Tyler Cowen: AGI is already here by the original definition, but human bottlenecks mean slow economic takeoff

it out. Uh, and we will get to the bottom of the conspiracy theory. Um, but we will do that later because our next guest is here. It's Tyler Cowan. How are you doing, Tyler? Which conspiracy theory? So, this conspiracy theory is that you'll actually like this one. Okay.

So, uh, are you familiar with the high uh, the ultra luxury watch brand Rishard Mill? Only by name. Okay. I don't have one. You may have seen it on a on a sailboat or something like that. Yes. So, Reishard Mill is famous for selling watches at 10 times the price of other luxury watch brands.

So, a Rishard Mill, the first one was $135,000 for a single watch. And at the time, $13,000 would have been more appropriate for that kind of tier of watch. And the conspiracy theory is that when Rishard Mill went to promote their new watch, they took out an advertisement in the Financial Times.

They sent over the copy for the advertisement saying, "Please uh please promote Rishard Mill. Our new watch is is available for purchase. " And the price is $13,500. And there was a mistake in the newsroom and they added an extra zero and they printed $135,000.

And what happened was instead of getting no offers, they got flooded with offers because people had to have it. a classic example of a vellin good. Um, and uh, and the rest is history and to this day the brand denies it. Uh, but what do you think?

Well, it sounds like the opposite of a conspiracy, just an innocent mistake and then it worked out well for the company and bad for the buyers. Yes, I believe it.

the the the conspiracy is that the brand now obviously denies it because it would it would not be a good look for them now to to be like, "Well, we actually wanted to price it at 13,000 and then we actually, you know, uh now that you're willing to pay 135, well, good deal. Let's do it.

" Um, but there'll be a deathbed confession on that one. One way or the other, maybe. Well, it's great to have you here. It's great to have you on video. Uh, super intelligence is Yes. Yes. Yeah. Yeah. Yeah. We had some technical difficulties last time, but I'm glad. Uh you look great. How are you overall?

Let me take your temperature on uh the last time we talked, you said AGI was around the corner. Later, two days later, uh 03 released. We now have 03 Pro. How are you feeling on artificial intelligence progress? Um overall, do you feel like we're accelerating, decelerating? Are you happy with the products?

How are you using them? Take me through kind of the state of the union. Everyone has a different definition of AGI, but the definition I grew up with, which is that it can pass a touring test and do at least as well as expert humans on most endeavors, intellectual endeavors, it clearly passes.

So, if you want to call it AGI, you can. I don't insist on the term. Maybe it's more misleading than useful. But 03 03 Pro, I use them every day. I ask them questions about history or music or travel. My wife and I did a trip through Europe, France, and Spain. We asked it about our hotels, our restaurants, the churches.

We saw everything. It just keeps on working. It's amazing. Much better than having, you know, a historian guide with you. So, to me, it's like AGI. It's better in fact. Are you What is your prior for hallucination at this point?

because that that Rishard Mill conspiracy theory that I just um relayed to you that comes from an from a chat GPT 03 pro deep research report and uh we're getting text messages that many people hadn't heard of it. I haven't checked the sourcing yet. Um so I'm not sure if it's real.

I haven't gone down and done all the factecking. But just in terms of your prior, how likely would you believe that it's uh that it's a a hallucination? Well, just ask it, right? If it's something important, you double check with another AI andor you ask and you find out.

So, hallucinations should not be a practical problem. They end up one because people are silly or they'll ask it questions where there may not be any answer at all. And that is by far when it's most likely to hallucinate. They'll just make something up to please you.

But if you know in advance, that's when it's most likely to hallucinate. Again, you take counter measures. So, it's a way overrated problem. I agree. I agree. And I just fact checked this and it does come from a 2021 GQ report, which again might be wrong. Hallucinated by the GQ never hallucinates, right? Take it out.

Gentlemen, do not hallucinate. No, gentlemen do not do not hallucinate. Um, what do you think of uh Dwaresh Patel's new formulation that there's sort of a a two axis diagram between you and him? Uh, AGI is here. AGI is not here. The impact of AGI economically will be massive in his in his mind.

Uh and the impact and the impact of AGI will be uh more muted in in his assessment of your assessment.

And I think I I think your general stance which is uh we should hold ourselves accountable to the definition that we set out to achieve you know to achieve you know decades ago when this work started and not constantly just be you know pushing out uh the the sort of requirements the goal moving the goalposts uh is is important and then there's this new definition super intelligence yes I'm waiting for super duper intelligence right giga giga intelligence ultra intelligence we already we're already past it.

Everyone, folks like us, we know what super intelligence is. We we we already have it. We're waiting for the next one. Maybe the best measure is how much of people's time are they spending with this thing and that's rising steadily, right? So that to me is quite impressive. Darkeesh understands my views perfectly well.

He and I use the term AGI in different ways, so it all ends up being confusing. I just think the human bottlenecks are so significant that it's not going to be fast progress anytime soon. That's my core view. It's not based on any kind of bearishness about AI.

It's that I've spent my whole life working in human institutions. And there are plenty of much simpler developments like don't be a jerk that would boost productivity immensely, but we still haven't gotten around to really implementing. Yeah.

Uh, but I mean on the on the topic of super intelligence, I think what people are kind of getting at is that it feels like we have 15minute AGI. It feels like we have spiky intelligence.

It feels like we have intelligence that can, you know, memorize every book in the English language and yet not tell you a funny joke or or put a novel connection together between two disperate disciplines. Well, and and potentially more important is this idea of of intelligence with amnesia, right?

It's like uh you know, we've been talking about this on the show and Darkeesh uh talked about it last week, which is this idea of PhD level intelligence that you know can't remember what you told it uh you know yesterday, right? And so it's difficult to sort of build real momentum and real uh agency.

I think it can do all those things. Now, it might require a little smidgen of work from you. Like you might need, for one thing, you can turn on memory. That's maybe not as much memory as you'd like, but you can just reinsert the dialogue you had with it and tell it to start from there.

So, okay, it's a slight inconvenience, but my goodness, people, they're so unhappy. Like, it's incredible what we have. And with modest effort from you, it does all these things. So, I completely agree. Uh I I love these systems. I I enjoy spending time with them even if they're not directly providing economic value.

Uh I find them entertaining. I find them just interesting. I I learn things. It's fantastic.

Um I guess the the the the question is like yes you can stuff the context window but increasingly folks are saying that uh there's some sort of fundamental limitation with the size of the context window that yes you can do needle in the haststack analysis with a big context window but even if you have you know pages and pages of all your best practices the current systems are more likely to forget some hard one lesson early in working with you than just an average employee.

And that feels uniquely unhuman. I don't know. I've worked with a lot of employees. This is one generous way I would put it. The things are not perfect. If you compare them to smart humans on intellectual tasks, they're basically ahead. They can't dribble a basketball. They cannot do most jobs. That's for sure.

It's a big reason why progress will be slow. If nothing else, those jobs intersect with the physical world. Robots are far away. And there you go. Um, I think it's a pretty clear prognosis. I don't really get why anyone disagrees with it. Well, if you need to raise a billion dollars, it's not super convenient truth.

If robots are far away, for example, specifically economic growth is far away. This is the Sachin Sachin Nadella uh uh position which I think he agrees with you on the economic impact that we're currently seeing and potentially the economic impact for the next few years.

And so he is underwriting a very different investment strategy in artificial intelligence. Yeah, it it says a lot in in our view that that that everywhere you look there's just blaring top signals in in the market.

both like literal all-time highs uh but then just other more more indirect signals and at the same time Satia is like you know doing layoffs being seemingly you know very pragmatic uh and and and I read that as more aligned with your view which is transformative impactful incredible wonderful m magic but also potentially incremental and um you know going to be transformative in the fullness of time but not next quarter Forgive me for not sounding like a complete capitalist, but this is all so important.

People should just be doing it. And to some extent they are. The labs are working very hard to make progress. You can debate how much the motive is profit. But look, these are driven people. They they get how important it is and they're doing their best. I think that's amazing. Yeah. Are they all going to get rich?

I'm not predicting that. I don't know. Yeah. Yeah. I I think the the tradeoff that people are debating broadly in the market is if if the next training run that's a little bit of a stretch for the market.

We were just talking to the f co-founder of coreweave and he was saying that like the capital markets just aren't really ready to absorb what it means to 10x data center construction and at the same time if AI researchers are saying that the the that by 10xing data center construction we will lift ourself off of the GDP baseline and we will see breakout economic growth and we will get through some of the problems that you've identified about actually you know restructuring ing the economy and speeding up economic growth.

All of a sudden like that question and that that how much should humanity as a whole invest as a portion of GDP that becomes a very important question for the overall market to solve. Well, let's go talk to United Arab Emirates. They can either buy some more parts of sports teams or do this.

I think they're going to do this. The growth rate, I think, will go up. I've offered the very speculative estimate of half a percentage point per quarter, but that's an enormous sum of money. That's massive. Now, you know, is it captured by the capital markets, the shareholders and so on? No.

Just the free time liberated by Chat GPT and the other services already is enormous money. So, remember Satia had that challenge. When will it earn $und00 million? Mhm. Well, it has many times over. It just hasn't earned it for him. So, people get their work done more quickly.

You add up those valuations of time, I'd love to see a study in the number, but it's way way above a hundred million right now. Well, on that point, there was a study last week about does uh does uh AI code generation tools? Do these tools actually improve developer productivity?

And they did a I believe a double-blinded trial with or or maybe just like a split up trial uh based on talented software engineers working on open source projects kind of trying to work on some of the harder problems in software engineering not just build me a new web page um actually advancing open source projects and they gave half of the developers uh access to tools like cursor and wind surf and cloud code and and then the other half didn't have these uh AI driven in uh development environments and they said that the AI enabled engineers expected a 20% speed up but they in fact experienced a 20% decrease or 19% decrease in productivity.

What is your take on that? Do you think that uh there's a world where we're fooling ourselves into believing that we're actually sped up by these tools or do you think that that's completely ridiculous and it's obvious that this that these tools speed us up? That's BS that paper. first only 16 data points.

Second, only one of those people had experience working with AI already and that person's productivity went up. It's the other fools who were just trying to get a handle on it at all where the productivity went down. So, we need to throw that out, get it out of our minds.

Okay, it's obvious from marketplace tests that AI speeds up developer productivity. Of course, you have to do it right. You need to learn some things, but it's working. Overwhelming real world evidence to that effect.

Do you think that there are particular tasks maybe that you've encountered where you've gone to AI and then afterwards realized like well for this particular task not just planning a a vacation or trip or something I was actually slowed down. Sometimes I like giving stupid dogmatic answers. I'll just say no.

I think it's a correct stupid dogmatic answer. No, it's never happened to me.

I love that because it's definitely happened to me when I've when I've said like generate a generate a a 2 by two chart and I'm like seven prompts deep and I'm like I could have done this in Excel or I could have done this and just taken a screenshot or I could have done this in Photoshop faster.

Uh it it it does it does happen to me. But I agree on net I I feel sped up and I and I feel very happy to use the tools and I'm I'm enjoying them. I use it to prepare for podcasts. Sure. It cuts my prep time in half. Sure. No real doubt there. I don't use it for charts. I this tell a human. Maybe the human uses it.

I don't even know. I get the chart back. It's great. I love the world. I love it. Uh what is the uh what is the labor economist in you think about the AI talent wars? Uh has have AI researchers been historically underpaid?

Are we in some sort of odd intellectual property theft ring where we're paying for trade secrets to trans to change hands? What is your read on kind of all the talent moves that we've seen at staggering numbers by comparison in what's historically happened in Silicon Valley? Yeah.

And then specifically, do you think that uh an AI researcher in the year 2030 will be getting a hund00 million signing bonus or do you believe that things will maybe normalize and and uh researchers will go back to being paid well but maybe not more than Fortune uh or MAG7 CEOs? Probably normalized.

Now, I'm an NBA basketball fan and what I observe is the top teams need some time to gel. So, you look at the Miami Heat, they got LeBron, Dwayne Wade, Chris Bosch, the first year they lost. You look at, you know, Chicago Bulls, Jordan and Pippen, it took them a few years.

You look at the Lakers, what they had Carl Malone, Gary Payton, was it Shaq and Kobe Bryant? They didn't win a title. So, it's not as easy as you think.

Not that those aren't great great players, but there's something about how it gels and the process and what part of people's trajectories you get them at that all has to fit together. I think it's a fascinating experiment. Uh I'm dying to find out how it's going to work.

What do you think of the missionary versus mercenary debate? uh a prior guest on the show uh highlighted Tom Brady in the sports world taking a lower salary in order to build a better team around him. Bit of an economic sacrifice for greater performance, longer legacy, ultimately reap the rewards economically.

But uh a lot of people have been saying uh you know you can't buy great missionary talent that will stick with you during the hard times and really advance the frontier of artificial intelligence. In the case of Brady, you know, he has these contracts for shoes, jerseys, footballs, other things.

He probably made a lot of money from that pay cut by staying more focal. So maybe for some of these AI companies, you know, we need more endorsements, a line of tennis balls or something. Uh, and then it'll be cheaper to hire talent because the outside market will pay them. We're not at that point yet.

Uh, maybe we'll get there. I think Yeah, I want a gameworn jersey. Here's what Illyu wore when he, you know, built the key parts of GPT4. Exactly.

Um, I mean, we've, we've been putting out these little trading card sports, uh, you know, when when any an AI research, whenever an AI researcher moves from one firm to the other, and uh, I mean, it seems like an insider joke, but there these posts are getting millions and millions of views.

like it really is turning into into sports in terms of the attention of the entire tech industry on these trade deals and maybe it's just the big numbers but I think it's also the fact that this is an important technology and people are uh excited about the developments and they and they want to know who who will win what what product will I be using in five years will it be open AI will it be meta who knows all of the above I hope probably how did you uh how would you try to value the Dutch East India company and the reason I bring that up is um uh uh Nvidia crossed a $4 trillion market cap.

Everybody's calling it, you know, one of the most valuable companies ever. But I think a lot of the people that run the kind of valuation analysis and inflation calculations and things like that believe that the Dutch East India Company was worth somewhere around, you know, 2x of that.

Uh have you spent much time um thinking about uh about that business? I look at all the numbers in general. I don't feel I can outg guess the markets. So the market is a better estimator there than I am. But look, markets are always wrong. That's why the price changes all the time.

So the fact that Nvidia is priced so high shows the market is paying attention to AI, which is good. But the fact that everything else is priced more or less normally, I think is evidence for my thesis of slow takeoff, that it will matter a lot over time, but not that much right away. Yeah, that makes sense.

Um, it is interesting something we we just had one of the founders of Core Wee on which is now a $70 billion public company and I was thinking about the advantage that he has as a public company where he doesn't have to raise capital from the private markets that are and and tell the story this like 10-year story.

he can kind of focus on the next quarter and just say like how do we deliver more compute to our customers?

And it's like this interesting change where a lot of the companies today, let's say you're a lab with a $10 billion valuation uh and you need to raise billions of dollars, you kind of need to tell the story around super intelligence.

Meanwhile, if you can just get out in the public markets and have billions of dollars of revenue, you can kind of go back to telling the story of, "Yeah, we're we're hoping to to bring more compute online, more data centers online, uh, and just actually focus on uh on the business. " Yeah, that can cut either way.

But look, if you're selling socks today, public market is fine. But as you well know, fewer and fewer companies want to go that route. So that does suggest to me the public markets are somewhat broken, overregulated, too much disclosure required, too much bureaucratization. Shareholders themselves can be kind of nutty.

So, you know, it's good to have the two compete, but I'd like to see us make the public markets easier to be in. What about Delaware? We saw Andre Horowitz moving to Nevada. Um, we've seen some other high-profile exits. Do you think Delaware will uh try to turn it around?

I know about a pretty meaningful amount of their uh their budget uh comes from uh franchise taxes, things like that. I say get out, send them a message. They're not feeling enough pain yet. Let's make it really tough for poor little Delaware. What do they call it? Small wonder.

Well, small wonder that they haven't gotten their regulatory apparatus into better order. Wow, that's a great thing. That's uh what uh what about in this I want to talk more about the public market distortions versus private market. It feels like Elon has kind of run an AB test with Tesla and SpaceX.

They're like within an order of magnitude of each other, maybe within like 2x in terms of valuation and now Elon seems to be building some sort of a Japanese style keretsu between all of his private market and sometimes the public market companies as well.

Um is like would would a hallmark of a well functioning public market be that there is just one ticker for Elon Inc. and all the companies are under one umbrella or is there a val is there value in the p in the public markets to having like pure plays?

I hear some public markets hedge fund investors are like I just want a pure play on the robo taxi. I just want a pure play on humanoids. And I don't know if that's actually like an economically valuable thing for our society to have or what would be best for Elon or the shareholders.

I'm just kind of struggling with all the different stakeholders. I suspect the pure play idea is useful. So you get a market price signal, what's working, what's not. So Elon obviously is become less popular. So the value of Tesla goes down. You see that quite directly.

If they're all bundled, the signal is far more muddled. So I think it's better separate. So ideally um if it was easy to operate in the public markets we would we would just see SpaceX and Neur and and companies would be going public once they hit like 5 billion or 10 billion something like that.

I'm not sure if something like Neurolink would be that's just weird to begin with non-leible. Yep. Who are the customers? What are the products? Yes, they're disabled but after that uh that I would think should be in private markets. Wait, are you sure about that?

because I feel like it's also in in many ways like a classic biotech drug, you know, FDA approvals. Once it's approved, it goes out to everyone.

It feels like that's one place that the public markets are particularly functioning as they always have where new drug comes goes through phase one, phase two trials, but goes public very early. We we're not seeing the stripe of biotech hang out in the private markets for two decades.

What am I ever going to buy from neuroch? No insult intended, by the way, but it's not transparent to me. I think it's amazing. I'm all for it. Yeah, maybe it has to be not so legible to public shareholders, and Elon wants to do it. That's great. But you might not buy a particular cancer drug.

I would hope that you'd never have to buy any of them. But that particular drug for some certain indication, some certain illness, could be taken through FDA trials. They could go public.

their stock price could trade on the basis of the FDA results and then when they when when they when they get approved they start selling that drug maybe they get acquired by a bigger drug company but then they live or die by the market size of that particular illness. Well, if it was a cancer drug, sure, I'd buy that.

But a brain computer interface, I don't know. I don't want one. I'm happy with it being there. And uh again, it seems to me that's probably not a good matter for public shareholders. Oh, okay. Yeah.

I'm more just think I mean yeah I I guess if you're thinking about valuing it as like brain computer interface for the masses that's a different story than uh it is a it is a treatment for the for the illness of being a a parapolgic or if you are blind it can cure blindness and that is a medical diagnosis and it is a medical product but obviously Elon's telling a much broader story about that so that's a different thing.

Were you surprised to see Uncle Sam put up a surplus in June or were you expecting that? I wouldn't say I expected it, but you know, month-to-month fluctuations can be so severe. Nothing like that should surprise us. The debt and deficit, they're still out of control. It's like a It's a startup.

It's a startup that that's like, yeah, we're we're we're profitable because they have one month free cash one month of free cash flow or one day annualized. Uh yeah. Yeah, that makes that makes sense. Um, how about stable coins? There's there's been new regulation. We have a a issuer that's that's now public.

Uh, there's a massive amount of excitement. It seems like a lot of the um use cases outside of um you know being you know uh as a um uh value transfer to facilitate investing in the US. But then the potentially, you know, more interesting use cases internationally.

But but how are you thinking about adoption and and do you believe that um the the US and specifically the traditional dollar will be a major beneficiary of the tech? It's been great. The rest of the world wants access to dollar-based systems. Our government makes it hard for them. Stable coins are a workaround.

I'm not sure what their long run fate will be because keep in mind regulations can change and other payments methods can innovate as well, but they're going to pass the Genius Act. It looks like I'm all for it.

It's not perfect legislation, but it will give people a chance to experiment with these things on a legal transparent basis and let it rip. Right. Why do you say that our government makes it hard?

I always thought it was an authoritarian government that or or or just a foreign government that wants the ability to inflate away their own debt and so they're keeping dollars out of their economy. They're trying to avoid dollarization. They don't always avoid it successfully.

But you look at the US government know your customer laws for instance. Uh it's a pain in the neck to deal with an American say who wants to open a bank account in Germany. I've tried to do this. That's because of the US government putting reporting requirements on the German bank. They don't want your money.

So a lot of the problem is USG, not you know Argentina, Russia, whatever. It's both but in fact if people just hold dollars that is usually tolerated and uh it's a workaround for both. Uh I want to talk about Google's acquisition of Windsurf. Uh Google paid a licensing agreement.

They acquired 50 or so engineers from Windsurf. The CEO and the share and the preferred shareholders got paid out by Google. Um roughly 400 or something employees were left in a ghost ship in the remain co. Uh they also had $150 million roughly on the balance sheet and so they were scooped up by Cognition.

Um, I'm not sure how closely you follow the drama, but my question was, is this a paro improvement or a Caldor Hicks improvement for Windsurf stakeholders as a whole? It raises the cost of capital, and we're going to need more covenants to limit this happening.

Like, say I start a company, raise money, and then just pay myself a huge dividend for all the money. Well, that's typically against the law because there's a covenant in the contract saying I can't do that. So, we need new covenants to cover these cases. Those will be harder to write, but the market will respond.

In the meantime, it's a problem. Do you believe that the the problem should be solved by employees demanding a new covenant, employees demanding more uh strict employment contracts? Where will this solution sit? It depends how you're doing the financing, but very often in other settings, it's done by the bond holders.

the bond holders want a covenant that you can't just drain all the resources out of the firm. The more the firm is based on human capital, the easier it becomes to do that because human capital leaving is not monitorable in the same way that a big cash dividend is monitorable. Uh but we'll figure out ways of doing it.

Maybe it will require arbitration, but this absolutely needs addressing. Otherwise, you just keep on starting up companies, ship the talent out somewhere else. In essence, sell it twice. Bunch of different people are screwed over for the next person who tries it with no intent of doing that.

The cost of capital is too high. And that's the the longer run problem. But again, markets are very good at solving problems like that with covenants. So, this is uh this will be solved by the market. This is not an FTC issue. I would not involve the FTC. I don't like them. I don't trust them.

their previous boss, their current boss. They're both bad news. Neither knows any economics. My goodness, they're a nightmare. Why would you call them on the phone? They're gonna make it worse. Well, thank you for the for the for the solid breakdown and being very clear about your opinion. This was fantastic, Tyler.

Thank you so much for coming on the show. Thank you for the update. My pleasure. And your wisdom. Good luck with the rest of the day. We appreciate it. Have a great day. We'll talk to you