Adam Jacob: Welcome to Infrastructure Frontiers, where automation experts run down all the AI news for infrastructure engineers. It's Monday, August 31st. I think that's the Monday it'll be,

Paul Stack: Amber first.

Nick: twenty eighth.

Adam Jacob: I think it's August 31st. So let's look at the calendar. Let me look. Yes, it will be Monday, August 31st, when you're listening to this. So grab a cup of coffee and let's get started. I'm Adam Jacob. I'm the CEO of Swamp Club. I'm joined this week by my swamp co founders, Paul Stack.

Paul Stack: Monday.

Adam Jacob: And Nick Steinmates.

Nick: It's me. Hello.

Adam Jacob: And our special guest is infrastructure engineer Jason Powell. Jason's a 16 year veteran of the game. He's from building data centers all the way up, you know, core networks, sans, virtualization. He's done it all. Also, Jason loves Voltron. And I remember the moment I collected all of the lions as a kid, and like my absolute joy when I finally had them all and put it put it all together and like formed Voltron. And so I was looking at it in the background and it brought me great pleasure.

Jason Powell: Yeah. Awesome. Thanks. Yeah, Jason. It's nice to be here and talk with you guys. this is what I talk about all day anyway, so it works. and fun story that Voltron it's built out of Legos and each cat is it comes apart like you fully build it. So

Adam Jacob: that's even cooler, right?

Jason Powell: I love yeah, that's my those are all Lego actually. So

Adam Jacob: Incredible. If you're listening on the podcast, he's got like, you know, Lego rocket ships and a Lego Voltron and a Lego space shuttle, right? A Lego Globe, right? madness. Okay, let's run down the news. so OpenAI has been building their own inferencing chipset called Jalapeno. And this started in mid twenty twenty four. And now they've got working chips in production, which is an insanely fast turnaround time, right? I think it was like 16 months from design to to building the first chips. So the folks at semianalysis.com, who have a test suite, I think, called inference X, got to test the actual chip. They didn't get to, they didn't get a chip, but they, you know, got to run their inference across it. And it basically just outperformed everyone on performance per watt. So tokens per watt output. And considering that the constraint right now on deployment of AI is power, that's the metric that matters the most, right? data center construction, data center all of that work is power limited. So more efficiency is hugely better. The article is super deep and comprehensive. it like walks through all kinds of interesting math if you're interested in it and sort of how you how to think about that performance. But I think it really points to our just being at the beginning of figuring out how to serve all of the demand. That we need to serve. And that the timelines for delivery here show that we're going to use AI to build new tools and better infrastructure at a rate that's crazy, right? This first generation leap of how we used AI, it's going to start to feed back in to like, now we're built we're using it to then build better chips, which then causes power to work, which then causes the whole thing to move. And so I think this is a sign of of of our ability to do that. and Of our ability to use that technology to solve the interesting bottlenecks that exist right now in deployment. Right. So when we look at like how are we going to get all these data centers out? What are we going to do about power? Well, like one answer here would be let's quadruple the efficiency tokens per watt. And suddenly we just need less power, you know? and that's one way to solve that problem. Jason, what do you think about jalapeno?

Jason Powell: I mean, it's awesome. It's gonna be neat to see. I think that the the difference in in the inference, making a big, you know, change, right? So I mean, NVIDIA has seen the lion's share of everything for a while now. So, you know, it's gonna be interesting to see compete and the fact that they're coming off these these fabs that are the most complex machines that are basically ever created, you know, imagined by us and we're pumping them out like it's nothing. So it's it's it's crazy. bottlenecks are moving faster than you know, we can keep up, which is the challenge now.

Nick: Yeah, the real

Adam Jacob: Totally.

Nick: the real question in my mind is are they gonna arm their competitors or not? And I'm I'm guessing they're not. You know? I think this is Ho Jalapena's open AI exclusive. And

Adam Jacob: No way.

Jason Powell: Yeah.

Nick: and it's and it you know, how can it not be if like the semi analysis

Adam Jacob: 'Cause they'll

Nick: people didn't have their own. They had to fucking use their other c you know, someone else's cluster to be able to test this out. I don't know, man.

Adam Jacob: Because they love money. And because because, you know, to prop up those valuations and to prop up the sky high insanity, and we'll talk about some of that later in terms of anthropic, like you being able to own more and more of your destiny and to and to prop up your neighbors, like is only to the good. And so like, yeah, one of the things that people did talk about with jalapeno was the idea that it was like that it was somehow open AI specific. And part of what this workload was showing is that it was not. They were running all kinds of models against it. So like it's not that it's fast because it's been tuned just for open AI. It's it is designed to run general purpose models. And I think that's what that's telling telling you is that they're absolutely going to wind up just selling jalapeno all over the all over the show.

Jason Powell: Yeah, Cerberus too, right? They're making the the big waiver

Adam Jacob: Yep.

Jason Powell: scale chips, right?

Adam Jacob: Yes. And they're already running Doom on Jalapeno, so before anybody needs to worry about that, that's already happened. So right?

Nick: Well I can't wait to put some there then, you know?

Adam Jacob: Cube's gonna have a jalapeno in the back of the rack. Okay. all

Nick: Absolutely. Yeah.

Paul Stack: 100%.

Adam Jacob: right, so moving on. So Duck Labs. So the folks who built DuckDB and a few of the other pieces around it, have joined AWS. So DuckDB itself stays in the open foundation that started To hold it. you know, first congratulations anytime somebody builds great technology. And then what happens is they, is they, they get acquired, you know, like what they deserve is all the praise in the world, both for building that technology and for, you know, finding a way that I hope they get paid and you know, find the home they're looking for. DuckDB is interesting because a lot of the monetization that was happening on top of it already was happening outside of the guys who built DuckDB. You know, like they built DuckDB. I'm and But you know, like Mother Duck and a few others that have become pretty good sized startups, sort of built on top of DuckDB. All of those people had nice things to say in the blog post that announced this, this move, the open foundation that they had for DuckDB, all of that stuff. you know, DuckDB, other embedded databases, I think are really starting to move and we're starting to see like legitimate alternatives to SQLite. DuckDB is already a very legitimate alternative for doing analytics. You know, what do we think about AWS as the steward here? There's always been some, you know, some concern around open around AWS as a as a steward of open source, as a as a you know, way to think about competition sort of within that startup ecosystem. Like, do we have worries here about that at all? Or are we just like, nah, it's a thousand flowers blooming, they found a nice home, you know, don't worry about it.

Paul Stack: I mean, is on a lot as well for open search and for that community. So I think they're getting better in this regard, right? They've learned a lot over the last, they've had some bad press with regards to some of the open source plays that they've actually made, but, you know, open searches is proven that they have a model that they can actually start to work with. And like DuckDB could definitely follow and, you know, and it could like spawn a whole ecosystem. There's a, like the fact that they've been working together for such a period of time already. sort of driving the project forward means that they have a real vested interest in what's actually happening here as well, which, you know, it shows that this isn't just, look, there's a cool project, let's go and buy it. They can really see that there's something in there and they can see something in that team because they're going to keep the team together as well as part of.

Adam Jacob: Yeah, for sure.

Jason Powell: That's good.

Adam Jacob: All right, moving on. So China's Z AI dropped GLM five three flash, which essentially delivers Opus level performance in a much smaller model at an incredibly lower price point. and they explained it basically as as hey, we're delivering this on top of Chinese AI chips, and we've tied together the architecture, richer multimodal. multimodal pre-training and then design of the inference hardware, that all three of those things being co-designed together are why the model's getting the efficiency that it's getting. I think that's interesting on its own, you know, in terms of the geopolitics of who's allowed to have GPUs and who are we allowed to ship silicon and how does all of that work. That's obviously like interesting on its own. More interesting than all of that, I think, is that they launched it in stealth as OX Alpha. And essentially made it free to try. And a bunch of people did, you know, a bunch of people went to open code and ran, you know, OX Alpha through their router and loved it. And then it was sort of secret revealed to be to be to be this model. That was a beautiful piece of marketing, I think. Jason, what do you think?

Jason Powell: Yeah, I mean it got the it got the stigma out of the way, right? Like you know, the the blind spots and some of the cool things that they're doing with models, like we're gonna see I hope coming with more of the open weights and more than out of the big foundries is that, you know, they're you know, smaller teams are becoming scrappy to take these models. Like you can even look at Muse, one of the big ones, that because of the way it introduces the multimodal, it's it you know, it's not as extensive and they can pack a lot more punch. So you can get a lot more granularity and sort of models that are specific for things. I think that's gonna be where we try to go 'cause, you know, we you don't always need the frontier, right? I mean

Nick: The distillation wars have begun. I love it. We're here for

Jason Powell: Yeah.

Nick: it. Driving the price down of AI is awesome. And also,

Jason Powell: Yeah. Right.

Nick: Adam, you made this point a little bit. The geopolitics of it are fascinating because what what actually did open AI just prove? That you can build a better chip and you don't need to rely on NVIDIA, man. And so, yeah, you distill the shit out of the frontier models. You build bespoke chips in China, which they can absolutely do. And like, yeah, all of these concerns about proliferating NVIDIA everywhere or whatever, having frontier GPUs, like they seem to be evaporating. So it's a good time to be a consumer.

Jason Powell: Well, I mean, I think I talked about

Adam Jacob: Yeah.

Jason Powell: NVIDIA yesterday, right? You know, the the the hugging face stuff. I mean, I

Adam Jacob: yeah.

Jason Powell: I was talking to a good buddy about that and I I hope that it's not a bad play. I hope that NVIDIA says, Well look, you know, maybe the future of inference isn't on the GPUs. The more that they get to it, the more that the math, you know, comes down, that the fine tuning and they can start moving to CPU infrastructures and some this ASIC and special built, you know, hardware. Well maybe Nvidia says, Okay, let me ensure that, you know, somebody doesn't take away some sort of a lead we have, right? I you know, I don't know. I I hope. Or I hope

Adam Jacob: Yeah, I mean

Jason Powell: somebody comes up and does the same thing.

Adam Jacob: I mean, that was a very natural segue into a into another news story, which was

Paul Stack: You

Adam Jacob: which was Hugging Face getting bought by NVIDIA, which which actually, you know, the first leak happened like Monday and then it got confirmed

Jason Powell: Yeah.

Adam Jacob: sort of midweek that they are actually buying it. And like, I think you're totally right to draw a line between those, between this story and that one, because I think what you see in the Hugging Face acquisition is so much of the control here isn't just the chips, it's the software stack, right? It's it's Okay, great. You built a model. It's got to run on top of some inference. Like, what is the software stack that runs that model? And the thing about NVIDIA is that that entire stack is connected, right? Like it goes from the hardware to the software stack to the way they deploy it. Like they're involved in every layer in that stack. And so I think, you know, the hugging face acquisition sort of tells you, hey, we know that what we're gonna that what we need is to control the software stack so that the deployment of these models. Whether it's going to NVIDIA or it's going to jalapeno or it's going to Chinese AI chips, like NVIDIA needs to be the place where where where that start of where that motion starts. and if it starts with Hugging Face, you know, it starts with here's the model, here are the weights, you know, and then it goes from Hugging Face to I deployed this directly onto your hardware, you know, you can kind of see that you can see the path of right, because they've been doing it for years. And so

Jason Powell: Well, NVIDIA's been doing it for years. Melanox. Melanox, right? They've just been slowly building their entire stack to own it, right? All of it. Bear Rubin

Paul Stack: it.

Jason Powell: was the first, you know, just ship it in, call it NVIDIA and be done, right? That's what they want. You don't have to talk to anybody else. You don't have to do anything else. I mean, that's their plan.

Paul Stack: And they announced this like the day after their earnings, which beat expectations in the market. You know what I mean? It's like,

Adam Jacob: Yeah.

Paul Stack: what a day to throw your money around and say, hey, we had a huge second quarter. We're just going to spend $13 billion bringing a bit more of the system into our control and ways that our customers can actually do it. It's kind of fascinating. The fact that that company got such a great payoff, such a great rise in valuation is absolutely fantastic.

Adam Jacob: Yeah. Twelve point nine billion. Twelve point nine billion

Paul Stack: Exactly.

Adam Jacob: dollars. Kevin Durant put in two hundred and fifty grand to hugging face. I think he made what was it, sixty million dollars or something?

Jason Powell: I mean open router, right? Like and Stripe acquiring that, like

Adam Jacob: Yeah.

Jason Powell: well the the cool I think one of the coolest things I'm seeing is that the big players see these the the the the FOSS community or the the the free and open source, you know, and they're like, well it's moving so fast that they know that's where the frontier of software seems to be, but how they control

Adam Jacob: Mm-hmm.

Jason Powell: it, right? 'Cause they want, you know, that's in their purview to control that. So they're now hopefully seeing that they can invest in that infrastructure and boost that and maybe reap the benefits of it that way. you know.

Adam Jacob: Yeah, and I think I think the fear that I've seen people talk about with Hugging Face and NVIDIA is like, well, maybe NVIDIA won't keep it open. Maybe it'll start to tie it more to NVIDIA's platform. And I I my bet is that they don't. My bet is that they also love money. And they know that like the that if they're competing on how easy it'll be to use their consolidated stack versus if you go piecemeal and you're using other chips, like they're gonna be happy to be the to be the layer that that emboldens you and take your money while you deploy not on NVIDIA gear so that they just get the opportunity to come back around again and cross sell you. You know, it's smart. yeah. And twelve point nine billion dollars, you know, Jensen's got that on the weekend. It's pocket lit. So

Jason Powell: They make the

Paul Stack: Swamp's

Jason Powell: what they They make the the GDP of small

Paul Stack: open for business, Jensen. Just saying. call us.

Adam Jacob: Yeah. Yeah.

Jason Powell: countries in a day. Yeah, saw that the other day on Bloomberg or something, right?

Adam Jacob: Sorry, what was it?

Jason Powell: The that NVIDIA makes the GP GDP of a small country in one day. Yeah, that's that that that's I mean that's just like I I can't even imagine. Like that's unfathomable.

Adam Jacob: Yes. Yes. Right? Bananas. Yeah. Yeah. The numbers we're talking about are nutty and getting nuttier. okay, let's let's go back to let's talk about Bill Gates. Bill Gates put out a new note called A Turbulent AI Era and Critique. so I love a good Bill Gates note. You know, every time Bill Gates writes notes, has been writing notes, you know, the his notes in Microsoft are legendary. I thought this one was particularly good. and if you're, you know, if you're listening to this podcast, you should go read it. You know, Bill basically sees three big risks. So, you know, many jobs are gonna disappear forever. AI is gonna empower people and perhaps other AI to do harm. and AI could stunt our kids' development and replace human relationships if we let it. And at the same time, He's very clear that what we do with AI can be very, very good. You know, like there's a ton of things we can do that will help society to move forward, that will that will improve the lives of everyone, that will create opportunity for for everyone. and he sort of calls for us to create a domestic and international framework for figuring out how to deal with AI across every sector. Basically, the sweeping potential for good and the real potential for for evil and for bad outcomes is is so real. That we need to figure out how to do this in a structured way internationally. you know, and he basically calls for us to do that across every sector from national security to employment to public lands. And he proposes that some jobs should be set aside for humans. He calls them human-reserved jobs, which

Nick: Ha ha ha

Adam Jacob: I really I that felt like a very Bill Gatessian thing to say. Like, I do not want to work a human-reserved job. You know what I mean? Like that just sounds.

Jason Powell: I can imagine his data center and he's got people standing there and you buy a reserved, a reserved human for your operation, right?

Nick: Yeah. Jesus Christ.

Adam Jacob: Yeah. I just you got a reserved person. I just like it just it sounded very dystopian to be like, I I work a human reserved human reserved

Nick: The the whole thing does, actually.

Adam Jacob: jobs, you

Jason Powell: Yeah.

Adam Jacob: know? he also calls for taxing robots, and AI so that we don't totally skew the economics in ways that only benefit the rich. I think that's actually very smart. I think the you know, right now we tax we we we tax you a lot on employment. You know, when I hire an employee, I pay a ton of tax. When I create a better software factory, I pay none. You know, like it's just that's pure RD expense. like I don't think that's gonna hold either. I don't I think he's I think that's smart. and so he sort of commits to using his voice and his time to get AI and equity to be higher on the public agenda. So overall, you know, whether you agree with this, and we're gonna talk more about it, I am sure. This one gets an A plus from me. Here's Bill's money quote, and then I'm gonna open it to what I know will be controversy on the floor. So you have a chance. His message to governments was this. He said, You have a chance to act now before unemployment rises sharply, communities are hurting, and public trust has eroded. You can make sure that your government handles the problem holistically rather than diving it up into multiple bureaucratic fiefdoms. You can make sure AI benefits everyone, and you can work with other governments to meet this national and global challenge. And I I just I think he's right. You know, I think that sentiment is correct. Go Paul.

Paul Stack: mean, we can, we can like open it up and make the government do it. And then we're just going to give the fucking contracts to our friends anyway. Right. And pretend that it's not a real thing. So come on. Like, I mean, you know what I mean? Like, I mean, sure you can do this, but is

Adam Jacob: Ha ha ha.

Paul Stack: it going to be done in a fair and like diplomatic way? Probably not. Like, I mean. Yeah.

Nick: Reg regulatory capture is one hundred percent a concern. I think, you know, what's interesting is I think Jensen is famous for saying, you know, don't become a radiologist, because all these jobs are going away completely autumn, you know, completely automated by AI. And now, you know, there's not enough because actually, as it turns out, that AI assisted radiology means that more radiology can be done. You know, it's Jevon's paradox applied to there. you you you extrapolate there, you know, like we've been 24 months into AI outcomes or whatever, and hiring is on the upward slope. You can't hire enough engineers. Matthew Prince, you know, very famously has said, Yeah, the shape of my organization is changing at Cloudflare, but now we can play offense. You know, now we have all of these teams that can go, we have infinite ideas of shit that we can do. And now what AI is unlocking is the ability to actually go play offense and do it. And so I don't know. I don't know about this like everyone's gonna lose their job. I think actually it's gonna be an era, you know, it it just is might just as well might have an era of like actually giant proliferation of people having a lot more work to do and a lot more interesting work to do as a result of this.

Adam Jacob: Well, I think I think that's I think that's Bill's point, is that we get to choose which one of those worlds we wind up in. You know, there is a wor there is a version of this story that doesn't go that way. And

Nick: Yeah, thanks Bill Gates for allowing us to be able to move in that direction or whatever. You know, like great, you know. I'm glad you have an opinion.

Adam Jacob: How?

Paul Stack: I I saw iRobot. It looked fucking cool in the future with the robots. I'm down for it. Like, let's see what happens, right?

Jason Powell: I mean I mean, I'm gonna say it right, like I bought a mini duck or a Yeah, I bought

Nick: Yeah, of course you did. Like who didn't?

Adam Jacob: Yeah.

Paul Stack: Hahaha!

Adam Jacob: Yeah.

Jason Powell: one. I ordered it this morning. So you know, the hugging face robot. My I have a daughter who is gonna be nine in September and she has just she loves science and she loves doing you know, like she likes doing circuitry and stuff with me and so I'm like, Yeah, awesome, let's do it, let's

Adam Jacob: Yeah. Let's go.

Jason Powell: go, right? Like I'm ready to bring her into the future, you know. Sha' d I'm not gonna hide her from it, but she's gonna do

Adam Jacob: No.

Jason Powell: it she's gonna do it intelligently, right? She's gonna know. So

Adam Jacob: Yes. And I I mean that's I have a sixteen year old daughter. obviously like navigating AI is a thing, but for for for her generation, it's social media, right? and yeah, I think I think every generation has challenges like that where it's like, here's something new. We have to figure out how to adapt, we're we have to figure out what's healthy and what's not healthy and what works for us and what doesn't. And I think we're we're we're sort of in that moment. And yeah, I mean, let's go for more robots, right? But maybe tax the robots. okay. Paul keeps like, do not tax the robots. This is silly. It's lunacy.

Nick: Who I mean, whatever.

Jason Powell: Yeah.

Adam Jacob: all right. So Paul wrote a great blog post about how we at Swamp have moved our CI flow into the agent and started building attestation instead instead. And we're already seeing real speed up from that. But more than that, I think it it it sort of Points to how much further it's possible to push in terms of saying, Hey, we have all these best practices that we came up with over years that really did serve us. We have to figure out how to keep the parts of those best practices that were that were beneficial and that that moved us forward as an industry. And we can rethink how those core principles are aligned in order to maximize our own experience today. And that, you know. that that process of reinvention and that process of like figuring out where we can move the needle is is like is an important one that I'm I'm seeing a lot. Paul, you wanna give us the rundown on on what happened when we moved CI?

Paul Stack: I mean, like the TLDR here in a nutshell is like, we know that CI for as long as it's been around has kind of conflated two jobs and verification and when things is actually ready to merge. And what we've kind of done here at SwampLine is we've changed that verification look outside of CI. It's actually local. When we actually do that verification, we produced attestations of everything that's been done against a specific commit shop. with SHA-256s of the actual files involved so that we can actually verify it when it goes into the central place to merge it down. And we've actually got a double speed of how we're actually merging systems at this moment in time. At this moment, as of Monday morning, it's in the region of about 90 to 100 pull requests have sailed through CI. And when I say CI, mean, it has sailed through a merge request, like first time without any... gates without any issues, without any attestation problems and have not actually failed UAT, know, actually verification against the binary on the other side. We're shipping up to shipping 20 CLI releases a day based on this new verification flow. And it feels swift. It feels that the agent is able to get really all of the context that needs to fix it without having to do that horrible erroneous feedback loop that was, we were basically meet proxies at some point for agents. like in CI because like we were pushing it up, we were getting comments back and we were pushing that back into our agent. And that just, it just felt awful. And we, we just had this trigger where we like, let's try it. And we had great results from.

Nick: I mean, I just like just just to take a little bit of a step back, you know, what you've done is actu is actually incredible. We were we were already operating at warp speed in a way that not many projects on the planet are operating, right? Like we're releasing 10 releases a day or whatever. Our metrics are ridiculous in terms of time to triage and time to ship. You know? And like you've doubled that. And not only did you double it, but like all of this you implemented from the time that you got serious about it to it being live and running through a hundred releases in a week. You know? And just incredible work. I was truly inspired when you wrote this blog post, just knowing like when you started, what the goal was, what we were already doing, like it's insane. So good job.

Paul Stack: The really funny thing about this thing is I expected to post this blog post out and be everybody like you are the worst person in the world. You're taking away my CI. You're taking away my verification to do this. But honestly, the amount of agreement that people have and going, we're coming to the same conclusion ourselves. We're seeing that the agent, the group is much faster than CI can't cope with. Now we have an added benefit here as well. The added benefit is, that we are no longer held hostage by CI being downed, you know, because we can still do the attestations. We can still do all this work. Yeah. Yeah.

Nick: I was just I was just thinking, man, we should tell everyone to do this so so then GitHub can stay up now, you know?

Adam Jacob: Yeah, then right.

Jason Powell: I was gonna say I I had the I had the conversations in work chats. I mean I can't tell you how many this week. let's move to GitLab. All right, fine. Just throw up a bunch of servers. We're gonna I I mean, I I threw up repo servers in each of our data centers for artifacts this week. So, you know, I mean I I I can luckily do that, right? I I have a a lot of flexibility in where we are, but we feel that pain hundred percent. on the infrastructure side, even of the CIs, right? Like It's it's being felt all the way through. All the way through.

Adam Jacob: Yeah. And I think you've been doing a lot of work, you know, to to to think about how these sorts of patterns that we're using for software, like how that emerges into infrastructure, you know? And you know, that that question of how do you play how do you play with those principles? You know, how do you play with the idea? Obviously we want to run all the tests, obviously we want to make sure that stuff passes. But then and and we've always been thinking of CI as attestation, right? We always think about the fact that it went through that process as a way to attest that the software was was was at a certain quality bar and that was produced by people we trusted. And I think, you know, Jason, do you think there's more to do? Like, do you have a thesis yet about how you think that moves into infrastructure or is it still too early to know for sure?

Jason Powell: I mean it it does so like I I've been a big non-fan of probably like Kubernetes for a long time, right? So just I have always felt that it is unnecessary,

Nick: Now of us.

Jason Powell: it's complex, that its entire design was basically so that the people shipping

Paul Stack: Thank

Jason Powell: the software didn't have to worry about anything else underneath it, right? we'll just give it to somebody else, right? That's not bad, but somebody has to deal with this, right? So

Adam Jacob: Yeah.

Jason Powell: I I maybe it was Paul's blog, I don't know, but somebody was I read it the other day and it was like, you know, we didn't we've been trailing in ops and infra, you know, behind the waves and it's time to not do that anymore, right? Like I don't want to say, Hey, yeah, let's let's throw this in a container, let's do this because it adds, you know, ten times more complexity, it adds latency, and then you get the well, what do you mean? Just it's just a container on there. Well, no. You know, it's all translation, it all makes a difference. If you strip out all the complexity and you get down to just running what you need to run, you get a lot more efficiencies too. Cause one of the things we're dealing with is hardware scarcity, right? these AI players are buying everything. memory, hardware, it's hard to get. And so

Adam Jacob: Yeah.

Jason Powell: you you like I think that something to reduce the complexity, which means we're gonna be able to get more out of what we have. is gonna let us run even longer on what we have too. So hopefully hopefully you know.

Adam Jacob: Totally. And I think yeah, I think that's totally right. And I think, you know, that that same technique that Paul applied to CI, like I just I really think we're gonna start seeing that get applied. And you know, it's one of the things that I'm I'm having the most fun with and hanging out in the swamp community is like we wind up having conversations where it's like, Yeah, okay, like, you know, what are what are we doing and how are you moving it? And like people are trying wild stuff and then seeing it stick and like like I think I think it's all there for for us to do in in AI.

Paul Stack: The thing here is like we're learning together as a community and like some of it's not sticking, some of it's sticking really well, some of it's resonating really well. like we have, this isn't just something we're like, maybe we should do this. Like we actually went and did this for empirical evidence of the improvements of our software because of it, which that like we literally can't do anymore. And sharing that with the community early, showing them the files, the plans, like it... So many people were positive about this move. So it's been pretty cool.

Adam Jacob: Totally. All right. So let's go back into like weird economics land. So Gary Marcus wrote a great substack about anthropics $30 trillion fantasy, which is that they're preparing to go public. And the word on the street is that the plan for their stated TAM will be thirty trillion dollars. And so just just so for for reference, that's more than the GDP of the United States of America, right? and meanwhile. This is a company that burned 42 billion last year, but is projecting 190 to 200 billion in revenue by 2028, which I think is completely rational. Like I think it's entirely possible that that revenue number is real and that they that they meet or exceed it. but a $30 trillion dollar total addressable market is insanity, right? You're essentially saying, like, our our our capture of the revenue that's possible for this market is all revenue, you know? of all people who have ever produced, you know? And like it's clearly a little detached from reality. And

Nick: Or it's just a statement on inflation. So, you know.

Jason Powell: Yeah.

Adam Jacob: Or it's a statement on inflation. I but, you know, this is a once in a generation company with like stunning results. There's never, you know, like you don't every, you know, everyone who builds startups or whatever, you're always trying to play the big game. Lots of people, when they go out and pitch, they're like, We're gonna be the next anthropic. And Rory O'Driscoll, who was on my board at Chef, once post made a post that was like, Stop saying that shit. You will not be an anthropic. There are plenty of great outcomes for you. It's gonna be great. But this is once in a generation business. Like it happened for them. That's it. It's not gonna happen. It doesn't happen. You know, we don't get 50 of them. You know, you get, you get, you get, you get one, you get two, you know? and it's interesting to me that a once-in-a-generation company, stunning results, ripping revenue, everything they need to go public and win is all here. And they still have to say insane things to make it look more impressive than it is, because that is the world we live in now. Like we've so inflated it that like you know, SpaceX was. almost as big. It was like it was 28 trillion or something. which was also madness, you know, like it was also obviously fiction. And so like I just I think it's interesting that we've heated the market up so much that even when what we have is this like incredible once in a generation thing. You know it's gonna work. There's, you know, like that is a company worth betting on. We still have to dress it up in insanity in order to just like make the headline be good. Right. It's weird.

Nick: I don't know, man. Anthropic

Paul Stack: They solve coding, so.

Nick: Yeah, true. You know,

Adam Jacob: Ha ha ha.

Nick: it anthropic loves to just say wild shit, you know? Like that is their marketing strategy in a nutshell. And sometimes they get slapped for it. Like when fucking Fable got embargoed or whatever, you know? And you get those

Adam Jacob: Yeah.

Nick: stupid prompts now when you like tell it to look at something and it's like, I can't do that for safety and security concerns. Like that was You know, that was the result of their overzealous marketing of just like, you know, saying the world is gonna end because Fable is released or whatever. You know, like and so I'm not surprised at all, you know, that in their S one they're gonna have some thrill thirty trillion dollar fantasy or whatever. You know, like this just sounds like par for the course, and at some point they're gonna continue to be the boy who cried wolf, you know, like It'll be interesting to see how the market reacts to them just being completely crazy.

Adam Jacob: Yeah, and an interesting question is do they is there is there a moment where someone sat in a room? I think there probably was, if I was betting, where where they decided what that number should be and they all looked each other in the eye and they were like, thirty trillion. You know? Like

Nick: Mm.

Adam Jacob: like I like when I th when I imagine SpaceX, what I imagine is there's a bunch of people in the room. Sorry, go ahead, Jason.

Jason Powell: Do you imagine they they were burning burning hundreds and I said do you imagine they like burning hundreds and their cigars and

Adam Jacob: Yeah, kinda. Yeah. Like 'cause I when SpaceX I imagine everybody in the room was like, no way. And Elon comes and he's like, 28 trillion. And then everybody's like, I don't know about and then Elon's like, you're fired, you know? And then and then you're like, 28 trillion it is. Elon is the greatest. You know, like that's how I imagine that conversation goes. but like, but when I imagine the anthropic one, like I imagine a bunch of very sincere people. Like Really sin because they've they've proven their sincerity over and over and over again in ways that are sometimes I find ridiculous, but like so sincere, you know? And so like, yeah, I imagine that they really looked each other in the eye and were like thirty trillion, you know? and it's like, Okay, yeah, we'll we'll see, you know? we'll see.

Paul Stack: And it makes me feel like there was a table like there was in the Wolf and Wall Street when the guy walked in and he's talking about the credit card bill and he's like 40,000 on sides. You know, like just these wild numbers that they're just making shit up.

Adam Jacob: Okay. well let's talk about their last article for the day. Drew De Vault. Drew is himself a controversial figure. He's the guy behind WL Roots and Sway and Source Hut. posted about how Source Hut Source Hut is a source code hosting company, you know, that they are no longer going to allow AI or LLM driven projects on their platform. and the post that they wrote lays out their values really clearly, right? They're essentially saying LLMs were built on theft. They ignore the obligations of free software licenses. SourceHUD is a thing that's built to further the aspirations of free software and sort of around people and these structures. And therefore, while you know LLMs might have some utility, it's dramatically outweighed by the impact it's having on the planet, on society, and by the people who use it. So Source Hut now prohibits original content written with or which facilitates the use of LLMs or other generative AI technologies. so what I like about this is that it aligns with their stated values and how they choose to interpret them. I felt the same way about what Codeberg did. You know, like I wouldn't make that decision. but I like that he laid it, that they laid out like these are the things that we believe, and this is why we believe them, and this is why we're making the choices that they make. For me, I I tend to think that we're going to find a path toward an era of more abundance and more equity by figuring out how to use these tools and working together to build that, to build that better shape and stay engaged in it.

Nick: You know, I when when Codeberg did the announcement, I actually got confused and thought that it was Source Hut in the beginning. because you could just basically interleave the two and it's the same. and like, you know, if they wanna keep putting their heads in the sand and building out their little carved out communities or whatever where they can complain about the ales of AI, like more power to At the same time. You know, they'll just continue to be irrelevant. It's fine.

Paul Stack: My biggest worry here is I hope that the companies who are taking this stance at this moment in time, at some point will put their hands up and say, you know what, we kind of got this wrong a little bit and we kind of like, you know, have to empower our people who use this platform in order to take advantage of these technologies. I love when people take a stand, I really do. But what I appreciate more is when people are actually able to put their hand up and say, maybe that stand wasn't quite right at that moment in time. And that's my hope that the people who are doing it and don't believe in it, that when they see that it is here to stay, regardless of whether we think it is or not, and it's taken over so much of how we build stuff and how we manage stuff and what we do, that they're actually able to be courteous and say, you know what, we kind of got this one.

Adam Jacob: What do you think, Jason?

Jason Powell: Yeah, I mean I hope so. You know, it you you would like to think that, right? Like that's the end goal is that somebody says, I don't know, or I was wrong, right? That's that's a hugely powerful thing. and you know, it just sometimes you don't get that as much as you want to these days.

Adam Jacob: okay. on that note, that brings us to the end of the news. So, Jason, five questions. You ready? Okay,

Jason Powell: Ready.

Adam Jacob: so of the stories we just covered, which one are you gonna be thinking about later on today?

Jason Powell: probably the jalapeno story. Just because I'm interested in I'm interested in seeing where that goes. I wanna see, you know, th the new models that they're putting on phones, right? I mean all the edge devices, all the specifics, right? Just if you need certain things like small models, I think are gonna start becoming more prevalent. So I wanna see some of that specialty come out.

Adam Jacob: Awesome. Which story do you think is noise? If you were like, we could have not talked about that one and I'd have been fine.

Jason Powell: Anthropic.

Adam Jacob: All right, anthropics valuation is noise, which I think is actually probably a pretty solid take.

Paul Stack: Thank you.

Jason Powell: I'm not getting a piece of it, so you know.

Adam Jacob: Yeah. Yeah, it goes one way or the other. What's a common

Jason Powell: Yeah.

Adam Jacob: belief in our industry that you think is just wrong?

Jason Powell: God. That's a tunnel that's a tru that's a that's a pitfall there, yeah. I don't know that there's any right answer. Or that there's one

Adam Jacob: There you go.

Jason Powell: right answer, I guess.

Adam Jacob: Yeah. Yeah. Okay.

Jason Powell: I've been in I've been in a lot of different places and nothing looks the same. So everyone's best practice, everyone's everything else, like that can only get you so far. So I think that that's what people realize in AI is that, you know, if you've seen something and you feel something and you know something, right? The engineering's not going away. It's not. So

Adam Jacob: Yeah, for sure. All right. What's one thing you wish you could teach everyone? That's such a good answer. All right. What's the what I mean, that's it. You can just mic drop on that one, right? what's the best and worst moment you've had working with AI agents?

Jason Powell: Mm and not really any bad moments, just unexpected, right? Disappointments but nothing nothing no crazy horror stories of it, you know, deleting my directories or, you know, anything else. No. No, no, no, no. No, no

Nick: You haven't deleted your root drive yet?

Paul Stack: You've not lived.

Jason Powell: RMRS. no and and you know, so it's not bad. Probably my worst has been on the receiving end of other people using agents like to get a a doctor's appointment and you know, you tell them, Hey, here's your symptoms and you can listen to them type in. Okay. Next, based on this, we need you to do this. I'm like, my God. You just put me through a prompt and now I want to punch you the face. Like

Adam Jacob: Yeah.

Jason Powell: that's that's the irritating part, I think.

Adam Jacob: Yeah, for sure. All right. Well, that's all the news that's fit for infrastructure this week. If you want to dive deeper into any of the things we talked about, you can find all the links and the summaries over at swampdash club.com/slash podcast. You can sign up to get it delivered straight to your inbox every Monday morning. If you're trying to figure out how to manage production infrastructure and applications or scalar use of AI within your organization, or how you're going to make good on all those promises you've made about how AI is going to transform the internals of your business. Swamp is here to help. Visit swamp club.com to get started using AI agents safely and effectively by building repeatable systems you can trust and verify. I'm Adam Jacob.

Paul Stack: Thank you.

Nick: Neck.

Jason Powell: Jason.

Adam Jacob: And Jason, as our special guest, you get to take us out. Is there anything you wanna plug?

Jason Powell: just skillble, right? Skillable that's where I work. so we're a the a lab platform. most people that do labs have heard of us. if you've ever taken s any kind of certification, you've probably run on my platform. So just

Adam Jacob: All right. You heard it from Jason. Skillable. All right. We'll see you all next Monday. Thank you.

Nick: See

Jason Powell: Awesome. Thanks.

Adam Jacob: you