Byrne Hobart on What The First AI Agents Will Do + China, Google Analysis

Yeah, so I think one of the things that we run into with a lot of these tools is that they can be kind of arbitrarily random in their outcomes, and the more thinking they're doing, just like people, the more quiet contemplation they have and the more chain of thought they do, the weirder the outcome can be. And so you want to deploy them in a useful way; you need some kind of real-world check on what's going on. And that applies to a lot of different areas of generative AI. You know, it's a lot easier to create synthetic data for code than it is for natural language because with code you can check and see if it runs and see if it produces the answers you expect, and with natural language it's just harder to check that. You need a person to actually read it and decide if it makes sense. And when I was talking to some people about just where do AI agents actually have that kind of feedback, and one of the things I realized was that if you look at tech support, you know, there are different levels of tech support, and it's basically the simplest kind is: this user did not actually read the documentation, so we will read the documentation to them. We will take their natural language question and answer it for them based on things that we wrote down in a place they could actually read them. And that has already been automated pretty quickly, and it's just a very natural use case for a generative AI tool where you dump the existing documentation into a RAG and then you query that, and then you're good, right? But there is that other category of help that is kind of similar but is basically: they're looking at the product, they're trying to use it, and they can't quite figure out how to do what they're trying to do, but it is probably possible to do within the application. And that's a case where a software agent that is just trying to use the application would be extremely helpful, especially if it does have comprehensive knowledge of what this application could do. So there's still room for human discretion and human judgment; the agent is not going to have perfect knowledge of what the user's actual intention is, and you may still need to step outside that context and really figure out what's going on.

I think it does help automate a different layer of what customer service people are doing, and it automates it in such a way that you can actually get a lot more data because if you find that a lot of your customer queries are asking to do the same thing, then you may find that there's a customer persona that needs to make this thing easier to do, or just that the application should make this thing more feasible to do. And you know, you could have your human customer service people take very detailed notes on every step of the workflow, but it could be like the agents are way more willing to just jot down notes on everything they do. And so once you've done that, you can actually see that as a kind of layer of intelligence gathering. And that gets to a broader question which we've talked about on The Riff and I've written about in The Diff, which is how quickly do customer service and product management roles start to emerge as a lot of the easy customer service stuff gets taken care of, and as a lot of the higher-touch customer service stuff actually becomes much more informative about how people use the product, what they think it should be able to do, and whether or not it can actually do those things.

Yeah, so this is a fun one. There was this sort of return on equity revolution in corporate finance, which I think stemmed from a lot of... So for a long time, companies were there was just not quite a metric they were thinking about, or was indirectly a metric they were thinking about but not something they were explicitly targeting. You had this state of empire-building companies in the 1960s where you'd have a high-growth company, they have a highly valued stock, they realize they can just mathematically increase their earnings per share. Like if they traded 30 times earnings, they use their stock to buy someone who trades at 10 times earnings, they've just increased their earnings per share. And so when they did that, they just kept acquiring slower and slower-growing companies because they're going for low multiple. And for a while they were high-growth companies that were acquiring low-multiple companies, and then when their own multiples came down because people realize, you know, this company used to be an electronics manufacturer and that was an amazing business, but now they're mostly doing steel and meat packing or whatever, and that's just not a hyper-growth business, so those stocks start to come down. And what we ended up with is this cohort of big companies where they had a lot of assets, they were in theory showing a lot of growth, but that growth was financially engineered, and now they're just not getting great returns on those assets. And so people started thinking a lot more about return on equity as just this measure of the quality of a business: you've put a dollar into this business, do you get back enough money to justify the uncertainty of the return? And when companies started doing that and started charging their subsidiaries for capital and ranking them based not on growth or size but on did you actually get a good return on the investments you made, that did encourage different business units to actually think pretty seriously about how they're using their capital and to be better stewards of capital. And that's all well and good in a world where companies are mostly funding their growth from their internally generated profits. But as markets have just gotten larger and more efficient and as we've built out this venture ecosystem, companies really don't have to do that. That is, there are businesses that do, but for businesses that grow, it is just very rare for your ideal growth rate as a business to be precisely the growth rate funded either. You have a very high-return business but it can't really scale, so think of like a service business that has one really exceptional service provider, like maybe it's a neighborhood barber shop, and the barber knows everybody, everybody's the barber's friend, so their cost of customer acquisition is low, they get really high returns, but when they open up a new location, it's just some new random guy who got here and just as good as Supercuts guy, but doesn't have quite the same brand for better or worse. Their marginal return on equity is low, so they have a high return on equity but they can't really grow at that return. And then you have other companies where their return right now is low, so think of the classic enterprise SaaS company where they spend a lot to develop the product and they spend a lot on a very long sales cycle. Once they get their customers, they're going to be getting revenue from those customers for a long time, that revenue is probably going to grow over time, so you almost want those companies to be losing money as fast as they can rationally manage. And I think it's Anshu from a16z talks about this with high-growth companies that it's not that they're totally insensitive to costs, it's that they're very, very sensitive to the unit economics of: here's what it takes to buy a customer, here's what the customer is worth, and how can we acquire those customers as fast as possible. So at that point, at any given time, those metrics are just a really messy snapshot of the ultimate economics of the business, and that company is going to be continuously raising money in order to expand as quickly as possible, until hopefully until they reach the point where that expansion is no longer justified, at which point their metrics are going to start drifting in a good direction because their existing customers are continuing to spend more money, they're not dumping so much money into marketing, they can kind of tamp down their R&D spending or at least not increase it in shins of revenue over time. So their accounting numbers start to look better, but their fundamental economics haven't really changed. So suddenly return on equity becomes this metric where it's really not an action-guiding metric except in cases where the company has a very low return on equity and is continuing to reinvest in the business. Like in that case, you say this company is just a money pit and they are continuously destroying value by, you know, if their cost of capital is say 9% and they can earn 5% on that capital, then every dollar they put back into the business is value-destructive and they should be trying to liquidate and return assets to their shareholders or should be trying to fix up their business. But with other companies, if their return on equity goes from say 20% down to 15%, that probably means that they have identified a set of opportunities that do earn 15%, which is good, and they're pursuing those as aggressively as possible, and they've saturated those opportunities. So you kind of want your companies to have their return on equity rapidly approach roughly their cost of capital, because that means they are deploying money to every opportunity that has a better return than their shareholders can get elsewhere, and that's kind of what you want. So it was a useful exercise to think through why did return on equity make sense, and it made sense in more of a capital-starved, information-starved world where companies fund a lot of their growth internally. And why is it that we can look at companies today and say like some of these companies have really high returns on equity, some have very low returns on equity, does it really correspond to our intuitive sense of the quality of the business? So if you look at someone like Apple, and this is another thing that is buybacks. So as companies mature and their revenue is locked in and very predictable, it is possible for them to borrow against that revenue and buy back stock. And it has for many companies historically been a pretty good decision. So you look at a chart of Apple shares outstanding, you'll see they've bought a lot of them back. If you look at a chart of their stock price overlaid against that, you'll see that they bought them back at pretty good prices, so it was the right call. But every time a company borrows money and buys back stock, it's reducing its equity but it's not really messing with the fundamentals of the business that much, and so it's artificially increasing the return on equity. But nobody looks at Apple and says they can definitely invest a marginal dollar in the business at the return on equity that they have on the historical investment; it's just an accounting artifact that the numbers look the way they do. And then you can take that even further and look at a company like TransDigm where technically they have negative equity; they have borrowed and bought back enough that in an accounting sense the company's value is less than zero, but of course the actual value of the company is the future cash flows and not the book value of the assets. So it's perfectly fine; nobody worries that TransDigm is going to run out of money just because they have a negative equity. And then, you know, rewinding a bit, you've mentioned that return on equity does help compare businesses across different sectors, and that is one of the great things about it because many, many businesses require capital, and even if they don't initially require capital, some of them will require it to grow. And you can sort of kind of in your head try to capitalize some of their growth costs. You would say, okay, this company spent a million dollars this year on marketing; we can treat that as an expense, and for tax purposes it is, and for GAAP accounting purposes it is, but really if they spent a million dollars to acquire customers who are going to last for the next five years, you should think of that as they spent 200k a year on marketing over the life of that customer, and we capitalize that and then we treat that as the asset and look at return on equity based on that, and we do get numbers that just let you basically ask what otherwise incoherent questions like: should we invest one more dollar in steel or banks or insurance or SaaS? Very hard to compare them, but if you do the accounting right and get the economic balance sheet measured, you can actually ask and answer that question in a useful way. There's this famous interview over 10 years ago now between Peter Thiel and Eric Schmidt where Peter Thiel is saying, hey, you guys are sitting on all this cash, it means you have no idea how to employ it, you don't have no idea what to spend it on. And you know, this was a time where he was most making his case of, you know, technological slowdown and using sort of corporate Google as an example of a company that has all the money and has invested in innovation and claims that it wants to solve all our problems but doesn't just doesn't is short of ideas of how to actually spend it.

I would look at tariffs from a macro and micro view. And the macro view is that tariffs are just affirmative action for American semiskilled manufacturing workers. Like there is a—you can just mess with the price points such that even if China is objectively more competitive in some category, if we put a high enough tax on things imported from China, then the US starts to be competitive and will do it instead. So it is indirectly a transfer from US consumers who are paying more for goods to US workers. And as with any kind of transfer payment like that, there is some deadweight loss, like those workers might be doing something the US specializes in more, they might be doing nothing, maybe they would impose some externalities if they didn't have jobs. But that is roughly what we're doing, and that to me is just—it's an economic question as to what is the cost, and then it's a values question as to is it better for this country to have a slightly higher employment rate and a little bit more manufacturing that is not really a competitive advantage but does keep people busy, give them a job, and you know, make them feel like they're contributing and maybe make them actually contribute materially to the standard of living in this country. Is there a fair trade-off between that and having a slightly lower standard? And that's the question you ask about any kind of transfer program because any kind of transfer is going to create different incentives for working versus not working or working in one place versus another thing. And we're always just asking about the trade-offs. And then I think, you know, most countries that can do it have some form of welfare state, and it's probably just better for social cohesion if the welfare state does involve actually working. And if you have tariffs, you are imposing some market discipline on that work. There are places that have historically had welfare programs where there was nominally a work requirement, and there's a book—I forget the title—but it's William F. Buckley's book on running for the mayor of New York City, which was not really a serious campaign. When he was asked in his first press conference after he announced he was running, he was asked what he would do, what's his first act if he wins, and he said the first thing he would do is demand a recount, and he definitely did not win. But one of the things he pointed out was that at the time in 1960s New York had a law on the books that if you receive welfare, you have to work for the city. And you notice that that law was just not enforced at all, but even if you enforce it, it's probably much more expensive to hire supervisory people and figure out, okay, what can we get someone to do that is like a fake job but has some kind of actual benefit. And everyone involved in this would know this is fake, we're pretending to work, and it doesn't really matter if we accomplish anything or not, we are getting the money either way. And that is, I think, a lot more demoralizing than just getting a check from the government every so often. But if you have a tariff such that there's a job that really only exists in the US because we are keeping more cost-competitive countries out, it is a job and someone can actually look at your profit, they can look at their profit and loss and decide whether or not to hire someone, and they can say that even if this job only exists because of these artificial constraints we've set up, you can be good or bad at that job and we can fire you for doing a bad job. And so it's more of a real job and less of just a pure make-work program. You know, every society sets up certain sets of rules that make certain things a job or not a job, and in some societies, drug dealer is just a non-existent profession; in some societies, it provides goods and services, and then there are externalities that we decide whether or not they're worth it. So we're always deciding in some implicit way what counts as a job, what counts as a job that we actually want to count in our system as a legitimate job. And I think of all the ways you could make that call, tariffs are probably one of the more pro-social ones. But it is going to be annoying for those of us who are not really in a pool of potential workers in American low-value-added manufacturing; we will just pay more for stuff. But we were already going to pay more for stuff in order to support a social safety net anyway, it's just a question of exactly how you arrange that. So that's the micro piece. The micro piece is that yeah, one of the ways you can use tariffs is you have an industry where it would work at scale, it won't get to scale unless it can generate some revenue now, but it has no hope of competing with foreign competitors who are already at scale. And that is the story of a number of countries' industrial development, is that they see what industries work in different places, they see if they can replicate those industries at home in theory, and then they put up enough trade barriers that they can actually replicate them in practice, and then they start to take down those trade barriers as the industry gets up to scale. So this is part of how Japan ended up with a steel industry and then how Korea ended up with a steel industry, is that it was supported and protected by the state until it got to the point where it was actually cost-competitive with the rest of the world. And at that point, it's basically paying back the benefits it got through the positive externalities of one, it's a domestic manufacturing business, it employs people, it produces profits, etc., and two, it's just useful for countries to control resources like that and be able to produce them domestically. Now, I don't think there's a very serious risk in, you know, the next 10 or 20 years that the US will just not have access to steel and really need it, but I don't think the risk is zero. And so I think it is useful to think about what strategic products we could hypothetically lose access to, what is the lead time for making sure that we can produce them in sufficient quantities at home, and then yeah, potentially taking the economic hit in order to do so. So you can think of it as state-sponsored equity or venture capital where you have this investment that only makes sense if you write a officially large check and does actually produce a return once you write that check, and maybe it's like ESG-flavored for a different set of E, S, and G, but it is like an ESG-flavored investment where you're saying this is not the thing we would invest in to get the maximum possible return, but between the financial return being better than zero and the social return being positive enough to make up the gap between that and the next best objective financial outcome, we can still agree that it's worth doing.

I think the Chinese internet industry is actually extremely impressive, and from what I understand, Baidu was pretty far ahead. There's an interesting kind of accidental linguistic protectionism that really helped earlier search engines. It's less relevant now because a lot of language models can translate pretty seamlessly across very different languages, but there was a time where if you have a search algorithm that works just fine in English, it's just not going to handle Mandarin especially impressively. And this was also true with Russian, where Yandex was actually better at handling some of the nuances of the Russian language, and so it was for a long time a competitive search compared to Google, not because Russia was really trying to support their domestic champion—that wasn't the main cause, although they certainly cared about that after a while—but really because you had to choose which language you wanted to go after, and of course the Russian-speaking engineers who founded Yandex, they picked Russia. So you do have some just independent development of different search engines in different places, and Baidu did a really good job. And then yeah, after a while, there was also a protectionist element to this where Baidu is the local company; if Google invests enough resources, they can win in China, the Chinese government feels like that is actually a policy decision, it's not just a natural outcome of these totally unstoppable forces, it's something that someone decides whether or not should happen, and they decided it shouldn't really happen more or less. Or I guess there's a near possible world where they decided that a Google that is as compliant with the CCP's wishes as Baidu is can be tolerated, but I think even that it presents a risk to them because that does mean that they need Google to always feel that way, and it's just very hard to get rid of a search engine that people are using more than eating one. To get rid of it on political grounds, just look at various efforts to dislodge Google, they don't work very well. Even when you tell people, you know, if they're buying a new Android in Europe, they get to choose whatever search engine they want, well there's one search engine that almost all of them are using and almost all of them prefer, and that's the one you choose that makes sense somewhat relatively. Other segue, you wrote about this idea, what did you call it, one big-shot approach to wealth creation, where you basically talk about how for some countries a lot of their sort of growth comes from specific technological development. Unpack this idea.

So the idea was not just—so there's a lot of national wealth does come from the deployment of these big general-purpose technologies, but a lot of the individual wealth accumulation is actually adjacent to those new technologies. So if you look at the railroads, for example, there were people who made money by starting a railroad and just slowly expanding it, but there's a lot more money made from promoting railroads or taking enough control of a railroad—this is 19th century—like taking enough control of a railroad that you can manipulate the stock and trade on the things that you were doing to make the stock price go up and down. Like that's where a lot of the big money was. And then in railroad-adjacent physical industries, there were, you know, steel fortunes that were made from selling steel initially to finance the railroad buildout and then shipping steel on railroads in order to...

Finance in order to supply things like the development of skyscrapers and steel in everything. Similarly, oil benefited a lot from there being a railroad network where Standard Oil was a very big customer for the railroads, was able to get very, very good pricing treatment from them. So the actual core business often doesn't create as much wealth as the consequences of that business existing. And if you look at the automotive sector, you do see something similar where yes, Henry Ford made a ton of money, but if you look at the aggregate wealth that was made from the rise of suburbia, increase in demand for oil from cars, the rise of suburban-centric business models like grocery stores that just swamps the money that was directly made from owning stock in a company that made cars.

And I think even if you look at more recently the world of software, there were a lot of people who — there's a lot more money made from noticing that chips get better and better and therefore we can make software as a complement to this thing whose price is headed toward zero. A lot more money was made from that observation than from actually making the chips better, even though again some decent-sized fortunes were made from that. And Jensen is doing just fine right now, even in Nvidia's case — part of their big part of their moat is the software side of things.

So you still — like if you find some really, really big trend where some kind of technology is improving rapidly, declining in cost, you think it will be deployed everywhere, the tempting thing to do is just bet directly on that, and then the wealth-maximizing thing to do is bet on the indirect consequences of that and just harvest the benefit. In fact, in that piece, one of the claims I made, which I don't take entirely seriously but is fun to think about, is that the success of value investing in the last half of the 20th century was partly a bet on cheaper information transmission — like cheaper communications, cheaper information processing. Where if you're Warren Buffett and it's 1951, you have to drive to the state insurance office to get the financials from these companies, and then you have to put an ad in the paper saying 'I will buy your shares for $30' whatever, and then you have to go door to door and find people who bought the stock 20 years ago and still have it, and buy it from them for cash. So it's a lot of work.

But now you can open up a stock screener and find everything that looks even slightly cheap and buy it, and it's very, very straightforward — you don't even have to pick up a phone and call a broker. So the drop in the cost of information was a subsidy to anyone, any investor who was willing to do extra work to get that information first. And that was like a one-time subsidy where markets were very inefficient because data was scarce, and as they got more efficient, the people who were exploiting those inefficiencies made this one-time gain from everything snapping into the right valuations. And now it's just a lot harder to repeat that.

So I would say there are two pieces of that. One is that if technology has a very low marginal cost, like software or communications, then there will be a period where it's being deployed more and it's being used more, but it's underpriced. And so at that point it doesn't contribute as much directly to GDP as it does indirectly through people using it in order to do more real-world measurable economic activities. And then over time it starts contributing more to GDP because it's fully deployed — we've saturated all the customers who could use it, and now our growth comes from getting more money out of those customers. So that is GDP-accretive.

So I think you may see something where software was deflationary until everyone had a Facebook account and an Instagram account and everyone used Google all the time and everyone had an iPhone, etc., and you know, everyone's subscribing to all these things. And then at that point the subscriptions keep rising in price and the ad loads creeping up and the prices for those ads keep rising, and so software actually becomes inflationary because the product isn't improving all that fast but the cost of that product — the economic activity that's driven through it — is rising. So I think that's one part of it.

Another part of it is that sometimes technologies will make a contribution that actually shows up on the input side rather than the output side. So there was a really interesting piece on the Great Stagnation that was written in early 2020 that talked about aspirin as an example of this, and it pointed out that aspirin, I think objectively, just increases productivity in the sense of you have the same set of people, the same set of capital equipment, but it is less likely that a headache means you can't work that day.

But when you're compiling macro statistics, you would look at things like number of days people work or percentage of sick days people take, etc., when you're figuring out what the labor input is. So if the net effect of aspirin is everyone has one more work day because they're not miserable with a headache on average one day here, then what you actually get is a roughly 2% increase in total labor input given the same set of laborers. So you have an increase in efficiency, you definitely have an increase in just human happiness — the headache goes away — but you don't actually see that as an increase in productivity. You see that as an increase in hours worked.

Aaron Levie, the CEO of Box, was on the More or Less podcast with Sam Lesson, and Sam basically tried to get Aaron to accept the idea or the possibility that AI might be deflationary in that there's no natural guarantee that it will lead to more usage. And so maybe if it just makes it cheaper, maybe there's actually less profit to go around. And Aaron was saying, 'Hey, there's no way we're at peak software,' you know, there's like making things cheaper and easier and just introducing new workflows will most certainly lead to more usage and thus more profits to be generated.

So I think there's a J curve with respect to revenue, and the inflationary impact is really tricky. So the J curve for revenue is: as you're deploying, say, AI customer service agents, and as you are, I don't know, replacing your direct-to-Kindle cover illustrator with just a call to Midjourney or DALL-E or whatever, you are reducing costs and you're eliminating some jobs, and so that's deflationary.

And then over time, my strong suspicion is that the vast majority of consumer-facing AI products will be primarily monetized through ads, and that those ads will just get better and better because the AI can constantly steer the conversation it's having in more and more commercial directions. It's collecting more data because you're just interacting with it more — you're giving it more tokens of data and just giving it more feedback. And meanwhile tech companies are just getting better at collecting data all the time, getting larger and larger samples. So over time they do generate more revenue, probably drive more purchases through these chatbots or, you know, through AI-generated search results, things like that.

But you have the question of: are those incremental purchases actually inflationary? And a lot of the things that consumers buy are just not very capacity-constrained. Like we can make more iPhones without the price of iPhones going up, and, you know, make more of just a lot of electronics, furniture, clothes, lots of consumable — like lots of consumer durables, lots of consumable products without an increase in costs. So probably — and in fact we can probably get some economies of scale. So it's entirely possible that it ends up being just accidentally deflationary in the sense of we have a lot more control over spikes and declines in demand because we have these tools that are very, very good at targeting ads.

And if there is something that's been overproduced and has a little bit of a glut of it, and so it's taking up very expensive warehouse space and we want to get rid of it, you can probably drive a lot more consumer interest towards that thing because your chatbot will steer you towards the idea of buying this thing. And so maybe that actually reduces the overall cost of manufacturing goods across any category where they can be marketed to consumers and ends up being net deflationary.

But it's also one of those cases where a world where these tools are fully deployed, where that is in effect, is a world that is sufficiently different from the current one that there's just going to be a lot of other uncertain questions. And there's also the question of: does AI end up being inflationary because it can convert electricity into economic activity very efficiently and very — maybe linearly or even super-linearly — and so we would actually see deflation across the manufacturing stage of the supply chain and across the transportation stage and storage stage and all that stuff, but then inflation in the raw material inputs, because some of those inputs are energy sources and we'll use a lot more of those.

But then you have the other open question of how much energy efficiency can we drive, how much solar will we deploy if we have just radically higher expectations for power demand, how much — you know, how cheap does nuclear get if we treat it as something that should be mass-produced rather than something where every power plant is this bespoke thing that has to be designed and redesigned over and over again at great cost and you can't break ground till many, many years of debate and discussion and creating giant reports have passed, etc. So lots and lots of uncertainty there, but that's how I think about the direct impact of AI on just macro variables is J curve, where at first it's destroying some demand but that's offset by increasing demand because you have to actually build the data centers, you have to actually hire people, you have to spend a lot of money on marketing to get these tools in front of consumers.

And then at some point, as they start monetizing more by driving purchases rather than through subscriptions or some other revenue source, they start to just increase the overall level of economic activity.