Vibes & Benchmarks Ep 05: Should the government own half of OpenAI?

Ali Rohde (00:02.316) Josh, welcome to episode five. Are you ready? Okay, big news this week. Actually, one of many pieces of big news. OpenAI confidentially filed their S1 to go public. They knew that it would leak, so they just said it. So it's not so confidential.

Josh (00:04.44) Hello.

Ali Rohde (00:24.362) What that means is that this year, three companies will go public at most likely more than a one trillion dollar valuation. That has never happened before in the US. It's happened once globally, but just once. So virtually unheard of.

Ali Rohde (00:50.082) No problem. Okay.

Ali Rohde (00:55.456) OpenAI just filed to go public, which means that this year, three companies, OpenAI Anthropic SpaceX, will go public, most likely all with one trillion plus dollar valuations. That has never happened before in the US, happened just once globally, but just once, so virtually unprecedented. Does this represent the pinnacle of American capitalism, or it's Achilles' heel?

Josh (01:25.376) Achilles heel. It doesn't make sense to have companies IPO so late because this prevents regular investors from capturing any of that upside. Like you can't go up a hundred X from one trillion, right? Google can go up a hundred X or even a thousand X. I don't remember what it started at, but you made a lot of money if you invested in Google or Amazon or one of those companies. You just can't get the same thing here. All those gains are already captured by the early stage VCs.

Ali Rohde (01:52.694) Anthropic though is actually a pretty young company. So it hasn't been that long, but they've grown at an unprecedented rate.

Josh (02:00.366) Mm-hmm. But it still most of those gains have been captured already. Like regular retail investors are kinda getting screwed here by not having the opportunity to invest earlier.

Ali Rohde (02:12.629) That feels like a whole other problem of like maybe you're saying that retail investors should get access to private companies.

Josh (02:19.854) Yeah. I mean, the whole reason this happened was Sarvanus Oaksley from the two thousand and one crash where a bunch of extra rules made it so that companies IPO'd much later. This was in reaction to a bunch of people losing money. So because a bunch of people were stupid and invested in things like pets.com or, you know, web van or whatever the heck it was, then like the there was enough backlash that

The government passed new rules saying like, no, you know, you need to have like proper accounting, et cetera, et cetera. Which in pract which sounds good, like, you don't want people getting scammed. But in practice, this is what it means. In practice, what it means is that only companies that like they just get pushed back so much further when they IPO that you can't really capture the upside anymore.

Ali Rohde (03:02.711) Who says that's the reason?

Josh (03:04.226) That is the reason that they IPO later is there's a lot more like crap that you have to do.

Ali Rohde (03:08.065) I think there's probably a lot of things that feed into that, including the dramatic escalation of private c capital available to these companies.

Josh (03:16.728) That was that was partly there because the public capital wasn't there. Like the reason why you have late stage funds is because you can't IPO. If you could just IPO, why not just IPO? Then anyone in the world can buy your thing, can buy your equity. The reason for late stage VC is those laws. Like, why why would you why would I take money from a late stage VC if I could take money from a random person on the street who doesn't care about anything about whether I make money or not, who doesn't take a board seat, who doesn't do shit? Like that sounds amazing.

Ali Rohde (03:46.423) That's not what going public is like at all. It's very annoying. It's much more administration and oversight.

Josh (03:50.243) Yeah. That's it is more inus administration oversight because of the regulations around it. It didn't used to be that way. That's my point. Like it didn't used to be the same thing. Yes, there's still, you know, shareholders, blah blah blah lawsuits, there's things, but like it's a lot more annoying than it used to be, which has pushed things back. It's just not as worthwhile to do it.

Ali Rohde (04:08.203) I believe there's a floating proposal to change it such that public companies only have to do quarterly earnings and quarterly reports once every six months. So trying to kind of lesser that burden. Does that matter? Will that help?

Ali Rohde (04:25.173) Anthropic doesn't feel like it quite fits in this box. It's actually quite a young company. It hasn't even existed for very long.

Josh (04:32.43) I mean it's existed for a good what five or six years at least now, right?

Josh (04:39.789) Yeah.

Ali Rohde (04:40.001) We Google it. I think maybe four or five. Was it started? No, it wasn't started before 2020. It was probably started like 2021.

Josh (04:48.684) I'm pretty sure it started in twenty twenty.

Ali Rohde (04:50.871) I'm just gonna look it up. Ha. I mean you know the founder, so you shouldn't know, but

Ali Rohde (05:01.345) Twenty January twenty twenty one. All right. So we were both right.

Close enough.

Still not that long.

Josh (05:10.378) Still not that long. It's it's it's rare for a company to grow to a trillion dollar valuation in that short of a time.

Ali Rohde (05:15.829) Are you saying you will not purchase equity yourself in anthropic and open AI when they go public?

Josh (05:22.83) Well, I don't know if I'll purchase more equity, but I might. We'll see.

Ali Rohde (05:27.595) So you don't like it, it's not good for retail investors, but you yourself might invest in this.

Josh (05:34.19) I might invest, it depends. It most of the other places where you can put capital right now are kind of trash. So it's it's not a statement about anthropic or open AI, it's a statement about the rest of the world. How much more gold or bullets do I want, you know?

Ali Rohde (05:51.254) Last week we talked a lot about policy and about people by and large shitting on or ignoring Senator Bernie Sanders' proposal to give the American public what did he say, like fifty percent stake, something crazy like that, in these companies that are going public.

Josh (06:05.432) Mm-hmm.

Ali Rohde (06:10.825) OpenAI has actually been putting out some other proposals and they've actually been talking to the White House, Sam Altman in particular, about the government possibly taking a stake in OpenAI when it goes public. Is this a good idea?

Josh (06:23.246) Mm-hmm.

That is totally different than taking 50% of the company. Hey, why don't you give us money at our super marked up, you know, IPO is not the same thing as, aha, we're taking 50% of your equity. These two things are this is just you being an investor at a super high valuation. That sounds great for Sam. Like, why does he care whether it's the government or Saudi Arabia or your uncle on the cap table?

Ali Rohde (06:48.907) Well, I think the US government wouldn't pay him for those shares or pay the company for the shares.

Josh (06:52.258) Are you sure? I bet that's what he's asking for. Take a stake like

Ali Rohde (06:57.705) No, no, no, no, no. The idea is that OpenAI could donate that equity. In fact, he's talked specifically about like this previously a few months ago, they put out some policy proposals around this like public wealth fund. So it would be open AI donating equity to the US government to cede this public wealth fund.

Josh (07:02.072) okay.

Josh (07:15.192) That's closer. Although I did see some calculations by someone on Twitter about how much would that actually be worth. It turns out even if that grew by some ridiculous amount, then it would it would only like barely offset the increase in the deficit from like the big beautiful bill. Like it just it just would not have a material impact on the like the federal budget.

Ali Rohde (07:38.155) That feels fair, but also more of a statement about like the government budget and government spending than anything about open AI.

Josh (07:43.063) Right, but right. But my point is like a few billion dollars to a trillion dollar company doesn't make a huge difference. And then being able and it's almost bad because then being able to say, like, you have a stake in it, blah, blah, blah, is like yeah. Yeah. I yeah, I think this is this is not something that we should do right now. I think you should say, Nope, we're gonna take fifty percent when we feel like it.

Ali Rohde (07:56.022) Yes, exactly. That's why I think this is brilliant for open AI.

Ali Rohde (08:09.591) Josh, are you even a capitalist? So the should the government be able to take fifty percent of a company just because it feels like it?

Josh (08:15.962) that is literally the purpose of the government. Do you remember Bell Labs or Carnier Steel? These things were not like, you know, like let's just buy into Carnier Steel or Bell Labs. Like, good thing we all have Bell Labs shares today. That's so great, because we wouldn't have had a fucking internet if we had done that.

Ali Rohde (08:36.895) I think my guess, and I don't know the specifics, push back if I'm wrong, is that the government ceded those companies and provided funding for those companies. They weren't just given equity in those companies.

Josh (08:46.088) No, they I'm saying they broke them up. They weren't given equity. They destroyed them. They took Bell Labs and they're like, No, you can't be this. You're now broken into smaller companies. Like that's not the same thing at all. Like that's the proper role of government here is to be like, you have a monopoly, we're going to destroy you and turn you into a competitive zone.

Ali Rohde (08:49.419) Hmm.

Ali Rohde (09:02.677) Okay. So you ideally the government breaks them up, but in the absence of that, the government should take a fifty percent stake in them.

Josh (09:08.482) Giving the no, this is the well, I don't think that's good. I'm saying taking a stake is bad because then you end up having conflicting incentives that prevent you from going and destroying the value, right? Like if a significant fraction of government income is coming from their shares accumulating, then you can guarantee that they're absolutely not going to go break it up or destroy this, right? So that's a terrible idea to do this. Yes. Yeah.

Ali Rohde (09:29.387) That's why it's brilliant.

I think from a PR perspective, it is one of the most brilliant things Sam Altman has proposed.

Josh (09:38.254) it's so it's so ridiculous. Don't don't do this, people. It's crazy. Stop, please. Can you not see where this is going? We can all see it. Stop. Let them let the government be responsible for doing what it's supposed to do.

Ali Rohde (09:52.545) I'm not so bullish that'll do that anyway. So you might as well take the money.

Josh (09:54.957) I think it eventually no, I think eventually it does. Like eventually it's like, well, we don't have any other way to get money, so let's take fifty percent or let's break it up. Okay, fine. That's the natural proper course of this, not well now we're all on the same side and like, you know, our our fake money is going up, so our taxes are going up.

Ali Rohde (10:15.551) In another proposal they put out last week, OpenAI said its broad goal was to broaden power, make it less concentrated. They stated three goals specifically. One, build an automated AI researcher, two, accelerate the economy, three, give everyone on earth a personal AGI. Does everyone on earth need a personal AGI?

Josh (10:23.874) Yes.

Josh (10:36.142) Yeah.

Before we get to that, those first two, how are those distributing power? Is everyone gonna have a recursively self-improving AI? Is everyone going to be gaining from this, you know, radically accelerating economy? Hint no on both of those. Okay, so now does everyone get their own personal AGI? What does that mean? That just means like a chat GPT subscription. Like, what? Yeah, what? Madness.

Ali Rohde (10:57.399) That the government pays for.

Ali Rohde (11:05.473) So no, people don't need their own personal A VA.

Josh (11:05.484) I that that article seemed ridiculous. I would encourage people to read it. There are zero specifics, and why would you trust OpenAI to stand by any of their commitments, even if there were specific commitments? They have stood by zero of them over time. we're gonna be open. Ha, never mind. we're gonna be nonprofit. Ha, never mind. Like what? What are we talking about here? That's why they don't even bother with specific commitments anymore. It's just yeah, it's pointless.

Ali Rohde (11:31.755) Let's talk about the biggest news of today. Claude Fable, i.e., Baby Mythos, was just released.

Hopefully you looked at the model card. You know, obviously it's too early.

Josh (11:47.862) Mm-hmm, mm-hmm.

Ali Rohde (11:48.437) Yes, thank you.

In other big news this week, Anthropic just released Claude Fable, aka Baby Mythos.

Is this a real model step change? Can there be real model step changes anymore?

Josh (12:06.754) Yeah, I mean, it seems materially better. There can be model step changes and there will be, at least for the next few families of or like few major steps of models. At some point, regular people will stop being able to tell the difference. I think that probably a lot of people asking your regular questions between Mythos and the next one are not really going to be able to tell. but there's still in the more complicated tasks, there's a pretty big difference.

Ali Rohde (12:31.115) Hundred percent. I saw a quote that I loved about Fable. Using this thing for regular knowledge work is like squashing an ant with a rocket launcher.

Josh (12:42.531) Yeah.

Ali Rohde (12:43.915) I'm actually not sure everyday users will see any improvement with this.

Josh (12:47.672) Well, it can count the number of R's in strawberry.

Ali Rohde (12:50.751) Okay. Are you sure that you test it? Is that your first eval?

Josh (12:52.822) Yeah. Yeah yeah. I saw some of the No no some of the some of the highest ranked tweets on the release thread were like, it's got that right. And it also can answer questions like if the car wash is fifty meters away, is it faster to walk or to drive? It's like or should I walk or drive there? So so you know.

Ali Rohde (13:07.381) Yes. I've seen this classic question. Wait, does that mean the last model didn't get those right?

Josh (13:14.646) Yeah. Yeah, exactly. So that's why I'm saying the next release you might not be able to tell. We don't have any like really dumb failure modes anymore. It can draw a pelican on a bicycle now, you know. Like it can pass all these basic tests, like it can do sort of the basic things. So you have to be paying attention if you want to notice a difference in the future models, I think.

Ali Rohde (13:34.581) It's interesting to start reading the responses to it and how our standards have changed. You know, it used to be like, can it code? And now the question is, can it automate multi week R and D?

Josh (13:47.384) Mm-hmm. Mm-hmm.

Ali Rohde (13:50.241) Can it?

Josh (13:51.599) I mean, it depends on how good of a researcher you were. If you were a terrible researcher, it can. If you weren't, it can't. Right. So that's what the the difference is now is in these long horizon tasks and these much more complicated tasks. So there will be still a difference for those tasks for the next few models.

Ali Rohde (14:08.225) So do you plan on using this model?

Josh (14:10.508) Yeah. But I'll probably use it for the more complicated tasks, right? Like I'm not gonna necessarily use it for every single thing I'm doing.

Ali Rohde (14:19.339) Yeah, I saw someone saying it's about five percent better and two X the price.

Josh (14:22.902) Yeah. Yeah. I think a lot of the core stuff that we do might still be driven by like four point eight. but yeah. So I'd probably use this if I'm doing like a big complicated refactor, doing some security hardening stuff, like cases where I'm doing this one off, I really care about it, it's really complicated, versus, you know, make me a dashboard for blah. Okay. Yeah. Opus is gonna be fine at that.

Ali Rohde (14:31.905) So then when exactly do you use this?

Ali Rohde (14:50.241) I'm personally quite sad because I use the Max plan. I haven't been kicked off yet. And Fable is included in the Max Plan for free only until June twenty second. Yeah, exactly. You won't see me for the next thirteen days.

Josh (15:00.846) Yeah, get all your work done before fourth of July.

Yeah.

Ali Rohde (15:09.567) Anthropic also put out a piece about recursive self-improvement called When I AI Builds Itself.

Ali Rohde (15:18.077) They part of what captured everyone's attention about this piece is that they talked about an option to pause development. Interesting timing. Right after they file for an IPO, they talk about this option. They explicitly rule out any one lab doing this. They say that's ineffective. So don't worry, anthropic is not pausing its work right before its IPO.

Ali Rohde (15:48.299) Does this matter? Like it t it's to me it's like this will never happen, so why are we talking about it anyway?

Josh (15:53.143) No, this this matters a lot. The purpose for this, this is similar to the open AI thing that we were just talking about, like giving a small percentage of the government. The reason for this is to get everyone on the same page about, like, it's dangerous. Like, yeah, we should all pause. What does pause mean? Well, it means we shouldn't be making new capabilities. Well, who's we? Well, we means only anthropic open AI and uh-huh. Yeah.

Ali Rohde (16:17.045) We means we're really tired of having to ship a new model every three months. It's really expensive. The compute for such models are really expensive to serve. Let's stop here.

Josh (16:22.594) We want regulated We No, it means we want regulated barriers. We want models that are too dangerous for regular people or for open source models. Look at those dangerous Chinese models. Look at those dangerous terrorists using their open source cybercrime models. no, we can't let them make something as good as our models. We need to make it so that these are regulated and only the big responsible people can use them. Then we don't need to make them better. Then we can just chill here and just take our ninety percent margins. So

That's where this is going. not that it's not about pausing. It's about preventing them from having to continue to compete. It's an anti-competitive measure. This is literally like the thing. This is collusion. This is like the the thing that anti-competitive laws are literally against.

Ali Rohde (17:06.719) You talk about a pause, and then we'll talk about a pause, and then we'll talk about a pause.

Josh (17:10.988) Yeah. No, a pause is just all of us have decided to get together and not compete. And then to make regulation to prevent other people from competing. That's fucked up.

Ali Rohde (17:20.695) Still seems unlikely to happen.

Josh (17:22.656) No, it seems very likely over time to happen. That's like, because how do their businesses make sense without this? Like they need to push for this. They need this story to be true because otherwise, that's what you're just talking about. The models advance, but the open source models are right behind them. And so if they don't have this like regulated barrier in between their models and the other models, if all the open ones are out there, then why would you pay for anthropic or open AI? You'll just use the really cheap free ones. Then their trillion dollar evaluation doesn't make sense without this.

Ali Rohde (17:25.206) Hmm.

Ali Rohde (17:53.986) They also talk a lot about recursive self-improvement, and they lay out three possible futures. One, the trend stalls, but today's AI capabilities are widely diffused. Two, AI labs continue to see compounding efficiency gains. Three, AI systems themselves become capable of full recursive self improvement and begin building their successors. Which of those do you think is most likely?

Josh (18:20.82) all three of those, and it doesn't matter. The first one of like the capability is sort of flatline is about event, like the question is when. and like they want we were just talking about, they want that to happen as soon as possible, right? Like they want this to be the case that we don't have to make bigger models. So it's the question isn't when does it happen from a modeling perspective? The question is when can we get this to happen regulatorily, right? Like, cause they don't want to keep going. That's too expensive. You can't keep making a hundred billion dollar training runs. So

That's one so that can still happen. And the other things can also happen. You can have most AI research done by AI systems and you can have compounding returns. But if the rate of compounding is low, if it's 1% per year, then it doesn't matter. And if you have models primarily doing recursive self-improvement because now you don't have to pay your really expensive researchers anymore, also great. Also doesn't matter, right? If you're improving by two percent per year, it still doesn't matter. It's like

All these things can happen and it still doesn't matter. The thing that really matters, the thing that people cared about before for recursive self-improvement, is making things that recursively self-improve at a rate that is meaningful and that is like accelerating. That doesn't seem super likely to happen at a like super accelerating rate. It might go a little bit faster, but as they point out in one sentence in that whole thing, like, then it would only be bottlenecked by compute. Well, guess what? Compute is already a huge bottleneck, right? Like we just don't have enough chips even to do inference.

So if you're bottlenecked by compute, great. All of us are bottlenecked by compute. We're all bottlenecked by physical reality. Like that doesn't change the game very much.

Ali Rohde (19:58.306) So you're saying we'll probably never get to this like very accelerated recursive self improvement.

Josh (20:04.482) Yeah, I don't think that you get these accelerating gains in a way that people were afraid of in the past because there are just too many physical constraints on it. Like you just need more B two hundreds and there just aren't that many. I wanna buy.

Ali Rohde (20:16.439) And you need that to build those models. So it's not that like it models can never get that good. It's just like we don't have the physical materials needed to get there.

Josh (20:25.666) Yeah. I wanna buy more Raspberry Pis. They're like mostly sold out and the ones that have enough memory are like, you know, super expensive now and like memory prices are not gonna come down at least for another year or two. It's just yeah, like there are other bottlenecks.

Ali Rohde (20:44.215) Apple just had their WWDC, their annual conference, actually Tim Cook's last as CEO.

They say Siri is going to improve.

Josh (20:58.606) Okay. Isn't that because now it's powered by OpenAI or by Google or something? yeah, sorry. Yeah, yeah. Yeah. Well I mean that makes sense that it would improve Gemini is better than Siri.

Ali Rohde (21:03.489) Google. Google won that. Yeah.

Ali Rohde (21:14.293) Is it too late?

Josh (21:17.0) something interesting about Apple is that they're betting on the like local, yeah, like people having their own hardware. Like you could run local models, et cetera. So I don't think I think that they could end up in a really good spot sort of by accident. Like I think they were just totally wrong and missed the boat on in terms of AI and like probably they should have been making data centers and getting all this stuff. or they should have been making much better local hardware and investing in open source models, but they just sort of didn't do anything.

But that might end up working out because now they haven't wasted a lot of money on it. And so if we do get a world where the open source models are relatively good, then like the best thing for running open source models today by far is a MacBook. So that's actually, and people are buying Mac minis for run OpenClaw, right? Like you can see a world where, like, if local hardware is the way that most AI inference runs, who is the best local hardware? That's Apple. That puts you in a great spot as a company. So maybe they're in a good spot, but I think it's mostly through luck, not through.

ingenious strategic planning.

Ali Rohde (22:17.537) Fair enough, though I will say them doubling down on this and doubling down on hardware. For instance, picking the new CEO, John Turnus, who is a hardware guy, I think you can say that was a good strategy. Once you already fucked up, then you like doubled down on your strengths.

Josh (22:29.046) Yeah, yeah. I think from from from here. Yeah. Yeah. I think up up to this point, it doesn't seem like they're being super strategic about it. I think at this point maybe that's a coherent strategy that hopefully they follow. I would be happy with that, but we'll see.

Ali Rohde (22:44.491) You said running models locally is the best solution, if you can. Why is that?

Josh (22:51.064) there's a lot of reasons.

Well, in a data center, you have to get all this like complicated infrastructure in order to dissipate the heat. If you run the thing on your laptop, the heat just dissipates. It's fine. Your house is like a little bit hotter. So there are natural reasons why you'd want to distribute this stuff anyway. Similarly, if something goes wrong in your laptop, you take it to the store, it gets fixed. Something goes wrong in a data center, you have to have a full-time person who goes there and swaps things out. And there's all this there's all the extra complication to like run it at scale. You're not really actually buying a lot. Like at the end of the day, you really just need a bunch of memory.

And some electricity. So if you can get enough memory and compute in your laptop, which we probably will be able to do in the next few years, then that's just like technically a really good way to make these models. Separately, for like privacy and security reasons, you also probably want to run them locally. Then none of your data leaves your device. Like it's just less, you know, less concentrated. There's less attack surface area. Like Apple's known to be relatively good from a security perspective. So

These are all like pretty good reasons to do this. Also, you're not paying for the model at all in that case, right? The model will be open source. So you pay almost nothing to the model provider. You pay some money to Apple. Okay, you were paying money to Apple anyway for your laptop. All right, great.

Ali Rohde (24:08.343) Are there any use cases or important use cases that the ability to run more things locally will enable?

Josh (24:17.824) In theory, it's a little bit faster. You don't have to make the extra round trip to the network. So you could imagine some agentic things being like a little bit better to run locally. you also don't need to necessarily have to pay for like an extra server that serves as a computer for the AI agent. You already have your computer right there. So that's another thing that kind of fits pretty naturally with with running it locally. similarly for running or for like storing your data.

You already have a good place to store it. It's already backed up to the cloud. All that stuff kind of makes a lot more sense locally.

Yeah, probably for gaming there's some interesting things there.

can imagine there being some enterprise reasons why they might want more security safeguards, et cetera, as well.

Ali Rohde (25:16.587) That'll make sense. It also feels like we're kind of going in the opposite direction, at least right now, in terms of, you know, everyone using frontier models as much as possible, and anthropic and open AI and others kind of building managed agents. So more and more things run on their servers.

Josh (25:31.875) That's the direction they want things to go. I don't know that there has been a big uptick in like the usage of say cloud code web though. Like most people that I know don't use cloud code web. Like if they do run remote agents, they don't really use the first party things. At least no one that I know. Yeah. And I would say that the usage in terms of tokens for open source models has also grown a lot, especially by startups. especially I've seen a few now that have switched off of models for cost reasons, right? Yeah. So it's not that it's all going in that direction.

Ali Rohde (25:46.634) And you don't

Ali Rohde (26:03.595) Let's talk about SpaceX.

Josh (26:05.9) And what what ridiculous thing did they do this week?

Ali Rohde (26:10.145) They're preparing to IPO as soon as possible. One thing that's interesting about their IPO is that they are at least reportedly reserving way more of it for retail investors. Normal IPO retail allocation is around 5 to 10%. This will likely be 25 to 30%.

Josh (26:30.23) Mm-hmm. That makes sense. They need unsophisticated investors who can't look at the price and be like, wait, so you have this much revenue and you know, this much cost. So how is that worth a trillion dollars? but you're going to space. That's so exciting. and you're doing AI neat. What are you doing with AI? You're selling your hardware to companies that actually do AI? Huh. Doesn't really make sense, but it's AI, yay.

Ali Rohde (26:53.449) AI space Elon the killer combination.

Josh (26:56.79) It's just a weird marketing company. It's amazing as a marketing company. It has nothing to do with business. Yeah. Now, there's one there's one case that I can see. Like, if yeah, the the bull case here is that, you know, I saw Elon talking about like, we're gonna make our own chips or whatever. Someone should fucking make some better chips and compete with NVIDIA and TSMC and everything. So if you did Yeah, so if you did that, then I can imagine things working out. The current business, most of the things that they're talking about make no sense. But

Ali Rohde (27:06.763) Tell me the bull case.

Ali Rohde (27:18.187) They have the scale and cash to do it.

Josh (27:26.7) That's one of those things sort of like Apple being kind of lucky and then, okay, great. Like if this is their plan is to like raise a lot of money pretending to do other things and like, haha, never mind, we're gonna make a chip fab. That's fine. That's a great, brilliant plan. That's not what's in the prospectus though.

Ali Rohde (27:42.962) I disagree that that this is similar to like Apple just getting lucky. I think this is what Elon does. He's an amazing marketer. He buys himself time and optionality to do to figure out what works. So, like, is he putting data centers in space? Of course he's not. But this is gonna help him piece together the AI and space story to do that IPO for $1.23 trillion to get all this cash that he can then probably figure out.

To put to some good use, whether that's acquiring companies or building chips.

Josh (28:10.232) That's what I'm saying. That's yeah. That's what I'm saying. Is if if we're talking about doing something that's not in the prospectus, okay, cool, fine. Like that could work out, but it just isn't what the current like that's just not what we're talking about here. And that that's fine. That's a fine strategy, but it's just not what you're saying you're gonna do. It's also actually a really interesting. It's

Ali Rohde (28:30.377) I think people know that. I think people investing in this are trying to invest in Elon or whatever Elon does.

Josh (28:37.92) It's it is interesting because this IPO is also unique in that the like voting shares are quite nonstandard. I remember seeing something about like e exactly, like that's very strange for a trillion dollar company.

Ali Rohde (28:48.703) Elon has like complete control.

Ali Rohde (28:54.634) Is that a good thing?

Josh (28:56.478) it's actually a good thing in this case if we're talking about them being competitive, because then they can be like, ha ha, never mind. Now we're doing blah. which is the only way they're actually gonna make any money because they're not making any money from launching things in space. They're not making any money from Starlink, they're not making any money from putting data centers up there. So they should do something that makes money.

Ali Rohde (29:13.025) I like this chip's idea. I hadn't thought about it before, but I think what Elon is great at is hardware. He is arguably not very good at software.

Josh (29:20.942) Mm-hmm.

Ali Rohde (29:23.233) This could be the plan.

Josh (29:23.49) Yeah. Yeah. it is difficult to make a chip fab. so good luck with that. I wish them all the best though, 'cause it'd be nice to have some competition.

Ali Rohde (29:33.953) We've talked before how Elon has this reality distortion field, similar to Steve Jobs. If someone has that, do you just back them no matter what?

Josh (29:43.924) No, I think didn't Elizabeth Holmes have that? Didn't you know Sam Bankman Fried have that? Like there's plenty of people who like

Ali Rohde (29:49.558) Hey, Sam Bakeman Freed would have been the most legendary investor in the world had they not thrown him in jail. Cursor seed, anthropic seed series A, maybe series B.

Josh (29:55.227) It's true.

Josh (30:03.406) Yeah, it's true. but he ran his whole hedge fund out of a Google spreadsheet and made up all the numbers for everything. So not really a legendary investor. But anyway, magnetic personality does not mean auto-invest. And I would almost say that it kind of it's a little bit anti anti-correlated in that some.

Times you have people like this where the thing is not real and there's a lot of promises, et cetera, et cetera, and then it all sort of falls down. Like we're not we're not out of the woods yet. There's no guarantee that, you know, SpaceX ends up being worth a trillion dollars. It could basically zero out, right? Like, yeah.

Ali Rohde (30:48.279) I think if you had backed Elon generally, you've done very well.

Josh (30:53.454) Yeah, but if that's right. If you had backed Elon Genley, you would have done well. But if you had backed every person who seemed charismatic and, you know, like they had a sort of reality distortion field, you all you may have done relatively poorly. So there's a survivorship bias here.

Ali Rohde (31:06.751) Maybe, but I think we talk about this, you know, when looking at companies and founders. I think we agree it helps more than it hurts. Like it's almost like you know, it's like that the the usual curve where like you want a lot, but not too much.

Josh (31:20.446) Yeah. yeah, I mean you definitely want s some of or some of this can be helpful and good. it's probably bad if you believe that, you know, fruit is gonna cure your cancer or whatever Steve Job believed. Like so there are downsides to this as well.

Ali Rohde (31:45.399) Drew Houston announced that he was going to be stepping down as Dropbox CEO after 19 years of running Dropbox.

Josh (31:53.752) Yeah, his companies old enough to vote.

Josh (31:59.031) No, you can vote eighteen.

Ali Rohde (31:59.24) no, we can vote when we're 18, right? Okay, yes. It just can't drink.

Josh (32:03.618) Yeah.

Ali Rohde (32:06.005) Do you expect to see more of these resignations?

Josh (32:12.151) I hope so. Yeah. I don't I don't think I would have yeah. Well, I don't think I would have wanted to continue running Dropbox. I don't think Drew was super happy running Dropbox. I think he's excited to do something new, especially with all this AI stuff happening and and things changing so much. So it does seem a little bit weird to like grind on your, you know, tax accounting software startup thing through the singularity. It's like I guess so. Yeah.

Ali Rohde (32:14.005) I hope so. Why?

Ali Rohde (32:38.175) Is there that much opportunity in AI? Maybe this is a broader question.

Josh (32:41.248) No. No, actually, this is an interesting question. I don't think there's even that much opportunity for starting a new company or even for doing that much related to it. But people want to be like cultural witnesses. Like they want to be at OpenAI and like see the new model and like be part of history. I think that most people work there for that reason, more so than anything else. Right. Yeah. People want to work on AI and build this stuff and be involved in it just because it's what's happening, not because it's strategic or irrational.

Ali Rohde (33:10.647) What is strategic or rational today?

Josh (33:17.068) It's kind of tricky to figure out since things are so complicated. But I don't know that it's it's definitely it's not clear that like, you have to work at, you know, one of the biggest companies or you have to work on AI. There are other things that you could do. it depends on what your goals are.

Ali Rohde (33:34.027) Maybe we'll see fewer other huge companies, but a rise of S and Bs. No.

Josh (33:39.062) Also seems unlikely.

Josh (33:43.522) Yeah. Maybe off ball of SMBs.

Ali Rohde (33:47.479) Fall of S and B is

Josh (33:49.453) Startups though. Yeah, like small it's like start like things like anthropic, right? Like what we're seeing, cursor anthropic, like we're seeing small companies grow it to be really, really large really quickly. And what that means is like it's not about being a medium sized company and growing revenue ten percent per quarter anymore, right? Like then you just die in the singularity. Like if something is just going exponential next to you, then you become irrelevant.

Ali Rohde (34:17.313) So the path forward is start something, keep it lean, lottery tickets, and then do everything you can to grow exponentially and super fast. And if your company is not growing exponentially, then you might as well quit.

Josh (34:20.589) Lottery tickets.

Josh (34:32.706) That's pretty much the game theory today, yeah.

Ali Rohde (34:36.684) Leak.

Josh (34:37.474) Yep. Lottery tickets as well. Yeah. Buy lottery tickets. Yeah, you know, make or join a startup. That's just a lottery ticket as well. Yeah.

Ali Rohde (34:48.087) Startups were always lottery tickets. It seems like you think they're significantly more lottery tickets today.

Josh (34:50.316) Yeah, there's so many more founders now.

Yeah. I think there's there are probably an order of magnitude more founders now than there were even a few years ago. It's it's crazy.

Ali Rohde (35:04.405) Outside of engineering, what do you think it takes for AI to go mainstream?

Josh (35:11.502) sorry, can you ask the question again? I'm not quite sure what you mean.

Ali Rohde (35:15.977) Outside of AI and then hold on try.

Ali Rohde (35:22.623) Outside of using AI for engineering, what will it take for AI to go mainstream for other uses?

Josh (35:32.056) Think that AI is going to be useful for people because it does engineering, but they're not thinking about it as engineering. Like if you're telling something, you know, go find all these companies that fit these criteria and it makes a script, like you don't care as long as it found the companies that meet those criteria. Right. So I think it's more about the accessibility to the engineering than it is about it actually doing something fundamentally different.

Ali Rohde (35:57.132) I agree when I'm coding, I'm not really coding, but every underlying everything is code. But I'm not sure we don't need something else for it to go more mainstream because yes, the capabilities are here. You as a marketer can use codecs to build whatever you want, but people aren't doing it. And I'm not sure that making it marginally easier will get them to do it.

Josh (35:59.791) Mm-hmm.

Josh (36:04.172) Mm-hmm.

Josh (36:18.286) Mm-hmm.

Josh (36:25.228) people are doing it though, right? There are some marketers. Yeah, yeah. But that but this goes back to the power law thing we were talking about about lottery tickets. Those people are going to be fine. The same thing we talked about with unemployment last time. It's about like a small number of people capturing a much larger fraction. So that person that is the marketing person who's like, how can I automate what I'm doing? How can I like make things at scale, et cetera? They can take their marketing agency and grow it 10x over the next year. The marketers that don't do this stuff and just kind of like keep doing the same thing as normal, then get surpassed by this, like,

Ali Rohde (36:27.188) Very few people.

Josh (36:55.072) much better, much faster way of doing things. So it doesn't matter that most people aren't using it because most people will be unemployed. The people who will be employed are the ones who are using it.

Ali Rohde (37:07.649) So companies that are, you know, mostly salespeople or marketing people, either they'll find some really good people and promote them and write on them or they'll just die as a company.

Josh (37:20.334) I mean, it depends on how important sales and marketing are to you. Most companies, sales or marketing is important, not both. So if you have bad marketers and you're a sales company, meh, that's fine.

Ali Rohde (37:31.191) Have you felt your own demand for software growing substantially?

Josh (37:35.18) Yeah. I have a a larger and larger backlog of things that I wanna build.

Ali Rohde (37:39.211) See are you building any of them?

Josh (37:42.41) yes. although I'll be building many more of them soon. It's like building up like layers, right? You have to like build this thing so you then can build the next thing so you can build the next thing. There's just

Ali Rohde (37:52.725) What layer are you building now?

Josh (37:55.063) our current product that we're working on is about making it easy to run AI agents over a longer period of time and with their own computer and in their own sandbox, making this much more accessible. Like now that I have that working, now I can do things like make an AI agent that's dedicated to watching the GitHub issues on one of my GitHub projects and making sure that it responds really quickly or flags things for me or makes new PRs when when people flag issues, et cetera.

And that makes it a lot easier for me to make new projects because now I don't have to spend all the time maintaining the old ones, or it becomes much more efficient to maintain them.

Ali Rohde (38:26.507) You're building you're building this infrastructure and then this system and then you're gonna throw your to-do list, your backlog at it.

Josh (38:34.828) I mean, yeah, once I can once I have a thing where I can just give it like, hey, here's the idea, like go off and make it. And then, you know, check in every once in a while, see what it's doing. Like once it's developing things a little more autonomously, you know. Once there's a development loop. my god, there was this whole thing on Twitter this week about loops. Okay.

Ali Rohde (38:49.907) I would that was literally my next question. I was like, let's talk about loops. I think I think it was kicked off by Boris Chesney talking about his workflow and how it was all about loops. And now there are like entire startups that have already popped up to help you build loops.

Josh (38:56.429) Yeah.

Josh (39:05.33) No, why we're in the dumbest timeline. This is absolute this is like bottom tenth percentile timeline.

Ali Rohde (39:15.051) Explain.

Josh (39:16.6) There's four loops. There's woo loops. We've known about them since probably Alan Turing-ish time. Like there's just what are loop loops? I mean, okay, there's two things that are stupid here. One, we've known about loops forever. two, loops don't magically solve anything. They just make you waste a ton of tokens if you don't know what you're doing anyway. So most people probably shouldn't be using loops.

Ali Rohde (39:37.311) Okay, but for the people who do know what they're doing.

Josh (39:40.599) some tiny fraction of people. Yes. We're talking about the like point one percent of the point one percent here who should be using loops. And then also it's gonna be super expensive. So you have to have something that's worth putting all those tokens and all that money into. There's very few places that are worth spending that.

Ali Rohde (39:48.289) Sure. Yes.

Ali Rohde (39:58.934) When you were talking about your own setup for your own workflows, it sounded like loops were a key part of that.

Josh (40:04.962) Yeah, but the reason that I'm not running those today is that they require a bunch of infrastructure to make the loops able to do something more useful, like s enough more useful that it's worth paying for it. Right. Like I can run it in a loop. In a loop, it will the agent will eventually make something that's like a little bit better. But if I spent 10 or 100 X on it, it's not worth the like the extra value is not worth the incremental cost, right? So the loop is only worth it if they can make progress that's worth paying for.

And so in order to get that to happen, you need to have like a goal that's worth optimizing. In order to do that, you need to like set up a bunch of scaffolding and infrastructure and things to like make it so that these systems, which are not perfect yet, can actually make meaningful progress on something you care about. So then you have to think like, okay, what do I care about? How would I know? How would I measure it? How would I get the agent to be able to verify that, et cetera, et cetera? So one loop that we have, for example, is like, does the build for this product work?

It's annoying 'cause it takes like, you know, ten, twenty minutes to like build the install or test it, blah, blah, blah. Okay, awesome. That's a great place for a loop of like overnight one of our engineers is always like, Okay, you know, go fix up the build, go go make it work. And then overnight it like slowly turns. By the time you wake up, the build is all fixed and all good. Okay, great. so that's a perfect place for a loop, but it just requires a lot of infrastructure.

Ali Rohde (41:19.831) That sounds like a loop that you're building now that is can that can fit with your existing infrastructure and can work with existing capabilities. What else could this enable in the future?

Josh (41:31.943) so when the capabilities get better, then you won't have to be quite as good at defining the loop. And so you might be able to do something like, you know, while true, make it good. Okay. I mean, I guess. Yeah. But at that point, what are you doing there? And you're probably you probably don't have a job anymore if your job is to say, while true, make it good. So that's kind of the problem. Is like if loops work really well, then what is the point of people?

Ali Rohde (41:43.073) That's the dream.

Ali Rohde (41:59.754) If I were an engineer wanting to now desperately build loops because Boris Chesney said to, but I want to do it right, what are your recommendations for how to do that?

Josh (42:10.562) Yeah, if you want to build loops right, then just make regular software. Like you can have a terminating condition, like okay, loop until when? How many times? Like just get into the details of it. It's just a regular, it's just a regular. Yeah. It's just regular software. Is the build passing? No. Is the test passing? No. Keep going. Right? But what the test passing is like a thousand lines of complex, you know, bash garbage and

Ali Rohde (42:22.005) I need you to be more specific because you mentioned a loop that you guys have built and it wasn't just like

Josh (42:39.426) GitHub CI actions and all this other stuff, just engineering stuff. Yeah, a lot of it was made by an agent, but like it's just boring, detailed stuff to tell, like, are you making progress on this? Right. Another loop, like, make the tests good in our system. But it's not make the test good in our system. Instead, there's a skill that's about how you identify problems. And there's another one for like when does this run? And then how do we review the things? How do we check them? How do we have the agent check them? It's just like a lot of.

detailed machinery and thought that goes into the iteration.

Ali Rohde (43:11.681) This feels just like earlier AI when you had to speci like remember when prompt engineering was a whole discipline and now it's like it's fine, you just put in whatever you need and so that's maybe this is the next step.

Josh (43:14.456) Mm-hmm. Yeah. Yeah.

Josh (43:23.308) Yeah, but this is the last step though. Because this is the last one. Because if you if you have something where it's where you can just say, make it good. While it's not good, make it good, then like what else are you doing? Like actually, yesterday was the first day that we had a PR that was generated where the change log was so long that I couldn't read it. Like the amount of changes created were too much for me to process. Like I'm a very fast reader. It was

Ali Rohde (43:26.133) The next and last step.

Ali Rohde (43:50.453) Wow. Josh Albrecht has hit the period where he's not reading all the code.

Josh (43:53.367) I have a I have a 4K portrait screen with eight-point font and scrolling on GitHub, it was two pages of the file listing of the change logs, like one file per change. It was insane. But I didn't read the change logs, but I read the code that it changed and I merged it. I merged it because it was on net pretty good. Gross. Terrible. Yeah. Imagine an AI system where it's

Ali Rohde (44:07.265) Did you read the whole thing?

Ali Rohde (44:14.955) And how did it feel, Josh?

Josh (44:21.814) Moving so fast you can't even tell what it did. Like what's what's the point of you at that point? Yeah.

Ali Rohde (44:28.789) know, the point of view the idea, is like what you want generated.

Josh (44:32.524) I know, but at this point it like I was joking with an engineer as we were reviewing it. It's like at this point we're basically just saying like make it good and then since we can't read the output, we're like, Did you make it good? And it's like, Yeah, I'm like, Okay, then I guess merge it.

Ali Rohde (44:42.463) Yeah.

Бо і десант.

Josh (44:48.393) yeah, yeah, yeah. So there there's still a little bit of us in there, but it is it is getting interestingly abstract.

Which does get into some of the self-improvement stuff we were talking about before as well. Like one of the things that's strange and interesting is companies talk about recursive self-improvement and they're like, there's gonna be like a loss of control someday. But they don't mean like, the AI is suddenly gonna get magically smart. They mean this. They mean that people are just gonna give up. Loss of control just means your engineers get too lazy to read the change log. So it's not really a loss of control so much as like deciding that we don't give a shit anymore. Yeah.

Ali Rohde (45:26.005) A giving up of control.

Josh (45:28.994) No, not really the same thing. Easy to fix, but

Ali Rohde (45:33.313) You mentioned infrastructure that you're building, maybe products around this. What can you tell me about that?

Josh (45:47.451) yeah.

Ali Rohde (45:49.461) Yeah. You mentioned infrastructure you're building to enable these loops and how you're building a product around it. What can you share about that?

Josh (45:58.041) Well, we're not building a thing for loops. We're just building my god. No. Yeah, this week and next week. but no. it's pretty straightforward. We're trying to make the long run version of what you want. Like we were talking about before. Eventually you want things to run local, you want things to run on your computer, you want the data under your control.

Ali Rohde (46:01.461) I really think you should brand it as such. You would do very well on Twitter.

Josh (46:23.628) I think eventually the open source models will be a bigger and bigger component of this. Those are run locally as well. Eventually you'll want to be smart about like which tasks you send to which models. Eventually you want to collect all of your data. So we're just making the infrastructure to make this relatively easy. Like, how do you spin up an agent in a dedicated workspace that can run whatever servers you want that can pull in your data, that can do the authentication and authorization and security stuff in the right way and keep all that stuff.

encrypted under your control and in a way that's like really easy for non engineers to build whatever they want.

Ali Rohde (46:57.559) Well I'm looking forward to trying it out.

Josh (46:59.342) That sounds great.

Ali Rohde (47:01.003) Josh, thank you for getting shut again.

Josh, we did it. Episode five in the books.

Ali Rohde (47:13.931) Okay. Well I'm looking forward to trying it out. Josh. One more time.

Josh (47:18.412) Me too.

Ali Rohde (47:24.587) Well, I'm looking forward to trying it out.

Josh Episode five in the books.

Josh (47:31.758) Thanks, Allie.

Ali Rohde (47:33.398) See you next week.