Chasing Trillion-Dollar Companies, Founder Ambition, Token Budgets, & Regulatory Capture

70% of France is still nuclear in terms of its power generation. 70%. Where are all the accidents and all the kerfuffles? Nothing. Nothing's happened. The US is at 18%, and we haven't built a reactor in 40 years. We had a safety lobby in the 1970s basically kill abundant clean energy for us. There are real outcomes where safety has hurt us, and the question is: Where do we want the spectrum to be on AI for this stuff? There are many worlds, many scenarios, many outcomes.

Elad, it's good to see you. It's been a while since we just got to hang out with each other.

Working on companies in DC, trying to take a day off. You?

I know. It's been too long. What happened? Where have you been?

There's just so much going on right now in AI. It's nonstop. These are very exciting times.

Are you chasing the next trillion-dollar company?

1. Trillion Dollar Companies Are Rare

Yeah, it's a really interesting point because, basically, over the last 5 years or so, we had 3 companies roughly go from close to zero to $1 trillion in market cap, right? Anthropic basically didn't exist 5 years ago. OpenAI was still quite early. I think GPT-3 had just come out, and SpaceX was trading at 80, 100, something like that. Suddenly, we had this massive inflection in terms of the valuations of these companies, and I think a lot of people now are assuming that there are a bunch of other trillion-dollar companies that will be formed in 3 to 5 years.

That's unprecedented in human history. Usually, it takes 20 years, right? SpaceX actually took since the early 2000s, and Google took since the 1990s, and these are usually 15- or 20-year arcs. Then we had this weird 5-year inflection, and I feel like a lot of people are now looking at different areas that are very exciting and very promising—robotics, materials—and everything in everybody's mind is going to be a trillion-dollar company.

Maybe some of these will be over the next decade, but it's unlikely that we'll see that many more in the next 3 to 5 years. I mean, there's 1 I can think of that could maybe get there, but not multiple. So, yeah.

What's the 1?

I'm not going to say.

Ugh. Elad, where am I going to put my money?

I don't know. It's like throwing darts.

Yeah. That's why we have the dartboard here. So you think it's actually just a very good, special point-in-time vintage versus the ecosystem always getting bigger.

Well, it's more like a punctuated equilibrium, right? If you look at theories of evolution, 1 of them is punctuated equilibrium, where you have a Cambridge explosion, and then you have consolidation, and things are kind of steady state for a while, and then you have an explosion.

That's kind of the history of technology, right? We had a big social wave, but there aren't a dozen new social companies all the time right now, and we had a SaaS wave, and then they kind of settled down. We just had a giant AI wave.

There's still more to come, right? One could argue the internet had 4 or 5 periods to it. It had the internet of the 1990s, social in the early 2010s, around 2012, SaaS, cloud, and big security companies. You had crypto as a wave. You had all these waves happening, and sometimes they had 2 pieces, right? Bitcoin had a couple of different cycles, and other technologies will have that.

AI, undoubtedly, will have some giant breakthrough in model capability, and we'll see another step in some of these startups again, right? But we see these moments in time where things go from zero to a lot, and then those things become consolidators. The question is what comes after that.

I think we've now seen at least some of the consolidators emerge. The question is how many more giant companies are coming in the next handful of years, and that's different from saying what happens over the next 20 years. Of course, there's going to be tons of interesting stuff over 20 years. Over 2 or 3 years, there are still things that will grow a lot. There are still a lot of $100 billion companies to be built, but multi-trillion-dollar companies are hard to get to.

I want to talk to the investors you're talking to because I feel like I run more into a failure of imagination of how much bigger or better something can be than the closest proxy market from a previous era. I think being able to rethink market size is still a key underpriced investor skill right now at any stage, right?

Sure.

Take some of the application companies that we have in common or that other people have invested in. I think there are a lot of investors who have intellectually recognized this idea of AI companies delivering services value, but they don't act like they believe it. They look at everything a little bit more linearly, right?

If you're looking at Harvey or Abridge or something, they think about a per-seat, per-lawyer, or per-doctor TAM, and they're not actually asking the question of what the company looks like if they can charge for outcomes. They're not actually thinking about what's happening in the coding domain, which is consumption and value, 100X from here.

Coding, I think, is a much, much bigger market than anyone thought, and I think both of us were saying that a year or 2 ago.

But now the evidence is out there. You don't have to be a genius to take that to other domains.

The evidence is out there, but it's also: What is a trillion-dollar market and what is a $100 billion market? Both of those are big numbers, right? I actually wrote a blog post in 2010 or something talking about how hard it was to get to $10 billion in market cap, which is now a seed round for some of these new labs.

Don't get me wrong. I think the reality is that a lot of these things could be $100 billion, but I don't think there are that many that could be $1 trillion. Those are just different orders of magnitude.

So then the question is: What are these things that could actually be a trillion-dollar market? You just think of the revenue basis that's needed for that, right? You need $50 billion to $100 billion of revenue pretty easily, with good margins, right? So then the question is: Where are the $50 billion to $100 billion revenue streams for single companies?

That's a different question than whether the TAM is really, really big, right? That's a huge TAM. That's a very small number of markets in the world. There are a lot of them—there are a dozen-plus companies that are there-ish—but how many more will there be in the next 5 years?

That's my question. It's not what happens in the next 20 years; it's what in the next 5 years will be able to get to $50 billion to $100 billion of revenue? That changes how you think about this, right? There are tons that can get to $5 billion or $10 billion of revenue, and they'll be a $100 billion company.

I think I'm looking both a little farther out, and I'd say I don't know that there are that many markets that are going to get to $100 billion of revenue in the next couple of years that aren't inference, right? Tell me what else you think could get there in that timeline. Perhaps some supply-chain- or energy-type technologies.

Maybe, yeah. You can make a list of 5 or 6 areas that seem promising. Part of it, too, is that if it's a physical-goods company—energy, robotics, et cetera—do you actually have the footprint to get there that fast?

Again, I'm not doubting the size of some of these markets. I'm doubting the speed at which you can get there. Yeah, 100%, and that's the issue. People are, at least in my experience, collectively investing against the fact that they believe the speed is there, which is different from the market size. People are conflating the 2 things right now, in my opinion.

2. Founders Fear The Labs

The other phenomenon that I think is happening is almost the opposite of it, which is I see some really, really excellent founders going after niche markets because they're now scared of the neo-labs.

I think there's much less head-to-head competition. If you look at the markets that Harvey, OpenEvidence, Decagon, or any of these folks at Sierra entered four or five years ago—or even three or four years ago, and Cognition roughly two years ago—they were big, big markets that could be in the roadmaps of these labs.

I feel like two things are happening at the same time. One is that, for the mid- to late-stage technology markets, people are continuing to invest as if there's velocity to get to $1 trillion for many companies, where I don't think there's that velocity. Again, I think some of them will get to $20 billion, some will get to $100 billion, and some, of course, will go to zero.

Then there's a separate thread of all the new stuff that's coming. How aggressive and ambitious are the founders relative to what the labs are doing? I think that's why you're seeing a flight to hardware companies: “The labs will never do this hardware thing, so we'll do that.” There's also American Dynamism, niche applications of AI, something that should be provided by an inference cloud, and so on.

There are a lot of these types of companies that I think are potentially going to be a bit more derivative. And don't get me wrong: There are people doing huge, amazing things simultaneously. It's not every startup. But there's more and more, at least in my perception, of people doing smaller, niche things out of fear of the labs. I think that's also a negative.

And you feel like they're being too meek—that they should just take on the head-on competition because you can create a much better experience and compete on the product, the distribution, or any of it.

I think so, for certain markets. Of course, there are certain markets where the labs will just eat it naturally, but there are a bunch of markets where they won't. I think people are staying away from both.

Well, we have companies in the portfolio that are going against pretty central premises, so I don't think all the founders are being too meek.

Oh, I don't think it's all of them. My point is that it's more of a trend line, and it's the newest stuff. I'm not talking about a company that's a year old or two years old. I feel like it's a trend line that's shifting. Again, it's not all of them; it's just enough of a subset. In a sense, it's a subset of the good founders. I'm not concerned about the median founder. I'm concerned about what the best founders are doing.

I'm more often disappointed right now that founders are being less ambitious than they could be. So maybe that's the trend line you're talking about.

3. Founders Need Exit Discipline

We were talking about when companies—when founders—should sell their companies. What is your thinking on it at this point in time, or what is your framework for it?

There's a handful of companies that should never, ever sell, at least any time in the near term. If you're Anthropic, you shouldn't sell. If you're OpenAI, you shouldn't sell. There's a handful of these things that should never sell.

Most companies in any given era should at least consider it, and there's usually a time-maximizing window where your best outcome is a sale within that window. It's usually a 12- to 18-month period when the company is worth the most it'll ever be worth. I think we saw one major exit where that was probably the case reasonably recently. I think there are other companies that should really actively think about it.

From a hygiene perspective, maybe what companies should do—I think Ben Horowitz wrote about this once—is basically have a preplanned, once-a-year board meeting where the discussion topic is, in a nonemotional way, “Should we consider exiting in the next 6-month period?” It's prescheduled, so it's not the founders pushing for it, and it's not the investors pushing for it. It's just a rational conversation.

The answer may be, “No, we should keep going. We still think we have X, Y, and Z ahead of us.” Amazing. But I think it's very useful for people to have that sort of conversation because I feel like, in this cycle, every year of AI time is like 3 to 4 years of a normal cycle.

Three years is like a decade, right? Think of what existed in AI three years ago from a model-capability perspective, from a vertical-app perspective, from AI roll-ups, from infrastructure—whatever. It's a radically different world from three years ago.

We're on an accelerated timeline right now where everything is moving faster. That means you should double-check your thinking more frequently because the underlying fact set is changing faster than it ever has. I don't know. What do you think? What's your approach to exits or not exits?

I agree with you that there are a set of companies that should never sell unless they cannot finance their future. If I think about it, maybe one principle that's new for this point in time is that I might ask at that board meeting—or at that meeting once a quarter or once a year, whatever you think is the right pacing today, and it's more often than it was a few years ago—

Yeah, it's every 6 months.

Okay, every 6 months. Great. Are you capturing value as costs fall and capabilities increase? Because if you're on the wrong side of this secular change and you can't get to the other side of it, you should, in fact, sell. You don't have good ideas about how to be on the right side of history. I think that's a question people should ask themselves.

If you think about our friends at Cursor, is the way you want to compete capital and compute access? Is it perhaps a maximally valuable point in time? That's an interesting question.

More broadly, I feel like it's a very personal and very interesting risk-management question. I do think people should ask themselves. The idea that there is pride around never considering this is nonsense. The situational-awareness situation is a good reminder that everyone has to stay alive to profit as well.

Hedge funds are different from companies. They have to survive to compound. But I think even the premise that you need to match your financing structure to your thesis horizon, and then be able to continually finance the company to the promised land of whatever you're trying to do, is important.

Yeah. I think the financing part, though, is going to be there because of what's happening with the rapid rise of 3 trillion-dollar-plus companies in a short timeframe. An enormous amount of venture capital is starting to get returned, which means people are raising bigger and bigger funds, and they need to put that money somewhere. They're going to put it against the trillion-dollar companies of the future.

I do think we're going to see an ongoing rise in valuations, most likely over the next year or two, much more than we've seen to date. Obviously, there'll be some great things in there, and there'll be a bunch of stuff that doesn't deserve it, but I actually think financing is going to get easier, not harder.

I'd view it less as a financing question and more as: What do you think is the true, likely expected outcome of your company? Not what investors are telling you, not what the press is telling you, and not what Twitter is telling you. Just sit down and run the math. Then remember that, at some point, you'll probably trade it at 10x or something—maybe 15x.

The question is: What is your thing going to be worth? Remember, eventually, things slow down in terms of compounding, too, and you can decide where that slowdown happens. You do that math. You account for future dilution. You look at your potential outcome and the years of work it'll take to get there, and you can come to a conclusion.

There are two types of opportunity costs—or risk management. There's risk management against the value of the thing you're doing, but the biggest opportunity cost is your time. Your most productive years of your life are on the line right now.

You can either walk away with a good amount of money and go do the next giant thing, now having done it before, with people willing to work with you again, or you can roll the dice. You can decide to roll the dice. That may be the right answer. Again, for some companies, absolutely, you should do that.

But for others, it may be, “Hey, actually, now is maybe the time to go.” A secondary is an intermediate option, which I actually don't think is always that great because it solves for some short-term needs, but it doesn't actually create a solution.

You see a lot of people from 2020 and 2021 still running companies 5 years later that aren't working. Think of that 5- or 6-year period when they've been locked up while all the AI change happened. What is the cost of that to a great founder?

I think there's that kind of cost that people don't really talk about as much, which I think is the real cost. It's your lifetime cost, right? You only live once, and it's a short life. Do you want to eventually be working on something that's going to continue to struggle, that's overcapitalized, and that has runway for the next 10 years or not?

That's where you end up. That's what happened with the 2020 and 2021 cohort. There are tons of people still running these companies that aren't working, and we forgot about them because we're talking about AI all the time.

That is a huge waste.

I think my point was really that even if there are lots of dollars still rotating into venture or being produced by these huge outcomes over now and over the next year or two, private markets don't have to be rational or right for long periods of time, right? And so being smart about your ability to finance a company is the equivalent of avoiding margin calls, right?

Some founders who are working on something that requires a technical point of view, for example, or even a structural point of view about how the market resolves, can get very frustrated because investors will believe something that they think is wrong or stupid for a long time, and it's just the job of the founders to navigate that narrative or that set of beliefs. And if they think it's untenable or if they think they're down some wasteful path of their time, as you describe, then they should sell the company.

But it could be worse. I mean, founders have the concentration risk, but they could be hedge fund managers facing retail irrational acts in the market and reductions next quarter, so it's just a different environment. But I don't think it's as simple as financing is now free. I do think it is going to skew, as you said, toward scale of opportunity naturally.

Or perceived scale.

Perceived scale. Yeah.

Perceived scale is important, I think. So the other thing a lot of people out here are working on or talking about is, if you talk to people at the labs, there's this enormous manic energy right now. We're 6 months-ish, or towards the end of the year, from being completely done with code as a solved problem. And then we'll probably hit some form of light RSI by the end of next year.

At that point, you'll have models training big chunks of the models themselves. I think it's probably more post-training initially. Maybe it could impact pre-training over time more quickly as well. And because of that, many people believe, “Hey, if I have a year, a year and a half left of productive work in my career, I should be working 16 hours a day because every week is 2% of all the time I have left to be productive before I get displaced by AI.” What do you think of that?

Do I believe it, or what happens if it's true?

Do you believe it?

I think the idea that the models can improve their own training—if the leading scientists working on this believe it, and it's an extension of what we're already seeing in code and math—of course you should believe it, right?

Yeah, but on that timeline.

The data I have is that a number of very smart and even very self-aware research scientists have felt that there was some knee in the curve on recursive self-improvement or ASI 18 months away every 18 months for the last 5 years. So how good of a predictor is that? It's not clear.

I think the extension from code to training code to data pipeline work is way easier to believe. The question of how you're going to gather that data for less verifiable, more complex domains, or whether you run into the actual constraints on the physical compute accessibility side—I think that's probably more of a limiter than this being algorithmically possible.

Yeah, I mean, the physical compute basically reinforces an oligopoly market because what it does is create a ceiling on the rate of progress any single lab can get if, effectively, you assume the compute is roughly pro rata across the ecosystem to the big labs. And so in the absence of a lack of compute constraints, you almost have an enforced oligopoly market up to a point, or at least you force closer competition between the players than would exist otherwise, which I think is an interesting, odd effect of this moment in time. And the question is, when does that lift, and what does that look like?

So, yeah, it's a very exciting time. I mean, I always wonder about second-order effects of that belief that it's 18 months away, because that does suggest there could be a burnout cycle in 18 months. I know some people at one of the major labs who, for a while, brought up with me whether they should get married. Do people want to get married? Should they get married because they don't know what happens in 18 months to the world? It's like, “You should get married. You should go ahead. It'll be okay.”

And so I do think we're living through this very manic, very exciting, very intense work period. So, yeah, it's really fun stuff.

I think it's kind of tragic, man.

Really? Why?

I think the reactions of some really extraordinary research friends to it feel a little bit tragic. I feel like it's psychologically most similar to if people think they're going to die, right? How would you spend the last 2 years of your life? Would you spend it the way you are today, or would you spend it in a very different way? That's a not-unrelated philosophical question.

And so you do have people who are like, “My contribution is a bit irrelevant given our ASI in the next 18 months. So should I get married? Should I bother to work? Should I travel? Should I only work?” And I just actually think it's a much more stable and satisfying state if people act as if they have, but maybe you think that's blind.

No, I just think there's a lot of second-order effects that are happening or are going to happen, and part of them are driven by this belief system and potential burnout over time. Part of it is going to be—one thing that I've noticed happening at some of the labs is that as compute becomes really the scarce resource, it turns out that there are, say, a few dozen researchers who drive a lot of—80% of—the results at any given place, which is a really interesting human power law, right?

If you actually look at it, in any field, there are at most a few dozen people who drive the field. You look at breast cancer research, you look at certain subfields of mathematics, you look at subfields of physics, you look at the entrepreneurial ecosystem and founders—there's a handful of people, dozens of people, who drive most progress. And that also happens in AI research.

4. Token Budgets Reshape AI Work

Increasingly, compute is differentially provided to those people, right? And so I know some labs have slowed down on their hiring of researchers unless they're above a very, very high bar, because the cost isn't the researcher; it's the compute associated with the person. That's really where the bottleneck is.

I think there's this broader concept of return on invested tokens, like an ROIT kind of metric. If you have a certain token budget, who do you give it to, and why? This is kind of like engineering back in the day: the internal tools teams at companies were always starved for resources because many companies, at least tech companies, would rather use the same engineers to build product than to build internal tools that would make other functions more productive.

That's why I think the death of SaaS is a little bit overstated, because why would you use tokens on a bunch of SaaS stuff that you're not actually paying that much for per year relative to the outcome of those same tokens being invested against a core product or against some massive margin lift or some other thing, right?

And so I think increasingly we've shifted from a world where people said, “Hey, everybody use AI and do whatever you want,” to, “Hey, we have to measure spend and move more things to open source.” And then I think the next wave is: What are the projects and people that should actually get outsized pieces of a token budget, and what is that return on investment? It's the sort of next shift that's coming. It'll take some time, though. Many people are still at, “Hey, everybody, try AI or whatever,” at big enterprises.

What do you think is the appropriate compute token budget for a business or a human being 3 to 5 years from now? Should I look at it like rent?

No, it depends on what the budget is for. People forget, too: Minecraft was—what was it?—5 people, 10 people when it was bought for billions of dollars by Microsoft. People keep talking about someday there'll be a multibillion-dollar single-person company. That was basically Minecraft, roughly. It already happened 15 years ago, or whenever that was.

So there are always people who can take outsized advantages of technology, and AI has accelerated that radically. And so at some point, it's like, why give tokens to people who can't do that on a relative basis unless you just run out of those people? And you may run out of them.

This is back to, well, all the engineers get laid off. Probably not anytime soon, but you could argue that at some companies, even before AI, there were a bunch of engineers who weren't that productive and could be let go, especially at some of the big tech companies. And I think a lot of those folks will be very coveted by GE, PG&E, or Hershey's.

So even if there is some displacement of engineers at some point in the future, I don't know when that is or if it happens, but if it does happen, there are lots and lots of homes for them because there are tons of enterprises that never had the capability set or ability to recruit these people, and they want the capabilities they bring. Even if they're mediocre in the context of a Google or Meta or whatever, they may be exceptional in the context of a certain subset of old-school enterprises.

And so I do think there's going to be this permeation through the enterprise landscape of engineering talent in an unexpected way.

This is probably many years away. I'm just saying I think that's probably a likely outcome.

I think, relatedly, if you are a researcher, 800, and you have not been allocated an outsized number of tokens to work with at one of the major labs, the opportunity to spend your energy on something where you have comparative advantage and understanding, and that should benefit from all of this—the supply-chain bottlenecks, domains that should accelerate, like bio diffusion into other valuable fields—seems a lot more exciting than being concerned about the downfall of mathematics and going on vacation until the world ends.

Oh, yeah, lots of places to go do stuff. I do think that's where a subset of the research community will end up over time, right? An interesting question is how many researchers you need if you're a top AI lab. How does that number relate to the number that you have now, and do you have the right number? Do you need 5 times as many people? Do you need half as many, but they all need to be above a certain bar because it's compute-constrained, and you want to map it against the best ideas, and the best ideas come from a subset of people on average? Not always, but on average.

So it's a really interesting question of how all this stuff falls out, and then where does the N + 1 person go? There are lots and lots and lots of places for the N + 1 person to go. They're still exceptional. They're still at the top of the bell curve. Again, I don't want to have that misinterpreted as the person not being amazing. It's just that at some point, people will do some cutoff on their power law.

I think you will appreciate this. Maybe you've heard it because it's an old ex-Googler joke. When Google was, I don't know, 50,000 people or something, the question was, “How many people does it take to run Google?” If you asked somebody within search and ads, they'd say, “Oh, like 20% of the people in search and ads.” If you asked somebody outside of search and ads, they'd say, like, 50,000 people, or whatever Google was.

So I think there are probably some varied perspectives on the concentration of contribution. I don't know, having never worked at Google.

Yeah, I mean, we definitely know that. I worked at Google. I thought it was a wonderful place.

In search.

I worked on mobile search a bit, and I worked on ads a bit. I worked on a bunch of mobile stuff, and then I worked on a bunch of AI- and ads-related stuff.

5. AI Faces New Disruptions

May I ask you a very different question? Can you think of anything that could disrupt us all right now? You could have investors—like a number of different players—collapse in their commitment on the CapEx side because the markets hate it. There's some sort of freakout about the debt and the returns profile. You've seen minor indications of that, but not real pressure yet.

And the last one is: do you think there's a technological disruption that's possible, like alternatives to transformers? Does that still matter at all? Is there anything that would make the landscape look really different technically? I think there are always technology unknowns, and then I think the idea of attempting to restrict usage of models we already have, or open-source models, to dramatically constrain the pace of progress is the other one.

Yeah, and I agree with the regulatory angle. What do you think is going to happen in California? They passed the billionaire tax, and the Democratic Party in California came out in favor of it. You're a founder of one of these companies that you've backed that's now worth $10 billion-plus. Is the founder going to have a forced asset sale next year, assuming it passes? Will dozens of founders have to sell big chunks of their companies?

It's not clear the regulators have thought through the execution and compliance of this, but I think the immediate effect is that a huge number of people attempting to create value or do new things in California choose to leave. That's already happening. It's hard to move an entire ecosystem very quickly. This is the fastest way I can think of to chase the entire ecosystem out. What do you think happens?

Yeah, I mean, the way that law is written is, my sense is, reasonably broad in terms of once it passes: they can reimplement it, they can lower the bar in future years, et cetera. And my sense is that in 2028, there's increasing talk about also trying to add an exit tax in California. So if you actually try to leave, they'll try to take a big chunk as sort of a penalty for that.

So is your prediction mass migration in 2027 to Miami? Miami finally happens?

I think it will take some time. I think the people who wrote the bill want the flight to happen. I think they want people to leave, and I think the 2 really negative signs for California were this bill and then the sort of ballot-harvesting initiatives. I think those are the 2 things that make it a potentially worse future for the state in different ways.

So I'm hopeful, as usual, that California figures it out, but I think if there were any alternative that was easy to do, a lot of people would—even more people would be leaving. I do think a lot of people are leaving. I know quite a few who are starting to go now or planning to go by the fall, next month or so.

What is your second-choice ecosystem?

I think there are a few different places that a lot of people are considering, and the question is, what does critical mass look like in 2 years at each one of those spots? A lot of these things self-assemble, and people talk about weather and all these other things, but the reality is, Boston used to be one of the main startup hubs. It still is for biotech, right?

They sort of lost competitively in the early '90s. In the '80s, Boston was sort of the counterweight to Silicon Valley, and the weather there is awful, you know? I think it's more about where you have enough smart people aggregated, working on common things, and then that's where these renaissances tend to happen.

I think it's been really exciting to see the amount of great technology migration and innovation in Texas around energy. That is really a reaction to the regulatory environment and demand, where I've seen a lot of people either move from Silicon Valley or move from other places because it is a place where you can experiment, and there is actually an ecosystem now that's super exciting.

Energy and hardware, actually. There's a really growing hardware corridor there as well, which was originally all around El Segundo because that's where SpaceX was, and then Anduril. Now SpaceX, and I think part of Tesla and stuff, moved to Texas, and so there's this new ecosystem kind of emerging around a part of Texas as well, in addition to Austin.

So I do think we are seeing these shifts, and these shifts are purely driven by regulation. They're not driven by whether Texas is a better or worse place to live. It impacts things, right? But it's regulatory shifts driving people out.

You didn't take my bait on architecture and technology.

What do you think about architectures?

I think we are going to, as an industry, consume all of the compute and power available, whatever the underlying architectures. So the idea that there's going to be a lot of pressure to find more memory- or power-efficient architectures is more interesting than ever, but catching up to transformers in scale and matching the hardware remains pretty tough.

I think people will make that bet as they get more desperate in terms of more experimentation. I don't think it changes the direction of the industry.

I think whatever it is gets copied, and then the labs do it, and they have all the compute anyway. That's the high-probability outcome. It's not the only outcome. There could be a lower-probability scenario where some neo-lab comes up with something, keeps it super, super secret, scales on it, and suddenly its model is better than anyone else's by far. Then it can afford all the extra compute and everything else, and everybody rallies around it.

You could always imagine a scenario like that, but you could also imagine a scenario where just 1 person from that team leaves for Anthropic or OpenAI, and the knowledge spreads, and the next thing you know, everybody has it, which is what's been happening so far in terms of these models.

Do you think, if the dominant thing is access to compute and it's an oligopoly because of it, you see the labs using that access to compute to control other verticals that they want to be in?

6. Safety Can Stall Progress

Or what is the safety that's really needed that's actually protective of people, right? What is the risk? What is the outcome? That's kind of the, you know, Jensen from Jensen Pharmaceuticals, you know, he's considered one of the best drug developers of all times.

He has these great videos on YouTube where he was interviewed 30 or 40 years ago, talking about regulatory capture in pharma. The reason things got so expensive and so slow is, number 1, regulatory capture, and the second is risk-reward scenarios where the FDA, in his mind—I’m not saying this is correct or incorrect—focuses too much on safety and risk and not enough on benefit. And so there’s no risk-reward. There’s only risk. That slows everything down because you’re only looking at one side of the equation.

One could imagine a scenario where, in the labs, a version of that is created as well, right? The safety burden is so high, even if the outcome is higher, even if the positive outcome is dramatically higher relative to the risk. This is back to: if you only focus on one side of the equation, you’ll always constrain things. If you constrain things but then push progress forward internally on an exponent, and a year is worth 3 or 4 years in normal time, then you’re a year ahead internally. That’s a massive advantage.

It’s this very interesting question of where we, as a society, feel comfortable on the risk-reward spectrum for different things. If my email gets hacked, is that so terrible relative to better healthcare through AI models sooner? That’s the trade-off. These are all things we’ll have to work through from a societal perspective.

I think part of the challenge here is that it’s not a comfortable stance for many policymakers to hear from technologists: you have to see what happens with the technology versus control.

We’ve always said that. That’s always been a tech thing. It’s always been throughout history: “Hey, of course, this technology could be used in negative ways.” Every technology has both positive and negative applications. Biotech: you could create a virus, but you can cure cancer. Nuclear: you could have free, cheap, abundant energy. You can also create weapons.

If you actually look at it, 70% of France is still nuclear in terms of its power generation. Seventy percent. Where are all the accidents, and where are all the kerfuffles? Nothing. Nothing’s happened. The US is 18%, and we haven’t built a reactor in 40 years. Japan is 25%. Very safe, very abundant, but we had a safety lobby in the ’70s basically kill abundant, clean energy for us, right?

Well, we’re making them now. We just need to make a lot of them.

We’re not making much. We’re not making much. I think we’ve seen real outcomes where safety has hurt us. It’s hurt us in power and energy production, it’s hurt us in aspects of medicine, and it’s hurt us in lots of places.

The question is, where do we want the spectrum to be on AI for this stuff? There are many worlds, many scenarios, many outcomes. Societally, we get to choose where we want to place that needle on the wheel of safety versus risk versus outcome.

Elad, before we go, what is something that you’re just excited about that is on the positive end of that wheel?

There’s so much stuff I’m excited about there. I think there’s so much we can do from a human productivity perspective, from an education perspective, from a healthcare perspective, from a daily-life and benefit-to-life perspective—self-driving, helping the elderly, everything. There’s so much good that can come of all this.

I’m optimistic about a lot of applications, and that’s why I’m cautious about where we should end up on that spectrum. I do think it’s always good to make sure that we have the proper safeguards societally. But I think that historically, for big industries, we’ve gone too far.

The reason tech has been so successful so quickly and has had so much human impact is because it’s been lightly regulated. I think it’s better to keep it that way than not. We’ll lose optimism, we’ll lose momentum, and we’ll lose progress. That’s what happened in biotech, and that’s what’s happened in a variety of areas over time. That’s what happened in energy for a long time.