Agents, DeFi and the Cost of Hype w/ Gerrit Hall

Rex Kirshner (00:00.12) stuff. Gerrit Hall (00:01.454) it's so nice. It's so nice. Rex Kirshner (00:03.346) I know, know, dude, as I was like setting this up this morning, I was like, maybe Garrett's smart. Gerrit Hall (00:10.882) I mean, no, no, I mean, like, I really appreciate you doing this because I know it's like important to like, you know, keep the muscle active in terms of like practicing podcasting as well as just like, you know, keeping our voice in the sphere, whatever that means. but it's so much work, like, and you edit it, which is like an even insane amount of work. Like llama party. did like the. Rex Kirshner (00:18.754) Yeah. Rex Kirshner (00:24.386) Yeah. Rex Kirshner (00:32.47) Yeah, I'm just like, I'm just so, so, so self-conscious of like looking like an idiot, but yeah, anyway. All right, let's get started. Garrett, welcome back to signaling theory. Gerrit Hall (00:47.394) Thank you, sir. Pleasure to be here. Rex Kirshner (00:49.806) So man, it's been a while since we've done one-on-one. I've got a few topics that I'd love to talk with you about, but before we get started, how are you? How are things been? What's been going on the last two weeks? Gerrit Hall (01:03.448) Things have been good. Let me think. In terms of last two weeks, as usual, a blur. I'm not sure. I'm sure as we talk, I'll zero in on specific things. But tons of vibe coding in the past few weeks, as you might imagine. I've been focusing extensively on So you've probably seen this where like it can get a little boring when you're just sitting there approving permissions for agents. So I've been trying to like experiment a bit more with this concept of like how I can give agents the most agency without like feeling like my systems or my entire operations or whatever are a risk. So like I've been focusing a lot in like stuff we've talked about previously on the show of what I thought was the next step. Like getting agents within kind of safe Docker eyes container. making sure that they have the tools that they need and then saying dangerously skip permissions, just run yourself every hour or every three hours or once per day or what have you and give it a fair amount of autonomy to just use its logical reasoning to attack problems. So I'm in the early stages of that and still trying to figure out exactly the best way to do it, but it's been very rewarding so far. Rex Kirshner (02:13.005) Yeah. Rex Kirshner (02:20.162) Yeah, you know what's funny is, you know, like, I've been just using dangerously skip permissions on and off, but like more on than off lately and without really protecting it. And the reason is because I went through this whole thing like a month ago about how do I set this up to allow my agent to run wild, but in a safe space. then, so did a ton of research. Turns out that Claude, code, the CLI tool, it can use the Mac OS's native sandboxing framework. And so I was like, that's perfect. Nobody will understand Cloud Code, how to use these safely than the company making them. And on top of that, it's leveraging OS level stuff. Awesome. So I went through all the effort to set that up. And dude, pretty quickly, it's It's bizarre because what Claude will do will run into an issue where the sandbox blocks it. And instead of asking me permission, it'll say, let me just disable the sandbox. it'll be like, I can't cause like, when you just look at the code, it looks like the same bash command, but the first time it failed, says, I'll disable the sandbox and then it'll run it again and it works. so I'm like, part of me is like, crap, like I need to be more proactive, like maybe I need to do Dockerized containers, but I also just think that if an intelligence who understands more about computers than we do has bash access, like there's really nothing you can do. Gerrit Hall (04:00.142) I mean, yeah, it's really insane what I found that the undockerized version, at least, is capable of. There'll be cases where it needs to try and authenticate to a website. I'll read my cookies and LARP as me using a browser. And I'm just like, OK, you do you. It's a, I mean. Rex Kirshner (04:12.866) Yeah. Gerrit Hall (04:20.076) Yeah, like we both are fairly comfortable with Claude and know that it's like generally operating in the service of like what we're asking it to do. So it's not like being like, let me like see if there's any money in your bank account or whatever, but like I should transfer. It is just wild, right? Like. Rex Kirshner (04:34.391) Yeah. Yeah, yeah, I... Gerrit Hall (04:38.604) And I think, I was going to ask because like I played a little bit a couple of weeks ago with an open claw install, which I think is supposed to be kind of the same concept, right? Like, like, but I didn't actually get very far with open claw before I decided to try the kind of self-built Dockerized approach. So I'm kind of curious to hear more what I'm doing wrong with open claw or like why I'm not getting that kind of aha moment because I know a lot of people who are smarter than me, like yourself and Mark Andreessen. talk about how OpenClaus, one of the big revolutions they've seen so far. And I'm not yet seeing why that in particular is better than just Cloud Code dangerously skipped permissions. Rex Kirshner (05:18.158) You know, what I would say is, oh man, so a couple of things here. But one, I guess I'll address like OpenClawn in a second, but I think that you and I have developed like a particularly good muscle of like, oh, this is total bullshit from crypto, where, you know, and especially like with our specific journeys where we... really got into it from DeFi because it seemed like there was so much real innovation happening. And then just through our experience, we realized that of that real innovation, maybe 50 % was actual scams, 40 % was like low effort forks that were like not scams, right? But like there wasn't really anything more there than like a group of people trying to create hype around a token and then. cash out and then, you know, like maybe 5 % of just like shit that could never work and like 3 % of something and like maybe 1 % is real. And that muscle that I developed, I feel like is really been going off in most things AI. And so when I look at the conversation around a lot of stuff, but open clause is a very good example. Dude, I think it's bullshit. I think the only people that it seems like have found a use for OpenClaw are the ones who are creating content about how good OpenClaw is. And it's like, my favorite one, because this is the guy who I first started watching YouTube videos on how to use Claude Code, his name's Alex Flynn, and he's like real big now, at least for fucking vibe coders, right? But he's gotten full, just like OpenClaw maxi, and what... On the one hand, you hear what he's saying, he's like, my God, OpenClaw is like, it's a one man company. It's doing all of his social media. It's his ideas, his scripts. Like, wow, this is amazing. And then you look at it you're like, this is all in service of creating content, which is fine, right? But like, I haven't seen anyone be able to use this stuff beyond that level that makes my crypto muscles like uncomfortable. Rex Kirshner (07:37.61) And so, love to get your reactions on that, but just to like close up real quick, I have like, I've got ridden of open claw, like for a few weeks or months now, because like, I just don't know what to do with it. Like I, the, the first unlock for me that was interesting was like, this is like an always on version of a personalized chat bot. And so like it's always available. I can reach a via telegram, but I don't know. Codcode is like that. just leave my laptop open and I can access it through the Codcode app. so like I don't get OpenClaw other than people understand there's so much money and like a gold rush in AI, but don't really know how to access it yet. And, you know, that's kind of my read on what's going on, but I don't know. How does that resonate with you? Gerrit Hall (08:24.11) Mm-hmm. Gerrit Hall (08:29.43) Okay, I mean, that's good to hear because I think, I mean, maybe I misinterpreted, I think, like, I remember you saying that you just, like, been playing with OpenClaw and were impressed by it a few, like, a month ago or something. So I was like, okay, like, you know, when people who are not, like, ex-influencers are telling me this, like, I should pay attention. And yeah, I just don't see, like, I'm kind of like with you, like, it seems kind of interesting, but I haven't, like, done more than the kind of first, like, go-around of things that I... Rex Kirshner (08:40.013) Yeah. Gerrit Hall (08:58.392) played with it on, whereas anything I could do within that, I feel like I could do it within Claude Code. So maybe it's just because my workflow is so heavily biased towards Claude Code. And that being said, this concept of an always-on thing, I think I'm kind of getting that by doing this dockerizing and trying to put these safety walls and give the agents just what it needs to work within this kind of, you know, Rex Kirshner (09:04.322) Yeah. Rex Kirshner (09:09.763) Yeah. Gerrit Hall (09:24.77) would I consider a fairly safe container and then turn it loose and just say go wild. And, you know, mixed results still with it, but it's like a lot closer to what I'm trying to accomplish, I think. So, and then on the DeFi point, like I think I won't go too far into that because it's not a DeFi podcast, except to say that please, please, please, friends, if you are using DeFi in 2026, just don't for a little bit, because I really do believe that AI is going to steal all your money or not AI, but hackers using AI. Rex Kirshner (09:36.195) Yeah. Rex Kirshner (09:40.75) You Gerrit Hall (09:53.774) Like D5 is probably going to be great in 2027 beyond just please, unless you're extremely desperate, don't chase those 20 % yields because you might lose everything. Rex Kirshner (10:03.948) Yeah, you know since we last talked and yeah, it was when we last talked after tarik left us and we were chatting about fire pan and all of your beliefs, I like my portfolio was like 95 percent bitcoin in eith and then five percent just Like I don't know like might as well like take a punt on something that could be interesting but since then I got out of all of those because it's just like Gerrit Hall (10:28.253) I mean that 5 % is probably safe. I just see these people who are like yoloing like, hey, monkey USD is making 80 % APY and it's like you are absolute fool. Like get out of like, know, like, I don't know. Rex Kirshner (10:42.51) Yeah, well, I mean, just while we're on DeFi real quick, like, so the big news in that world, right, over the last week was Drift Protocol and Solana hacked for 200, 250 million. And it's like, yeah, man, I am with you that AI is changing the threat landscape so that nothing is safe. But then you look at that hack and it has nothing to do with AI. it... Gerrit Hall (10:54.83) Mm-hmm. Gerrit Hall (11:06.466) I know. Rex Kirshner (11:07.79) And it's like, for those people who have a life and are not on Twitter reading about crypto stuff, it's looking like North Korea sent these like sophisticated hackers to conferences like six months before we're developing real relationships that even deposited a million dollars into the protocol as like to show that they're legitimate and blah, blah, blah. And I still don't understand how we got from that to they reached a point where they had enough control over the system to pull out 250 million, but like point is that nothing to do with AI and everything to do with like international geopolitical espionage. Gerrit Hall (11:51.79) Yeah, I definitely agree with the take and it's crazy. and I, you know, it does seem like there was like, and I'm, you know, I haven't like done a full like deep dive into it. seems like that their contracts, like the fact that like, it, they were brittle enough where like, they could so easily like allow like one or two people to be able to pull the funds out. Like, you know, is it probably a big thing? And hypothetically, if they've been using something like fire pan, they might have been like, you that down a bit tighter. I don't know. can't say. But this just points to like, you know, the core thesis that like, I had never heard of Drift. I don't know what they were offering, but like, you know, you can get 3 % in banks and they're not, they're probably going to be pretty good. Like don't get greedy here people. Rex Kirshner (12:41.196) Yeah. Yeah. I, and I just, you know, last thing on crypto, unless you have anything else to say, but I just like when all my crypto skeptical people are say stuff like crypto is just for scams and it's just for crime and all that stuff. You know, there's part of me that has been in the trenches and worked at Lido and worked at BlockWorks and like, you know, friends with like Dow founders and all this stuff and I see them and I like genuinely believe there's people that really believe in like some weird interesting shit that are trying to build technology and like I think that's good. But like at this point and then this moment after this North Korea situation like it's impossible for me not to acknowledge that when you look at the effects of crypto. The effects of crypto are not in this like DeFi new primitive stuff, right? Like it's essentially like you can split it into like three buckets of like mutually consented crime, scams, which is like not mutually consented crime. And then like legitimately funding the North Korean nuclear program. And like, I'm not the kind of guy who like way, you know, I don't believe in like. super utilitarianism where you take the good and you weight against the bad and that's how you decide if something's good or bad or not but like when it's like 99 to 1 you know i don't know man Gerrit Hall (14:08.59) Yeah, I mean, so, you know, I guess like my take on it is like fairly nuanced because, okay, so like, I I'd say that like, basically, there's like two competing paradigms that we've seen. There's the code is law and there's law as law. And especially in this era of AI hacks and all these other like ways you can get exploited. Code is laws proving to be a horrible place to like keep your money. Whereas it turns out like the social compact of laws, like, you know, there's all sorts of problems with the legal system, as we know. But it's a little bit safer if you're trying to weigh the two on balance. Because if I put $1,000 in a bank account, the bank takes it. It might take a long time. It might not even be worth it to fight it. I'll probably be able to go to, OK, not the $1,000. If I had $1 million and put it in a bank and they stole it, I could probably eventually get that back after a lot of lawyer fees or whatever. It just gone in DeFi, though. And that's an uncomfortable place. at least at the point where like we can't secure these smart contracts. Now that being, well I say it's like my take is nuanced though is I'm with you on the innovation and I do think that like the power of DeFi is the innovation because if you look at, and if I had to handicap kind of what's going on in crypto as of 2026, like the institutions are here. The institutions are mostly playing with these kind of like quasi like hype. hybrid DeFi, C-Fi like things that are like really nicely catering towards bringing institutions on chain. Like a lot of RWA type things where they kind of can split the difference, where they can kind of like dip a toe into this DeFi thing, but the assets are still held in a bank. if like they get stolen, know, the North Korea is stealing nothing and the dollars that we're backing these like tokens are still in the bank so that they get their money back. And you know, like what... Rex Kirshner (15:56.44) Mm-hmm. Gerrit Hall (15:59.907) The unfortunate implication of this is that if you are not like Franklin Templeton or BlackRock playing with DeFi, if you're just like a random schmo putting your money into DeFi, what's the innovation? The innovation is like you are risking your funds so that people can test out all these kind of crazy new formulas, because banks can't do that, right? Banks are not allowed to just be like, yeah, we'll just throw up some random equations on chain and see if you lose your money, and if so, too bad. Banks can't do that, DeFi builders can do that, so they're just experimenting with your money at the moment to see if these equations work. And if they work, the banks are going to use whatever works in this safer context. So basically, at the moment, if you're playing with DeFi, basically more or less like you are the... the experiment victim to allow institutions to come on chain. So if you like being, if you're willing to lose your money for that, great, go for it. The innovation I do think, where I say it's nuanced, is I think that whatever we see come out of this, I don't know how long it'll take, maybe six months, maybe a year, maybe faster, maybe slower, who knows? It'll also bleed into DeFi, and we'll at some point see a fairly safe DeFi ecosystem where we can, as end users, Rex Kirshner (17:09.443) Yeah. Gerrit Hall (17:28.078) interact with these things and have better products that might not necessarily touch like the institutional rails. But just at the moment, like the way I see it is that the institutional guardrails, like whatever problems a lot of the cypherpunk people within crypto have with it, like they look safer to me in my risk analysis. Rex Kirshner (17:46.339) Yeah, you know, I think I agree with basically everything you said and I just to add like a little more flavor to it. It's not that if you're in DeFi now, you're experiment, you're the experiment. And if it works out, then like institutions are going to come and reward you for taking that risk. Right? Like what it more looks like is. No, yeah. What it more looks like is that that you put your money at risk to prove that like. Gerrit Hall (18:06.06) No, that's the thing, they don't reward you at all, exactly. Rex Kirshner (18:15.666) Uniswap works right that like XY equals K works And then the bank goes and implements XY equals K. It doesn't like buy into uni token and so You know, I I think I think probably if there's anything legitimate to crypto It will look in the end like swift, right and swift isn't this like complicated financial system where you can Take leverage and do like like it's literally just a Gerrit Hall (18:26.614) Right, exactly. Yeah. Exactly. Rex Kirshner (18:44.578) wiring system and like that's probably the The end state for crypto right where it's like I know Celsius was a scam but like that Effectively is what it's gonna look like it's like you put in your assets to like centralized institution they might go use DeFi might not whatever but like I Don't know man. Like I don't I don't think defies for people and if it's legitimate at all, like I think it's for agents, not for people. Gerrit Hall (19:16.398) Yeah, and I think there will be something to it, I think that I'm excited to see where DeFi ends up out of all this. I just think that this is just too treacherous at the moment for my taste. Rex Kirshner (19:27.042) Yeah, yeah. Anyway, back to AI and agents and like this new frontier that we're on. yeah, man, in terms of OpenClaw, like I'm with you, that I just, it was a very big unlock for me in terms of evolving my thinking of like these are chat bots and programming tools into like, okay, this is what an agent might be. And like how that is a real evolution of. instead of chatting with something in this interactive way, this is something that can do complex things on your behalf without your supervision. But once you make that leap, just let's be real, OpenClaus is some janky, open source, inefficient, insecure implementation of frontier tech. And that is something that we're very familiar with in DeFi. Gerrit Hall (20:21.29) Okay, yeah, I get that. So question for you, like if I need to like hire a contractor for something, like to let's say I need a plumber all of a sudden, is OpenClaw the right tool to like be like, hey, OpenClaw, like my pipes are burst, go find me a plumber and I don't want to spend like top dollar. Would that be like the target application for it or the use case? Rex Kirshner (20:46.318) I mean, not really because like, why wouldn't you just use even just like the, web chat vault version of Claude or chat GPT, right? Like why bother setting up. Gerrit Hall (20:56.558) Right, but in theory, according to these influencers, I can start an open class session. will call up plumbers. will masquerade as my voice or whatever. It will negotiate. It will get them in bidding wars to drop the price or whatever. Is that just completely vaporware? Rex Kirshner (21:15.502) I just, guess like as you're saying this, if you told me like you'd give me a hundred thousand dollars to set that up, like I think I could do it. I think like, okay, 11 labs will allow you to take any text and turn it into speech. Right? Okay. That's like one piece. The proactive like searching and filtering and finding the plumbers like, okay, that I know that Claude can figure out. And like, I think all the pieces are there to do something like that. I just, um, You know, I think this is like the 1990s version of like, yeah, you can build your own PC. Like you, and you can get it connected to the internet or something, but it's like, we're so early and this stuff is so janky that like, I don't think the juice is worth the squeeze, but I don't know. Gerrit Hall (21:46.103) You Gerrit Hall (22:00.398) OK, so that's not the primary use case. What would be the primary use case for a situation where I might try using OpenClaw for some random task, as far as you understand the narrative? Rex Kirshner (22:13.302) My, again, I think that Cloud Code has put in so much effort into importing the best features from OpenClaw to the point where OpenClaw is not particularly special, right? I think the interesting thing about OpenClaw was being able to talk to your computer via Telegram, but like you can do that now through Cloud Code. My understanding of the workflows that make this always on like customized agent work for you are essentially cron jobs, right? So the interesting-ish ones that I've heard is like every night at 3 a.m., I want you to look at all of the trending tweets on this topic from the day before, and I want you to write a executive summary with like a summary, a paragraph on why it's significant, a link to the tweet, and like three links to related tweets, something like that. Every day it does that for you and like I can imagine that being something useful to wake up to. like frankly, I have never once actually read a newsletter. So like I'm concerned I would set this up and then like have another newsletter that I ignore. But, you know, when you think in those type of terms in those types of use cases, like I can see it. I could see like, if, if you and I were, what's a business that requires cold calls, right? Like any business that requires cold calls, I can see Cloud every night at 3 a.m. like gathering up our CRM data and doing Google searches and figuring out like, okay, this is the next 10 people you should call them the next day. Like that kind of stuff seems interesting. I don't think it has anything specific to open Cloud at all. And again, I think that these types of things are like pretty easy to envision, but. I think it would be pretty hard to set up and then you would be like forced to ask yourself the product that you're getting out of it was that worth it? Gerrit Hall (24:20.49) Yeah, no, makes sense. makes sense. So I mean, for my sake, I'm going to probably put open claw as a concept on ice for a little bit. Just not worry about it, which is a shame because I got all my old laptops out of the basement and started trying to get them restored. Rex Kirshner (24:31.566) Yeah, well, and I'll kind of like use this as a little bit of a janky segue, but like for me, what I did, and I mentioned this with Tarek a couple of weeks ago. So I'm at a point where I'm on two Claude Max plans and I have one set up on my laptop and then one set up on my Mac mini over here. And so when I like essentially burn through my MacBooks, quota for the week, I just like SSH into my Mac mini and then use that account there. And that's pretty useful because what I've done is on my phone, I'm logged into the account that's running on my Mac mini. And because my Mac mini is running 100 % of the time and I have the app, I essentially have like the ability to like open up new sessions, like do coding work for my phone. And Like that works for me. So I don't know. mean, you pull out all your old laptops, like you have a computer running all the time. Like I still think there's interesting things that you can do with the new remote aspect of these tools that OpenClaw brought up. But like, if I'm just being honest, it's really like allowing me to continue quote unquote being productive, i.e. like feeding my productivity addiction, like while I'm out for a walk. or like doing grocery shopping or whatever. Gerrit Hall (25:58.19) Fair enough, fair enough. Now I think I've at least kept the fairly good habit of trying to be like attentive to other people and keep my phone down when I'm out in real world. Because I can see that being a lot of fun. Rex Kirshner (26:10.636) Huh, what's that like? Yeah, no, I hear you. So, yeah, I can't figure out how to do this segue gracefully, but the other thing that's been going on the last week or two is there's been so much conversation on Reddit, on Twitter, about how the terms of our deal with anthropic has kind of changed out from under us in terms of. how much usage we get for 20, 100, 200 bucks a month. And so many people complaining about how all I said was high and it used 10 % of my weekly quota or all this kind of stuff. So I wanted to ask you, one, is this something that you've been tracking or is for whatever reason my algorithm locked into that and it's like a silo thing? And two, Have you noticed like a degradation in your quota, let's say? Gerrit Hall (27:13.998) It's a good question because I did get close to my 100 % last week, which I usually am trying to be light enough touch and I'm trying to be thoughtful with my usage. So was surprised to see I got close to 100%. But it was also at the same time that I was experimenting with phasing in these hourly cron jobs, which I'd expect giving an hourly cron job of dangerously skipped permissions could chew through the quota. So I kind of like been monitoring it. I definitely noticed Codex. I barely used it this week and it like chewed through the usage very quickly. So I definitely noticed Codex was, you know, surprisingly like stingy. So I guess I don't know. I'm paying more attention to it this week because my quota reset and I'm on a new week and I'm trying to like keep a lighter footprint just to see. Rex Kirshner (28:08.076) Yeah, you know, but so and in terms of the amount of chatter you've seen about this on Twitter or Reddit, like are you seeing what I'm seeing, which is like every single day, like at least 50 % of the posts about clogged code are whining about the quota. Gerrit Hall (28:27.318) Yeah, I do see the chatter. And then I see some influencers talking about how the reason that we even have the cloud code at such a nice price point. I know people say $200 per month is fairly expensive. But a lot of people also try and break it down and say, well, actually, that's actually fairly cheap. They're subsidizing it. It's a lot of VC money because they just need to have a nice IPO. Rex Kirshner (28:50.222) Mm-hmm. Gerrit Hall (28:54.658) you might see that price creep up or the amount you get ticked down. I do know that when I set up my Dockerized bots to run this, it's probably not the way that they want you to do it, but instead of paying to access the API directly, I'm running it as cloud code, give them authentication to my cloud code account and let them run there. that being said, the monitoring of it doesn't realize that, it's like always reporting to me like the cost in dollars of usage for it, which looks way off to me because it's like very much being like, okay, that run costs $2.50 and here's what it did. And I'm like, if it was $2.50, like I would have like three days worth of credit or whatever. if that math is accurate, which know, cloud code remains terrible at math, so who knows. But if that math is accurate, like, Rex Kirshner (29:40.374) Yeah. Gerrit Hall (29:49.144) then it is very clear that they are giving us a massive discount through the Claude code plan relative to their API pricing. Now, that being said, yeah, go ahead. Rex Kirshner (29:59.137) Yeah, well, that No, no, like, so for that, like objectively, they're subsidizing us at least 10 X, but like maybe even 20 X. And I've been like pretty annoying about tracking this. there's this tool I used to talk about a lot earlier in the podcast. It's called CC usage. And it's a, just a command line tool. And all it does is go in to the metadata files that Claude code creates and like counts up how many. input tokens and output tokens that you use in your session and then gives you the dollar amount that like it would have cost you had you gone through the API as opposed to using, you know, like the monthly subscription and like just for my own personal use, like I can see like it is not that hard for me for over a month to get $2,000 worth of tokens out of my $200 plan. Like for sure. So like they're Gerrit Hall (31:00.93) Yeah, but. Rex Kirshner (31:02.274) Definitely being super generous. Gerrit Hall (31:04.494) Okay, that seems to map, guess, then with the math that I'm seeing. So, like, I do, like, tend to think that they're, and, I don't actually know anything about the raw costs or anything. Like, I do, like, my hunch is that the, like, API is marked up, like, crazy, is kind of my read on that. Like, I don't think that it was $2,000 worth of tokens cost them $2,000 worth of infrastructure costs. But I also don't know, because, like, I see, like, they're investing massively in infrastructure and all that. I just know that, like, Rex Kirshner (31:19.448) Yeah. Gerrit Hall (31:32.844) Whenever I try and run things against the API with a credit card per token call, the quality jump between a deep seek model and Claude is not 10x, but the pricing is 10x. So I really think they're marking these tokens up extremely, extremely high to fuel the bubble is my hunch. Rex Kirshner (31:56.047) Possibly I think the other so I very much agree. I started experimenting with local models and Because like like I said, I'm already at two Claude max plans. Like this is outrageous. You know, I want to be able to offload some stuff and On the one hand, I definitely think okay If I have all of my fancy state-of-the-art models build out the plan Like can I just have cheaper? Models execute it and like it'll probably work out right and Then the other side of me thinks like look the thing with code right or or in my case like enriching datasets is like you do it once and then It's there's no Ongoing cost like I don't have to continue to pay opus once it's built. It's built, you know, and so I kind of look at it as Okay, I could use deep seek and pay one tenth in order to get it built, but like, do I have the same faith that that is like good, reliable, robust, resilient code? And like, if I'm to just pay for Opus and maybe that means that I have to shift some of my usage at the end of the week to the next week, but like, I know that like it's good, high quality stuff. Like to me, that trade-off just seems like, why would I mess around with the cheaper models? Gerrit Hall (33:00.109) Nah. Gerrit Hall (33:16.182) 100 % for building code. I'm only using the most premium, premier tier model I can, because I've already, we were early-ish to the AI boom, let's say. We were using these coding tools six months ago, and it was just nothing but frustration to try and get it. It's worth it for me to have, it's the difference between back in the day paying for a top tier senior developer versus outsourcing it to some sweatshop overseas. Rex Kirshner (33:29.976) Mm-mm. Rex Kirshner (33:44.984) Yeah. Gerrit Hall (33:45.038) So for code, I'm willing to pay any price. When I'm talking about using DeepSeek, it would be for a customer support type chat, let's say. Do you get that much more juice out of having a really good, because I'll grant, Claude is better for an application if you're trying to chat with someone about resolving customer service issue. Claude is gonna be better. Is it 10x better than letting DeepSeek handle that chat? Rex Kirshner (33:54.637) Yeah. Gerrit Hall (34:14.662) Probably not. It might be 2x better, but then if it's like for customer service call like How important is that 10x like can you get away with like a little bit inferior? For something that's not so mission critical Rex Kirshner (34:29.998) Well, I think this is a fundamental question for all business leaders right now, right? Which is, okay, let's say I can implement my customer support using DeepSeek and let's say it costs a dollar to deal with any complaint. Or I can do it for Opus and it costs $10. And really the rate at which you get a good result out of Opus, let's say, only 2x, right? But you're paying 10x. And on the one hand, you're like, well that... Gerrit Hall (34:47.181) Mm-hmm. Rex Kirshner (34:58.094) The math there doesn't math, like that's stupid. But on the other hand, if it's like, okay, DeepSeek is going to 75 % of the time result in the customer being pissed off and like you don't care and like this is a solace company who just has a chat bot and they leave versus Opus that only happens 50 % of the time. And like your cost of acquiring a new customer, let's say it's like $50. Like, I don't know, man. Maybe the math works out where it's just like. It's worth it to invest in customer service. But we live in America in 2026. I don't think anyone thinks it's worth it to invest in customer service. Gerrit Hall (35:30.082) yeah, absolutely. no, but yeah, Like the way you're approaching the problem is, think, the exact right way to look at it. And thus far from what I'm seeing, like, you know, I've tried to run as best I can, like, A-B tests on this. Like, what happens if you use Claude versus, like, DeepSeek and something? Like, game is, like, generally real and visible with Claude. Sometimes it's not. Sometimes, like, you know, like, for certain things that are like, I picked customer service as a random example. There's so many cases where you need to inject, a little, like, intelligence into it that's not going to... affect the bottom line. And in some cases, you do see meaningful improvement with Claude, but it's not in order of two or three X. It's like a 20 % or 10 % boost against what you get with DeepSeek. So if it's like a 10 % boost, but it costs a 10X difference in tokens, that's a real decision. And a lot of the applications are things like, you know, just like super basic things like you want to like check the tagging on something to make sure the tags are logical. you know, you're not, there's not even like a meaningful way of saying what's a 10x better tagging taxonomy that gets created with Claude code versus deep seek. if you're just talking about, like, you know, replacing like the title of something to be like, you know, like, know, when you're like chatting and chat GPT, like it gives you that like smart title for like your thing that's Generate AI like. Rex Kirshner (36:52.95) Yeah. Gerrit Hall (36:55.01) How smart do you need it to be? is it really worthwhile to use like the Ferrari to drive that? Will like the jalopy will get you there just to find? Rex Kirshner (37:04.886) Yeah, no, man, I hear you. I think, you know, what's, yeah, I don't know. I think that's like all very specific to the applications you're building and the choices that you make as either the developer or the business, like, you know, commercial, like decision maker or whatever. and I don't really know where to go with that, but going back to the like, though essentially whining on Twitter about the usage quotas. Just like taking a step back, I just want to like talk you through how I see this whole situation and like get your reaction. And right, so like my read on what happened was that like, Anthropic and ChatGPT were in this like kind of rivalry, let's say three months ago where there wasn't, you know. ChatGBT had bigger brand awareness and more individual people using it. looked like Claude was like more going after enterprise and like whatever, right? And then something very specific happened, which was this whole like dust up between the Department of Defense and Anthropic about can we use this for autonomous weapons or blah blah blah, right? And then we all cheered Anthropic because they stood up to the Pentagon and were like, we're not just folding, like this technology isn't ready, no surveillance on US citizens. And then the Department of Defense was like, okay, well, you're like both a terrorist and critical to US defense. So like no one can ever do business with you. And then what happened was that so many people were so happy with an anthropics business decision that they moved into Claude. And... Then what happened is too many people moved into Claude that they don't have the bandwidth to support all these new people. And so, like they had to adjust quotas, right? Like they can't be as generous because there's too much demand. And so like, I do believe that we're getting less quotas. I know from my workflow that our quotas are smaller, but like you put it in context, that's why this happened, right? And now I'm watching people on Twitter and on Reddit being like, Rex Kirshner (39:22.934) Anthropic is shit, they don't know anything about business. Like, I can't believe I did this. I'm canceling my cloud code subscription today. I'm moving over to chat GPT. And it's like, hold on. You guys are the same ones that a month ago left chat GPT because they like essentially like got on their knees for the department of defense. And like, you thought that was not okay. But like the second that you realize like, this is going to have implications for you. And like, you are going to have to give up some of your like generous quota in order to like make room for all of the people that think anthropic did the right thing. And now you're like, Fuck it, let's just go back to Death Star. To me, the whole situation reads really bad. I don't know how deep you are, not at the implementation, the coding level, like we normally talk about it more on the business drama level, but I just wanted to throw that take out there and see what you thought, that resonate, you think that's wrong? What do you think? Gerrit Hall (40:29.134) Well, what I'm trying to model in my head is trying to make the math math. And I'm trying to figure out like that, because this just doesn't seem like it, it seems like it would be like even swap or almost like the anthropic would get the worst end of the deal in terms of the. Okay. So like how much Pentagon slash department of defense or department of war usage or whatever. Like they lost all that and at the same time they gain some percent of the customers from that. But like when I try and model it mentally, I would think that the amount of new business they got would be like more than, I think they'd be losing more business because like, I mean, the Pentagon. Like their budgets are huge, right? Like, you know, they like they are what a sixth of the entire economy. So like how much random B2C usage would they need to get to offset losing that much business? And I just don't think that that would add up to me. I would think that they would be on net losing a lot more usage than I mean, I know that like the echo chambers are really strong on Twitter and Reddit and social media and all that. But It doesn't seem credible to me to because like on top of it like the Department of Defense is not going to be like Paying for like just max plans will be paying for like bulk ultra max plans So I don't know like I mean there's I'd be interested to see the numbers on that it just doesn't quite seem credible to me that that would that that the amount of incoming new customers wouldn't be offset by the loss in the Department of Defense business Rex Kirshner (42:06.156) Yeah, I don't know. I mean, I see what you're saying. I think at the end of the day, this it all like, my understanding was there had like a one to $200 million contract with the Pentagon and like, yeah, but I think it was more because they were like piloting things and figuring out how to do the technology and the idea of being like, once they have products that are, you know, like powered by AI, then the spending would grow. And so I think it's more like Gerrit Hall (42:18.382) That's it. Rex Kirshner (42:35.532) they lost a decent current client with massive upside. so maybe the amount of compute that they had just for the Pentagon was the equivalent of, I don't know, 50,000 consumer accounts. So maybe the math works, maybe it doesn't, I don't know. But I was really struck by how everyone seems so upset with Anthropic a month after Anthropic was the good guy. And Anthropic was like the one we were cheering because they had like morals and like stood up for something and like are not willing to surveil us. But like, welcome to 2026. Like the news cycle moved so fast that now like that is completely forgotten. And the entire conversation is around like how stingy, how unfair, how like bad business Anthropic is. Gerrit Hall (43:31.438) Mm-hmm. Yeah. mean, my internet was cutting in and out there for just a little bit, so I might have missed some of that. But I think I got the key points. the actual numbers aside, it does seem like... I don't know much about the CEO of Anthropic Dario. I do know a bit more about Sam Altman because he's kind of been around Silicon Valley scene for quite some time. And like my read on him was always that like, you know, he is like a cutthroat business person. So when I saw this whole thing happen, it kind of didn't surprise me that he was jumping at the opportunity to do whatever it would take to get the business of the Department of Defense. So it just sort of tracked with what I knew about his capabilities. But I think the largest thing that I think about from all this is that really the switching costs do remain very small. Rex Kirshner (44:30.157) Mm-hmm. Gerrit Hall (44:30.306) So at the moment, I've still been using Cloud Code a lot because I really like it. Even though I think, like I've said before, Codex is a better coder, and I do tend to use both. I still tend to use Cloud for most of my workhorse tasks, but try and use both. But I was using Cursor late last year. It turned out that Gemini proved better for my applications, or something new came along. it would not be a complex for me to switch to something new. So I don't think there's actually that much moat and a lot of these companies are like, it's a tough fight basically. Rex Kirshner (45:09.954) Yeah, so I have been messing around with Codex, ChatGBT a lot more because I don't like paying for two CloudMax accounts and so my thought is like, can I offload some of the, and I will say that right now the quotas on ChatGBT are like undoubtedly so much more generous. Then they aren't, I'm paying for $20 for chat GPT and I get like similar amount of use that I do out of the $200 plan. I don't think that'll last. Yeah. I don't think that'll last, but, I will say. I, I have a few just like templated prompts of code reviews, right? Like so specific, specific type of code reviews. And I always, you know, just like when I hit major checkpoints on projects, I'm like, all right. Gerrit Hall (45:39.33) Right. yeah, same here. Rex Kirshner (46:01.721) Claude, can you just run this code review, see what you find? And in general, it finds like, small incremental things that happen, but like, it always says my code is pretty good. And I started running that on ChatGPT, and like, not only is it finding things that Claude didn't surface, but it's like pointing out that like, hey, like the Claude.md file is, it said the word it's lying, right? Like. It says that these patterns don't exist, like it's lying. Cause I found these patterns here. And I honestly, I think that the competition is so like fierce and these people are so petty that like it's totally possible that chat GPT reads something out of Claude and it like there's an automatic snark meter. Like I believe that that's possible, but I am a little bit concerned that why Claude. feels better and nicer is like, it is not able to understand when it's like just doing stupid shit. Gerrit Hall (47:10.956) Yeah, I don't know. mean, it's, so you're with me though, like Codex does seem to be the better coder, right? Like, but yet, Cloud remains a better experience in a lot of ways. Rex Kirshner (47:19.584) Yeah. So I, I, if you asked me this a month ago, I'd say, I don't understand how people are saying that because the first time I use codecs seriously, I had it build like a full application from, you know, greenfield to like having a working UI and the UI that came out of it was like, it was so bad. It w it looked like 95 garbage and, and I've since learned that apparently for whatever reason chat GPT is very bad at UI stuff and... maybe that's just like a weakness that it has. so like rely on Cloud Code for the UI, like user facing stuff, but like Codex can do all the backend stuff. Maybe that's where we're at. I don't know, but to me, it's just impossible to ignore that right now, if I have Cloud Code and Codex audit all of my projects using the exact same prompt, Cloud Code will be like, this is as close to perfect as possible. And Codex will be like, Wow, this is a disaster. Gerrit Hall (48:27.158) I mean, I do kind of like that about Codex and that it is kind of more like no BS. I think they have a mode where you can be more sycophantic. But Cloud Code is excessively sycophantic. And I think that's probably why people just gravitate towards Cloud Code, because who doesn't want to hear that they, oh, you're such a genius. This is the best idea I've ever heard. Codex is much more likely to be like, OK, I can do that. Rex Kirshner (48:50.583) Yeah. Rex Kirshner (48:54.646) Yeah. Yeah. Yeah. I, I don't know. I think to me it also like it's moments like those that where I'm really just like forced to put things back in perspective on like you use these tools and while you're using them, it feels like we're in the end game, you know, like you're watching it, right? Like 30,000 lines of code just in front of you and something comes out and it like kind of works and Then it's moments like these where you're comparing against two different ones and you realize, man, we're in the rickety popsicle sticks and glue phase of this and kind of nothing works that well and it's up to us to piece these things together. Back in the 90s, you had to build all your own shit from scratch and yeah, it could look kind of good or could not, but it's not the same as having Squarespace today. Gerrit Hall (49:53.314) Yeah, it's true, although I do think that like by and large, like the amount of improvements we can get in terms of the coding abilities of these things is gonna be more incremental. I think the past few releases were fairly significant jumps. First it was like, it can kind of produce usable code. And then as we've gotten a couple more releases, it's like, okay, it can actually produce pretty good code. But now I think we're actually gonna start to, I'm sure, see the influencers complain that, the newest model came out and it's not that much better. Because I think it's probably like the 60, 80 % level. And then we're just going to get to 90%, 95, 90. There comes a point where there's only so much more you can squeeze out in terms of gains, in terms of code quality. Because it is producing reasonably high quality code that's usable in a production environment at this point. Rex Kirshner (50:51.67) Yeah, I do think that's true. Speaking of models, do you catch like that whole drama about how like Anthropic leaked something? I'm not really sure, but in it, there's like this new model mythos that is going to like, I don't like that. It's so powerful. They have to hide it or some shit like I don't even know. But I these. Yeah, the way they're talking about like the next gen is like we have a full on Jarvis. Gerrit Hall (51:02.228) yeah, the April Fools, maybe? Gerrit Hall (51:21.678) Okay, so I'm a little confused about what exactly is was the situation here? Maybe you've like read a more and like can inform me so like my understanding is that there's like a big quote-unquote leak of some sort of Claude source code and people like were a lot of these influencers especially were like being like here's ten wild things I found from going through the source code and then like this came out like very very late March and Rex Kirshner (51:42.773) Yeah. Gerrit Hall (51:51.252) on April 1st or after like the Claude announced that that was actually like their April fools joke was like that this like massive thing that was released was actually just like, you know, some vibe coded, like, like monster, like, repository, with a lot of like jokes baked into it and It might, I don't know, if so, that's actually kind of like a weirdly brilliant April Fool's joke if that actually released this fake repository before April 1st. But I don't know, is that actually what the story was? You tell me. Rex Kirshner (52:32.81) I yeah, I mean, this is definitely a weird one, but here's my understanding. So it was like March 30th when they accidentally published all the source code for Claude code. And like we have known for, they talk about it openly that Claude code was built from Claude, right? Like it's all like they're not writing code anymore. It's all Claude, which I don't have any problem with. Like that's what I'm doing. So that's fine. but. So they believe that their secret sauce of like why cloud code is so good is in the like the CLI tool itself and That is in contrast to open AI and all these other companies who all of them including open AI they have they've open sourced their the harness so like codecs is open source and What so anthropic when they publish cloud code they what's the word they like flatten all the JS so that like you, what's that called? Like minify it or whatever, so that you like people cannot understand what the code is. And then what they accidentally published with that is the, the manifest that allows that says like, okay, this is like minified code is associated with this actual function call. And because that stuff is needed internally to do Like bug reporting and that kind of stuff and so like they accidentally published that From that people were able to rebuild the entire repository like the source code and so now there's like repos out there that you can find from that are the clod code source code and Then my understanding is like two days later. So on April Fool's Day Anthropic posted that like actually it's April Fool's like ha ha ha But my understanding was that was the April Fool's joke, that they were saying that their leak was on purpose was the fool. Because at the same time that they said that, they've been incredibly aggressive about DMCA complaining every GitHub repo with the source code in it. Rex Kirshner (54:54.232) So like, it's a mess, man. Like, I don't really understand or care, but I do think it is like kind of a signal of how fast everyone's moving to the point of recklessness. And like, it's one thing when you like leak the source code of your coding harness, okay? But like, you gotta wonder what else is falling through the cracks. Rex Kirshner (55:29.784) Yeah. Rex Kirshner (55:45.551) I see the same thing too, it says slash buddy in like the bottom right corner of my terminal, but I just, I can't bring myself to do it. stop like injecting fun, know, like fuck, like this is a coding tool. I don't want a Tamagotchi pet in it. don't, to me, this is just like so like, it's like these people are watching Silicon Valley, the HBO show from the 2010s and like, Instead of realizing like, they're making fun of us, they're like, this is good. Like, let's do this. You know? Like, I want to be like these guys. Rex Kirshner (56:43.916) Yeah. Yeah, well, it's just like, why are you like, this is what I want from you. Like, please stop forcing me to download like whimsy, your definition of whimsy. Like I don't, I don't want to see this. This is not, and it's fine when it's like, I'm sure like one 30th of the package and like, won't really use that many tokens. And like, yeah, sure. Like we're allowed to have fun. That's fine. But this is just not where our I think like I or most customers want to see them put effort into. Make their product better. Stop doing little toys. Rex Kirshner (57:34.477) Maybe, yeah. But if that had been the April Fool's joke, I would say fine. Not my cup of tea, but fine. But this was also paired with a clearly mistake of publishing the source code of Cloud Code. And I'm like, were you guys so focused on the Tamagotchi that you had to rush things through and that's why this was missed? I don't know. Rex Kirshner (58:20.728) Yeah. Rex Kirshner (58:39.062) Yeah, you know, just to like to bring this home, going back to crypto, to me, this like smells a little bit like the beginnings of something like bear a chain where it's like, okay, like I, I understand whether or not I think it's going to work. I understand this idea of proof of liquidity and how that could be different. And maybe there's something interesting there, but it's like when increasingly more and more of your time until by the end of all of your time is focused on like NFTs of bears smoking bongs or like drone light shows in Dubai or like having the biggest, baddest party at every conference, you start to wonder like, what are you guys here for? Rex Kirshner (59:38.784) did I lose you? Are you muted? Rex Kirshner (59:54.188) There you go. no, back off. Rex Kirshner (01:00:01.347) That's okay. Just respond to the barrage chain thing and then I'll close this off. Rex Kirshner (01:00:24.61) Yeah, well, I hope their slash buddy command helps JP Morgan realize that they're a strong financial investment. Rex Kirshner (01:00:56.566) Yeah. Well, I will say like the best Halloween party I ever went to was 2015. Yeah. 2015. I was living in New York and my friends are all in the tech industry and we went to the WeWork Halloween party, had like, it was at least a thousand people. had like major tier one celebrities doing music. And I just remember thinking like, this is really cool, but like, why are they doing this? Like why? Why are they throwing party for like a thousand random people in New York on Halloween? Like what is this? like, look what happened with WeWork. Like it turned out that there really wasn't anything there other than just like some guy's like delusion and then like ability to convince investors that there was something there. And I think that's like really the story of Bear a Chain as well. like, I just. I smell that a little bit in Anthropic right now and I don't like it. Rex Kirshner (01:02:04.852) I guess, but at the end of the day, John Gier still sells tractors, right? We weren't really selling anything. Bear Chain definitely doesn't sell anything. Rex Kirshner (01:02:31.394) Yeah, well there's something like definitely true and very depressing and like broken in that. But anyway man, we've been talking for an hour and I, now we're talking about WeWork parties so I think this is good place to cut this off. But Garret man, it's always a pleasure. Thank you very much and I'll see you soon.