thing on commoditize your compliments. And, you know, that’s what’s happening in in AI right now. Is saying, like, what are the compliments to what I’m doing, and how do I commoditize them? It’s interesting. Like, I the reviews of NVIDIA’s model are pretty mixed, to put it mildly. I think you’ll certainly see people trying to figure out, like, which layers of the stack do you need to own and, you know, what do you have to commoditize in order to, really protect your core business?

Maybe we’ll always say this, but it really feels to me like we’re gonna learn so much in the next twelve months. You know, like, the to me, the key question that until we resolve, we’re all just guessing is sort of how good are the GPT-five era of models? If they’re, like, wildly better. If it’s as big a gap between four and five as it was between three and four, that is a very different world from if it really feels like we’re eking out these tiny improvements despite having invested truly colossal amounts of money.

What do you think would be a needle moving shift? I think the one that people are talking most about, which makes sense to me, is just agency. I mean, like, GPT-three wasn’t good enough really to do anything remotely agentic. People have obviously built agents on GPT-four, but, you know, people doing that have said, like, you can build great demos. You know, you run it a 100 times, you pick the best one, it looks amazing. But robustness is very, very huge problem, and and and they’re not reliable.

But I think if GPT-five were good enough that you could get much, much more reliable agents, I think that would feel like a qualitative change that was really impressive.

Not one that where, like, many millions of instances can be run and they can operate at, like, nearly human level. No. I don’t think that’s what GPT-five will be, but, like, one of the theses I’m really interested in right now is I think we’ll get really good agents in the next five years. And I think if you look at parts of the economy where we already have a lot of automation, it’s pretty clear that that requires a lot of infrastructure, requires protocols for agents to interact.

I think we’re gonna need that for AI agents. We’re gonna need to build great infrastructure to maximize the economic value of these agents. I’m sorry. What does great infrastructure mean to maximize? Well, let me give you an example. There there are an area of the economy where there’s already a lot of autonomous agents is high frequency trading. In trading, a lot of the work is now done by, I think, you can only describe as autonomous agents. But we don’t let them just go wild and, you know, like, call their broker and I mean, physicists.

But, like, if you look at the complexity of the infrastructure we’ve built And

Yeah. So I’m certainly not saying that we can just copy what we have, but I think it gives you to me, it’s like an intuition builder that, like, let’s say we’re gonna send out lots of agents to do economically valuable work. We need to govern things like, how do they interact with each other? What are the rules for that? How do you handle the edge cases? How do you observe what they’re doing? How do you govern what they’re doing? How do you let them keep to the rules?

And so it strikes me that whoever builds the protocol that for key slices of the economy actually allows agents to come together and do business together, that’s gonna be an extraordinarily valuable company. Should government be the ones to do that? I think government should, like, set the standards, but I just don’t think government has a great track record of building important scalable software.

So. I think it’s like it’s the equivalent to building an operating system or a protocol. It’s basically saying if you want agents to transact with each other, we provide the tooling that allows you to observe that, govern that, turn them off when they’re not working, handle the edge cases. But imagine that a large chunk of the economy over the next decade ends up being transacted by autonomous agents. I don’t think they’re gonna interact in the way that humans interact. There’s gonna have to be infrastructure that allows them to interact, and whoever owns that infrastructure, and that’s gonna be a huge asset.

I think one of the good things is people benefit from technology even when they don’t understand it. I think that’s, like, one of the big lessons in, you know, the history of the last two hundred years of capitalism is that technology is is the engine of prosperity. And, you know, like, people whose jobs involve very little technology today make a whole lot more than their ancestors two hundred years ago and have a much better quality of life because of the benefits that technology brings. I do think it’s true so far, and the broad trend is that technology is what, I guess, economists call skills biased.

It benefits people with more skills more than it benefits people with with fewer skills. That is a big challenge for governments. It’s probably a big opportunity for entrepreneurs. But I think in general, the story of technology is is an extraordinarily positive one. I’m not saying that people shouldn’t, including policymakers, look at this and say, right. How do we plan for this world where relatively few people understand this, but it’s a huge factor in everyone’s lives? But I worry about that because I worry that you start to get the instinct of, like, maybe we should try and slow it down.

Maybe we should try and stop it.

My view would be, and I think you see this already, I didn’t vote for Brexit broadly. I’m not a fan. But, you know, probably the first issue where I’m like, maybe there are some benefits to this, is just looking at the difference in approaches to AI regulation. I mean, I do think the EU AI act is a mistake. My basic model of how that legislation gets written is like a bunch of old dudes sit in Brussels and say, let’s imagine everything that could go wrong. Okay.

You’re not allowed to do that. What’s the biggest problem with the EU AI act? Basically this, that it basically tries to anticipate the future, labels a bunch of things high risk, and creates an enormous burden for companies trying to innovate in those areas. I think, actually, this is an area where The UK has done a lot better job than Europe, where we’re not regulating to oblivion. In fact, on AI, The UK has less regulation than any other country that has a significant AI industry. So, you know, right now, there is no specific AI regulation in The UK.

I think that’s one of the factors. I think the other is just the extraordinary talent base. If you go back to the thing we start this conversation talking about, let’s say I’m right that there’s gonna be, like, a strong return to ideas, to new ideas over the next few years. I think it’s very likely that of people in Europe who might have those ideas, they are disproportionately disproportionately in in the The UK. UK. Do you think

Let’s say you really believe that you could build, as I framed it and you quoted, like, Anthropic or OpenAI scale company in terms of ambition. I actually think one of your biggest challenges building that in the Bay Area today would be building and retaining the AI research talent that makes that. And, you know, people always talk about hiring. Much less often talk about retention. But, like, the packages that Sam Altman will offer to your best people the minute you get your first, like, you know, demo, you know, it’s really hard.

In London, it’s much more plausible that you build a truly world class research team and keep them together for long enough to really see the impact. And, you know, like, here in London, we have DeepMind. We have the London office of we have the European office of Anthropic, the European office of OpenAI. We have Wave. We have, you know, Oxford, Cambridge, Imperial, UCL.

I don’t know. Like, I’m gonna do a spoiler. I know one of the questions you’re gonna ask later is what do you believe that people around you don’t? And I thought a lot about this, and I was like, oh, maybe I can say something about AI. And I realized the thing I really believe that almost no one believes is I think The UK can go back to being pretty much the richest country in the world we’re captive. I really truly believe that. I think we can be at least as rich as The United States on a per head basis.

I think we’ve chosen not to be, and we just need to choose to be. We should be the obvious place in the world to build and scale technology companies as So we can what is that rate limiting factor? I think the big things are actually levers that we can choose to pull. You know, one of the things that I found just, like, really dispiriting when I was working in government last year was that I mean, it an enormous privilege, but I, you know, I got to go and meet Satya Nadella and Andy Jassy and, you know, these people and negotiate with them on, you know, on on this summit.

And, you know, I was mainly talking about AI policy, but, you know, obviously, they wanted to talk about the full tech agenda with The UK. And the number one issue that these companies had, and I’m I’m really not exaggerating, was we want to invest billions and billions of dollars to build compute infrastructure in The UK, and your local county councils in the middle of nowhere keep vetoing our data centers because they obscure a view from a railway bridge. That is a choice we have made as a country, to not let the most sophisticated companies in the world invest in sophisticated technology infrastructure in The UK.

They want to. So all we have to do is decide that we want them to. How do we do that then? Do we just over all county councils? Well, no. There I mean, there there are various ways you can do it. Parliament can legislate to make data centers. There’s a category of infrastructure project that doesn’t get to be overridden locally that, you know, like national government can say this is a priority. We can choose to do this. I I don’t wanna make it sound simple. It’s hard, and it will annoy the hell out of a lot of people.

But, you know, planning is one, energy is the other. You know, I’ve spoken, again, like, kind of with last year when I was working in government. You know, I I spoke to people that wanted to build enormous clusters of GPUs in The UK and couldn’t because of the limitations of our electricity grid. So what do they say then? Should we give up? We don’t need to give up. These are all levers

Well, you know, if you read the labor manifesto, they say that they’re committed to growth being the number one priority, and they talk about things like planning. They talk about things like compute infrastructure, so we’ll see. There is a window of time where this country has to decide what it wants its future to be. My very good friend, Ian Hogarth, wrote this great essay very presciently, actually now five years ago, called AI Nationalism, where he talked about, like, if you believe that AI is the defining technology of, you know, our generation, which I do, then at some point, countries will have to decide whether they wanna be AI makers or AI takers.

And if you’re an AI taker, it doesn’t mean that your country won’t use AI. It just means you’ll be shipping dollars to California to use California models on Californian compute. We don’t have to choose that. We can choose a different future, and we have to decide. Or we can have beautiful views from railway bridges.

But it’s so funny. Right? Because, you know, as you said, we’re sitting here in London. And I would say, you know, within three miles of where we are, you have a top three global AI lab, arguably the one with the broadest research agenda, DeepMind. So if I’m right that ideas will matter, then DeepMind’s, you know, right up there. You have probably the top embodied AI company, private company in the world in Wave. You have, like, probably the best life sciences cluster in certainly in Europe. You’ve got I’m super biased because I’m the chair, but you’ve got probably the most radical government r and d funding agency in the world in ARIA.

We know how to do this stuff already. We just gotta believe in it.

I think they’re right. The way I see it is this is really a question about who benefits. It’s not actually a question about whether we can build. What I mean by that is I would love it that when, you know, a 20 VC company or an EF company goes public, that the pools of capital that own that company are my mom and dad’s, you know, West Yorkshire pension fund. Right now, that’s not what’s gonna happen because they’re gonna IPO in places that the shareholders are probably American.

You know, it’s Vanguard and BlackRock. Right? And honestly, for you and me and for the entrepreneurs, who cares? It’s all cash. But again, if you wanna yeah. From a UK, like, long term growth perspective, I believe that so much of our growth, if we make it happen, will come from entrepreneurship, big advances in in in science and technology. I think frankly, capital markets are very efficient. Good things will raise capital. The question is, who’s capital? And I think I think from a UK PLC perspective, we should want it to be that it’s pension funds, you know, with UK pensions, UK savers that are benefiting from that.

So that’s a big problem. What

Well, we were joking before. You know, I I’ve raised more money in Ann Arbor, Michigan than I have in London. And I won’t name the name of the pension fund, but one of my biggest investors is a sort of state pension fund in Michigan. And, you know, they were one of the first pension funds that that that we pitched. So I sort of came out of it being like, these are the people we should be pitching. And I the scheme in Michigan is a $70,000,000,000 pension plan.

And I remember coming to pitch a comparably sized plan in The UK and saying to them, what have you done in venture capital? And they said, we’ve heard of venture capital. But their allocation to venture was zero. And I just think that if private companies if if the most ambitious entrepreneurs are saying that they want to keep their companies private for longer, you know, if they IPO at all, they wanna do it when the companies are extremely mature, then most of the benefits from growth will come to pools of capital that invest in in venture.

And if UK pension funds aren’t doing that, how do they expect to be able to return to to meet their promises to their pensioners? I don’t know the answer to that, but

think that’s what I would worry about.

My certainly my best angel investment so far. That was a very lucky break. Well done. He is

That’s a very Alex thing to say. I don’t think AI today is more powerful than nuclear weapons. I don’t think GPT-four is Have you tried perplexity? Yeah. Well, yeah. Have you tried a hydrogen bomb? I would say, as a as a slight side note, I think nuclear war is really underrated as a thing to worry about. If you’ve not read it already and if your readers haven’t read it already, the best book I’ve read this year, by some distance, is nuclear war colon a scenario by this amazing American journalist, Annie Jacobson, and she describes it as a nonfiction thriller.

And what she does is basically just walk through step by step what would happen if a rogue country launched an attempted attack on The United States and all the things that could go wrong and lead to disaster. And it’s chilling. I don’t think I’ve ever had an experience with a book before where I had to keep taking breaks because I was like, woah. But, you know, she she she didn’t make this up. She went and she she spent thousands of hours interviewing the top people in nuclear security in The US and, you know, like, former Soviet people.

And it’s it’s an extremely compelling book. So, like, nuclear war, underrated. But to answer your question, I think AI changes everything in the future of war. I think that if you look at what do soldiers do, how do we think about their value, and how do we think about the harm that comes to them? When

An open question. I, you know, I hope your listeners might might send us their views. I think it’s really interesting question. How does AI change the relative power of offensive and defensive weapons? It’s certainly plausible to tell a story how being able to build cheap, smart drones benefits rogue actors, you know, and nonstate actors a lot more than it benefits states. You know, if you have very large important assets that you don’t wanna be attacked, like aircraft carriers, you can make the argument that the ability to deploy very, very inexpensive but pretty smart explosive drones is, you know, like, really asymmetric.

Like, there’s not much an aircraft carrier can do about that. One of the reasons I’m very interested in defensive technologies is I think we’re gonna have to think about problems like that and get ahead of them. I think there are a lot of scenarios where it’s not obvious it actually benefits the current, you know, established powers.

So I think probably, like, the worst idea in the world that it’s acceptable to hold in polite company is, like, degrowth. Like, the idea that, like, oh, you know, like, the real problem in the world is, like, too many people, too much staff, too much consumption. We should shrink populations. We should shrink the economy. I could not hate that idea more. All human progress has come from growth. I want to see technological progress. I want to see people building amazing things and as using more clean energy to do more amazing things.

I don’t think that it’s therefore inevitable that all technological progress automatically leads to great outcomes. I think you just have to look at nuclear weapons to see that, yeah, so far we’ve done pretty well, but there’s some pretty terrifying stories about near misses. Russian relatively junior soldiers ignoring false alarms that looked like incoming strikes and, as a result, saving the world. That’s not to say that tech progress, you know, inevitably leads to disaster. I don’t believe that at all. But I do think it matters in what order tech is built, and I think it matters who builds it.

I think technology is ultimately, like, values laden. And so when I look to the future, I see, like, paths to the most amazing futures. As you probably tell from the discussion about UK, I find it very easy to be optimistic. But I think the only way we get to those futures is if we prevent very powerful capabilities, nuclear weapons, for example, fall into the hands of people that would do us harm. And so, you know, one of the reasons I’ve really pivoted my time within EF to working on defensive technologies is I think there’s an amazing commercial opportunity there.

But I also think it’s what the world needs. We need to build the technologies that mean that the negative consequences of the technologies we build are mitigated, not by banning them, but by building great defensive capabilities.

Look, I I think our generation, if I’m allowed to put myself in the same generation as you, Harry, I think there is a real risk of complacency for us that we have, like, grown up in a world of relative peace where basically there wasn’t really any sort of, like, geopolitical threat to our value system. It’s clearly no longer true. I think we absolutely need to think about how do we make sure that we have the power to defend our values and our way of life, and I don’t think it’s a given.

We’re all

This is something I’ve changed my mind on. I actually think more and more the synergies matter, but ultimately, I do think entrepreneurship is one of the paths that most relies on the peak performance of the highest performer. So there are some walks of life where what really matters is average performance. You know, like, how well do you do on, you know, the median day? And there are other walks of life where it’s like, how well do you do at your very best? You need people in every company who are in the former category who can day in, day out turn out excellence.

Let’s use Helsink as an example you already brought up. You also just need the peak performance of the best day from the top talent in the company to be truly exceptional. And so, like, yes, synergies do matter, but probably the best predictor is how good is the best person at their best.

So I think that’s true, and I think a lot of that is about self discovery. As you know, a lot of the time when that happens, it’s actually about those founders learning what they really want. You know, like, I think one thing I’ve learned having now followed hundreds of founders from kind of often before they even knew they wanna start a company, a lot of the time, like, especially first time founders, just can’t really predict what that’s gonna be like. This is very hard for most people.

And and and, actually, I I don’t think that the best founders are the ones who predict best what it’s gonna be like. I think they just take it as it comes. But I think a lot of the time when founders don’t work out in companies that do, it’s that one of them realizes this really wasn’t, or at least the bit that’s to come is not for them. Do you think everyone has the ability to be an entrepreneur? No. I think entrepreneurship is best to think of it as like there’s nothing like medicine, but in the same way that I don’t think everyone should be a doctor or could be a doctor.

I think you should think of entrepreneurship as like a high skill extremely high skill profe you know what? I’m the sort of person that’s inclined to wanna read, like, the academic literature on something and on the basis that surely, like, with something we know and we can you know, there’s some alpha in lining. I would say in general, the academic literature on entrepreneurship is not helpful for either entrepreneurs or VCs. Might be interesting, but it’s not it’s there’s no alpha in it. The paper I think about all the time called forced entrepreneurs, what it looks at is in certain markets when there’s a recession, the dominant employers of high skilled people hire fewer people.

Obvious example in The UK is financial crisis, the banks hire fewer bankers. And unsurprisingly, in those periods where the level of hiring of kind of highly skilled individuals into finance falls, more people start companies. That’s kind of, like, probably quite obvious. Everyone would sort of intuit that. What’s nonintuitive, and in fact, think maybe even counterintuitive, is that not only do those people then start companies, they do on average better than the median person who was starting a company in a year without that recession. In other words, the forced entrepreneurs, the ones that go into it because the bank stopped hiring, do better than the people that chose the previous previous year.

I realize this is, like, on the verge of being, like, rude and controversial is that in most ecosystems, it is not by default the very most talented and ambitious people that become founders. And it probably is in the Bay Area. The reason that I think there’s a huge opportunity in The UK and in Europe and in India where where we work is that you can find those truly exceptional people, and they it’s not obvious. It’s much harder in the Bay Area in a way because there’s so many more opportunities for, you know, investors to meet those people.

But the point I wanna make is even today in The UK, even with all the progress has been over the last decade in the ecosystem, I would say the most aspirational job for a Cambridge computer science crowd is to go work at Jane Street and be a be a trader. I think that’s the number one job.

Eye watching amounts, even for you, Harry. It’s it’s pretty extraordinary. Like, a million bucks? In year one. Yeah. But, like, by the time you’re a few years in, if you’re good, many, many multiples of that. So this is the point about, like, this paper is what it shows is talent is somewhat fungible. It’s really tempting for those of us that live and breathe the industry to be like, there are founders and there are nonfounders. But actually, what there is is there’s talent and there’s ambition. And if that talent ambition is applied to trading, then those people will be great traders.

And if that talent ambition is applied to entrepreneurship, those people will be great founders. And this is the secret of the Bay Area. Like, why did Silicon Valley work? Because the most ambitious and talented people in that area come founders. If the same thing happened in The UK, in Europe, in India, we’d have very similar outcomes.

You know, like, there there is this sense that, like, if you’re a Cambridge computer science grad five years out, probably the richest person you know is the one that went to Jane Street five years ago. That’s probably not true in the Bay Area. And And so, like, there’s just It’s Reid Hoffman, always. Yeah. No. But I mean, even your peer.

And so just but but that’s still not true here. That’s the bit I think we still need to see change.

Which is why it’s a lot better than it was. It’s a lot better than it was. When I started visiting university campuses properly when when Alice and I started EF, I think it was Imperial West, 60% of computer science grads went into finance. That number is not that anymore. I actually don’t know what the number is, but it would I I wouldn’t be surprised if it was, like, a fraction of that number. So it’s changing.

For sure. And and, you know, I I’m sort of deliberately being sort of reductive about it to money. And clearly, part of the reason why actually really exceptional people do start companies in The UK is exactly that. They’re like, yeah, I’m sure they could go work at James Street or wherever, but they but they wanna do something. Whenever, you know, I’m in a conversation where we’ve offered someone, you know, a place, a EF, and they’ve got an offer from James Street or or similar. I mean, I’m like I’m like, yeah.

You could be an extremely highly paid crossword puzzle solver, which is sort of how I think about James Street playing a zero sum game, or you can, like, change people’s lives. Like, it’s up to you. And actually, a lot of people, exactly as you say, are not just, like, maximizing next year’s income. That’s where it comes from. My point, though, is that it’s less at the individual level. I’m saying at the system level, our biggest challenge, I think, in Europe is still talent allocation.

You know, one of the challenges of this business, and you may feel the same, is that the feedback loops are so long. And so, you know, you do something, you observe a short term metric like, do they raise a seed round? Do they raise a series a? And then years later, you find out whether the company’s any good. And so I think one of the things we got wrong was just overreacted to short term data early on. So in the first couple of years, we largely funded people who were straight out of university.

And then as the brand grew and the track record grew, we got more and more applicants who were, like, in the sort of, like, thirties. We’re like, wow. This is great. We’re able to, like, move up the the the experience curve. And, of course, the problem was they were more experienced, but they weren’t the very best 30 year olds.

One of the things that I think we made a mistake on is that as we scaled, and particularly as Alice and I stopped making every selection decision, you got into the thing where it’s very hard if you’re new to what we do to compare the CV of, like, a 31 year old who’s actually good but not exceptional, and a 21 year old who’s exceptional but has no experience, and choose the 21 year old. But you actually should choose the 21 year old. What’s the graduation rate?

So we have two phases to what we do. So, like, we have a phase where it’s basically common seat. We put you in a room with what we hope is, like, a really exceptional peer group of people that are equally ambitious, equally talented, and say, like, you know, take ten weeks to figure out if there’s something that you really wanna work on or someone you really wanna work with. And if there is, awesome. Like, let’s do the next bit. And, you know, that bit’s the sort of unique bit of EF.

It’s, like, not like an accelerator. It’s It’s really like about peer group curation. And so, like, let’s say for every 100 people that join at that stage, something like 80 of them will pair up with someone and come up with an idea that they that they’re excited about. But but we then do a cut where we say, like, do you really wanna spend the next ten years of your life doing that, and do we really think that’s the best use of your time? And so we cut about half at that point.

Everyone who joins us has, like, a huge opportunity cost. Like, they we’re very lucky to have, like, thousands and thousands of applicants to be able to pick really great people. And so, like, I just think, like, the the worst thing we could do is say, like, yeah, do that. Do that for the next ten years. I don’t feel guilty at all about people who wanna see whether they should be found or they come and they find out it’s not for them. Well, it’s fine. They go back and get it up.

The only people I would ever feel guilty about is people that come and spend seven years of their life doing something that could never be big enough to fulfill their ambitions. And so probably at that stage, we we say we’d like to fund about half of the combinations that have come up. These are teams that are ten weeks old. You know, we now put in 250 k at that stage, and I would say of those, typically, like, two thirds to three quarters would would raise a seed round.

Well, I was gonna say Charlie Songhurst, and you’ve already name checked him. I think He’s also an angel. Yeah. But he’s probably deploys more than most seed funds, I would say, every year. He he could easily put a wrap around it, and he would be a fund. The reason I would say Charlie is, like, I think Charlie has two things that I think he is very plausibly the best at in the world. One is I think he’s just a great talent spotter. I think he’s just really able to see the best version of every founder that he meets and decide how good that is.

You go back to my idea that that actually what matters is the peak performance from the peak person, like, has an incredible way of, like, very rapidly building a model of, like, how good is this person at their best? That’s an exceptional thing to because he invests so early. That’s, like, an amazing thing to be able to do. The other thing he does better than anyone else I know, he pitches back to the founder the most ambitious version of the idea they’ve just pitched him. He’s just exceptionally good at that, and he learns so much from how the founder responds.

A lot of people in Europe, and, you Charlie invests all around the world, but, you know, one of the things that he and I discussed a lot is part of the reason it’s hard to do investing here is that often by the time you meet founders, they’ve conditioned themselves to, like, reduce the ambition in the pitch because they’re just being so used to people being like, well, that’s not very realistic. So then what Charlie will do is he’ll pitch back to you. He did you know, he was he was one the first investors in the EF.

Like, he’ll pitch back to you, like, what it could actually be. I think he’s basically saying, does that excite you? Do you say, like, yes. Finally, someone who gets it, or do you flinch? And I think it’s an incredible skill.

a stranger. I think a lot of the conventional wisdom on who you should start companies with is just survivorship bias. Just that historically, it was almost impossible to start a company with a stranger. So the only people that the venture community saw were people who founded companies with people they’d known for a long And therefore, you know, of course, it looks like that’s the only way to do it.

Twelve months ago, if I’d been on here, I would have told you it’s too late to build an AGI company. Like, that ship sailed. There’s gonna be, like, three or four of them. They already exist. I now think the most ambitious founders today should seriously consider whether whether that’s what they can do.

So I’ve got two little boys. One’s six, and one is three and three quarters, as he very proudly says. I do think fatherhood has been, for me, a real sort of exercise in in humility. One of the advantages of founding something is, at least at work, people sort of listen to you, and you sort of earned in the sense that you started, but it’s sort of on a day to day basis, it’s somewhat unearned. Kids do not care. Like, my sons have no, like, oh, well, maybe just means I’m a bad dad, but there’s no, like, sense in which, like, oh, well, dad’s talking, so, like, he must be right.

And I I don’t wanna sugarcoat it. I I think fatherhood is really hard a lot of the time. What’s the

The hardest part is that there are no shortcuts. Like, these are humans, and if you want to have a great relationship with them, if you wanna be a great dad, you have to invest in that in exactly the same way you’d invest in a relationship or a friendship. And early on, they’re less fun than your friends, frankly. A lot of people seem to like maybe I’m not very good, but, like, it’s this thing where you truly love them in this very deep way.

It’s hard work to invest very deeply in this sort of relationship where, you know, whenever I travel, which is a lot, I, you know, I sort of have this idealized fantasy of when I come home and they’re gonna be like, daddy, and we’re gonna have this, like, deep conversation about, like, where I’ve been and what I’ve seen. And instead, it’s like yeah. Basically, it’s like, you know, I’ve pooed on the floor and, like, I’m gonna scream because I wanted a ball and you give me a pink ball and, like, it’s it’s hard.

You know? Like, it’s it’s not this, like, linear journey. But what’s amazing is it does compound, like, so many of the best things in life. And the moments when you realize you’ve built enough trust that they can find something, something that’s clearly been going around their little minds and they’ve not almost not being sure about how to articulate it, then they say it to you. Usually, in our case, like, when they’re going to sleep, going to bed, you’re like, wow. It’s really totally worth it.

Very few people ask me why I like writing immersive murder mystery games, but they should.

Most of us have this itch that is very hard to scrap, which is to be in these completely alien environments where we get to be someone else and do something else in a in a way where no one is in charge. I I you if know, you look at things like secret cinema and how popular that’s been, I kinda feel like it’s really interesting. People love dressing up and, you know, doing the thing, but, like, it kind of like, you dress up, you watch the film.

I think most people have an itch to see, like, what would it be like to actually have to, like, live through this scenario but with no consequences? I’ve written a series of historical murder mystery games where you don’t know it’s not predetermined who or if anyone will be killed. There’s 12 characters. You each have a set of goals and relationships, and the evening just unfolds and you decide what happens. And I love playing them. I love writing them. I think it’s a sort of human itch that there are very few ways to scratch, and it’s like some of the most fun that I think you can have.

See them on my website, Harry, for your next dinner party.

Lucrative. I can tell you.