The Standup with ThePrimeagen - Whats really going on with AI, Expert weighs in

Episode Date: March 13, 2026

AI researcher Dimitri joins the show to talk about the real impact of AI on software engineering. We discuss token costs, AI coding tools, the future of developers, and whether engineers are heading toward a world of reviewing AI-generated code instead of writing it. It’s a grounded conversation about the hype, the risks, and where AI might actually be taking the industry.

Welcome, everybody, to The Standup. We have an extremely special episode of the stand-up. If you have been living under a rock, you did not realize that Casey has started his own podcast, a competing podcast. Competing. I know, Casey. How could you do this? How could you? Yeah, yeah, yeah, yeah, yeah, yeah, yeah, yeah, yeah, yeah, yeah, yeah, yeah, yeah, yeah, yeah, yeah, yeah, sorry.

He and Dimitri are talking about and having a real conversation about what is the actual effects of AI and all that, and you're probably asking, well, why would they do that? Well, Dmitri turns out to be a legend when it comes to AI. He knows a lot about it. It's been working in it for a very, very long time. And just like Casey is significantly more competent than T.J. and I combined. And so therefore, their conversations are actually useful and good to listen to. And we thought we'd bring them on and ask some of our own very own questions. Kind of a podcast melding.

There's a lot less cars references and jokes in their podcast. So if that's what you're looking for, just make sure you're still coming back here. but if you're, you know, just wanted to make that clear where our sort of competitive edges is the car's references. Specifically, the cars references and that is it. Exactly. Oh, they also do like technical skill and excellence and all these other things and have deep reasonable conversations. Well, we weren't listening what they did better. I was listening what we did better.

Cars references. Focus on the positive. Yes, right. I'm more of a glass half full kind of guy or like glass one-tenth full kind of situation here. Glass has water kind of guy. That's the person I am. All right. In all seriousness, though, the podcast was actually

Dimitri's idea. It wasn't even my idea. It should have been my idea because I agree that I should ditch you guys and go with some higher quality co-hosts here. I mean, that's obviously clear. But it wasn't because Dimitri, you know, you were actually

the one who was kind of like, I want to talk about AI because I'm just, basically you weren't happy. You were like, I don't like what's being said. It feels like it's misleading. Did you go on Twitter by accident? Is that what happened? You went on Twitter. But wait a second.

We need a podcast immediately. Yeah. So let me add a little bit to what Casey said there. So actually, I did not set out to have a podcast. I set out. So I've been working this a long time, like 20 plus years. I have lots of friends who are not programmers or any kind of, like, lawyers, accountants, doctors, whatever.

And they're frequently asking me, You know, first of all, can I use this in my job and other things like, how does this go to effect? Like, I have kids in college. Should I stop sending them to college, right? All of those kinds of questions. So at some point I thought, I should just try to put this stuff together somehow. And first I started writing stuff, but that felt awkward. Then I thought maybe I'll record something, but I'm not really a recording personality.

And then Casey and I've been talking about AI on and off for years. And my suggestion was, what if we record, like, one session? and I had a list of things that we could talk about. And Casey said, well, there's so many things here that we could just turn this to a podcast. So that's how we ended up here. But let me have one more thing to that, which is, so I know that there are a lot of people out there who are, you know, I see Casey, I click personalities. I am that kind of personality as well. So part of part of what I want to do here is to facilitate having more Casey out there on stuff.

where I can be the wingman, right, that Casey understand. And one of the things that I appreciate about him quite a bit is that he wants to know what he's talking about before he says something, which, you know, puts him in the top, top 20% of podcasters at least. At least. At least top 50% of this podcast. Wait a second. Hey, is that HTTP?

Get that out of here. That's not how we order coffee. We order coffee via SSH terminal dot shop. Yeah, you want a real experience. You want real coffee. You want awesome subscriptions so you never have to remember again. Oh, you want exclusive blends with exclusive coffee

and exclusive content. Then check out Kron. You don't know what SSH is? Well, maybe the coffee is not for you. Terminal coffee in hand. What I hope comes out of this in total is that we get kind of Casey culture-aligned commentary on AI. And I mean, Casey can, I don't want to put words in his mouth, but we see software in a very similar way. I'm similarly skeptical about the quality of anything coming out of big tech these days, about questionable ethical business practices and so on.

And so on general software, not AI software, KC and I see I do I on most things. And so what I'm hoping is that I can be the wingman and give a platform for more KC culture commentary, specifically on this crazy AI thing that's taking over everyone's life, including my own. As I, in our inaugural episode, I mentioned that my life used to be much more quiet. And nobody believed AI would work, and we were just quietly doing fun stuff. and in a way, things were better. I mean, it's good that things are working now or sort of working. And the sort of working actually is a big part of what I care about,

trying to make it better than sort of work. So, I don't know. I've gone on long enough. Maybe, Casey, you should put a bow on it. I think you're being way too generous to me. Yeah, I'm the wingman on our podcast because I don't really, like, I just don't work with AI stuff. So to me it was just like a good opportunity to, like, have somebody on who, like,

I trusted to give solid perspective on AI because, like, it's really hard to, you know, the only real way I thought otherwise that I would be able to give any commentaries, I'd have to go spend a ton of time with it, right? And I just didn't really want to do that. So it's been great, and I've really enjoyed it. Obviously, Dutri and I have recorded a couple things that, you know, we haven't posted yet. So, but, yeah, it's been really great having you on. And I should kick it back to Prime and Teage.

So, like, we don't want to just talk about our podcast. here. So like you guys had some AI questions you wanted to talk about. Uh, or I don't know, take it. Take, take it. We are here. We are here at your disposal. I did want to say part of it is to have both of you talk about the podcast because I said Casey, you should come on and then talk about the podcast with Dimitri so that people can hear about it. I really enjoyed episode one. So we at least wanted to make sure everybody knows that this thing is happening. So don't feel bad about talking about it. That's why. part of the reason that we're doing this today. So that's number one.

I really liked it. Prime, I don't know anything you want to start with. Otherwise, I've got a few questions. I do want to start with some things, which is, Dimitri, can you please, we didn't give you really like an actual qualifying intro other than your legendary? Could you please define? I should downplay that a little bit.

They say this about me all the time too, Demetri. Legendary, it's very embarrassing. It's very embarrassing. It's very embarrassing. So there you know. Legendary for what? If I talk like Trump as just a warning, everybody's legendary. Everybody's the best.

You've been hearing some very good things about. me, Prime. It's like, some beautiful things. Best game developer. Beautiful code. Everybody agrees. You are the best game.

Every American. Amazing code. Totally corrupt, but beautiful code. Beautiful code. All right, sorry. Keep going. Yeah,

I was just going to say, I appreciate the kind of words. Like, I consider myself relatively competent. But, I mean, there are, like, big names in the industry who are either friends or friends of friends. And it's useful for me to keep, keep myself calibrated relative,

relative to people with much more, at least public renown. So, yeah, I don't know. I guess in terms of introduction, I've been doing broadly AI-related research for 20-plus years. I shipped to many people, my very first custom design AI in 2025 at Google. It was one of their first. 20-2005. Sorry, 2005. I was going to say that's like 20 years off.

That's like banana if you're shipping that one. Okay, okay. Yeah. So anyway, I've been doing this for a long time and seen lots of the ups and downs in the business. And also I study the history of the business as well. So I know about the ups and downs for like the preceding 50 years, like going back to even the 60s. And so one of the things that really interests me is how this is going to to affect other people, like people that I worked with, for example, who are now, I mean,

in the first episode, we just talked about the impact on junior engineers and separately probably should talk about the impact on senior engineers, right? Because they face a different problem, which is if you're, like, let's say you're whatever, 45 years old now, you've been doing this a long time, you have a good, stable career, making good money, and now you're wondering what happens if they, like, crack a lot? the nut and in five years I'm useless, right? And I know people who are wondering this myself.

Mm-hmm. And that's an especially bad time to be useless, right? If you're late mid-career, what do I do if I'm 50 years old and I don't know, like, what do I do next, right? And there, it's not anything like the problems that the junior engineers face where it's

I mean, again, big problem for the junior engineers but they have their whole lives ahead of them and they can't, they have time to try to do something else. But if you're 50 years and don't have, you know, 15 years of building up a new career ahead of you, what do you do? So that's not something that I see people worrying about a lot as well. Yeah, I don't know. I guess maybe talk too much about myself.

You can't talk too much. We're used to hanging out with Casey. All right. It's fine. There you know. We like it when people talk. That's why we bring smart guests on. Yeah, so I guess another thing I will add, and this connects back to,

kind of having a Casey style perspective on the industry is that I've seen a lot of how the sausage gets made, both in terms of the technical side, but also the business side. And this is something that we talk about in later episodes about how much can you take certain claims at face value and how to evaluate which claims can be taken at face value and so on. So what I'm in the same way that Like you can't take whatever Microsoft claims at face value about improvements in Windows performance

Right You can't same same way out This is going to be the best windows yet This is going to be perfect and flawless I was promised It was going to be the last Windows ever Exactly as well I was promised the last windows Well I did

They are still working on making it the last Windows ever That's a good point The year of the Linux desktop is what TJ is trying to say. They put that into the AI, and it gave them a plan back of how to make it the last Windows ever. They're executing on that flawlessly. Yeah, and they're like, we got this thing called Windows 11, and it will ensure that Windows 10 is the very last version of Windows. All right. And if that doesn't work, we have Windows 12.

Yeah. So I've just got a few topics. thought would be fun. We can just see where, like, they lead us and what happens. I think there's similarly to on, on the episode of weeks that I've listened to, you start on a topic and there's like 95 different branches that we can go off of. So I'm not, if we don't get to all of them, it's whatever. I think it'll be fun.

But one of the things I'm interested in, and like, it's just hard to get anybody's, like, unbiased, everyone's got to bias, whatever. just like a more rational take on like this current state of token cost and what that's looking like going forward. I don't know if you have any thoughts, Dimitri, about like, is it going to be that Sam Altman was right and we're going to get 10x cheaper every year or whatever? I don't know. That's what Tim. Sam has it on the record. He said it twice. He said it twice and made projections into the future saying it's 100x cheaper.

GP2 5.2 high will be 100x cheaper in 10.000. Two years. Okay. So a lot of that is infrastructure development and algorithm development that is trade secrets that I can't evaluate. And even like even an open AI insider, it would be illegal for them to to evaluate in public, right? I don't know what they're cooking up there. I mean, they have an extremely talented team. A hundred X cheaper seems hard to believe. but I guess wait and see.

So one of the problems with these, with evaluating claims in this business is that the timing matters as much as the content of the claim. So like, Musk is a good example just to take Sam out of this for a moment, where Musk was saying, you know, we'll have, I mean, we'll have, you know, reusable rockets in, I think he first made the claim in 2005 or something like that.

It took many years before that actually worked. Later, when was the first promise of Tesla full stuff driving up? It was like 2016 that he was saying. And then every six months after that. So I would, I found it useful to try to separate out the content of the claim from the timing of the claim. So I would not be surprised. I would not be surprised if eventually we can get 100x cheaper token cost. whether or not they can do it now, that's beyond my knowledge, because that's, that's,

partly that's infrastructure, partly that's, like, custom hardware, partly that's, can you get cheaper electricity? There are so many things that go into the, into the business that it's hard, hard for me to claim one way or another. What I can tell you is that it has been, as I'm sure you've seen, it has been a land rush mentality, right? So everyone's building as quickly as they can, whatever they can. Put it out. Does it work? Does it barely work? Okay. Move on to the next thing. And some of that is maybe recklessness, but some of that is just market forces, right? Because right now there are multiple really big players and probably there won't be as many big players in 10 years. And so given how much money they've all put into this already, they really don't want to be the ones caught out. Right. And so the, like the arms,

race is rational from a business perspective, whether or not it's producing completely rational engineering artifacts. So all of that is to say, I'm sure that there is very substantial opportunity for improving efficiency in the current stock. And I don't know if they – so my biggest question would be, is the internal stock stable enough that you can optimize it now, right? Because like say you spend a year optimizing it to get the token costs down and then there's whatever the next architectural innovation is, does that invalidate your optimization or not? So I don't know, right? And so you will hear me

say, I don't know, perhaps more than most people talking about AI just because it's, there's a lot that's unknown even to people who are deep, like deep inside these companies about what's going to happen in like two or three years, right? I would add one thing to that that, you know, it's not adding something as in information, adding something as in sort of like a thing to think about. And that is that it's worth noting that if the costs were really going to get that low that fast, I think I would have expected Google's costs to be very low already. And I know it's a weird thing to say, but if you assume that there's any hardware component,

like if you think that all 100x is going to come from software, then that seems, okay, maybe that could happen. But if we're counting on a substantial portion of that 100x coming from infrastructure builds, it feels to me like Google has kind of been building AI-centric stuff for quite some time and has gotten their thing. Like, their TPUs are very much just like, we built machines who's only people. purpose is to do this job, as opposed to, like, say, for example, Nvidia, who's not doing that. Like, the, Nvidia cards are kind of, like, still in a weird, like, hybrid state where only the most, only the very latest Nvidia's could be said to be focusing on AI, really.

And even those are still a little bit hybrid feeling to me. So I could imagine if someone said, well, you know, if all we had was Nvidia to look at, maybe you could believe that, like, you know, there's a very long lead time on hardware. So, you know, we don't know what they, you know, when they shifted probably four or five years ago to going like, oh my God, AI is going to be the biggest thing. We need to completely redesign everything to just focus on that. That stuff will only be, you know, will only be seeing the end of that pipeline, you know, now or something like that, right? So I could see that, but I feel like Google's kind of, I mean, am I wrong about the stream? I feel like Google's been doing this for a long time.

If there was a huge, if there was huge gains to be had by doing just AI hardware, I would have thought, they'd have them already. Is that, am I way on there? I'm not sure they don't, right? So the first, the first time I heard about TPU, I think the first time I heard that they were starting to work on it was like 2014 or something. So that's just when I heard about it, right? Like, I am not a Google insider, so I'm sure, like, knowing how Google works, I'm sure that they were talking about that, like, three years before I heard about it in public, right? But when I first heard about TPU development, it was somewhere around 2014. Yeah.

So I guess I have not been studying the, like whatever internal economics reports Google puts out. It's possible that they are already substantially cheaper than... I would think. On operating costs substantially cheaper than opening, I don't know. But I mean, this is something that's a common folk belief in the industry, which is Google is sort of the quiet monster in the business because, like, they, if you compare them for a moment to, say, Open AI, they have had, they have

been infrastructure leaders forever. They have been AI leaders forever. Unlike OpenAI, they have, they have, I'm trying not to say the word surveillance, but I'm just going to say surveillance. On the entire internet, right? So they have just a constant flow of potentially trainable information. and XAI has that flow that is their own proprietary flow of data. So many people have this feeling of Google was kind of slowed on the product side. I emphasize on the product side with chat LLM products. But there is many people have this feeling of possibly Google is just going to quietly leverage all those advantages that nobody else has. And XAI is the only one that's sort of close

in having that capability. It's interesting to note that as far as I've seen in public, XAI, they're merging with Starlink. They're talking about, sorry, with SpaceX. SpaceX. And they're talking now about orbital data centers. And something that I thought was interesting there was that the designs that they've talked about are GPU-based, not TPU-based. And I don't know if that's because they don't think they can build it up as well as Google did, or they don't think they can license it. So I don't know, right? But the thing I will point out is the designs I've seen, I mean, I guess I haven't looked in like three months,

but the designs I've seen were estimates based on, you know, NVIDIA-style GPUs and not TPUs. So they must know something because this is, this is the kind of thing that they do, right? Like the big, the biggest competence advantage that I would put in the kind of musk category of companies is

the relentless execution on building and optimizing and streamlining physical stuff. That's something that they've been doing really well for a very long time. So if, anyway, that caught my eye that the designs I saw were based on GPs and not TPUs. And I do know that they're working on their own

their own AI chips. I don't know that much about how far they are there. But it's interesting that when they were talking about the orbital data centers designs, they were talking about conventional GPUs as we understand them. I got to say orbital data center is like the coolest sounding thing to build.

Like it seems really not practical. I get there saying, oh, we're going to have robots do it and it'll be outer space and it's really cool. But I got, I mean, I just feel like Everyone's got to admit orbital data center is probably the coolest thing to say you could work at in software development. I think, like, yeah, based on what I've seen for orbital data centers, it sounds like it's, it is either much cooler or much less cool than what you're thinking about, depending on your perspective, right? So, excuse me, if you're imagining, like, a giant building in space.

Like, it sounds like the kind of more leading designs that are like more plausible. They're more like Starlink. They're more like satellite clusters where you just have lots of little things that talk to each other over light links basically for high speed communication. Well, and giant, giant radiators, right? And huge, yep, flat. It's almost all radiator. And then there's a little like nub in the center that contains. That contains the GPUs, right?

Yeah. That doesn't sound as cool as what I had in my head of a big skyscraper in outer space where we've got racks of GPUs and you can walk around in it and it's like really cool. But that's okay because when you tell your parents that you work at an orbital data center, they won't know. So it's fine. That's what they imagine. Yeah, yeah, yeah.

I say it's exactly like Star Wars mom and dad. Trust me. Trust me. Unfortunately, like heat is just very, I guess, as the kids say, problematic in space. Like, you don't have anywhere to put it. Like, it's very hard to get rid of heat in space. So this is a problem.

As Dimitri was saying, you end up having to have a lot of surface area per chip, like per heat generating Nvidia of like GB200 or whatever. I mean, it won't be a GV200 by the time they get these things launched. but the uh you know that's the thing i didn't i didn't get i don't what is the advantage then because in in like i get that it feels like if you talk to like a regular person about putting in an outer space they're like well it's cold in outer space so that's going to be really good for GPUs but it's the opposite i think right i mean i so i don't get it's not that it's not cold in outer space cold is the lack of available you know like not having energy right there's just

There's no density, so you have nothing you can. Yeah, I meant more like they said cooling would be super easy. I feel like that's like the normy thing. If you say put something out, it's like, it'll be cold. Sick, it's like a huge refrigerator, bro. But that's not how heat works. So I get that part. So then what is the advantage?

Why this is, we also don't have to go way down this drought, but I have been wondering this myself. So what, like, why do they want to put it in outer space? Incredible investment, TJ. Like the amount of investments that are coming in will be it. The TAM? The tam for outer space is big. Think about how many data centers

we can put in outer space. Well, it's free power and difficult but free cooling. Okay. So if you can make the physics work, and I don't know, right, they have really, really good thermal people. And you can imagine they have very good thermal people

at SpaceX. So I assume that they have done simulations that suggests of them that this can be made viable. But yeah, so it's free power and also like, more free power than you would get for the same panel on Earth because there's no atmosphere eating up your atmosphere or weather for that matter.

Yeah, no clouds. Right. So free power and tricky but free cooling. Gotcha. So it's like when you build a really complicated thing, but then it's passive in a video game. I understand. This is, this mechanic makes sense to me. And maybe one thing I would add, so this is entirely business strategy speculation.

So whatever. Like take, like, poor big... It's financial advice, got it. Exactly. You're investing right now. If it works, if it works, and you are Musk's AI team, and you just get to launch your AI into space. And everyone else is, like, fighting local governments and can I build a hydro plant here? Oh, how much you have to pay you to, right?

And you'll just be like, we're going to space, right? Like, you losers can hang out here. we're going to space, and by the way, if you want to go to space, you have to come through me, right? Yeah, yeah, yeah. Yeah, he pretty much owns the passage to space, so it's a pretty synergy. I can't ask Katie Perry. She's the nod of space. Well, and also, like, it's pretty trivial for, like, when you look at the projections for these sorts of things, it's all about launch cost.

Like, whether or not it makes sense to put a data center in space, a quote-unquote data center in space. It's like, here is the launch cost. Like, at this launch cost, we think we could do it. And that's based on the failure rate of the thing, how long it's expected to be able to be up there, all those sorts of things. And so at the end of the day, if you look at somewhere like SpaceX and they're like, well, for us, we could get our launch costs down to $50 a kilogram or whatever the magic number is. We just charge everyone else $75 a kilogram and then it's not worth it for them to put it in space or theirs will always be worse than ours. It's a very good position to be in at a strangely profitable vet for SpaceX if it turns out that these are actually useful in the future. So I think we should probably attempt to get a little bit more on a more practical approach to AI.

I do think orbital data centers does sound amazing. You don't want the title of this video to be orbital data centers. AI in space, Prime. How many clicks is that going to generate? AI in space. AI and space. Go ahead. That's my rock opera I'm writing.

That's crazy. How did you know? Okay. Because I am also writing a rock opera called Alien Space as it turns out. Oh, I was sick we could collapse. Yep. That's not going to happen. He's just going to take it all from you.

Okay, so I guess we should. We should probably talk. I think the number, I get a lot of concerns from people. And one of the most frequent questions I've been getting lately is, is my future just viewing code. Yeah, so quite possible. Let me try to be measured in what I, like, how likely I think that is. So, look, one thing that I've been, I have discovered in my own work and that I've been

saying publicly for a while now is, so I try to work on, like, repeatable and reproducible AI results. So, like, the bucket term I use for that has been reliable AI, but it turns out that there's a company called reliable AI and has a lot. So anyway, the thing that I think you can do right now, like completely hands off and not have to review anything, right? So this is what I mean by reliable, right? That like, if I ask you teach to whatever, go implement, like, download this thing from this HTTP endpoint and like save it into the database and give me a dashboard that whatever, calculates the averages of some things, right? I can just tell you that and expect a result and it will work and I don't have to talk to you about it again, right? So that's what I mean in the list of the completely hands-off style. Right now, I think the limit of that is somewhere around a couple thousand lines of relatively standard junior-level code. I want to emphasize it. Some people have justifiably pushed back on this a bit. I'm not saying you can't do more by having oversight

and doing multiple tries and having algorithm of test suites. I'm not saying you can't do more. Right now, what I'm talking about is what's the limit of what you can do hands-off and you can mostly just trust that it happens. I think at a minimum, we are going to go. We are going to go through a review heavy phase because the businesses are, as I'm sure you know, the businesses are trying to shove this into everyone's workflow, whether they like it or not. Right. And I get lots of people, lots of people, even in the AI business, you tell me, like, I hate how much they're trying to push us to use these things, right?

I just want to do my job. But like, they're monitoring my token usage. Right. Yeah. Anecdotally, too, not just like literally, like, I have friends who are like, we have to use this tool. Part of my KPI is like we use, we have to use Claude Code

only. That's, and my manager has a dashboard with my token count for the month and it is a thing we talk about and review. Not even like, oh, you can use whatever one you want. We just want to see you guys experiment. They're like, you have to use cloud code. You have to do this. Yeah,

it's crazy. Yeah. Think about Amazon Kiro. That's a lot of people are talking about that right now with the whole Amazon accidentally having several Sev-1s. Yeah, they're like, you have to use our stuff to take ourselves down. If anyone's going to bring it down, it's going to be me. Okay, that's Amazon. But we want to make sure our site goes down and our AI looks bad at the same time.

We don't want, like, someone else's AI to look bad. No, that would be mean. Okay, if it's going to, if it's going to go down because of AI, it's our AI. They're put in the AI and AI, okay. There you go. Very good, yes. So, yeah, Prime, on that point, I think it's very likely that we will go through a review-heavy, phase. I think

that at some point the workers will object to this enough that we're going to have to find something more acceptable, right? Because I don't know if you have tried it, but if you're just reviewing stuff like, you know, claw dumps another thing on you every, you know, five minutes. You're like, I can I have to review this and approve and merge, whatever.

Or, I mean, you can just yolo, right? But we've seen what happens if you just yolo. So I do think we will, at a minimum, go through a, like, reviewing, mostly reviewing phase, and I think people are not going to like it. That's my expectation. I don't know how – I don't see what the alternative is right now because – so the either – the two possibilities I see right now are – they're going to be monitoring your token count, and so you need to be consuming and generating tokens and then that has to go that has to get linked which you know in some places they

actually link the tokens to PRs right and so you need to be generating lots of PRs you need to be consuming lots of tokens how can you do that other than just generating many many PRs with AI and reviewing and like hoping that you're not taking down you know taking down by by missing something in your review right so I don't know how we get out of that other than at some point the business side says, okay, we can cool it, right? We're going to stop monitoring your token use, which I mean, I personally feel like that's not that different

from a desktop monitor or like a keyboard monitor or something. It feels very invasive to me. Like, hey, you know, I understand you're asking me, I think it's fair to ask me, hey, can you use this tool? Can you try to do better work? I think that that's fair, right? Like, I'm, ultimately, you're being paid to do something for the business.

I think it's fair for them to ask, hey, we'd like to be using these tools and we think you can do whatever, like 20% more work if you can, if you use this. I think that's fair. I think the micro tracking feels worse than the past. It's really easy to game too, right? You just like sequentially look at every single file and it's just going to be like dang, dang, dang, dang, dang.

I'm crushing the tokens. I know people who are at big techs who are working on stuff and they have to do, exactly what you're saying and they're saying I like I do I do this just to keep up appearances for my AI use and you know by the end my output changes maybe 10%

right I do like how it's literally just setting money on fire and they're like this makes me look fast because so that actually before we got distracted on the orbital data centers which I'll repeat coolest name ever super cool command and conquer I was I was interested in you know some

of your thoughts about sort of the tokenomics side of things because right now it feels like people are burning a bunch of money trying stuff out and exploring and willing to just be like oh yeah you just spent $5,000 this month whereas before like you couldn't get like a $45 like online course approved and they're like oh but yeah yeah $5,000 of tokens for nothing yes we're totally down for that do that every month because I was interested if you know uh like if we're going to see a big know, or your thoughts on if we're going to see a big push on like generating less or like more focus on like we're going to use more tools to check what's being generated. You know, because at some point, you know, Claude, Claude just released their thing about

code review. And they're like, yeah, usually it's like 15 or 25 bucks or review. And you're like, well, that adds up kind of fast. Especially at the rate people are saying I'm shipping 10,000 lines of code today or I made a 300,000 line Ruby on Rails application for my blog. that's Gary Tan by the way, Casey. I don't see that tweet. Okay. Hey, he's boiling on a lake and then he's going to boil the ocean. All right, that's what he told us.

Yep. Side note, I cannot tell if he's the most genius rage baiter of all time. But we can circle back to that on a different episode. Yeah. Yeah, so because I'm interested in sort of that angle because I do find like some tasks I can get done like a lot faster with letting agents work on it or like agent spin on something like forever. It has a really clear outcome. There's some stuff. I already have a bunch of patterns in my code base spin. But I'm like, is, is everyone just going to be chill forever with me

spending $500 of tokens overnight on this? That doesn't seem like what a business would like. I don't think we're going to get $500 of value back from this feature. So I don't know if you have any thoughts like in that vertical or yeah. So this is speculation. Yeah, yeah, of course. Speculation. A lot of this is like principal agent problems for the business. So there are for sure like CEOs and like one step down from CEOs who one way or another, their personal benefit is tied to claiming that they did a lot with AI. Yes, true. Yeah.

And as you know, like in any big organization, you never just evaluate something for whether it's good or not, it has to be evaluated through some KPI, right? So there are people, I'm sure, like, I haven't seen this personally, but I'm sure that something close to this exists, that there is someone with many millions of dollars worth of bonuses tied up to, can I get our token usage to double this year, and can I get our PRs to increase by 20% as attributed to that token increase, right? Right. Like, extremely bureaucratic thinking about using this new technology, right?

So I'm sure that there are many people like that. And so to those people, they're not spending their own money, right? They're spending the company's money. They're telling you to spend the company's money to do something that makes them to go up that gets them a bonus, right? Yeah. So I don't know how long that's going to last, but there's certainly right now a very large premium at multiple levels of the kind of the business stack where people are like,

you know, why isn't your token usage double, right? Right, yes. Yeah, yeah, no. to some people that sounds like a joke, I have friends who literally, they say, my boss came and asked me why hasn't my token usage doubled,

right? But you only used $300 of clod last month. Honey, you barely touched your clod last month. What's going on? Is something wrong, sweetheart? Okay, so I do actually want to follow up on that because this is the, you know, this was one of the big things

that has been going around, is this Amazon Kiro thing, taking down everything. And now they're saying, hey, we're going to make it so that juniors and mid-level people who use generative AI must have a senior

sign off on this. Can I do a quick question on that as well, Prime, which is, was there a policy before seniors didn't have to review junior's code? Just in general? I assume it's like a lot of companies, which is that a junior could have a mid-level person review the code and say, hey, this is not really good. Like, not every change needs to have, you know, the same, because you effectively will exhaust and lose your senior population if they have to review every single PR. And so kind of, that makes sense. So excluding the personnel problem that will likely develop from this with Amazon, is this? like the first sign of people realizing that token measurement isn't necessarily the best metric,

because I assume that's what's hurting Amazon is that they really push super hard as token metric is the greatest thing that has ever existed and you must only push on this one thing. And now they're feeling some of the effects of maybe moving too fast. Is there a world that's going to exist where people are going, okay, instead of trying to double metrics, maybe we should double some other thing, say, I don't know, uptime. maybe like uptime could be the thing that we're like you know that we're valuing and then that can be that could be like is this an actual like path forward or are we from your perspective are we just going to see continued push on double your token usage double your token usage because they both

sound super appealing I guess from a purely theoretical point of view of like oh more features we could know we could get all the features or no let's be stable again I think so as we as we discussed with Casey and Casey you should you should jump in on on this part like we've seen these cycles of like everything has to be you know like 25 years ago everything had to be online and then everything had to be web 2.0 and then everything had to be like social local mobile app and then everything had to be crypto right and that right so we are in that phase absolutely right now we're separate from any utility that the technology is offering, there's just this social mania of everything has to be this right now. I don't know when that fever will break. I think certainly events like the Amazon

Amazon event will help with that. And I do think that at some point people, someone is going to say, okay, look, we just can't be burning this kind of money all the time and also like setting our infrastructure on fire and driving our engineers crazy. At some point, at some point, at some we should actually get a benefit from this instead of hurting ourselves. You can only pick two points on that triangle. Drive his year's crazy, burn a lot of money, up time. You get to pick two of those points. Yeah, true.

But I think this could last for years. And the reason I say it could last for years, I know people will not like that. I think the reason this can last for years is it's easy to underestimate. Like, you guys are like people who watch this, you know, watch the stream, how far up the power lock curve of early adopter you are. And I just meet lots of people all the time who still conceptualize AI as like it's a better search engine, right? And I mean, it's easy to forget that actually the vast majority of

people don't know much about this and aren't really trying to use it. And so all those people somehow will have to be carried through the fever swamp. And I don't know, like, I think that could take years. I would agree with that for, and add the sort of the chaser, which I think we did talk about on that first episode of the podcast, which was, I don't see it as mattering whether it takes down Proud all the time, either, is the problem. Because, like, at least my experience for the past 20 years has been, it doesn't matter how bad the practice is.

if it's just something that is in the zeitgeist, then people do it. And you can show them clearly, like you can literally demonstrate, look, if you had done it this way, this is how much better it would be, they don't care. Because the best practices is that you do it this other way.

The best practice is you use Python, then we're using Python or whatever it is. And you're like, look, if you just written it this way, it's 100 times faster. And you're like, we don't care. That's how this industry has operated. And so it would be very weird to think that in the special case,

of AI that anyone would care one way or another whether it even was better, right? So you can almost take your, take your guesses about how good AI will be in practice about generating code in the first place. You could even put that aside. I think the fact that it has this much momentum is enough to know that this will be here for a long time, the current way we're doing things, even if it never gets any better than it is right now. Even if it's state exactly as good as it is right now, no improvements literally at all. I don't think that would change the outcome. Honestly, I really don't. Well, there's like a trillion dollars invested in a two, right? So they're like, well, exactly. Even if even and it's, I mean, it's not a really

even sunk cost fallacy. It's just like literally sunk cost truthyce. Which is like, we really, we really need this to work, guys, or we're really underwater. Right. It's not like. Yeah. So it will become the workflow. I think that's inevitable. And like I wish that it I wish that we could say like well if the AIs don't improve substantially then people would like rethink these. I don't I don't think

that's true. Like I think whether or not they're able to push the quality up I mean hopefully they are because you know everyone is putting a lot of money into pushing the quality of the AIS so hopefully they do get better but even if they don't I don't think it's going to matter personally. So I actually have a bit of a follow-up to that one, Casey. And I'm no offense to meet you. I'm actually very curious on Casey's thought of on this one. In 2024, GDC started doing this survey asking about the perceptions of AI inside of the game development space.

And in 2024, 18% of the developers held a negative view on Gen AI. In 2025, 30% of the developers at GDC had a negative view on AI. In 2026, it's now at 52% having negative... view and only 7% have a positive view. Is there, is, is, is this like a special unique thing to games? Because games does contain a bit more art than the average programming, shall we say. Is this like a special games thing? Or is this actually a momentum change that people are feeling due to, hey, you must use

this as the greatest thing ever. You have to use it to therefore. And then you start using. And you're like, okay, well, it's great. Here, it's bad. You know, you feel this giant disconnect from Twitter. Because I have my own opinions that are largely formed because of how much I hate what Twitter has to say.

I mean, obviously, I can't speak for people who are taking that survey. I mean, they know what they were thinking when they checked the boxes or pushed whatever they pushed. But I, you know, I think, as I've said before, I think it's very difficult to separate out people's opinions of Gen A.I. The technology from Gen A.I. The people and the companies. And I think when we look at things like negative surveys, I'll say I'm right there with them. I have nothing but disdain for this sort of current moment, but that's not based on an objective look at generative AI and where I think it's going. It's based more on the behavior, the legal behavior, the attitude.

It's based on, you know, like many things, when, you know, we're humans, we make decisions largely based on what we think the intentions are. of the people who are in charge of things. Because at the end of the day, like, that's, you know, if we have a really good feeling about someone who's running something, we are less concerned about that thing than if we have a really bad feeling about it, right? And so I would suspect that a lot of it comes from there

and that, you know, especially, like you said, there's more artistry in games. There's a lot more people making art assets, a lot of more people making music, making sound effects, making narratives. and, you know, the behavior with respect to that of the AI companies has been utterly deplorable, right? I mean, I've said this many, many times on Twitter. You know, it's pure disdain for the law, no concern about what they were doing.

They basically just made a decision. We're just going to do what we're going to do and will deal with the lawsuits later. It was clearly what they decided. I mean, I don't have evidence that they literally decided that in a meeting, but their external behavior makes that clear. If they had been at all concerned about whether they were breaking the law, they would have done a lot of things first that they didn't do. By the way, Casey, what's it called? The Facebook actually was caught with email saying,

this is illegal, right? Yes, this is. Don't do it on a company computer idiot. Multiple of them. Now, if you look in their in-court documents, multiple of these AI companies are on the record as having known they stole things. Not about fair use. training that like just literally pirated.

Yeah, they weren't like, this is transformative. No, no, no, no. There's none of that. There's like, there's like, we, we are literally taking an illicit step to acquire data and we know there's multiple companies that are now in court records. Like, they have that evidence, right? So. Yeah. So I also know like, like GitHub could have just filtered out non-permissively licensed stuff to train co-pilot, for example.

Like that's the easiest one in, that's one of the ones where it's easiest like, you owe. the search thing that you're doing to find the code. You could have clicked your own checks. It's not like, oh, we just crawled the internet and we got back 10 million things. I mean, like, one of the primary things that these LLMs are good at is understanding language. They all now are at the point where they could analyze their own training data set and assign a probability of legality to each document. That could be done today. For all I know they have done this, right?

100% legal for my company. As long as you don't To steal it from us, it's legal. Okay, but it's restricted from us. So I think that's, you know, from an artist's perspective, you can see why this is a big deal because it's not just, you know, it's not just looking at something and saying,

I'm afraid that this thing is going to take my job, or I don't like using this thing because it doesn't feel like art to me when I just, it feels like management, because I think that's another thing that, you know, we haven't talked about that much, but an artist wants, to directly manipulate the thing that they're doing sometimes,

you tell them, oh, well, you know, now you don't do that. You just ask it to make this thing for you. That doesn't feel like art to a lot of artists. Some artists are fine with that. There's technical artists and there's non-technical artists. And a technical artist might be more used to that because they're used to kind of going through this like technology intermediated process to produce what they're doing.

But a non-technical artist, they're in that business because they like the direct tactile work of me. making art. And so not only is it bad in that sense where they don't like it because they're like, I don't really want to use this and I feel like it's threatening my job. That's obviously bad. But then you add on top of that, oh, and by the way, it was your own artwork that they used to take your job from you without paying you anything extra and without asking you first. I mean, how do you not check the negative, like, it's obvious that that's just bad. Nice about it. Also, the CEOs get there and they're, hey guys, guess what? Guess what?

We're putting all of white color workers out of work. Am I right, boy? Well, hold on. TJ, TJ, you are, to be fair, TJ, that was a bit of a misinterpretation, okay? What actually happens is Dario first makes a really painful face and goes, I'm so sorry, but we're going to have to have some economic struggle, okay? We're just going to have to lose some jobs, okay?

Some are going to have to go now. And when I say we, I mean you, because I'm not interested in a struggling. at all. Okay, I'm going to buy the cookies. You're a broke boy. Yeah. I mean, what I would, I have proposed before I was like, they always say, like, we're going to have universal basic income.

My thing is like, look, if everyone at AI companies demonstrate that, all of you go on UBI first. Like, all of you go on a fixed low salary first, because if that's what you believe the future is, let's see it. Let's see you guys all be happy with that first and then sell it to the public, right? But they're like, no, actually, we would like all the really high salaries for us, and everyone else can have the UVI. Like, it's just, it keeps getting worse every time, you know, they step into these things. So, I don't know, my opinion on that is like, it's obvious why someone at GDC would check that box.

I think that's probably going to be true for a lot of workers, not just game developers. It's just that game developers are, you know, more exposed to this technology currently because they tend to be more technologically savvy than, say, someone working at an accounting. firm randomly right now who maybe hasn't really seen this stuff all that much. Their firm hasn't started adopting it, whatever. The average worker may not have been encountering AI to the degree that a game developer has. They're further ahead on that curve. That would be my answer prime. Dimitri might have a different take on this, but that's what I would say. I completely agree. So the two points I would pull out, starting from the most recent one, is the accounting world has not been dragged through this yet

and they will. The HR world so I happen to know someone who works on apps for HR surveys and stuff like that, we've all seen these very tedious HR service right there. But all those were made by someone who had a job

which was like a HR survey designer that's probably not surviving at least not in the current form and they don't like that world doesn't even know that this is happening yet. Right. So all those other industries like the reason they're not yelling yet is because they just haven't been under the gun yet right and then the other thing was I think Casey made this

really good point about if you are if you're the kind of if your position in the art production pipeline is more editorial and composition so you're not you're not making the assets yourself you are you have artists under you who are making the assets and then you're doing editorial work or composition or something like So someone's, I don't know if art director is exactly the right word, but this is parallel to, you know, the people in tech right now that I think see the biggest buff, boost from what's happening right now is a technically trained project manager or product manager.

And I know a couple of those guys who are talented enough to review their own code and understand when something is obviously stupid. And they seem to be doing plausibly good work. and so whereas they used to be managing three or four junior engineers, they just manage 20 clods or something, right? So anyway, this idea of if your position in the pipeline is editing, assembling, composing, like the big picture structure of somebody else's work, I think those are the people right now who are least disrupted and most benefit.

I think you look like you're about to say I'm waiting for Prime to No no no no I'm just there was a stroke I'm considering I'm considering his position which is that I'm just playing through my head like the biggest thing that I really enjoyed doing was not in fact managing juniors

but in fact writing the code and I've also had some of the hardest transitions in an AI world because I still just like writing the code and I find it less you know so I'm like oh yeah I see myself in what you're saying is that there's a bunch of people And now this is where things, this is where the, I call it the, I call it the classical bullshit thing, which is that people say they're good at this, but I don't think you can review code and actually be good at it, beyond a certain amount of lines per day.

So it's like, if you're reviewing more than like 500 lines of code a day, I just don't, I just don't believe you. Like at some point, there's like some line that exists somewhere where you're just like, oh, no, I really have to think about this. It's higher than 500. Okay. A thousand, right? I mean, but people like, I can I think what T's trying to say is like he can do, you know, a lot more lines of code

than average is what I think T.E. There's a study. Maybe it's 500 for you, Prime. It is 500 for me. For T's like 10, 20,000. Yeah. No problem. 10 to 20,000. There we go. That's closer.

No problem. That's why we call, you know, JT and T.J. They're just mirrors of each other. Okay. He reviews code. the way. By the way, JT's Gary Tan, if you're wondering, even though the G-T. Okay.

But, uh, let's see. Oh, that was a stretch. Yeah, do you manage to get a new power cable from the Best Buy? How's he doing? Did he, uh, did he get in line there? He got his USBCs figured out. Is he able to get that power, that MacBook powered up? Because the whole future of civilization is riding on Gary.

And if he doesn't get that power cable, I think we're on a lot of trouble. If Gary's list, don't keep listing. Yeah. I don't know if you guys ever saw there was a fantastic movie starring Ice Cube called War of the World's that came out recently. I don't know if any of you saw this movie. I watched it in five separate small clips. All right.

Well, as you know, you think I'm joking, but as you know, the fate of humanity itself can very well hinge on one delivery of an Amazon thumb drive. So I don't think what I just laid out is all that far fetch. Just go watch the movie if you don't believe it. I heard it was. That's what I heard too. I just am super proud that Ice Cube was like, give me that bag. I don't care what's about to come on the other end, just give me the bag.

More power to that guy. How many people do you think they asked before Ice Cube who they read the script? They're like, no, I can't do this. Oh, I cannot do this? What are you talking about? No, in the meeting, they were like, do you think we can get Ice Cube for this?

Nah, he's never going to do it. Just ask. What's the worst going to happen? And he was like, yes, I'll do it. They were like, oh, my God, thank God we didn't have to go to Daniel Day Lewis. All right. I believe I have generally forgotten

what I was going to say, but I would just at least... You were saying there's a limit on code review. You're saying you don't believe. You're calling bullshit is what you're doing. Okay, okay. I forget where the professor was. I forget her name. I think it's Susan something or another.

I just read it just two days ago. She's coming up with the term called... Yeah, I know. Cognitive debt. It is... And so instead of just being like tech debt, which is here's part of our code in which we know is really, really bad,

but we've kind of covered it up. And someday we're going to be able to come back to this that actually make it better. We can't add more features on this one side of our product. Cognitive debt is when you have like a senior leave your team. And then there's just a portion of your code base that nobody knows how it works. And so her whole argument is that these people effectively claiming that they're reviewing, as T.J said, 20,000 lines of code a day.

Really all you're stacking up is positive debt. I want to make it clear. I'm not saying I can review 20,000 lines of code. Okay. Yeah. All right. 50,000, 60,000, whatever. Right.

That's way too low. That's way too low for me, guys. But you're getting this idea of cognitive debt, which is that we're stacking up something in which nobody knows how it works. And you now are only like your fate is directly tied to the performance of say Claude or chat Djipity or any of these programs. Because you can't like me as the human eventually like I can't even go in there and make a meaningful impact in a day anymore. Because it's just like it's like me starting a brand new job at, you know, Google. It's like, okay, figure out the code base. You're like, I don't know.

There's millions of lines at this point. And so I'm just curious what the future is going to look like here in six months because there's a lot of companies that are currently stacking cognitive debt. And we don't really know the outcome. I had a really good insightful question here, but I don't remember what it was. So that was my best swing at what I was like. I'll just say to really quick prime that I was reading somebody talking about this problem. And one of the things that I thought was really insightful about it was they were talking about how like maybe even if 99% of the time, the LLM can. understand all of it and it works really good. Like there's so many catastrophic outcomes that can

happen. Like for example, let's say you're running a SaaS and someone depends on your site and something goes wrong. And then the LLM gets stuck in a loop and cannot solve. Right? Which like everyone who has done this, it happens. It happens. Even on some type of stuff where you're like, bro, it's literally, it's right here. Like I'll just make the change. I'll make the change. Put the token in the bag, bro. These hands were not meant for any. anything besides prompting, but I'll still do it. These things were built for Excel, okay? And so I'll, you know, I'll tipy-typey in the editor still occasionally, you know, and it's like,

but so, okay, so then you imagine that. And then now you're like, okay, I need to run my agents in a loop. They have to do all this crazy stuff. No one can understand the code. And then it's like three weeks of downtime. Like, you'll just lose all your customers. You'll just all leave. So there's like so many things where I feel like people are not a plot.

like a proper level of like even just simple common sense game theory you can call it whatever you want to make it sound cooler orbital data to center um and or orbital data knowledge if you will um to this to this problem where it's like yeah you can get to that point but then what happens when something goes wrong you like actually need to think about that if it's your business and not your hobby blog for ruby on rails boiling the lake you know i don't know sorry demetri you were going say something? Let me just jump in really one quick second. Sorry, Denise. There was a paper just released. Just released right here.

The classic guest experience. Sorry, me and T.J. are talking, everybody. Hey, hey, hey, hey. Well, that's it for the stand of this week. Sorry we didn't get to talk to the guests at all, but maybe they could come on again next week. That's actually how we're doing this. We did clip a section out to put on YouTube and Bash happened

to not talk a lot in that section. Like, guys, that was so rude. Bash never spoke that episode. And I'm like, She spoke a lot later. We had a whole segment for her, guys. Relax. But anyway, sorry. Prime, continue on the thing on.

Sorry, I was just going to say they did release this paper. There are some holes with which models they tested, but effectively the floor rate, the bottom rate that if you give it a document and can it retrieve answers out of it, its hallucination rate was or its incorrect response rate was at 1.19% at the most bottom rate. And so this is out of these, like, and it gets worse as the context window gets open. So like, you know, just imagine what that means compound. in a code base at some point you could find yourself in a situation where you'd like tj describes where you're just you're beholden to you somehow conjuring the correct context and the correct prompt until you can master this or else you're just spinning tokens and i personally had this happen on a problem so it's just a very exciting thing and it was very simple solution all right sorry go on demetri you're

very very so i'm just going to say that on on exactly that and uh you know connecting to what teacher was saying about the 300,000 line Ruby on Rails blog. I think people are easily kind of duped by confusing. Is this thing like a software tool or is this thing like a personality? So sometimes I like to, I suggest this mental exercise. Whenever you get an AI output, I want you to ask yourself, if someone told me that they hired someone on Upwork to do this and they delivered it, how would I feel about it? Right. So, for example, if you hired someone to make a very simple, like, 2010 era blog, and they showed up with 300,000 lines of rail, rails, you'd say something's wrong here, right? I probably...

But if there's 150,000 lines of tests, would you say there's something wrong there? Quite likely. Yes. Yes, I would. There's something wrong up here. There's something wrong up here, Brian. Right. So this thing is a, it's an inhuman artifact, right? So, like, in the discussion with Casey, I was making this joke about, like, a 4 million line app to order pet food, right? And it was an exaggeration, but it's only like a factor of 10 exaggeration, right? Like a 300,000 line Rails blog is like at least 10 times the size it needs to be. It's made to do this. Right, it's amazing.

Exactly. The original thing was Rails new blog or Rails. Yes, exactly. I was going to say there's probably just a built in thing in Rails to like have this happen. This is how people were like, holy cow, Rails is amazing. Is DJ? built a Rails blog on stage. I'm pretty sure.

I'm pretty sure that's the initial thing. So you have this inhuman artifact, right? The 300,000 line blog. And you also now have to rely on an inhuman tool to do anything to it, right? So you're like, first of all, you could be screwed because the tool can't do it, right? As you were saying, that like it just spins

and doesn't solve a very simple thing. You could just be screwed and you're locked in. Right. So you may not be locked into exactly the provider. Not the good way. You are locked in somehow to using an AI tool to deal with it Because like what engineer let's say that the that the blog goes down tomorrow and it's like it's a big tragedy Lives are on the line somehow

And you go and try to get an engineer and say like we're going to pay you anything you want We need you to dive into this 300,000 line rails blog and solve this subtle problem Who's going to do that or at what price? Quick question Casey. what would it cost to get you to go solve the 300,000 line rails problem? So the answer is very simple. I'm going to message DHS really quick and ask him how much he would charge, and I'm going to add about 50 bucks to it.

And I'm going to do it as a pass-through. Smart. I love that. This is a really interesting strategy for monetizing contracts. Get someone who knows how to do it and then charge slightly more. That's my motto. A lot of strategy. That's my motto.

But I'm going to go ahead and guess that he will charge a lot of money. I'm guessing he won't be like $30. Like I'm guessing it's going to cost you a lot of money. Sorry, I interrupted you to ask Casey that. Go ahead, Dimitri. So I'm just riffing on that. I think one of the things that, so there's something I partly, I do with partial success in my own work. And I think many people are going to be doing this in the coming years is working on,

driving AI outputs to something that might be human scale, right? So getting it to optimize for terseness or optimize for legibility in the human sense. I don't know of any well-established practices for this. Like I basically just mess around with it until I get it into a state of, like, okay, fine, maybe I'll accept 15,000 lines of rails for a blog. I'm definitely not accepting 300,000 lines, right? So I think that that's definitely That's something that is coming because it has to

Right I don't like we're not going to be shipping 300,000 lines of blog Like every single blog is going to be 300,000 lines of Like not just 300,000 lines of code But 300,000 lines of completely custom code right So it's like one thing if it's I can't remember how big WordPress was I saw someone do like a side by side of this thing versus WordPress And WordPress was like, there are lots of side modules and stuff,

but they said that the core was like 150,000 lines, which, okay, fine, that's bloated as well, but whatever. But it's WordPress, and the whole world knows that code base, like metaphorically, the whole world knows that code base. So it's not, it's not 300,000 lines that not only does the whole world not know it, not even one person knows it, right? Not even Gary Tan knows it, right? So I just don't think that that can work. Long term.

And, you know, to play the other side for a moment here, like, everything here is new. There are growing pains. People try weird stuff. Right. Like early in any technology, or even, like, in games, like the early 3D games all looked weird and played weird until people figured out roughly how to get like stick controls and cameras working. It's working reasonably. So some of this is just, you know, people are working through the initial weirdness. And, yeah, it's weird.

So the part that I find a little frustrating is we could be a lot more honest about that, right? Yeah. We could just say, hey, look, I made this thing. It happens to work for me, this 300,000 line thing. But also, obviously, this can't be the path going forward, right? Or at least can't be the best thing possible, right? Like probably the blog would still work at. 200,000 lines of code.

I don't know about that, TJ. Okay, this is the block. I guess I would add something to that, which is, this is, like, if I had to summarize my primary frustration with AI currently, it is sort of just the condensed version of the things that Dimitri was talking about, which is basically like, it's being adopted too soon. That's, like, because it's a gold rush and everyone fears, being left behind and like you've got to get on the AI train and all that stuff.

The result of that is people are using it prematurely. So they're tackling things with it that should not be tackled. Like they should be tackled in an experimental sense, right? So to be fair to Gary Tan, just something I'm never going to do again. But to be fair to Gary Tan this one time. Thank you. No one gives a crap what he's doing, right?

So like the, you know, he makes a blog, it doesn't matter. That's not going to take down AWS. right? We haven't seen how much rail code there is. Also, we don't know. Maybe Kiro depends on the blog now, so. So it's just like, fine. Like, that's actually, Gary Tan making a 300,000 lines of code blog is actually very harmless

to the industry, I would argue. So it's actually not that bad. You know, it's fun to kind of poke at him on this podcast or whatever, but, like, actually, I don't have a problem with that. I really don't. Great. You know, he's, playing around with this thing and we're learning something from that and hey the results weren't that great but and maybe he's not ready to admit that yet but like you know that's we learned something from that right and that's all fine um the the worst thing is when yeah like when amazon

starts using this at scale and that was a bad idea and now people suffer the consequences because the infrastructure gets more brittle uh and all this other stuff and so really that's most of it it's like i would much prefer if we could have done done this in a way where we were validating that the AI could really do this better than humans for whatever the task was, and then we roll it out, that would have felt a lot more comfortable. And instead, we're very much riding the curve where like, oh, can an AI sort of maybe do this at all? Then we're doing that.

Like, in production as the actual model. And I understand why people are doing that because of the Gold Rush sort of, or the FOMO, of a problem. They don't want to be the last company to adopt. I get it but the problem is like I think we're going to pay a lot of, I think we're going to suffer a bunch over the next two or three years

for no reason other than people if they had just waited, like each task, if they just waited another six months or another year to like vet these things out and let the AIs get better at the same time and all that, we could have prevented. They're like they're riding just

below where they should be. instead of just above where they should be, right? I don't know how you feel about that, Dimitri, but that's kind of like the thing that I see being the bad part currently. So it is perhaps worse even than what you are imagining right now. Okay. So multiple times. Good.

So you're talking about within the tech industry. I very frequently work with industries that are not tech, right? I have worked with people who are in charge of various kinds of regulatory committees. And for example, they have a dispute about how a particular category of product, I can't say what it is, but a particular category of product should be regulated. And they'll say, well, maybe we should get the AI to write the regulation. And I have to talk them down and say, like, look, wait, like, stop, you know, stop right where you are. You obviously have an extremely dangerous idea in your head, which is that this thing can,

that this thing is smart in the way that you understand a human to be smart. And that when you ask it to do, write me, you know, 10 pages of regulations on the safe use of this device, that it's executing like human value judgments the way you might, right? And they don't know that, right? These are like smart people, like lots of degrees, 30 years of industry. These are not stupid people,

but they just have very dangerous mental models of what these things are and what they can do. Right. So like, however bad. Well, they're getting told all the time. Yeah. This thing is, this thing is literally God. This thing is literally God.

It's impossible for it to make a mistake. It knows every book that's ever been read, every scientific fact of all. time and it can't be wrong. That's because me, Dario, me Sam, or me Elon or whoever, we're saying, yes, and I created God. Yeah, and they say that exact thing to me that, well, all the, you know, like all of the federal register, all the regular federal rules, right, all the federal register is in the training data, right? So it should know all the rules. And I can just ask it to generate. And it, like, I, I spend a lot of my time talking people down

from these kinds of ideas. And the, the AI companies absolutely do not make that any easier. because they're obviously highly incentivized to say the opposite. Right. Like it's a lot harder. And partly there's a not a prisoner's dilemma is a Mexican standoff, I guess, where the like if whatever Dario would come out and say, you know, actually let's talk about having reasonable expectations. This isn't going to solve all your problems in life.

It's not going to get you a girlfriend. It's not whatever, right? that person would almost certainly be then at a disadvantage, at least in the financial markets, you know, compared to what, you know, Mosker Alvman or Hussabies at Deepind might be saying, right? I didn't mean to single out hit. Like, I think the problem is pervasive. Equally, for each of them, they have this exact same problem. And they also need boundless amounts of capital to execute on what their current vision is. no malice required just they think the way to get really awesome AI is to build a trillion dollars of data centers well they don't have a trillion dollars on hand they have to raise the capital so they can't like it like it actually makes sense it follows yeah oh that's everyone is motivated by the same thing everyone is worried about being cars references in second third place fourth place whatever right so if you're if you're terrified of that and you think that it will justify you know that that that

justifies it, then, you know, the AI companies are behaving in that way for the same reason that the other ones are. What is, it's not that funny the car's references. No, it's just, it's being motivated only by car's preferences. But we know that about him. We know that he's only, like, Twitter chat, just, Twitch chat noticing my joke always makes me laugh. Sorry, Casey. Okay, so I was done anyway. one thing out there that I think one thing that is kind of, I would argue as being massively understated in this is that it's not that I think all these AI companies and all these people are thinking that they're going to be losing their job and they have to be keeping up with the

times or else they're going to be left behind and in the permanent underclass. I do think a large amount of people just don't perform that hard on their jobs. I've worked with plenty of people who they just want to really arrive there and do three hours of work, right? We've seen the TikToks. We know this just simply exists. Sure. And if you have something that's going to take the 30 minutes you would normally spend on writing an email and trim that down to three minutes, you can just compress your four-hour workday into 20 minutes and you're just like, why would I not press the easy button? They're going to press the button.

And so for me, it just seems like a very simple, like a large percentage of this is just like, oh, this just makes all the little like mundane things I do not in software world super, super easy. And for those that just don't care. I mean, we've all worked with plenty of people that can produce amazing code, including myself. and then you produce this amazing and you're just like, oh, dude, I can just easy button this one thing out. Easy button, easy button it. But I don't think that's on an individual base.

That's like somebody individually who decides to use AI and they're really not going to be the problem, quote unquote, because the problem comes from institutional adoption, right? Like that person is not really an issue for anyone other than like maybe, you know, if you're getting slightly lower quality work because the AI isn't doing as good a job writing the email, but it probably is.

So, like, I don't know that I'm that worried about even that. Because if someone was that checked out to begin with, like you say, like, I mean, how much worse is the AI really going to be? So the problem is more when you have this institutional adoption because then what you're doing is you're forcing people who, the people who were taking their job seriously to take it less seriously effectively, right? Because you're forcing them to use this tool

and they may not be able to get the, results that they want from the tool and you're forcing them to use it and all of that stuff. And so if that's premature, that's, to me, the problem. I don't see it as being a huge problem if the lazy person just wants to use AI to make themselves even lazier, because they were already kind of a problem to begin with and, like, you know, how much more of a problem are they really becoming? I don't know. To me, anyway, that would be my gut reaction to that.

But, Prime, maybe you, I don't know, maybe you disagree there, though. It's not that, like, I mean, obviously the institution is the catalyst, but when, you know, I don't know what percentage of people at these, like, super large companies tend to have a very large amount of checked out people. Yeah. It's just like a reality. But they were already non-contributors, sort of, is my, is what I'm saying. But they're also non-they also didn't make a big impact. Now they can move at a rate that is like unknown to them previously.

And so now they won't. Because they're lazy. They can make bigger impact than they've ever made before, right? They're like, I'm 10x more productive. And maybe I don't want you to be 10x more productive. Like right now, I can tell you that I've had. had more problems on YouTube studio than I've ever had in the last five years I've had in the last a year. It's almost like every single time I use that product, a new thing is just completely

broken constantly. And it's just like, I don't, I know Google is like historically a place where you go not to work. That's just like one of the, a big thing of the general software. And I know there's some places in Google that are very rigorous and very difficult to work. And I'm positive that exists. But I've met enough people to know that there's like a lot of coasting going on. And now there's a lot of changes happening, but everything feels broken. and I don't know what to attribute this to, other than like, hey, we've given a bunch of people easy buttons.

Now everyone's moving at a rate faster than they normally would, but they still have the same decision-making level that they did before they went into it, which is like, to me, the dangerous part. It's like, I don't want you moving two-x faster. I want you moving more. I want you move the same speed. You could be right. That would be terrible.

If a bunch of people who don't really care are now suddenly doing a lot of, quote-unquote, work, that would be very bad. That is, I don't know. What appears to be happening. Okay. then I agree then I agree with you that's pretty bad if that doesn't even have to be they don't care it's just they may not have even the skills to recognize like demetri your point of saying like hey I thought the AI could solve this problem well but I don't it actually can't it actually can't write a good rigorous spec for this thing for 10 pages or like I'm a junior dev I don't know that these five things are actually hidden foot guns that I shouldn't merge that are going to turn into gigantic foot cannons in PR two three four five but now I could just merge those in and they're like a cascade of, you know, effects because nobody is putting the stop to those things. That part, you know, does strike me as kind of dangerous.

Demetri, I do want to circle back to something you said. You said you're working on trying to minimize these outputs with like, okay, hey, we have 300,000 line Ruby blog, but really that should be, you should be able to get that down to 50,000 lines, 100,000 lines, right? You're trying to work on how to reduce the entropy, not grow it. how like is there actual meaningful progress being made in that region to the point where year over year you're saying hey this would have taken 300,000 line ruby blog post but now with our work we're at 275 and is there any sort of like direction that you're saying hey we're actually improving

this kind of stuff because that's to me that's very appealing like I'd much rather see it play code golf than play whatever it's playing right now so that's a this is a small slice of the industry where where I'm working and I'm not aware of any I know very few people who are even talking about trying to solve this problem problem, right? On my side, I do this all through

sort of meta-practices of, I mean, just the most basic, starting from the most basic thing, which is I have some project, let's just say hypothetically, a blog, right? I know roughly what the components of the blog should be. I'll make an architecture diagram,

metaphorically speaking, and then I'll say, hey, like this thing should be ballpark, lines of code and they'll have it generate that and separately evaluate that right and then write out an interface file and then separately generate another module of okay we've got whatever like the back end and the orm that's one thing it's standalone it's done right and then i'll go and do whatever like here's the low balancing layer right if you have like a super high performance blog or something um i found that's

uh generally my the thing that i would like to encourage people to do is like it instead of trying to solve the biggest, most hypothetical game you can right away, just try to maintain quality and increase your productivity by 10% and see if you can do that, right? And that's, so like people are trying to go from, you know, like whatever my current pace is, I'm going to try to go 10 times as fast. And, you know, I mean, that would be great. But actually, in the history of computer programming, like a 10 or 20% productivity improvement is already enormous. Right.

Like, there are very few things that actually be great. Right? So what I encourage people to try to do is can you just get 10% better using AI? And if you frame it that way, I approach things very differently, which is I sort of spec out, this is how I would do the project myself. Here are the components. Here's how they talk to each other. Write out a spec for the interface and then generate each thing separately. And so I have like 20 AI generated modules that are a couple thousand lines of code each. And then I just stitch them together at the end with like $500. lines of code that I wrote using the interfaces that I spec, and then it filled out. And there I'm able to generate things that are ballparked the same size of code that I would generate. AI is more

robust than I would be, but I come from a mathematical computing background, so my style is very, very terse to begin with, right? I think you can generate quite reasonable artifacts that way. I don't know of any standard organized push to try to do what you suggested prime, which is is, okay, I got the inhuman artifact, the 300,000 lines. Can we now shave it down? I haven't seen anything like that that works. Now, I've tried here, you know, take these 50,000 lines, and can you reduce it down to 40,000 lines?

And it's extremely unreliable. I haven't done that test in a few months. So you've probably heard the meme that, like, everything changed December of 2025. And there's some truth to that. It's not as true as some people are. are suggesting. Oh, can I follow up with that after you're done with what you're both

saying? I actually do want to ask you about that. Yeah, I'm done with this. Oh, just, yes. Okay. So I have a hypothesis that the reason why people are seeing the big change that they're seeing in December 2025 is not because the new model was just like 20x better, but because we've been investing nonstop effort into all these harnesses.

And they just have gotten better and better and better over time. So this nominal improvement in the model actually made something that perceivably was a larger improvement, but it actually is just the stuff around the model that has so vastly improved over the course of the last year and not the model itself. Yeah, I mean, I would agree without like 70%. I mean, there is substantial improvement in like 4.6, but also you're absolutely correct that I would attribute most of the excitement of the like post-Christmas 2025 meme is sufficient. maturation of these matter practices like, you know, agent loops and like agent teams and

stuff like that. So I think broadly I agree, although there really was a meaningful improvement with 4.16. I've got one last bigger question that I wanted, or like topic that I wanted to talk about, that I wanted to make sure we covered from the podcast. Prime, did you have anything else you wanted to add or Dimitri? But first, Dimitri and Casey, are you still okay going because we were a half an hour over our projected amount of time? And I don't want to take your guys's time if it's too valuable. I should probably go in like five minutes. Okay.

But he just has one quick big topic. He wants to talk about. Really quick big topic. Okay. I got to get up a whiteboard too. I thought that was a fake one. Oh, no. You got to have one.

That's a green screen. And then one that's a real one. That's the trick. That's the trick. Okay. It's not see-through though. So I have to look away when I'm ready to get. So I'm just going to draw this really quick.

Just kidding. That side's dead. There we go. Supply and demand. I have a quick question about this. There you go. That's all I need. Okay.

Simple topic. That was a really, really good diagram. Thanks. It's one that a lot of people aren't familiar with, so I wanted to make sure that we drew it out for those that aren't out there. There's two part question, just so you know where I want to go with this. The first one is sort of like, I don't feel people meaningfully talking about, like, let's say we 100 or 1,000 X or like a million X demand for tokens,

which would have to be the case if everybody's using tokens for all their jobs. aren't we going to like 100 million X or 300 million X? It's going to be a big number. Yeah. I'm using way more tokens than like my dad. My dad like occasionally asks chat GPT something. He doesn't even know about deep research mode. Right.

Like so I can burn a million tokens before even notice. Right. So it's like, so there's that part of the equation where if the demand is going to go crazy on one side is I'm not going to push the price of tokens up higher even if we have technical things. But on the flip side of that of this like we're getting tokens cheaper. and all this other stuff. And you guys talked about Javon's paradox. So I just want to at least give a spot for you to at least bring up this topic of some things like, hey, if we're going to get tokens a bunch cheaper, but then we're going to keep pushing up demand for tokens.

Is that going to make it more expensive? And what does that mean for like everything that people want to do with tokens? Do we want to do agent loops? Everyone can't possibly just have an agent loop running 100 million tokens a second. So it's a big broad question. I'm just interested in your thoughts sort of on like those market pressures where you see that and like how that affects people's usage of AI if it actually does get broadly adopted. I guess I'll say two things. One connecting to the previous conversation of the 300,000 line blog.

Obviously that's generating, that's consuming and generating a lot more tokens. Yes. Than you might think would be required to write a blog in Rails, right? Right. So there is that. there's this question of how long will we be in a period of subsidized growth with multiple players so that no one can really afford to try to squeeze the market for margin. And as I said with Casey, the big thing everyone is worried about is somehow someone gets a position in Gen AI

that is as dominant as Google had or has in search. right so Google for a long time had like 90 plus percent of the market and the next player was like 8 percent or something right so it was like 90 percent Google 8 percent uh Yahoo or then later being and then like 2 percent random other stuff right at least in the U.S. market so I should say that but almost all my experiences in the U.S. market right so I mean notably once Google got that dominant position they were able to eat tons of the margin in advertising right and And so without some competing search engine to force Google to reduce margin on ad clicks, well, they just sat and ate almost all of the margin in the online advertising business.

If that happens, it could be very scary economically for anyone who's relying on AI, right? Because right now we're, right now we have not only this VC honeymoon period of, they don't really have to make the money back right now. Right. And that phase could last a long time. Like Uber was in that phase for like 10 years roughly. So we have this, like at some point the honeymoon will be over. And possibly also at some point there will be a Google of Gen A.I.

And it might even be Google. Right. So like the Google of Gen A.I might be Google. We've seen what they've done with a Google style advantage in the past in terms of how much of the value in the market they were able to capture. Right. And so I think, like, we won't know anything about that for years, but I think that's the biggest factor. Right. I do think that the open stuff is getting better and better, and I think it will be good enough. I mean, it's already good enough soon, actually, that you can use it as a substitute in many cases. The main, excuse me, the main thing that is missing in the

in the open stuff is the relatively polished product wrapped around it and like all the extra features that go into the chat. I understand like you and I are using it through the API but like the the mass of the market that still has not moved through the phase of hey what the heck is this thing and can I

can I use it? They're going to go through the chat interfaces first right even if they eventually end up using using it through the API So yeah, I guess the open models are really quite good and getting better. So at some point, that will have to come to a head where people will say, okay, look, like, yeah, Claude is great, but, you know, between Deep Seek and Quinn and Mistral and Lama, I can stitch together a pipeline that does 90% of what Claude does, and all it costs me is electricity, right? Right.

And that's so far on the other end of the economic structure that I don't have good predictions. I frequently set up such systems for customers where we just do like everything is open because like whatever. For regulatory reasons it has to be on premises, nothing can leave. There can't be some third party involved in the deal or whatever. So that's so far on the other side that it's hard to compare directly because setting it up is also more complicated. but I think probably well not probably for certain there are people right now who are trying to build

kind of a like bring your own open model and they'll just host it and like pipe stuff together for you something like that has to succeed in the next few years I haven't seen anyone who's like has a strong lead there but something like that I think has to exist in the market if only as a as a check on

on how greedy the commercial players can be. I don't know if that answers your question, but sorry, Pram, go ahead. Just a quick question. If, let's just say all things are given the same, I don't know much about kind of the infrastructure overhead that OpenAI or Cloud currently has, like how much room do they have to be able to grow? But if everybody were to turn on a Cloudbot tomorrow, and they all just ran 100,000 tokens through,

which is not a very hard amount to run through, or maybe 250,000 tokens. And just in America, I mean, that would be hundreds of trillions of tokens that just weren't there beforehand. Like, how can that, in this theoretical situation, would we see, I would assume the infrastructure would be so overwhelmed that there would,

like we would have to skyrocket price? Would that be the only path forward? So if something doesn't come out of this, then price does have to go up if usage goes up a whole bunch. Or throttle speed. Or throttle speed. Which effectively will reduce the usage. Yeah.

But that's why everyone is fighting for rights to build data centers and get electricity and hydropower. Because you're thinking of it the same way they're thinking about it at least in public, which is like we are maybe 10%, 20% into market penetration of AI use. So they're hoping to get to 100% penetration, right? So that's like five to 10 times where we are now. In order to do that, they need vastly more physical resources just to run the thing. never mind making better models, anything else.

Just freeze, like freeze right now everything you have and just serve it to 10 times as many people doing, and not just 10 times as many people, but like 10 times as many people doing maybe 10 times as many tasks, right? Because maybe right now we're only using it for coding, but later we'll use it for, you know, I don't know. Well, like cursor's got remote agents.

It takes a video and sends you a video of the output to see if the thing actually still runs so that you can know, Like, did your front end change actually work? Yes. Like, that burns a lot of tokens. It's super cool feature. But, like, that's a ton of tokens.

Dmitri said he had to go. Yeah. Dimitri, thank you. Oh, sorry, yeah, I have to let him go. Where can everybody find you at? What's your most, like, favorite place that we could link you at? It's just attached to Casey's regular thing at computer. It has this waiting through AI.

That's Casey and me roughly once a week. And then if you want to find me specifically for whatever reason, It's Dimitrispanos.com. Sweet. Dimitrispanos.com. We'll link that in the description for people later as well. Yeah, I just want to make sure I get it. PhD.

Yes. You have Ph.D. I'm, I am generally anti-credentialist, but yes, I do have a PhD. Okay. Well, that's like, yeah, me too. If I had a doctor, I'd be against credentials too, but then I would say it like I'm cool as well. I'm not against credentials. I'm against credentialism.

Yeah. All right, well, thank you very much, Dr. Spanos. Yes. Thanks, Dimitri. We do need to get you back on here. Hopefully, our style, which is very different than Casey's, which involves a little bit more joking in orbital data centers. Hopefully you enjoyed your time here.

We certainly enjoy you. Chat was wild. They want you back. A lot of people that bring back Dimitri, they really, really appreciate you. So, thank you very much for joining us. I appreciate it. I appreciate it.

Yeah. All right. I'm out. Thank you as a man. Bye. Bye.