How AI is Changing the Way We Build Companies (ft. Guy Wuollet and Noah Citron)

Transcript

For the first time, it's easier to think about token spend more like headcount. The question no longer becomes how much work can we do, but rather how much money can we allocate to the token budget in order to get work done. >> The absolute numbers have gone from maybe a few thousand a month to a few hundred thousand a month and in theory could go to a few million a month. And so how we think about that is super interesting. People have speculated since the AI started getting good that we're going to see these like billion dollar companies run by like one guy bunch of agents. [music] Maybe we'll see variations of that where we see much leaner companies. We also might see businesses that are like much more horizontal. >> If everyone has access to these models, this will be an incredibly [music] democratizing force where you basically

slam shut this dispersion in IQ. And it probably becomes much [music] more important not to be particularly smart, but to be particularly gritty or agentic or to have strong desires for things. Hello and welcome to the A6Z Crypto Show. I'm Robert Hackett, an editor here at A16Z Crypto. And I'm here with two of my colleagues, Guy Wlette, GP at A16Z Crypto, and Noah Citron, head of engineering. And today we're going to talk about AI and the way that it's changing business. So, I'd like to toss it to you, Guy. You know, everybody sort of understands that AI makes things more efficient. Maybe it makes people more productive. It makes things cheaper.

That's sort of like the basic surface level understanding, the the debate that's going on right now. But you're thinking also about how the nature of the company might change in this new world. Tell us a little bit about that. >> Well, Noah and I were having a discussion yesterday which was a lot of the inspiration for this and and we started by talking about uh how many tokens each person on his team on the engineering team are consuming. Uh which I think was the a good jumping off point for basically discussion on which sort of work is productive and how to think about uh like firm formation. you know, we as a venture capital firm want to invest large amounts of money in companies that need a large amount of money and will then, you know, hopefully be fantastically successful and uh

occasionally will be abject failures. I think there's a whole new opportunity or space in the market for a sort of company uh that needs variable headcount that can be run in a much more asset and headcount like fashion today because of AI models. It does not need the equivalent of something like venture capital is in fact not a particularly good fit for it because the terminal outcome the success case is not a big enough success for most venture outcomes to to make sense. And so we're talking a little bit about the financing models for that. Is it something like convertible debt? Uh do these end up looking a bit like a modernized form of private equity or a bit like a builder form of pod shops as opposed to a quant and and a trading form of that? But I

think an interesting place to start is just by asking the question of for a long time it was hard to spend a lot of money on AI inference on models. It has become much easier to spend a lot of money on AI inference and we were talking a bit about that shift for for your team and essentially how much everyone is spending which I I I thought was like an interesting number [laughter] and that's because a lot of the leading model there are newer models being released that allow you to use a lot more computation. and they're a lot more computationheavy and so they are a lot more expensive but they do a lot more and there's a lot more that uh complexity they can handle. I think the fundamental change was that, you know, a year ago or even less than a year ago,

six months ago, we really didn't even have to think much about token spend mostly because our ability to scale tokens was limited. So, we were able to, you know, spend, you know, almost within our subscription limits for the most part. Uh maybe going a bit into the API pricing, but you know, not really a significant spend. But in the past maybe 3 months, we've noticed that, you know, our ability to get more work done has basically started to get more linear or at least linear for a little bit longer with uh the the amount of money we're willing to burn on tokens. So now the question no longer becomes how much work can we do, but rather you know how much money can we allocate to the token budget in order to get work done.

>> And you know this is this kind of changes the paradigm a lot. So I I think for the first time like it's easier to think about token spend uh more like headcount, right? where an engineering organization can scale headcount. You know, uh assuming you have like unlimited capital, you can continue to scale headcount and barring any organizational inefficiencies, you can just keep uh keep scaling that and get more work. And then you have to, you know, answer the question of how much work to do is valuable, right? For like our given business units like you know where can we allocate allocate resources to actually increase revenue. This was not the way we used to think about it but now it is absolutely the way we think about it. So, you know, we see, you know, for the first time, like I I I, you know, talk to my team about how

much uh they spend on tokens and every once in a while I I hear, you know, on some given week where there was some some heavy work being done and I go, "Oh, that that could cause a problem." I had never been put in that situation before. Um so, you know, now we're trying to grapple with you know, both how do we, you know, manage those costs, but also how do we decide how much we should spend when when is it too much? Yeah. I I think there's a very interesting concept here which uh like the concept of a P&L mostly comes from finance that someone is like running a book of business and they have a profit and loss statement and it seems like that that was never a concept for an engineering team like even if you were an engineering manager you didn't run a P&L

you like got a certain amount of headcount and you were asked to do a certain amount of things and now in many senses if you are the one controlling token spend like you are running some form of P&L you have at least a loss in the form of the tokens that you're spending and and maybe there's a way to attribute ute the work being done by the token spend to some you know profit or income stream for for the business itself and I think like that shift specifically is very interesting uh there's been a whole discussion about return on token like if I'm spending this much in tokens what am I getting back and I think that's been because exactly as you as you were saying it was very hard to spend a lot of money on models so no one really cared about the

return on token or the return on investment now people do because the absolute numbers have gone from maybe a few thousand a month to a few hundred thousand a month and in theory it could go to a few million a month. And so how we think about that is is super interesting. And it also then like implies uh we don't do a particularly good job I think like in the economy broadly quantifying the ROI of each individual person's contribution or or labor to like the broader organization. I think in in most tech companies it's very hard to AB test like what if we had this one person and we took them out of the org for a week. How would that impact the overall P&L and the overall productivity of of the company? So many tech businesses are fundamentally network effects businesses where you get

to a certain scale and you almost kind of can't screw it up as opposed to more traditional financial businesses where they often explicitly like require somebody to take a few weeks off to see how their book and their P&L is affected. Often time this is done for you know regulatory reasons to see if they're marking their book correctly. But uh it has like super interesting implications for basically AB testing the quality of token spend or AB testing kind of the quality of of work so to speak. And so I think this is obviously interesting in the context of an engineering team trying to build something but also very interesting in the context of uh like these sort of micro businesses where maybe it's a sole proprietorship or a small team building a company that they entirely own largely a software product but not going to be a

venture scale outcome like how do you finance that token spend I think is a really interesting opportunity for for crypto and for blockchains if you're able to underwrite like a specific return on token and then make microl loans or you know uh sort small convertible equity or or debt investments like that I think is a super interesting space is probably a very interesting like emerging manager space from an investment perspective is very much not our focus you know we have a sizable fund and want to make large investments in companies that will either conquer the world or or ultimately not succeed unfortunately but I think there's like a whole sort of liinal space for uh lots of smaller businesses and and this is a lot of what we were talking about yesterday >> you're right in like the description of a finance company where you know a team

might have a book and have some some P&L and then you can like pretty easily quantify how how good of a job they do. But then when I I look at how we do on on our team, you know, I kind of have a reasonable sense for what our token spend is on, for example, a given week or given month, talk to everyone on my team and see, you know, what they've been doing and can make kind of the the qualitative assessment of, you know, was that spend worth it? >> Unfortunately, like this is for a large swap business is probably going to be the way that it does work. Um I I I think like the the way I I would kind of segment these two things is there is you know the category of businesses where you can kind of look at your token spend

like headcount >> and for those category of businesses you you kind of have a very fuzzy sense of return on token you you have to do what I do but then there's a separate category of businesses where uh you don't look at token spend like headcount but you look at it as like um some some other kind of cost in your business unit that you scale up or down quite rapidly. >> And in that case, you know, I I think like you said, there there's probably like a more limited subset of businesses where that makes sense for, but you can do uh some interesting things. And I think some of the things we were talking about is, you know, are there like subsets of, you know, businesses that look more like the headcount business where you can actually like start to try and quantify return on token for like

more limited uh subs. So uh you know, example we had talked about was like an AB test. So you know even though we may have like for example some network effect business like a social media company uh these businesses love to to to run AB tests and you know their their goal is you know we will try two different features and just measure you know how does retention return get affected by that and uh you know now we can instead basically point tokens at something and say okay let's uh let's allocate a token budget for you know AB tests and you know we'll build some potentially like very autonomous uh agent that is just constantly making hypothesis of like okay what what can we AB test next and implement it deploy it you know potentially without even a

human in the loop and then measure you know how how do your metric change how does your retention rate how does your turn change uh in those different AB tests but also now the the additional unit you get is you know how much did it actually cost you to implement that feature and we can start running those pipelines and just you know continually me measure is this single pipeline actually going to be continuously improving the business And I I suspect this is maybe a good way for us to kind of pipe in this, you know, more P&L like analysis into a business that like looks more like like a business where you you view your engineering work as as headcount.

>> Yep. There's sort of like two categories here of being legible to tokens or like legible to capital as a proxy via tokens. One of which is the classic like CAC LTV concept of if I put dollars in, I get dollars out. And I I think a lot of for acquisition cost and long-term value. >> Yeah. Uh essentially like if I acquire a customer, this is how much it costs me and for as long as they're using my product, this is my payback period. So, uh in that sense, from the the investing perspective, you kind of have like a implicit irra on on that. And a lot of classic web 2 businesses that did very well had like wonderful CAC to LTV ratios and you could, you know, throw a ton of cash at them and so they raised a lot of money and they they grew very

quickly. I think that's like the classic notion of being sort of legible to capital from a product sense. I I think the extent to which you're now like legible to tokens would be the ability for the product itself to change and the like CAC to LTV ratio to improve with the number of tokens you're spending on improving the product. uh and so in some sense you've made this even more quantitative and I wonder uh the extent to which this is just like a whole new domain of of statistics like in in some sense astrophysics and sports analytics and like quant trading are all the same job and I wonder the extent to which software engineering is now just like

another domain of the quant [laughter] I think if that's true there are lots of interesting implications both for you know very big firms and and for for small firms uh especially for smaller firms. I think maybe you now have this kind of invention of the software engineering pod shop where you do have small engineering teams that like do actually own a P&L for a specific product and get a slice of uh the finance way of saying this would be like get some balance sheet where you get money that you can spend on tokens to improving your product trying to improve your cact and then you know put putting money into that itself. I feel like you you did not like this idea or this analogy of like the software engineering pod shop. So tell me where that breaks breaks down and where that's wrong.

>> Also just break down what is a pod shop for people who aren't familiar. It's a concept borrowed from the hedge fund world. Yes. Maybe you could just elaborate briefly on that. >> I think in finance there's kind of this classic evolution of the principle or the the like person that starts a firm is very good at one thing. So maybe you're like a good trader to start with and you trade a small set of assets and you like move to trading a wider set of assets and then you sort of graduate not from doing the investing but to picking other people that can do the investing and this has been the path of of many successful managers and and hedge funds in the last couple of decades. You sort of like start with a game, you move to a specific, this is like the classic, you

know, poker trading manager selection. Like now I run a pod shop. And the notion of a pod shop is basically there's one overarching fund that picks a bunch of portfolio managers or you know individual investors who each run their own teams and they try to come up with alpha or signals essentially uh that can be traded by the firm overall. And the compensation model is a bit different from classic 2 and 20 investing where the pod shop or the PM will uh will get in many cases something like half of the returns from the the alpha the signals they they generate.

You know the overarching firm or team or company gives you often data supporting infrastructure uh a talent pipeline oftent times mentoring and and sort of help and training uh but also importantly a balance sheet. They give you money to to trade. And one of the questions is if I have $100 and I have five PMs each running their own pod like who gets how much of that that dollar? Uh and this is a very simple straightforward obvious question when there's like one manager uh either running a software engineering or like one manager running a P&L in a trading context. This is like then its own sort of meta domain of investing or trading or engineering where you have to allocate resources amongst people that can use them effectively or not. If you

believe that building a software engineering product is becoming more quantitative and is software engineering products that are legible to tokens uh will do very very well then that that seems to be a super important skill that regardless of if you're at a big company or a small company in the same way you used to get allocated a headcount budget and your headcount budget was relatively static. That's not true with tokens. You can like drastically increase your token budget or reduce it overnight. Yeah, it's sort of in this direction that that's why I think there's an interesting analogy between something like a pod shop and uh a software engineering org if you know you sort of believe that a lot of the spend in a software engineering org will go to tokens and engineers can be you know

significantly augmented and improved with the AI models which I feel like is just true and everyone takes for granted but maybe someone will disagree with me. So the pod shop model is effectively you put a bunch of teams on the field, you say go win by whatever means possible and then the ones who do win you give more resources toward and it's a little bit evolutionary kind of Darwinian in that way. >> Yeah. I I think there's specifically the interesting uh pattern that someone starts with like a a lower stakes game like they master chess or poker and then they move into a bit more of an amorphous structure like trading and they become very good at that and then they sort of graduate to not doing the thing themselves but selecting other

people that are very good at doing the thing. And I do wonder the extent to which being an excellent software engineer or software engineering manager will in the future be more of a talent business where you are really good at selecting people that are good at spending tokens well uh as opposed to uh spending tokens well yourself or even writing code yourself. >> It reminds me very much of the same reckoning that came for the advertising industry. There's that famous line, you know, I know I'm wasting half the money I spend. I just don't know which half. But then you got digital advertising and it became very clear, you know, you have you can uh target hyperarget certain demographics and actually see um how much return you're getting on your spend. It seems like it's that same kind of concept now being applied to

engineering and tech teams. Yeah. But I I do think I have maybe some comments and push back on the pod shop analogy is that um I I I think one is that depending on on the business it's very hard to you know you may be able to measure like your kind of your gross P&L but it's not necessarily easy to understand how much each kind of pod would contribute to it because you know unlike a pod chat where it's you know very obvious like everyone gets a book they run their own trading strategy and then they get a P&L from that. when you kind of centralize it more to being that you have each of the pods but are all you know working on different parts of the same overarching product you know what does each one contribute to the to the P&L and how do you measure that so I

think there are maybe you know businesses where or engineering organizations where each pod can actually be a separate business unit in which case this makes a lot of sense right like you can imagine like almost doing like a AI engineering PE fund right where you're you know buying or building and scaling a bunch of like small normal businesses and you know each pod is is its own business in which case like measuring the P&L is really easy and then you you can do this thing where you kind of direct the the token funds in that direction but if you're instead saying you know I have one product that has you know like one engine that generates revenue but there's you know 100 different pieces of that product that work together to keep that uh that revenue generation engine healthy you know it's it's much harder

to actually measure that which makes the allocation problem harder and I I think this is why like It keeps going back to the headcount problem, right? like where you know I I I think that uh engineering organizations already have this like similar dynamic to pod shops where you know you have a bunch of engineering managers in your organization and you know you need to decide you know how big each of their teams are going to be and what their budget's going to look like and you know we qualitatively look at the performance and decide okay you know this guy's great so he's gonna be promoted or or we're going to allow to hire more people or give him a higher token budget right I think most businesses are going to be stuck doing it that way potentially The the more interesting thing is not um the

legibility of return on token because I I don't think like the existence of AI makes return on token that much more legible but I do think it allows us to scale up and scale down our token usage uh much more aggressively which which may in turn cause us to be able to to kind of have slightly better uh analysis on our return of token because we can do these things like AB tests where you know we can like very briefly be like all right we're going to allocate this guy token see like what he does with it which you can't necessarily do with headc counts without you know potentially breaking organization. >> Yep. Yeah. I I think there are probably lots of businesses that have the natural uh requirement or path of being like five people and then want like needing or wanting to be 100 people and then

like once again needing or wanting to be five people. That specifically probably influences a lot how you think about the shape of a business and also how you think about financing a business. Lots of people raise money because they need to higher headcount and you're sort of committing to a specific headcount for a long period of time and the there's no idea that like that headcount will reduce after we've built something especially in a venture context like a startup is momentum where it just kind of keeps getting bigger and the ball keeps rolling and that that is very much the idea but I think there's maybe a different sort of company that uh needs for a brief period of time a lot of money to spend on tokens and then builds a very repeatable retentive engine that

you can, you know, put money into in the CAC LTV sense and receive money out of. [laughter] And I could totally see there being either individual companies like this or, you know, holding companies, fund structures to build a bunch of these different things. And that seems like a skill set that's probably translatable across these these businesses. So I I think that itself is a very interesting opportunity for you know uh new funds, new managers, new new investments and I also think it like maps perfectly to what we see working on Jane today which is the ability to do very granular smallcale either loans or investments uh in a programmatic sense to you know companies or agents that uh are doing

something relatively deterministically. is very much like not a Wall Street model and it's very much like not a classic Silicon Valley venture capital model and I I think that specifically is is super interesting to to me today. >> I think this brings up an interesting point though that maybe the better analogy to use on you know how tokens affects business is not to look at pod shops but to look at consulting businesses right because you know you you talk about how you know some businesses you know might need a bunch of energy devoted into one area for a period of time and then need to scale it down. This is of course like really hard to do organizationally which is you know people argue all the time about why is it that businesses hire consultants and you know it's a big big question but I I

think uh one of the like standard answers is like I have this like very particular problem that I need to solve and I I need to solve this problem right now and I don't and it's a onetime problem right and I'll I'll throw the consultants at it right and I can said say like maybe I'll just throw the AI at it. I think there are like two reasons that management consultants really exists often and I'm sure there are things outside of this. Like one of them is justification. Like the core team already knows they want to do something, but they need some in theory third party unbiased, you know, entity to confirm what they already want to do and they're willing to pay a very large amount of money for that to to exist. It's like, you [snorts] know, I'm actually willing to pay a lot for consensus that we get

everyone in the company to agree to do this. Sort of like >> a highly paid scapegoat. >> Yeah. 100%. like the there's a whole rant on Renee Gerard in there. There's a whole rant on like monarchy that we'd rather have a bad king as long as everyone agrees that he's the king. So, that's actually super valuable. And then I I think there's also this like management consulting idea of of sort of talent laundering where if you have a very boring business that no super highly credentialed person would want to go work at directly out of college, but everyone wants to go work at McKenzie because that's very high status. So those people are willing to take a high status job that actually probably pays them a little bit less than going to work for the low status company directly. And it's essentially like a recruiting arbitrage. I I think if you

live in a world now of being able to use the models directly, you like kind of don't need that recruiting arbitrage. it it's much more interesting to just like go build the business yourself coming out of some uh you know incubator call it whether it's a a college or or another company that's known for developing talent. It's like one of the classic things like oh we hire a consultant to help us justify why we need to do a layoff. But if you ask a consultant why you know what what their job is it it is not that cynical answer. It's more aligned with the the answer I gave. Right? So regardless of why I I don't I actually don't think that's a cynical answer at all. I I think like one of the hardest things to do is to get people to agree. And I I would

actually argue like if the management consultant literally doesn't do anything other than getting everyone to agree on what was probably the right path anyway, that's an incredibly valuable service. So I I I would like push back on this being cynical. But I I do agree there are some like consulting services where it's like yeah, we need a point solution for a thing for a time and we hire a consultant and that works very well. >> Yeah. So let me rephrase that then rather there there is a a different version of consulting which maybe some consultants do um which is the the boring answer of like I have a problem I need to solve so I'm going to hire these people to temporarily solve it. They're you know kind of contractors.

>> Yeah. Well I think there's also tax are in there like depending on how these people are are they employees are they not employees like you know that's also certainly a part of this. But if you view it as like a valid thing that businesses need is like the temporary workforce which maybe consultants provide, maybe contractors provide, you know, a new entrance into that that space is that that's something that tokens provide. >> Yep. You were trying to make this point for like 10 minutes and I missed it which is that like >> you know the consultants are variable labor and tokens are variable labor and so now we can like scale consulting.

Actually maybe the entire world is now consulting. >> Exactly. And you know consultants have the maybe one problem that if you look at the archetype for a consultant it's like all kind of like this like finance managerial style of person. So it means that when you hire a consultant there is a a pretty like small slice of like you know these point problems that you could attack with them you know but uh when it comes to AI that that slice is much larger right because it's it's much more general intelligence. So now we can maybe expand actually like the TAM of these like consulting style businesses to being like it's not clear you would ever hire McKenzie because you have like an engineering project that you need to get done. Maybe you'd hire hire a dev shop.

>> I mean I think management consulting is a lot of what I was saying and like systems integrators are a lot of what you were saying and we call them >> but the the it's kind of the same problem in the case of the the variable labor case rather than the you know scapegoat case. >> Yeah 100%. I don't think tokens help you get to agreement at all. uh other than maybe like people are willing to outsource their thinking to a model like you know I ask >> I think saying that the machine god says do this is actually maybe a valid way to get some consensus. >> Yeah that's yeah fair that's maybe a cultural or like a social question as to how quickly are people just willing to outsource their thinking to one of these models >> apparently quite quickly.

>> Yes. Yeah, [snorts] that does seem to be empirically true. I I would have thought it would take longer. [laughter] >> So [snorts] I mean Noah with what you're describing like this whole world becomes consulting. Are you are you seeing an opportunity for new businesses to arise in this domain or are you just saying that like the AI models as they exist today eat those businesses. It's possible that there's new business or new ways businesses get run. uh you know people have speculated for you know since the AI started getting good that we're going to see you know at least like billion dollar companies run by like one guy with a bunch of agents maybe we'll see that but maybe maybe we'll see like variations of that where we see much leaner companies we also might see businesses that are like much

more horizontal because uh you know something I noticed in my own engineering work is that it's no longer efficient to do things uh particularly single threaded some of the best engineers were are people who are able to focus on kind of one problem for a very long time and think very squintally Um and then the best managers are people who can who can kind of focus on on many problems and constantly context switch. I think uh in engineering that's changing because I spend if if I were to do it sequentially I'd be spending a lot of my time sitting waiting for the uh for claw or codeex to kind of turn through a problem and that's just not a very good allocation of my time. If I have like a you know intelligent vet may be you know as as good as me but much

cheaper than me uh why would I wait while the you know the bar is kind of spinning. So instead I I find myself like literally working on multiple like very different things at the same time. In fact, it's sometimes you know if you can nicer to to work on different projects at at the same time because at least you know you can avoid you know them overlapping a little bit too much. I mean you can still obviously you know work on 10 different things in the same project just like any engineering organization does but you know it's becoming much more viable for like one person to think about 10 you know totally different uh projects which I I think means that like you can have you know the thing I was describing before is like this like PE style thing where you have like one company that might own

10 like boring businesses and now you can have like one person who's really good with AI who is basically working on all of them in parallel, which I I I think is is a little bit different and it changes the archetype who's going to be good at that thing, right? Cuz before you have like one person who should devote all of their time to being the best that they can at running that one business and now you're going to have one person who is really good at managing TLM. It's turning I think engineers are are more and more so becoming I I wouldn't say engineers are becoming managers because like there is still like a lot of at least today there's still like a lot of sort of technical knowledge that goes into sort of like managing your your LLM but uh it's like a totally different style of

thinking now where it used to be sit and and think very sequentially and very deep and hard on one line of reasoning to becoming thinking on many line of reasonings in parallel. I think it's an interesting question the like there all these AI transformation PE things now and we've been talking a lot about like some form of private equity. So it's like what catalyzed private equity to begin with especially back in the like the late '7s early ' 80s. I think it's some combination of like Lotus notes or just spreadsheets made it a lot more legible to in a day with a computer analyze a business instead of you needed a floor of people in 2 weeks to figure out whether or not you could buy a business. the like classic LBOS's also used a lot more leverage than than you

use today. Uh and so I think like uh just in the same way that that was probably the time for this sort of thing to happen and their transformation in many cases was taking businesses that were not well managed and putting like the classic professional managerial class in there and just trying to run them more effectively. I think today people think of private equity as being very like we buy a company and then we trim the fat. uh that was not like the original sort of orientation I think of private equity. is much more so these businesses are now legible to someone that would like to buy them because we have spreadsheets and what if we changed slightly how they are operated and we moved from uh sort of the original operators or maybe their kids to sort of

professional managers and I think that worked very well for a time but it's you know we've been doing this for like 40 or 50 years now and it's probably time to move beyond that in in the same way like that time period probably catalyzed this specific sort of movement it seems relatively straightforward to say that now that you have LLMs that will make a whole new set of businesses much more legible either to acquisition or to formation to to begin with like maybe the modern version of private equity is not buying existing companies but you know using an LLM to very quickly recreate a very similar product and and compete with them and then I think there's the question of like what is the operations transformation where you know maybe in 1982 you would have hired someone from Harvard Business School and like you know had them run the the

company now today it seems like the modern version of business school is getting AI psychosis It's like how the sort of person that actually can say this is my return on tokens and very effectively use those those tokens. I think like that's a super interesting opportunity or skill set and will have like lots of ramifications on financial markets. And my hope is that because the velocity of spend uh both on tokens and in dollar terms is much much higher. This will end up being something that just kind of by default happens in stable coins that if I want to change my subscription if I want my agent to act economically, I will just end up doing so on a blockchain almost uh you know at least in my opinion doing so in stable coins. And people will not think of this

as crypto. They will think of this as like AI transformation. But I I think under the hood a lot of what has been built in crypto over the last many years will you know very likely be used to enact this AI transformation. So I'm curious your your take on this and then I have another question which hopefully we'll argue about. >> I think stable coins are really like the new financial stack. I think we're seeing something interesting happen where in the past a lot of times with the businesses we try and look at are you know businesses that like weren't possible without crypto or like are sort of like incredibly uniquely enabled wi with crypto. I think now we're seeing something slightly different happen where there there are more and more businesses that aren't actually like you

know could use some traditional financial rail or could use crypto and it actually wouldn't necessarily matter but then you actually just look at you know making that choice and be like it's actually easier for me to use crypto right I can kind of very quickly integrate a blockchain get some stable coins on it rather than get some making partners and go deal with you know terrible uh AC bulls and the the like. Um, so I I think, you know, we're going to start seeing more and more like just it it not being like, oh my god, there there's this new business that is like so cool because of crypto, but rather being like uh we have a new business and we need to do some financial stuff and like, you know, crypto is the way we do this just cuz it's like the the new

stack for for the same reason why like when you go to the cash register, you don't just like pay in cash and have someone calculate your your change, but you tap your credit card. It's just like the, you know, obvious choice. it's not the it's not the thing that made the bodega uniquely able to to to scale. It was just it's what they do because it's the obvious answer. Um so you know how AI plays into it. I mean, I think, you know, it's going hopefully I think makes it easier for businesses to kind of do that transformation because obviously the new businesses I think will make the choice of like let's use stable coins, let's use crypto just because it's it's the obvious answer, but you have these businesses that have like entrenched relationships and and code bases that are built on on the old system. And I

think AI is probably going to be useful in us swapping that out just because again, if it's the obvious answer um but we already have something that works almost as well. Mhm. >> Then, you know, it's not worth spending a lot of money to rip it out. But, um, if the tokens are cheap, then maybe it is worth it. Especially if it's like, oh, I'll get like a tiny improvement on my margins. Um, but if my business isn't at some like massive scale where it's like worth totally redesigning something, you wouldn't do it. But totally redesigning something is increasingly becoming free. Yeah. I think there's the old quip of uh like science advances one dead scientist at a time. Maybe like the accounting department advances one retired CFO at a time. And so existing companies will probably take longer to adopt these things, which is uh I think in many

cases why especially new newly formed small companies or teams that want very high leverage and headcount terms will just end up using stable coins and and onchain finance. I I think there's some analogy here to how in many parts of the developing world they they like never used cash. They never used credit cards. They just moved immediately to contactless payments. And yet in the US like you still carry cash around very occasionally and you almost certainly like tap your credit card uh quite often. So I I think there's probably a dynamic like that here as as well. So I I think there's a very interesting question. Uh, many people may have seen, I think this is 2 or 3 months old, the Financial Times graph of GDP growth year-over-year

in the US going back to like 1850. And it more or less looks like uh 2% year-over-year. There's some years where it's 4%, there's some years where it's negative slightly, but it basically looks like a linear extrapolation over the last 150 years. Uh and then the forward projection is either like GDP skyrockets or it falls off a cliff and goes to zero. It's either like utopia or death which I think is an interesting way of considering how public markets are, you know, thinking about AI. Uh but the implication to this is why has GDP growth not inflected or not fallen off a cliff as a result of the AI models? It it seems like there's relative consensus

that we got, you know, a form of AGI at the end of 2025. Uh the models have been very performant for a number of years now. And I I think the null hypothesis here is that, you know, GDP growth will increase with time. It just takes time for technology to diffuse throughout society. Uh I do think there are other interesting, you know, uh hypotheses or conclusions for that. Maybe we'll just start with a question of like when do you think we will see uh macroeconomic indicators begin to show increasing productivity from LLMs?

>> I think when it comes to how many years it's going to be, who knows? I have no idea. But >> yeah, I'll pick pick a quarter and a year. >> Totally guessing like five years, right? Because I I think uh >> I think these things take time because you know there are many reasons why you might see productivity grow from AI. one one is that like a business that exists like with the labor it has starts consuming a bunch of tokens and they uh they produce more value right and and that's GDP growth I hope that we see that right but there's you know I think a open question of like how much can a given business scale right like you know if you give a social media company 10 times more money you know how can they sell us 10 times more social media I

don't I don't think so one explanation for this is what causes GDP growth what causes the ability to do more with less, it's technology. So, you could look at roughly 2% GDP growth year-over-year and you could say, well, that is the pace of inventing new technology. Or you could say that is the pace at which society is able to integrate new technology and that maybe the barrier has not been we can't create more new technology, but that 2% year-over-year in economic terms is just roughly like the level of change that society can adapt to at any given point in time. that is more of like a human or a societal or an organizational thing as opposed to a technological thing. You know, the idea that 2% year-over-year is uh very similar to

Moore's law in that like we've had all of these incredible wonderful inventions and technological changes that have continued to enable 2% year year-over-year that that's like uh you know a modern miracle and I will just like continue along the 2% year-over-year and I think that's originally uh an Andre Karpathy argument which is like very intellectually consistent and and I think interesting. I I think there's another form of this which is one reason you wouldn't see uh drastic changes in macroeconomic data would be there are a lot of people that are being more productive at their corporate job but where that job actually was not productive at all to begin with. I would be quite skeptical that large corporations will actually

make significant headcount reductions because I think the incentives just don't make a lot of sense if you're sort of part of the professional managerial class. Like if you're the CEO and you fire half of your company and you have a bad quarter or two, like you will probably get fired. Uh and so I would suspect you only see the sort of drastic headcount changes where there is like a a sole proprietor of the business. I think it's not happen stance that where this happened at at Twitter and X with Elon like he owned the business outright. No one was going to fire him. Uh, and you know, maybe you'd call this founder mode, but I I just think the incentives are very unlikely that we will see like drastic layoffs at sort of big corporations. I think it's just much more likely that uh you will have

stronger competition from newer firms that are run in a different way and then eventually you will have like this new class of uh transformation or I think you would just call it private equity by shorthand, but that's probably not going to be correct that like ends up transforming those larger businesses. I have two more theories now on this, right? Um my my my first one is it could also be that um we're AI has massively transformed software businesses um because it's it's like the easiest thing to to start scaling up on. Mhm. >> Um and maybe it is just that there is like not that much more output to squeeze out of the software side and it's going to take time for us to figure out how do we apply AI to massively

scale businesses that still have productivity to squeeze out of them right so you know maybe it's simply that you know the software companies are you know doing the most they can do right they can't sell us more social media there are other sectors or new sectors that we have much more work to do right like robotics right that is going to take much more time and maybe Maybe we'll see, you know, 5 years from now as AI gets deployed across much more broadly and in more cases, we'll we'll see the GDP growth explode. Y >> um my, you know, second theory is that, you know, maybe GDP isn't a real number. I think it's like uh what's the the the the kind of saying about, you know, once something uh becomes a a measure, it no longer makes any sense.

>> Goodart's law, I think. >> Yeah. Goodart's law. So, you know, I think like GDP, so GDP and inflation are like the things that central banks look at and they go they pull their levers to make sure that both those things are are the right number and uh maybe it's just that like you know the actual productivity of the economy is is super high but the central bank pulled the levers to make sure it's the it's the percentage that they feel comfortable with and like no amount of of technology is going to make that measure change because you know you have so much power by being able to change rates or or change policy. I think that's correct for the CPI and for inflation. Like I I just I think that that's totally fake.

If you look at any of the inflation adjusted numbers, I don't know that there's a better measure than than GDP. Uh and I'm I'm like I've become probably for worse, but for better or for worse, just like maniacally obsessed with this concept of productivity like everyone is very interested in the idea that AI will make you more productive. a niche. One of our partners had a a really interesting post recently that most people don't want to be more productive, don't want to save time, they want to spend time, which is like one interesting conclusion for uh AI making everything more productive. But the other question would be like even if you do want to be more productive, how do you define productivity? Is that in terms of more dollars created? And then to your point, does monetary policy make

it very hard to understand like [snorts] whether more dollars being created is actually good? [laughter] uh like does this mean more production of information or technology like I think that's very hard to measure you could look at you know number of patents number of PhDs granted uh like all of these different metrics which I I think are probably skewed like we we have I think it's something like 10 times more physicists if you just looked qualitatively I think you would say physics has not advanced you know very much after World War II this is really a yearning for a better definition of of productivity like how to say at a broader macro perspective like what is the value of the level of intelligence

that we have created like we turn sand into thought which is a miracle uh but sand into how much thought and how are we using that thought [laughter] is it also possible too that you know maybe we are actually like in a massive recession we just have no idea because AI picked up all the slack [snorts] right like it could be I mean I I've I've heard this argument before that you know that the economy is still like wrecked by sort of the after effects of shutting it down for a little while from COVID and that there's like a bunch of measures that still don't do very well and maybe it's that like hey I picked up the slack and like the you know large swaps of the economy are not doing very well so it looks like 2% but it was actually that should be down 10% and

it's up 12%. Yeah, I I mean I think a lot of the the vibe session this idea that if you look at the economic indicators they seem to be doing pretty well but if you talk to people they seem to be sad about the state of the economy. So I I think this is just like a second order effect of the internet that people are like deeply mimedic animals and if you can see how all of the people on earth that are doing the best are doing at every you know second of your day all of the time you will try to compare yourself to them which is a horrible idea and then you will feel sad and you should be you know inspired by the classic quote that you know true nobility is not superiority to your fellow man but to your former self like

that this is just sort of a psychological thing that you have to grapple with for for better or for worse. I do think it's interesting that if you look at why have AI models gotten better, they've gotten better because you can throw a ton of data and compute at them and you know the entire world has learned the bitter lesson that the size of the model is directly proportional to the performance. Uh and like what are the inputs to you know data and compute the most important input is probably energy. In the US for roughly 20 years, call it between 2000 and 2020 to make the numbers course. Uh there was a very slow growth in energy demand from you know call it 1920 uh through depending on how you define it the 80s or the '90s something like that.

You had like 7% year-over-year growth in in demand or demand for energy or energy generation. And so I think one implication of this is like we ran out of things to do. If you look at many of the productivity measures in society, things like self-reported happiness or infant mortality or GDP, like they're strongly correlated just with the amount of energy a society consumes. And there was not kind of a marginal buyer for megawatt or kilowatt hour. Like we didn't have anything to do with the power. [laughter] Uh and now we do, which I think is an incredible change and essentially means we can deploy as much new energy generation capacity as as we want. And like this is what has been happening in in the US and in the west broadly is just like building as much generation as

we can as many data centers as we possibly can. And that I think is very positive. But you could totally say that just the capex spend on AI has been you know papering over or covering up other areas of the economy not performing the the way they would have wanted. I think my argument for that would be until you have a new thing to put the money into, it won't change [laughter] that >> [snorts] >> uh like if someone invents fusion or you know we have time travel or or or whatever like your next sci-fi technology is like you then would probably would see like a narrative rotation and the unit economics of of like AI would probably become more important. But it seems relatively unlikely in a short term that we'll like have a narrative rotation. And so yeah, I I I think we're probably in a pretty

good position there. We probably have no way of measuring this, right? But you can, you know, if you don't like uh GDP as the thing you measure. We should like measure like the gross intelligence in our country or planet. Um, >> but how do you measure intelligence? I'm not claiming that this is is something that is easy or possible to measure, but simply that like it imagine you had that measure like it could be somewhat informative that like it could just be that you know you were in our little bubble and and we see AI everywhere but it could be that um the intelligence in the world has actually only grown by a couple percent because you know we just haven't scaled this enough and if the intelligence only scaled by 2% maybe the GDP should always scale by 2%.

>> Yeah. I I mean I think it's very hard to measure intelligence directly because the world seems to want things at given points in time and like people deliver them and sometimes we I think make the mistake of saying that like the person and their incredible intelligence is the reason this happened. I I would point to the creation of the touring machine and the lambda calculus at almost exactly the same time by two totally different people who never met each other all the way across the world. That that's just like the world wants this to happen at a given point in time. and you can make whatever spiritual, religious, like information, theoretic conclusion you want for that, but I think there's like a time for things and then people materialize them. Uh, so I think it's very hard to quantify intelligence in any of those like outcome factors. And

then I don't know if you believe in IQ tests, but uh they just don't seem to be particularly effective. [laughter] >> Yeah. >> Like how do you quantify intelligence? >> So So I I am not saying that like we should try and quantify it. I'm saying that like you know imagine there is some theoretical quantification of this and and it could be that like we just haven't increased the amount of intelligence that we have right like you you know you make like a very simple model that like you know for economies that are like you know full of skilled labor that like we kind of just convert intelligence into GDP maybe and then like maybe it's just that like we haven't increased our amount of intelligence very very much and yeah we may totally never be able to measure that or it may like there it may not

even like mean anything to measure that. But if you imagine the madeup measure, it could just be that that that hasn't increased by very much because we have a whole bunch of human brains and not so many data centers. And you know, at some point we'll have more data centers and brains and and at that point we see a GDP explosion. We've talked about this before. If if you look at the correlation between IQ and earning uh IQ and and income like the correlation goes away above roughly 125 or 130 IQ. I would argue the models are definitely smarter on an IQ test basis than 125 or 130.

>> 56 soul ultra measured at 136 according to Twitter yesterday. So there we go. I think a super interesting question is like as the mart as the models get smarter do they uh become more economically productive and if you look at people your projected earnings as a result of IQ do not go up if you get smarter than the models currently are. So like what is strongly correlated with higher earning potential? I I think the shortest answer or my argument would be some form of grit or tenacity or agency which I think has become an overloaded term which inherently like these models do not have because they do not have desires. It's just math. [laughter] >> I'm referring to intelligence as like a gross unit, right? Like when you talk about IQ increasing, it's like one

person gets smarter. I'm talking about like you can think of like one person versus 10 people being like approximately 10 times the amount of intelligence assuming they're all kind of just as smart. And like you know the AI can be like part of that measure of like it you know the more like each token like gives you some amount of like intelligence more tokens equals more gross intelligence in the world. >> Yeah. One is definitely good if everyone is smarter. That seems obvious. And two if everyone has access to these models and I I think they they will this will be an incredibly democratizing force where you basically like slam shut this dispersion in in IQ. And it probably becomes much more important not to be particularly smart but to be you know particularly gritty or agentic or to

have strong desires for things which is strange and weird because the western economy has been selecting for intelligence in like the leadership and elite for hundreds of years now and maybe overnight uh at least in the the scheme of time uh we will like start to select for something else entirely. [laughter] I think that is like that's a very interesting question. If now the amount of intelligence we have is just a result of how many data centers we've built and not the number of people we have or the quality of the education system or something else. Uh yeah, it's going to be a fun time.

>> So you had started by posing to Noah what is the year the quarter that you're going to start seeing these macroeconomic indicators show up but you didn't answer it yourself. >> I think it would be very hard to say. Uh I'd say like Q427 Q128 something like that. Uh, I don't have a good justification for that. That's just that's a that's a vibes based assessment. The other thing I'll say which we probably should have caveed to begin with. I I think like the most obvious counterargument to the discussion we've just been having is like that uh the capex spend from AI like obviously shows up in GDP. I think that was something like half of GDP growth in 2025. Uh so like yes that is that is true. The question is much more like when do the when does inference

from the models themselves not you know the dollars we're putting into the ground show up in in GDP growth but uh like the technology itself enable people to to do things more more efficiently and I think the answer probably is just time like it's probably a smattering of the other things we're saying which are more interesting but it's probably just things things take time and if we started to see this you know really happen at the end of last year it doesn't seem unrealistic for it to take you know 2 or 3 years for for that to properly diffuse throughout society. >> One final question for you guys, guy. When you were describing this opportunity that AI opens up, >> you described it in relation to traditional venture bets where you're trying to swing for the fences and create the biggest possible business and

outcome possible. But in contrast to that, you're saying AI enables these much smaller, nimler sort of opportunities for businesses. You call them micro businesses. Maybe they even get funded by microloans through crypto. Does that not seem though like a lessening of the ambition of business building? >> Like how Yeah. How do how do you I still think there will be like these giant you know network effect uh very successful technology businesses and that AI gives you much stronger leverage and the ability to build those those sorts of things. Uh the the set of comments was not uh excluding that. It was more I feel like there's this category of

things that is not talked about enough or a lot. There seems to usually be like a default path for smart interesting ambitious people in the western world and uh you know for a long time that was finance and that has definitely become tech and it seems like the default now if you want to start a new thing is just to raise venture capital. The meaning of venture capital has definitely changed in the last 20 years. Uh and so the the comments are all more just to really say like what is venture capital? venture capital is the idea that I need a lot of money to do something very ambitious that if it works which is very unlikely will have an incredibly you know successful payoff both in terms of how it changes society and how valuable the company becomes and

that like that's not the only sort of business that you can build that you know the American economy is composed of a very large number of very small businesses that in their aggregate are incredibly important and there are really interesting opportunities to build those sorts of things even if you come from like a you know Harvard, Stanford, Google etc. background like you do not necessarily have to go try to build something that's worth a trillion dollars. Like you can have a very fulfilling and interesting and like a creative to society life building something at a smaller scale that probably wasn't possible before. Uh and I I think that's interesting from the perspective of how you think about the landscape of investments and investment managers. I also think that's

interesting from the perspective of a lot more people being in charge of their own destiny instead of trying to get, you know, uh, a job at a big company or build a big company yourself. You have a much greater distribution of, you know, smaller companies and sole proprietorships that all interact with each other well. And I think that is the sort of bottoms up distributed decentralized world is very much the one that we've been excited about and building in for a number of years. Uh and or investing in not you've been building in. Uh I I used to be an engineer and definitely am not anymore.

So I I think I think it's more that just the idea that uh if you believe in a world of smaller firms, more firms, more coordination between them, higher velocity of commerce, like it it all rhymes very well with kind of a a change in how people think about starting new businesses. And the small businesses that people have been interested in starting previously, it's like usually you come out of some like M7 MBA program and you buy an HVAC installer. I would just like advertise to you that they're all on a spreadsheet and it's probably not a good idea anymore and like you should spend 18 hours a day with Claude and that will probably be a better outcome for you and your business and for society. I also think like we should actually be stoked about, you know,

these things that that that you were calling unambitious businesses, having more of them and being able to to do it more efficiently, we should be incredibly stoked about because I think there you can split the economy in two pieces. You can think of it, there are businesses that work very well under like what one might call like the perfect model of capitalism and there are businesses that don't. The businesses that do, there are businesses that are small, there are many of them, they have no product differentiation, they have no advertising. And when you have those businesses, capitalism does one thing very very very well. What it does is it makes the prices go very very low. It it makes the the margin be be tiny. And we actually we love it when there are business like that. This is

why food is so cheap, right? Because like food is for the mo or you know like corn or rice right it's for the most part like undifferiated. There are like many small producers of course there are some very large producers as well that have some scale effects but that portion of the economy is constantly like squishing itself to make prices lower and we should be excited about potentially like more of the economy looking like that. On other hand we shouldn't be we shouldn't say that there the other category is back. over the the part that almost to a certain extent breaks capitalism because they have like massive scale effects and and network effects and they are widely differentiated from one another and these businesses maybe rather than cynically calling them the businesses that break capitalism you might call them like the innovation economy they're businesses that are like net new and we

see these incredibly large companies kind of grow out of them >> now I think we need both these things right we need the innovation economy to have new things and then we need an economy of businesses that are commoditizing themselves so that people can buy things for cheaply. Now what we want to do though is that as the innovation side of the economy grows and we have these new massive businesses, we want this second thing to happen is that we actually want some force that converts some of these kind of new businesses that are breaking capitalism in and you know are therefore able to extract high margin into businesses that kind of go back into that perfect model where they break up into many businesses and they their margins get compressed, they get commoditized, maybe less excited to invest in but you know we'll

we'll continue investing in the the the part of the innovation economy as these companies get created, which is, you know, I think what a VC's job is. But then we also, you know, need some force that converts those businesses, you know, once they've grown and, you know, return to their investors and all of these great things and and hopefully change the world. We want to convert them into businesses that have low margins and, you know, provide things to people at a low price. And I think people leveraging AI is actually a great thing and we should be super excited about that. >> Very well put. Uh, capitalism is a process of creative destruction and uh, you need both of those in equal measure.

We covered a lot of ground today. There's a lot to chew on. We started with token spend, went into different models for companies, you know, pod shops versus the sort of like PE style stuff that you were talking about, the nature of consulting and how that's going to change, GDP, whether it even means anything and what happens next there. So, this has been great. Guy Noah, thanks so much for coming on. >> Thank you. >> [music] >> Heat. Heat.