Vibes & Benchmarks Ep 04: Is anyone actually working less because of AI?

TOPIC 1 - Bernie Sanders and AI wealth redistribution

Ali Rohde (00:02) All right, Josh, ready for Vibes and Benchmarks episode four?

Josh (00:05) I'm ready. Do we ever talk about vibes or benchmarks? I'm not sure, but it's fine. Okay, sure. It's an aspirational title.

Ali Rohde (00:12) No. Not yet.

Ali Rohde (00:17) Yeah, we can talk about more vibes. Benchmarks? I don't know. How much do you care about benchmarks? Yeah, that's my question. Hmm. All right.

Josh (00:19) No, no, it's fine.

Josh (00:23) Do they matter? Yeah.

Ali Rohde (00:34) Senator Bernie Sanders just announced the American AI Sovereign Wealth Fund Act, a bill that would give the public a fifty percent ownership stake in the largest American AI companies, including OpenAI, Anthropic, SpaceX, AI. Do you agree with this proposal?

Josh (00:55) I did I forgot that it was SpaceX, so that actually changed it a tiny bit. SpaceX is not an AI company anyway. Yeah, we can talk about that more in a second as well. I feel like you don't want to do that because then it's gonna look bad when the value just plummets to zero. It's like, man, we like, you know, took this thing and now it's worthless. But and let's let's just talk about Yeah, let's talk about anthropic and about AI. Well, then there's another like one of the things that I

Ali Rohde (00:58) I don't I'm not sure it's specified. Well, it's like XAI.

Ali Rohde (01:13) Okay, if it were just OpenAI an anthropic, then what would your response be?

Josh (01:22) remembered when this came out was like, haven't Sam and Dario talked a lot in the past about like, yeah, like in the future there'll be UBI and like maybe even citizens could like, you know, have some equity or whatever in the companies. You know, guess who is not going to be saying a lot in support of this bill? It's gonna be either anthropic or OpenAI, right? Now, to be fair, the bill's kind of dumb. Like taking 50%, I don't know, that does that is that like really the right place to start or whatever. I don't know. But I want to see counter proposals. Like I don't think it's authentic or serious to have think pieces that are like, yeah, we'll totally be redistributing wealth. And as soon as someone, you know, does this, you don't engage and make a counter proposal. I mean, that to me signals like you're not really serious about this, which we didn't think you were serious about anyway, but then maybe shut the fuck up.

Ali Rohde (02:23) Knowing about Bernie Sanders and how he moves, my guess isn't that this is a serious proposal. This is a statement, like an ideological statement. So I don't expect those companies to engage either. One, because it's just like a very sensitive time given Anthropic just filed to go public. But two, it's like it this has no chance of passing. But you say you want counter-proposals. So what do you want?

Josh (02:50) Right. But like they have literally proposed making UBI, making, you know, a sort of sovereign wealth fund, et cetera, type things in their previous like in in multiple previous pieces and posts, right? And you can quote them and you can look at them and be like, Yeah, that that's you just saying it right there. Like there was one that I pulled. In fact, I had Claude go find the quotes for me because I was too lazy and I was like, Yeah, yeah, I got one in my post draft here from

Josh (05:03) Yeah, I mean Dario and Sam, OpenAI and Anthropic, have put out a bunch of different pieces about, you know, what if we had UBI in the future? What if we had, you know, some kind of, you know, equity that people had in in these companies or whatever. So what I want to see is like a more serious, like, okay, but how would we actually do that? and so I had Claude go pull some quotes from Dario and Sam, et cetera. So like here's one from Sam, for example, from one of his so from Dario actually from one of his very recent Things. the natural policy response to an enormous economic pie coupled with high inequality due to lack of jobs or poorly paid jobs, et cetera, is progressive taxation. The tax could be general or it could be targeted against AI companies in particular. Obviously tax design is complicated and there are many ways for it to go wrong. I don't support poorly designed tax policies. I think extremely, you know, extreme levels of inequality predicted in here justify a more robust tax policy based on basic moral grounds, et cetera, et cetera. Great. So what does that mean? If it doesn't mean giving the cup the country or the world a 50% stake so that like democratically elected leaders would control what happened, then that's fine. But what does that mean? Is it like a token, you know, everybody gets a dollar? Or like, is it you get taxed? It's not 50%, it's 40%. Is it like 5% per year? Is it based on revenue? Like there's lots of options, but nothing isn't a serious option if you're gonna like continue to say stuff like, we're gonna be redistributing this. in any in any way or like, you know, counteracting inequality in in some way, as like this big company that makes lots of money. Like somehow we're gonna like help. I I don't wanna see what is that.

Ali Rohde (06:43) You mentioned UBI and they've mentioned UBI, I think Sam Altman in particular. If they were to seriously work on and work towards UBI, is now the right time or might it be too early?

Josh (06:57) Don't think it's too early when you're making a company that is one of the five largest companies in the country or in the world. Like trillions of like there are not many that many companies that are worth trillions of dollars. Right. So like yeah, but they're I'm

Ali Rohde (07:11) There are some they didn't introduce UBI when they went public or when they reached those heights. Nvidia didn't introduce UBI.

Josh (07:16) Yeah. Yeah, but they didn't grow from zero to a trillion dollar. NVIDIA grew over the course of many decades and has a relative like cap and also doesn't control like really that much at the end of the day. They they make the chips, but like the data centers are the ones who actually own the chips. The people who like make the models are the ones who sell them. Like they make one component in this very large system. They don't make a system that influences people and how people think about politics. They don't make a system that controls all the information that people in your d democracy see. This is a very different set of things, right?

Ali Rohde (07:52) Okay, what about Meta and Google who do just that?

Josh (07:54) Well, also people have called for regulation for those companies for a long ass time, and there's a good reason for that. Yeah. And I would say that that is another good thing that we should reevaluate. There's been a well, like, if you think about it, companies they they want to grow. That's what they're naturally supposed to do. They're supposed to make money, they're supposed to grow. Okay, that's great. But the role of government from all the way back from Adam Smith and the Wealth of Nations is to cap them and when there are monopolies, to break them up and do something about it. Okay. And so we did that historically. But the problem is that our antitrust laws are about consumer harm from prices. And so today's monopolies don't charge consumers anything. So the price is zero. So there's no price discrimination. So no monopoly, right? No, that's not how that works. But that means that we

Ali Rohde (08:36) Anthropic and open eye, unlike Google, do charge people stuff actually.

Josh (08:41) That's right. But I'm saying that the I'm I'm saying about Google, you're asking, like, should we, you know, have similar things that apply to Google and Facebook and similar sorts of laws? Yes, we should. Like they are monopolies. There's no question that Google was a search monopoly for a really long time, right? And so it would make sense for the government to do something, but it has not done anything. Because we need new laws. But the problem is that now companies have enough influence over what laws come into existence that those laws are not really the ones that anyone is pushing for or talking about.

Ali Rohde (09:12) Perhaps. I don't think it's anthropic and OpenAI's influence on on the administration. I think Yeah. But I think like this administration is just not interested in that generally. Are you implying that OpenAI and anthropic are monopolies or duopolies?

Josh (09:17) I'm saying Google and and meta. I'm saying Google and Meta, not Facebook and OpenAI. Yeah.

Josh (09:26) Yeah, exactly.

Josh (09:33) No, they're not a monopoly yet. Although if you look at the terms of service and the product and the performance and the pricing for anthropic and open I and Gemini, they're basically fucking indistinguishable. If I swapped one out underneath you, you wouldn't notice. So it's not that they're a monopoly, but they're a lot closer to a oligopoly or something where it's it's not really c competition in the way that you want. Because if you had competition, you'd have competitive pricing, they wouldn't be making any money. They'd be competing on price. But instead, they're starting to make tons of money because it's like, great. Like we should not be competing with price, competing on price with each other, because that's just bad for all of us. But you can't have that kind of coordination when you have tons of people. Like bakers don't get to be like, no, that the loaf the price of a loaf is this. Cause someone's like, price of a loaf is $30. I can make a loaf for less than $30. And then it like things take care of themselves. But when you have only three, you can all just kind of not do that and it sort of works out.

Ali Rohde (10:33) Do you think that they're coordinating? I mean, it feels like the opposite right now.

Josh (10:35) No, you don't need to co you don't need to coordinate is what I'm saying. Like when you have such a small number, there's there's no defectors, right? Like no one else is being like, Well, I'll just do this for, you know, half the price or whatever. There's no pricing pressure. Yeah, but it's not the same thing. Like I'm saying the companies that have the like leading closed source models are not competing on price. There's only three of them.

Ali Rohde (10:46) I mean, is that what open source is?

Josh (11:00) So there's no competition, which is what the government is supposed to do is cause competition. That's why capitalism is good, is from competition. But there's

Ali Rohde (11:07) How would the government cause competition here? Fund its own AI lab?

Josh (11:10) By making regulations, p taxing things, making regulations, and sa incentivizing open source. I mean, you could incentivize people to do science, you can incentivize people to, you know, you could have fucking protected the resources that went into this in the first fucking place, instead of just, you know, yeah, I guess it's fine to steal all the books and then resell them to everybody. So there are plenty of places that you could intervene. It's not like a impossible thing that could, you know, to do.

Ali Rohde (11:38) The other part about UBI is that it's supposed to help people once they can no longer work, the once those jobs aren't possible. So they just don't have the opportunities to make a living wage. Do you think we've gotten to that point yet?

Josh (11:52) no, I mean we're clearly not at a point where like nobody can work yet, but we have seen AI layoffs. We talked about that, you know, a week or two ago.

Ali Rohde (11:59) We have seen AI layoffs, though I will say we have not seen that show up in overall unemployment stats. Like they have not really moved.

Josh (12:07) Yeah, I mean, I'm not certain that that that's gonna be a pretty lagging indicator. I think if you wait until that's the case, then like it's probably not the right way to go about it. Like what what kind of power differential will there be when it's like, yep, well, turns out twenty percent of us are unemployed and like the rest of us are really worried about, you know, losing our jobs? Like, is that the time to like, you know what, we should probably make better protections for unemployment and stuff? Maybe not. Maybe now is the time to do that.

TOPIC 2 - Product-market fit and AI token spend

Ali Rohde (12:36) This reminds me of another question. There was an article this week or last week by Simon Willison. And the title was something around Anthropic and OpenAI have product market fit. Do you think they found product market fit?

Josh (12:52) I mean, they found it a long time ago, right? The question is about like the the question are these larger like policy and world and regulation questions, not a like technical question. They're it's useful, it's been useful to use language models for a long time. You know, that's why Google switched their search thing to be like, we're gonna have language models do stuff. Like people recognize that these are useful for at least a good, you know, year or two or three here. And then we also That's like the early version of it. That's a small piece. And then you have all the coding models, right? Where it's like, yep, lots of people making code are like, Wow, this is really useful. I like using it. I like spending lots of money on it. So yeah, that seems like product market bit. And they can sell it for more than that it costs them to make. So that's that's good.

Ali Rohde (13:36) That was gonna my next question because I think we've talked about you, like if you were in an anthropic seat, if you were an OpenAI seat, you would be worried about open source, about Google long term, about these models commoditizing, about these models reaching some quality point after which it doesn't matter if they get better and therefore you can switch away from frontier models. Doesn't that contradict what you're saying about product market fit?

Josh (14:02) I mean, you can have it now. Does that just having product market fit doesn't mean that your company is going to be valuable for forever? Right. So I think that's there's a question of timing. There's a lot of uncertainty, right? Like, is it that things stay this way? Or, you know, we've seen things about companies accidentally spending half a billion dollars. Does someone at the company go, huh? Maybe we should like get these at a discount or like stop wasting so much money. Are we gonna really getting a return from this? So it's not clear that that continues forever into the future, right? And then you'll have competition from open source models. You have, you know, if you don't release mythos and open source models catch up, like there's a lot of other things that can happen. It's a tenuous place to be. It's not a locked in forever kind of win.

Ali Rohde (14:43) Okay, so they have product market fit now. You think they won't in a few years?

Josh (14:46) Mm-hmm. I think part of it part of the you have to like separate the usage for these these models. One piece of it is consumers, like in the Simon Willison piece, he was like, if I look at my tokens, I s I got like, you know, $2,000 worth of value and I only spent like, you know, $100 or $200 on the subscriptions, right? Okay. So enterprises are paying that $2,000 price, right? But users are not. So most of our perception of like, how good is this product? Like, if you had to pay $2,000 for it, most people are like, I'm not paying $2,000 for that. Fuck that. Like, so

Ali Rohde (15:22) I don't know, isn't aren't you and Bue paying mostly API costs?

Josh (15:25) No, no, no. Most individual users. Like individual users would really balk, I think, at paying $2,000 in a month. Right. Now, so okay, but when you're at a company, the company pays for it, spending $2,000 of your company's money on it, well, that's not so bad. Right. Then it's like a business. Exactly. Then then it's like a business, you know, cost benefit decision. And so the question is like, where does that land for businesses? But they're gonna like as you start to pay the same amount.

Ali Rohde (15:40) The beauty of enterprise sales.

Josh (15:53) in headcount and tokens, you're gonna start to think about optimizing these costs and pushing it down, et cetera. It's just we're at the very beginnings of that. So how much of that do people do? Like, do companies switch to open source models? Do they just restrict how much their employees spend? Like, I don't know. We'll see. Yeah.

Ali Rohde (16:10) One thing I liked about the piece was its reaction to all of these headlines in the last week, basically, kicked off by that Uber COO quote about how they're not seeing true ROI from their token spend. And we're seeing a lot of organizations start to limit their token spend. We're seeing Microsoft revoke Claude Code licenses. I'm sure to the chagrin of Microsoft employees who are now expo expected to use Copilot.

Josh (16:35) Mm-hmm.

Ali Rohde (16:39) Actually Facebook Meta put out something, a memo or something saying we're done with token maxing, stop token maxing. Turns out just usage alone is not leading to ROI. Shocker.

Josh (16:51) Mm-hmm. Yeah, exactly. So

Ali Rohde (16:53) And I think a lot of people are taking those headlines to be like, you know, these companies are in trouble and now people are cutting spend. And Simon is like, no, no, no, the opposite. The fact that companies are now having to have these conversations, the fact that expensive employees are using them constantly and using a ton of tokens is actually an indication of how powerful they are and how they do a product market fit.

Josh (17:18) Kind of I mean, I think it's somewhere in between, right? Like n just wasting a ton of money on these things in the short term is fine when you're experimenting as an enterprise. Like who cares about a few hundred million here or there as a as a huge company, right? But that doesn't mean that you're gonna keep doing that forever. So yeah, it does need to pay off. Uber is a good example. We talked about it before.

Ali Rohde (17:39) Does need to pay off. But I think the point is like the fact that we are having these conversations and the CEOs are now having to decide if they need to limit token use is kind of a sign that like they have these companies have won won. And like we're really this is a huge expense. and

Josh (17:57) Yeah, but it has to be worth it though. It has to be worth spending the money, right? And there's a lot of stuff that's getting made and created that's just extra cost.

Ali Rohde (18:11) It seems like AI the best use case of AI is almost always AI. Like cybersecurity, right? We need to spend a ton of money on cybersecurity and like securing our systems with AI from AI. The hiring space, recruiting. We're having to invent all of these solutions for going through resumes and applications, because now it's so easy to send in resumes and applications with AI.

Josh (18:20) Yeah.

Josh (18:30) Yeah, exactly.

Josh (18:45) Ki kind of. Yeah. Well, I mean, for hiring, maybe people don't hire as much. So some of those things go away, right? Yes. You're not just gonna keep spending more and more infinitely on blasting out your resume to more and more people and on, you know, better and better filtering tools. Like there's a there's a game theoretic optimal there. Yeah. Cybersecurity was was interesting. I saw a thing yesterday that some company spent like a million dollars or whatever finding, you know, a dozen or two bugs.

Ali Rohde (18:47) Push back. What what's off about that?

Ali Rohde (19:02) Sure, but what about cyber?

Josh (19:15) You definitely could have just hired some people to find those dozen or two bugs. Like, was that worth finding those dozen or two bugs? The only reason you did it was because you're like, ooh, look at this exciting new model. But if someone had come along before and like, hey, we could spend an extra million dollars to find more bugs in security, you'd like, So fuck off. Like, we're not spending more in security. We don't need to do that. So it's, you know, is that gonna continue being something that people spend lots of of money on and time? Or is it more of a one-time thing? Like, yeah, you know, we'll go find all the easy stuff for a little while. And then call it a day and go back to not caring about security like we have in the past.

Ali Rohde (19:49) Of why I mention that is because even for companies for large enterprises that are not finding great use cases of AI, I kind of expect that engineers in particular, and maybe all knowledge workers are going to need AI in the future because they're like letting their skills and knowledge decay. And so now just to do the work that they used to do normally, now they need AI and to spend tokens. And so like AI is creating this necessity for more AI.

Josh (20:09) Yeah.

Josh (20:17) Yeah. And I mean that that's what we've always been saying, though, is like, and that's what I'm saying with like the labor replacement stuff is like if you're just using a bunch of tokens, like what are you doing for the prompt? Like eventually the the harnesses get better, the skills get better. Eventually you're like, make it good. Okay, cool. Well, I can also just make a thing that says make it good at the top, and then the AI does the rest. Like, why am I paying you to say make it good? So that's that's the replacement of labor. And that's that's why companies don't care. Because companies see, like, cool, like we can, you know, take this thing, offload all like the knowledge that people have into this system. That's then we're certain is going to become much, much cheaper over time. So amazing. So we've taken what used to be a one hundred thousand dollar a year headcount and we've turned it into a two hundred thousand dollar a year headcount plus AI. And then we're gonna lay off the person and then we're gonna wait for the hundred thousand to go to ten thousand over time because of the economics of these tokens. So that's an amazing deal. That's just investing in the future. As a business.

TOPIC 3 - Domain knowledge, unemployment, and smaller companies

Ali Rohde (21:18) Testing in the future. That's depressing as hell. There was actually a tweet about this that I responded to. So this YC partner, Tom Bloomfield, tweeted, imagine replacing 90% of your employees with a team of geniuses who have no idea how your company operates. Total chaos. Nothing works. That's what AI feels like today. The missing piece is extracting all the domain knowledge from people's heads and providing that as structured context to the models. Paul Graham jumps in.

Josh (21:19) Yeah. Yeah, exactly.

Ali Rohde (21:47) This problem will naturally tend to go away as companies are grown from the start using AI. Then you don't need to extract any domain knowledge from people's heads. It will never have been in people's heads to begin with.

Josh (22:00) That's right. So we'll either have Yeah, exactly. That's what I'm talking about. This is what AI employ unemployment looks like. It's there's two things. There's one, there's starting companies and making smaller companies where like you're really highly leveraging the things so you don't hire those people in the first place. We see that at our company and at other companies that we work with. and then the other one is like enterprises spending tons and tons of money to get the knowledge out of people's heads into these more legible systems. And then, yeah, you wanted to retire a little early. Okay, great.

Ali Rohde (22:02) that's so bleak.

Josh (22:29) Like just letting people go.

Ali Rohde (22:33) One thing I've been hoping for, I don't think we've seen it yet, but maybe it's to come, is a rise of companies, bootstrapped companies. Because you used to kind of have to raise venture capital and have like a significant sum of capital before you made profits because it just was very expensive to hire engineers. Now you can build a lot yourself. And so you don't need to be targeting one billion dollar plus markets. You can target much smaller markets and just take home the profits.

Josh (23:01) Yeah, I think that's people should be doing that. People should stop making AI tools. They should start making things that are actually valuable, but things that are only valuable on the order of like, you know, a hundred thousand to a few million dollars a year. Because Anthropic is not gonna automate that because they don't care. And so just go make those. Yeah. There was a there was a guy that we talked to that did like, you know, storybooks for kids and like had a reasonable business out of it. I was like, Yeah, just just do that. Like, don't worry about the rest. Like, just just do that, you know?

Ali Rohde (23:16) Yes, yeah, that's a side benefit as well.

Ali Rohde (23:28) Yeah. I feel like we look at so many deals that are like this is just not a venture outcome. It's not a venture market. And it's like, can you just not raise venture and just like build it? Like you know it's still a life changing outcome in terms of wealth and like huge value created. It's not as sexy though. Make bootstrapping sexy.

Josh (23:34) Yeah. You just don't need it. Yeah.

Josh (23:43) Yeah. Do that.

Josh (23:48) That sounds great.

TOPIC 4 - Craft, AI slop, and Jevons paradox for knowledge work

Ali Rohde (23:50) Is AI going to make us much stupider?

Josh (23:52) Yes. B because of exactly this thing you talked about about like, you know, extracting the knowledge from people's heads, right? It's like it's just easier to ask the AI to go do something and it's hard for somebody else to verify it. And it's like, did people at these big enterprises really care about the output anyway? Sometimes no.

Ali Rohde (24:04) Really is.

Ali Rohde (24:09) Is there any hope of guarding against this? If I'm an individual worker, an individual employee, an individual,

Josh (24:20) I mean, the the weird thing is like as an individual, you can be in a good spot for a really long time because it's not that everyone is gonna lose their job. It's that the people who are just saying, like, just make it good are gonna lose their jobs. So the Yeah, I mean, met so two two classes maybe. One class is like people who are mostly just using an AI tool and not really doing anything anymore, and like either the like

Ali Rohde (24:36) Like the managers.

Josh (24:47) SaaS that they subscribe to it has AI or they're just using Cloud or whatever, and they're not really adding anything to it, those jobs are going to go away. Another class is then the like recruiting management, the things that are associated with having headcount, those also will go away as you lose other staff. But it's it's annoying because as an individual, you can keep your job and actually get a much higher paying, better job by actually being the person who understands how these systems work and how to like build them and how to build up your own AI systems and how to automate things, et cetera. Then you can get this amazing high-level job for a little while longer, but it just kind of like accrues to a smaller and smaller number of people who are or orchestrating bigger and bigger like changes, right? And like that's what we're seeing just across the board. It's just like that's the trend, right? It's like, I'm just saying like companies, individuals, like it's already sort of aggregating. Like there's tons of wealth accumulation. If you look at the like real wage gain for people in the past few decades, it's kind of flat.

Ali Rohde (25:29) You're already seeing that.

Josh (25:44) but GDP growth, amazing, average salary, amazing, because like a smaller number of people are taking at home a larger and larger fraction of the value. And I'm just saying that trend I think continues and accelerates even. So this is actually bad because then the smartest people are like, well, I can survive by, you know, not fixing the system, but just like being slightly better than, you know, being slightly faster than the other people who are run also running away from the bear. so it's just like a really it's a bad place to be.

Ali Rohde (26:15) What are you doing? You're running faster. Yeah.

Josh (26:16) Run faster. Yeah. I I also don't want to be eaten. Yeah. I mean, it's that's the problem, is the individual game theory is bad. That's why going back to the Bernie Sanders thing, I'm interested in like what are the real the real counter proposals. Yes, his thing is not a real proposal, right? But okay, what is a real proposal?

Ali Rohde (26:37) I don't know you saw, I tagged you on a tweet on Twitter. Rune posted a tweet. The new state of things does not encourage deep worker craft, except in people who have a really Benedictine self discipline. I responded, so you've met Josh Albrecht.

Josh (26:41) Mm.

Ali Rohde (27:00) Rune posted a tweet. The new state of things does not encourage deep work or craft, except in people who have a really benedictine self-discipline. I responded, So you've met Josh Albrecht.

Josh (27:13) I mean, I did have my headphones on and I was hacking since six AM this morning. So yeah, but that that's the thing that I'm saying about people. Like some people will be able to continue using these systems and like benefiting from them, but not everyone wants to or should be doing that. Like it's kind of annoying. You're just like, No. Or I mean, like we don't want a world in which that's the only thing that people can do. Yeah.

Ali Rohde (27:31) Not everyone should be doing that. Isn't that the only option they have?

Ali Rohde (27:40) If I'm an individual engineer and I want to like build actual skills, be a craftsman, but I'm being hired because I'm young and AI native and expectations are changing in terms of how fast I write code and ship features. What do I actually do? What is your advice for someone who wants to be a craftsman?

Josh (28:06) Yeah, I think it's it's really about whether you invest in yourself. Like, do you take the time to learn things and understand how they work? Or are you just like racing and trying to get more and more stuff done? Just like back off the racing a tiny, tiny bit and learn, like, okay, what are those technologies? Like, let me look at the code, let me s read stuff, let me see what else is happening. There's a lot of stuff happening and a lot of stuff changing, but the core things are the same as before. Like you need to write high quality code, you need to have tests, like you need to have good architecture. What is that? Like Let me look at this. Let me build an intuition. Let me look at what other people say. Like all those kinds of let me make some side projects. Like there's you can use these AI tools to learn so much. That's what I'm saying. Like people can you can use them for good or for evil. You can use them to get better or to do less. Like it's really just up to you and how you use them.

Ali Rohde (28:55) It is, though it feels like you do need this Benedictine self-discipline to use them the way you're saying, of course it's possible, but the incentives are no longer there. So in the past you had to learn these things in order to get these jobs and do this work. Now it's optional and human beings fundamentally respond to incentives.

Josh (29:07) Mm-hmm.

Josh (29:15) Yep. So I don't know, set up a good incentive for yourself. Have somebody watch your screen and, you know, take all the money from your bank account unless you use it properly. What whatever you want to do, just yeah, set it up so that you're doing this the right way.

Ali Rohde (29:29) Can you use an AI to watch your screen?

Josh (29:30) You could. Mm-hmm.

Ali Rohde (29:35) One more question on this topic. he doesn't want me to say that. Where do you think token spend levels out?

Josh (29:45) That's a that's really interesting question. There's a lot of factors that go into it, which is like capacity on the data center and the GPU side. There's how many tasks that are even worth doing. There's how efficient the models are. There's there's all these things. So it's actually really hard to say. In fact, I would almost say that it's better to take tokens out of it and think about like knowledge work. Like, There's a certain amount of knowledge work. Even if you could do an infinite amount of knowledge work, it just isn't worth doing some of it today. Like we just really don't care enough to be willing to spend the money on it. So it's like, how much money can you spend for how much like intellectual work? And there's like a cost demand curve from this. and it's, I forget the name of it, but it's that thing where like when the costs come down, actually there's a lot more demand. So that's what we're seeing. You could probably plot it. Yeah, exactly.

Ali Rohde (30:38) Devon's paradox.

Josh (30:40) So it's Jevon's paradox with intellectual labor. Like as the cost of intellectual labor goes to almost zero, how much of it do you want? Probably a lot is my question. That's why every that's why the AI stuff matters so much. Like I bet many orders of magnitude more than where we are today in terms of tokens used.

Ali Rohde (31:02) I tweeted, is literally anyone working less because of AI?

Josh (31:09) Ha I've never met that person, no.

Ali Rohde (31:11) I've never I haven't met a single person. We love to talk about the three day work week. Show it to me.

Josh (31:18) No, not even close. there are some people who are working less because of AI. That's right.

Ali Rohde (31:22) Because they're fired.

Josh (31:27) Yeah. No one no one who wasn't fired, who's working less because of AI.

Ali Rohde (31:27) We

TOPIC 5 - Microsoft and AI independence

Ali Rohde (31:33) Let's talk about Microsoft. We've spent a lot of time talking about Google. I'm curious your take on Microsoft. Today, Microsoft Build 2026 begins. The information framed it as Microsoft's AI Independence Day, a public test of whether Microsoft can show its own model team, co-pilot stack, GitHub strategy, and developer platform are no longer just downstream of downstream of Open AI. Do they have any chance?

Josh (32:01) I didn't even know they had models. No. Do they? Are they not just wrapped OpenAI models? Like

Ali Rohde (32:03) Mm-hmm.

Ali Rohde (32:07) Yeah, they have their own models.

Josh (32:09) Okay. Are they any good compared to OpenAI and Anthropic models? No, yeah, I don't think so. Yeah. Yep. That's my point. It's it's really just the the three companies, Google and Anthropic and OpenAI, right now. although the open source models are getting better. I was looking at some of them relatively recently and they're they're getting better. Yeah.

Ali Rohde (32:14) Probably not.

Ali Rohde (32:32) Are you bearish on Microsoft over time?

Josh (32:36) it'd be pretty ironic if Microsoft died because of AI on account of them funding OpenAI. Like that would be just really, really bad negotiation. And like bringing them back after Sam almost, you know, like after the out off the brink, like you could why not just let them die? Then the technology exists and you can benefit as a huge company. Like just annihilate your competitor.

Ali Rohde (32:42) Yeah.

Ali Rohde (32:55) They looked brilliant for so long because of that. But perhaps they were building towards their own demise.

Josh (33:02) Yeah, exactly. So I'm not I'm not bullish on Microsoft.

Ali Rohde (33:07) They do have insane distribution.

Josh (33:09) They do with Windows.

Ali Rohde (33:16) Microsoft Suite.

Josh (33:17) right, 'cause everybody uses Excel instead of Google spreadsheets now. Everybody uses Word. I wonder was the last time I opened Word.

Ali Rohde (33:20) Word Excel. You are not the target audience. I think this is something you d dramatically underestimate because you haven't ta touched a Microsoft product in thirty years. Enterprises.

Josh (33:32) Yeah, yeah, yeah. But who are all right, sorry, sorry, that's right. Who are those users? Uh-huh. And what were the enterprises gonna do with their employees again? That's right. And so then after you fire them, guess what else you can get rid of? Your Microsoft Word subscriptions.

Ali Rohde (33:40) Fire them.

Ali Rohde (33:47) Yeah, and Microsoft Word, all of these many of these products are just crap with AI agents.

Josh (33:53) Exactly. They're just in the way. You just want a markdown file, right? So

Ali Rohde (33:56) Exactly. Markdown and then export as PDF, export as a word file, if needed, hopefully not at the end.

Josh (34:02) Yeah. Just post it straight to Twitter.

Ali Rohde (34:06) Yeah.

TOPIC 6 - Google TPUs, compute scarcity, and space data centers

Ali Rohde (34:09) Let's talk about Google. There was an announcement that Google and Blackstone are turning TPUs from mostly internal Google infrastructure into a financed cloud product with five billion dollars committed for five hundred megawatts of AI capacity beginning in twenty twenty seven.

Ali Rohde (34:42) Google and Blackstone are turning TPUs from mostly internal Google infrastructure into a financed cloud product, with five billion dollars committed for five hundred megawatts of AI capacity beginning in twenty twenty seven. Are TPUs becoming a serious alternative to Nvidia?

Josh (35:04) TPUs were already a serious alternative to NVIDIA. You could already Yeah, for Google. But then Gemini is you using TPUs to do AI, right? So Gemini TPU, yeah, doesn't doesn't really make a huge difference. You can run most open source models on TPUs or GPUs these days as well. So they stand to benefit if open source models do a really good job, which is why I keep coming back to like Google. I think long run Google is in a fine spot, right? Like they have good hardware.

Ali Rohde (35:07) For Google.

Josh (35:32) They have lots of money, they have existing distribution, like it's they are fucking it up pretty pretty badly, but still they just have so many advantages, they have to mess it up for so long that I they might they might be able to make it, you know. Yeah, I think also if you look at the absolute numbers, like 2027, 500 megawatts is like not that much, I think, in terms of overall deployment. So it's it's nice to have an alternative to NVIDIA that isn't AMD. yeah. And is it one of these newer the newer ones? I think we'll see more competition there. There's just too much value for NVIDIA to be allowed to capture all of it forever. Yeah.

Ali Rohde (36:13) Do think we'll run out of compute?

Josh (36:16) like in a sense we've already run out of compute a little bit, right? Like Anthropic was compute constrained there for a while until they got their SpaceX rentals. Yeah.

Ali Rohde (36:24) Because they didn't have the contracts in time. Not because the compute didn't exist, like the ability to generate that compute.

Josh (36:32) I mean at large, like if they hadn't got it from SpaceX, it's not like there's a lot of it laying around somewhere. So yeah. So we're we're kind of always I think you can produce

Ali Rohde (36:38) Sure. I guess how much more do you think can be produced? Like there's not that much now because, you know, we didn't expect this explosion in usage, but presumably we can create a lot more. Like, is there a limit to how much we can create and do we expect to hit that?

Josh (36:54) Well, there is a limit. Like there's only so much copper and silicon and you know, all the other rare earth metals that go into it. But those are pretty far away. Wait, like eventually you do hit those kind of limits. Like the thing is you're always hitting limits. Like we recently hit limits with like, you know, the amount of RAM that OpenAI bought, right? Or like the amount of silicon wafers or the amount of like there's just like TSM C capacity. Yeah, exactly. Yeah. So

Ali Rohde (36:59) Fair enough. Do we expect do you expect for us to hit that?

Ali Rohde (37:17) Memory. Memory is yeah.

Josh (37:21) There are lots of we keep hitting different limits. The question is how elastic is each limit? Most of them have like a response time that's actually kind of long. Like to get new silicon wafers. I remember a while ago, like that was the limit. It was like, yep, we just don't have enough wafers like print this stuff anymore. And then but there's only like one or two companies that make them at the right purity and that they didn't want to like overbuild them. And so there's this slow ratcheting up of up of all of the like limiting factors.

Ali Rohde (37:44) So it's slow, but it doesn't sound like you're that worried anytime soon that we will hit limits that we can't build around.

Josh (37:51) It's hard to say. I haven't seen anyone saying that we're gonna hit fundamental limits in the short term.

Ali Rohde (37:57) Except SpaceX. Then we have to put it in space.

Josh (38:04) Wait, that's not a limit. The limit isn't how much space do they take? They take almost no space. They do not need to go to space. We have plenty of power. They take a small fraction of our overall power and almost none of our water.

Ali Rohde (38:19) Ben Thompson had a piece on SpaceX. Very skeptical of the claims. But he said long run compute demand could get so large that space becomes the only remaining place to build.

Josh (38:37) What? No.

Ali Rohde (38:37) That's what I was hedging. Like, he's very, very, very skeptical. I think he's trying to say, like, is there any path to this being true?

Josh (38:45) okay. I'm sure logically, yes. Like a thousand years from now or a hundred years from now, or like some at some point we're like, well, we covered the whole earth with these things. Like there's no more space. And we've also built up and down, and like there's nowhere else to go. I guess you could go to space. But like it's just, it's just not the limiting factor that matters in the short term. It's not space. That's so ridiculous. But like. Value density in a data center, like how much it costs per square foot of data center. Like that's that's crazy. It's crazy dense.

Ali Rohde (39:22) Ooh, someone just ping me this.

Josh (39:30) The only reason it's not more dense is that it's hot. And so you just don't put them closer together so the heat can dissipate. 'Cause real estate is cheap. That's yeah. If you're if you're covering the earth in that, you're making it really hot.

TOPIC 7 - Cognition, Devin, and independent agent labs

Ali Rohde (39:44) Cognition just announced a one billion dollar raise at a twenty six billion valuation for their product Debian. Do they have any chance of remaining independent long term?

Josh (39:58) I mean a billion dollars can go a long way, so I guess so. Yeah. I don't know. when I yeah, all I've heard for Devin is that like they got a bunch of salespeople. They're mostly a sales org. That's it. So seems good to sell AI when people want to buy AI. It's not really a tech company.

Ali Rohde (40:19) What has to be true for an independent agent lab to be durable?

Josh (40:30) Really don't know what the durable agent moat is. Like over time, it just feels so commoditized and competitive. It it's like the definition of a commodity of like, well, it's this substrate where I can tell it to do X and then does it do X? Great. So it's just like it's just about cost. Well, if I was gonna tell this thing or that thing to do it, like how much is it gonna cost? And it's it's just gonna get so commodified. So You can have so then if you think about your competitive modes, you can have economies of scale. That's why I like Google. You can have brand, which I don't think Devin wins on. And can have your like, you know, we know how to do it better type of a thing. yeah, I don't think that switching costs matter as much for the agents, for for this particular type of agent, right? Because if it's just

Ali Rohde (41:15) Switching costs.

Ali Rohde (41:22) Yeah. Turns out developers are very good at at switching products.

Josh (41:27) Yeah, if you just say, it's gonna do whatever I say, then that's pretty easy to switch to something else. That's a easy interface to change.

TOPIC 8 - Memory, company brains, and agent context

Ali Rohde (41:36) talk about memory, not memory in sorry. We're seeing a ton of startups in the memory or kind of company brain space and trying to help these systems learn, like with continuous learning and context learning, because I think right now people are frustrated that they have to go back and forth with agents and teach things again and again. Do you see products in that category as durable long term? Why not?

Josh (42:09) It doesn't really make sense. the amount of knowledge that you have as a person in your head can and that you could write down can be contained in memory on your phone. And so it's a trivial technical problem. Like the the hard problem is getting it out of you and into a digital system. Once it's in the digital system, it's very easy to transfer it to a different form, to export from SQL and put it in JSON or to move it to your from your local thing to the cloud. None of that matters. All the formats can be easily indexed. It you can just put it in all the formats at the same time. It just doesn't matter. And it's going to be easy to export these types of things. I think the whole reason people are doing this is like, like we could lock people in, we could build this platform. It'd be like, have all your memories. Cool. But like it's gonna be pretty easy to get them out in most cases. Or whatever you did to put them in, you could just put them in somewhere else as well. So and similarly, people and companies approaching these products. have a perspective of like, well, why would I like put all my data into this thing that I can't get it back out of? That seems kind of weird. Like what if they go out of business or whatever? So then you want to have it in a way that you can export it. Okay, great. Well then yeah, there's no protection at all.

Ali Rohde (43:25) I think it's a convenience thing. You're right, you could export it, but you know, if I'm using a lot of different systems, if I'm constantly switching between Anthropic and Google and OpenAI, then I want that information centralized somewhere that's steady, especially if it's company information that is not any of those specific providers.

Josh (43:43) Yeah. I think we'll see much better software for this over the next few years. I think one of the reasons, in fact, we're working on some stuff related to this. One of the reasons why I think this hasn't happened before is that in the past, this kind of like data accumulation stuff for individuals is just not really worth like the amount of software infrastructure that you had to create to do it. Like this stuff is always changing, the formats are different. How do we synthesize them? What fields do they have? Like it's a lot of work for an individual person to do. But now with coding agents, it's so much easier to go and have them make some janky data export scripts for your chat GPT histories and your cloud histories and your local cloud code sessions and whatever. Great. Okay, eight one agent for each of those, go find it all, pull it all local, done. So those things are starting to get really easy to make. I think people making stuff like that, that type of technology will let people not have to worry about the stuff being all scattered. You can try out new tools as long as you can pull it back into your like one centralized brain. Great.

Ali Rohde (44:42) So the software should exist, but it's gonna be commoditized.

Josh (44:45) Yeah. It'll just be a bunch of scripts, a bunch of open source stuff, a bunch of small you can have small companies here. That seems fine. But like I don't think there's like a big defensible new tech company for that.

Ali Rohde (44:59) It just sounds like just too early.

Josh (45:03) Yeah. I mean I think it's happening now. That's why we're also seeing these companies, right?

Ali Rohde (45:05) Yeah. I guess you could set these things up, right? Like if I were an individual, okay, I am an individual who wants Codecs and Claude Co to know me better, my preferences better, and learn from iterations better. What should I do?

Josh (45:09) Mm-hmm.

Josh (45:15) Mm-hmm. Mm-hmm. I mean, you can just tell them, like, put all your memories in this file and then have another agent that also sticks a bunch of, you know, data from all of them into some big repo and just make a skill, which is like, if you need to look at previous conversations, look here. If you need to look at my emails, look here. If you need to look at this, look there. Okay. It'll kind of work. It's going to be kind of janky and buggy and everything, but people are working on making it better. Yeah. But it's not, it's not actually going to be that slow. Like you can use, there are good open source tools for like,

Ali Rohde (45:37) And slow.

Josh (45:44) indexing this kind of information, et cetera. So you can give a little CLI that allows the agent to quickly search over this. yeah, it'll just check everything and see what's relevant and then yeah.

Ali Rohde (45:50) Every for every query, it just checks.

Ali Rohde (45:56) What about overlapping, contradicting stuff?

Josh (46:00) It's great. It's amazing. You just dump it all into the L L context, pay for way more tokens, and then it can figure out, that one seems outdated. This one seems like the real new thing. I'm curious about this one. So yeah.