TWiT 1093: California Sober - Kimi K3, Qwen3.8, & China's Open-Weight AI Gambit

[00:00:00] It's time for TWiT This Week in Tech. We've got Alex Wilhelm here. We've got Harper Reid. It's going to be a fun show. All of the new AI coming from China. What's China up to? What's America up to? And how's it going to respond? We'll also talk about the giant bill a few people received from Amazon Web Service and the new AI teacher in Salamanca, New York. You won't believe your eyes. It's coming up next on TWiT.

[00:00:28] This episode is brought to you by Black Hat USA. If you listen to this show, you go deep on the technical detail. Well, so does Black Hat. For nearly three decades, it's been where the security industry's most rigorous research gets presented and pressure tested. More than a hundred hands-on trainings taught by practitioners who actually deployed in live environments, not lecturers reading from slides. And hundreds of peer reviewed briefings that go well past the overview into the room.

[00:00:57] Real work across the four areas defining security right now. AI and autonomous threats, cyber conflict, systemic resilience, and identity. This year, Black Hat's briefings pass includes all keynotes and main stage access, plus business hall entry. You also get breakfast, lunch, Arsenal live tool demos, on-demand session access, and admission to the Midnight in the War Room screening.

[00:01:23] Black Hat takes place from August 1st to the 6th in Las Vegas. If you want the depth this show gets into in person with the people doing the work, this is the room. And we'll be there too. Use the code TWIT for $200 off your briefings pass at blackhat.com slash US-26. That's B-L-A-C-K-H-A-T dot com slash US-26.

[00:01:52] Podcasts you love. From people you trust. This is TWIT. This is TWIT. This Week in Tech. Episode 1093. Recorded Sunday, July 19th, 2026. California sober. It's time for TWIT. This Week in Tech. The show we cover the week's tech news.

[00:02:20] And as usual, when I have Harper Reid on the show. Hello, Harper. Great to see you. Technologist, entrepreneur, hacker. Nice to see you. He runs a little AI doohickey called 2389.ai. Always wonderful to have Harper on. He's been traveling a lot. We've been trying to get him on. I'm happy to be here. And when we have Harper on, we often don't. We decide to have a smaller panel because. Well, when I get shut up. No.

[00:02:48] That way you're like, man, Harper will not shut up. No, no. There are people. Michelle Mare and Nameless who come on the show who I can't get to shut up. You're not one of them. But you have such great things to say. And your rapport with everybody is so great. Remember, last time you were on was just you and Amy Webb. This time, I'm going to introduce you to a new character I think you will like. Alex Wilhelm is also here. Alex has a wonderful newsletter. That is going, you know, great guns, cautious optimism.news. Yes, sir.

[00:03:17] And you also hear him on this week in startups. And it's great to see you once again, my dear friend coming to me from my old backyard. Indeed. And Leo, I just got, I just got to say of all the times to come on the show, the week after the AI policy fracas is the right time to have me on. So I'm excited. I'm excited to have, I mean, like everything that I want to talk about is what's in the show notes today. So thank you for having me as always. It's a real treat and I will not burn down your house. I, please don't. I, that is an option. I suppose it's not your house. It's your, your house.

[00:03:47] It isn't my house anymore. No, it is not. So you can burn it down. You have permission. I'm not on the deed. Ladies and gentlemen, it is true that this is the, this is the week that the AI stuff kind of exploded. We do have a full-time AI show. I like that. This is the week that the AI stuff exploded. I don't know if y'all have been awake for the last, I don't know, two years. Like this is, this was no different than the rest of the week. No, it's true.

[00:04:16] And people, and people email me and say, well, can you stop talking about AI? Isn't there a new phone out? And I say, no. But if it were to be out, it would be all about AI. It would literally be all about AI. It is. The AI phone. It is. It'd be like the new foldable iPhone is full of AI and it's horrible or whatever. Yeah. I do have one phone that has no AI on it, which Jeff Atwood sent me. It's a e-ink screen. It is the Mudita. It has nothing. It's got nothing on it.

[00:04:45] It's just nothing. It's just the simplest, purest. And who wants a phone? This is text and phone. Oh, wait a minute. It has some apps, but I can't imagine they're very good. Calculator, calendar. Can Jeff Atwood send us some? I will send you this one. There is a camera. You don't use it? You know what's funny about the camera is it takes full color pictures, but they're, but they're, go ahead and go to the three shot real quickly so I can get a good picture of these guys.

[00:05:15] They're, oh, it's so slow, but they're, they're, they're, whoa. Oh, I love that. I love that. I'm familiar with this. They're monochrome pictures. Oh man. We've never really made e-ink screens fast, have we? The Kindle one is pretty fast, but other than that, most of the consumer ones are not as fast. But I think the point of this is it's a phone that you don't want to use. So you don't have to worry about getting sucked down the Instagram hole.

[00:05:40] I think, I think the problem that I have is that I, I love and I fantasize and I dream about this type of phone. But I think if I started calling all of my friends on the telephone, they would stop being my friends. You don't want this phone. They would just be like, please do not call me. You have to text. No, no, no, no, no, no. People that complain about phone calls that you know are wussies. I just call everybody. In fact, some people, I just aggressively FaceTime them without calling ahead of time. I'm just like, that's actually one of the best things ever.

[00:06:07] And the unannounced FaceTime, especially at odd hours is so good. So horrible. It's so good. So horrible. You get to see your friends waking up from a nap there in bed holding their phone like a baby. Exactly. What? What? What's wrong? I'm like, no. Well, first of all, there's no, you don't know that it's going to be a picture, right? You just think your phone's ringing. And so you pick it up and there's Harper? It says. I mean, that's true. That's true. That's. What are you doing here?

[00:06:37] I love it. I love it. I think it's great. I'm sticking my face into your house like this. I'm just going to walk into the door and just put my face in your bedroom. Hi. You know what? It does. It does show you which friends are like always down and which friends you need to put on the calendar first. Like if somebody smiles broadly and says, Alex, then, you know, this is, this is a true friend, even though they're still wearing curlers. So you're right.

[00:07:02] I should point out this is no really special week in AI, except it sort of feels like. We're still alive. That's what makes it special is like the things about AI that's cool is we haven't all been killed yet. But that's good news. Right. The fast takeoff hasn't happened as everyone promised it would. It feels like it's happening. You don't think it's happened? I don't know. I oftentimes wonder there's, there's a couple of things that I wonder.

[00:07:30] One of the things I wonder a lot about is as these models get more powerful, you think about both our friends that have worked at some of these big companies and they say not like these, but like, let's say Apple or Google, and they've been using some technology for a year before it gets to us. And obviously the cycles are much shorter now, but like what is, you know, open AI or, or anthropic or Kimi, what is their six months early look like?

[00:07:57] Like they're like, what is, what do they have access to that we as just boring consumers would never ever dream of? Well, I think we find out pretty quick because one thing that I always find really frustrating is when they drop a new model, like the open AI GT, GPT 5.6 family. And then everyone that I know on Twitter is like, oh yeah, I've had this four weeks. It's fantastic. I know this is frustrating. Okay. Well, how did that happen? Who are you? Yeah. Why do you have it? You have seven followers and you're annoying and I don't like your face.

[00:08:26] Uh, have we figured that out? I don't know how that works. I've been trying to figure that out too. It's, it's, there must be some like club that we're not in, but I will say that I don't think there's stuff that's that production ready. That's so far advanced of what we have or is coming down the pike. Like, I don't think there's a special access program of like, you know, special jets from open AI or whatever. I don't think the models are that much better. Cause I think you have to, given the cost of creating them, get them into the market with a reasonable timeframe to recoup the money you spent training them.

[00:08:56] Um, but I will say the cadence that we've seen from AI in the last six to nine months to me really does. I think show that the breakthrough is an agenda coding. We saw at the end of 25 with, um, was it opus 4, 5 at November 24th, a day that will live in my heart and infamy. Yeah. I was concerned, it was like, okay, the cadence is faster. And I think we're seeing AI models either accelerate humans to do more, more quickly.

[00:09:24] Or I think as we've seen the last couple of major releases, you know, sparks of recursive self-improvement or whatever. I think we're starting to get to the point where that's having a very modest impact, but I think that that plus the accelerated humans does feel like a new gear. Have y'all noticed more features being released and other software you're using? Oh, that's interesting. I have noticed more bugs. So I think that's part of it.

[00:09:49] This week, Tuesday, Microsoft shipped a patch Tuesday with a record number, more than 500 fixes. And if you include the fixes they applied to Microsoft edge because of the Chromium project, it was over a thousand. And that's AI, right? That's, that's mythos. Yeah. Well, I meant the other kind of direction of literal features. Yeah. Yeah. I mean, it's one thing when you have all of these things. I have and have you?

[00:10:14] So I've, I've been starting to notice a few things and I, and it's not enough to, to like have a list, but I'm, I keep thinking like, oh, that's new. I wonder if that was just something they shipped really quickly. Like I keep thinking this as I use an app, like a Gmail or I'm trying to think what else app I use, but like there, it seems like there's some, a quicker pace. One would expect that actually.

[00:10:41] Uh, instead, what I see is a lot of, uh, vibe coded crap, to be honest with you. GitHub is overflowing. And by the way, I have contributed. I all of a sudden have 20 GitHub repos and I'm not even posting everything I'm doing. So the explosion of consumer software in that since Leo has been so, uh, fun to watch through the lens of a niche. So I'm a, I'm a poker fan and I hang out on the poker subreddit.

[00:11:11] And there was this period of time when cloud code took off before everyone ran out of compute capacity in which they literally had to tell everyone no more sharing your free poker app that you built because there are too many here. There is a, because everyone was like making the, you know, GTO trainers and all this cool stuff. And it was fun. It was cool. Everyone was pretty happy, but it was just every other post was like, look at what I made. I follow a subreddit called RSS, which was a fairly quiet, not, not much was happening.

[00:11:38] All of a sudden, every single person on there has written, including me, their own RSS reader. There are an infinite number now of five coded RSS readers out there. Uh, so that if, if you include that Harper, yes, there are more features, but in commercial software, I don't see a great. So just start paying attention because this is something I'm kind of working through. We should see it.

[00:12:03] I think, I think I really do think there is something happening. Um, just because I have, I have, I'm going to create a list and then next time, whenever I see one, I'll create a list. I'll make a note and I'll just start doing it because yeah. Well, you can't have the AI do it because it's a lot about feel, right? Like it's a lot about like, oh, this is interesting. This is a new feature that I wouldn't have expected. Cause I guess what I'm trying to suggest is that it's not just features.

[00:12:32] It's that features that are not that, that you, as if you like go through a quick product manager audit, you'd be like, oh, why would they prioritize this? This is interesting. It has a little too many flourishes or it's a little too fancy or it's a little too on the edge of something or another. But I think it's great because we are seeing this internally to my company, like the fact that we can ship anything very quickly. So I would love to see it more at some of these bigger companies. It shows that it's working. You're looking for a move 37 feature. Do you know what I'm saying?

[00:13:02] When I say move 37, uh, when alpha go was playing Lisa doll in go, there was one move move 37 where all of them go masters said, no, that's, uh, uh, what they said. That was a terrible move. It was a non-human move that turned out to be a winning move. Yeah. And that's what you're saying. There should be some move 37s out there. Stuff that one wouldn't have expected. Maybe.

[00:13:32] I think that there are, I think that there's a thing that I keep thinking about. Like we, everyone talks about software factories in my world. Like that's the thing they talk about. And there's a couple of things that I don't really understand why we are talking about that. The first one, it's like, it sounds like a good idea. It sounds hard, but everyone's acting like they're done.

[00:13:55] Um, and, and I feel, I feel like there's a, this thing where if you had a real software factory and it was truly effective, absolutely working. And you were able to ship things with like, you know, just let's say you 10 X or 20 X all of your humans. Why would you tell anyone about it? Oh, you're looking for signals of a rumble from someone else's earthquake. That's beneath the surface. I see. So you're not looking like, here's the corporate blog post.

[00:14:23] Because everyone who's saying like, we use AI to do a billion things seems to be lying. And there seemed to say that in a, in a, in a coordinated way, because they're, they're then going to do a layoff of like 15% of their workforce two weeks later and just be like, Oh, it was AI. And it's like, no, you just over hired during COVID. Like, this is a very clear example of you hiring, you know, 40% more people than you need. It's AI washing. Yeah. It's AI washing. And I feel like that's been for the last couple of years, but I guess my assumption is I'm seeing

[00:14:53] we've seen, and our company has been changed forever with some of how this stuff works. So I am assuming that a bigger company that is doing similar stuff to we is going to have the same examples, the same outcomes, the same, et cetera. How many people are at your company? Six. Yeah. That's why it works. I think, I think that's why when we see these layoffs and I'm not even being facetious, people say like, you know, we're going for smaller teams. You know, the Amazon two pizza rule has become the Amazon one pizza rule. Everyone wants to move faster. Yeah.

[00:15:20] I think the lack of bureaucracy layers, size, scale, enterprise SLAs, and so forth is why you can do that. But what I am seeing that I think matches what you're saying is I'm shocked at how much stuff companies now drop in a major release. So like when Databricks put out their genie, Omni, whatever their new agent is, they had like, it felt like 15 different posts about new stuff they put out. Microsoft during build, they dropped more things than I could literally keep in my head to write about. And so this is, this is kind of what I mean, right?

[00:15:49] Like these, I think are signals that it's working for some people. And I don't think you have a, I don't think it's of interest or there's no strategic interest to tell people you're using these tools effectively. Yeah. Because it, it will alienate hiring. It'll cause hiring to be harder. It will maybe give your competitors some sort of heads up to say, oh, like maybe we should be looking at that a little more seriously. Especially if you're linking it to actual outcomes.

[00:16:19] And so this is where I'm like trying every time I'm, I'm using a big company's app, I'm a little bit looking, like trying to read tea leaves through that interaction to be like, oh, is this new? Did they always have that shadow there? Is that, that seems like a little bit of a flourish because I do think that what we'll see is we'll see parts where people are showing off where they're actually explicitly, you know, because like you're saying it is a smaller team thing, which allows for more specifics. Like a friend of mine here in Chicago built a, you know, NBA jam, the video game.

[00:16:48] And because it was such a small, tight team, the Easter eggs are literally the developers heads on the players. Right. And so I think we're going to start seeing some of those kind of Easter eggy personalizations again, because the teams are so small and tight. And so that's the flourishes that I'm kind of looking for. What is this little thing? Oh, when you hover over this, you, this happens, or you double click on here and you get the death star instead of a planet or whatever it might be. It's my guess though, that the bigger enterprises all are working as quickly as they can to create

[00:17:18] internal software that you're right. They probably don't want to tell anybody what they're doing for a variety of reasons. Um, I, you know, I've heard anecdotally in our AI user group, we have a number of people who work for a big enterprise that very commonly in big enterprises, you have multiple teams. There's not one, you know, monolithic AI strategy. You have multiple teams working on multiple different things. And I also, from time to time, see stories like we replaced our CM CRM, our half million

[00:17:46] dollar a year CRM budget with a vibe coded software, or, you know, where you really see it, customer service. Yeah. Yeah. Um, right. Uh, yeah, yeah. People are admitting that, that the, in fact, it's kind of interesting because while you and I might say, God, we hate those phone trees and we hate automated customer service in general, it seems to be a good experience for customers that they're triaging the easiest things fast so that the humans can handle the hard things.

[00:18:16] You know, it, it happened in Linux, Linux world this week, Linus Torvalds surprised me a little bit. He on the Linux newsletter said, I realized that some people, uh, or rather mailing list. I realized some people really dislike AI, but this is an area where I'm willing absolutely to put my foot down as the top level maintainer. Linux is not one of those anti AI projects. And if somebody has issues with that, they can do the Orpus open source thing and fork it

[00:18:44] or just walk away. AI is a tool just like other tools we use. And it's clearly a useful one. So this was seen by, by the open source world as an endorsement from somebody you might say is probably anti LLM. Not at all. I, I mean, I, I think it's interesting to fight this battle a little bit because what I have found is, um, it doesn't seem like this is a bell that we're going to unring. No, that's for sure. Well, okay.

[00:19:13] But we are going to talk about some headwinds because there are some serious headwinds. But yeah, tell me. All right. Well, no, no, I want you to finish the thought. So if this, if we have already crossed the Rubicon and AI is, is done, it's going to, it's going to happen. What? Well, I don't even mean it's like done. I don't, because done have offers some idea of completion. I just think that we are, we are things have accelerated. Like I've done a handful of companies and every time we had the same pattern.

[00:19:43] And I've said this before on the show, like, you know, we used to, we used to very quickly hire 20 people and we have very quickly not hired anyone since we started the company. Right. And we have had so much output. And I do contend that like, I agree with Alex, like this is because we're small, we're able to do this, but the amount of output that we're doing is more than I've ever seen in my entire life. It's almost more output than my entire career prior. Like it's so much output.

[00:20:11] There's a focus issue there, but well, let's not talk about that. But, but I do think that there's a, there's a lot of like, this is hard to contend with. And I think if you build in the process to handle it, then the benefits are so big. You have such huge benefits, which I obviously Linus is seeing, especially for something as important as Linux or, or what have you.

[00:20:37] You know, there was a friend who was mentioning this, that when the US government nuked Fable, the project Glasswing or whatever was still doing reports because you were still seeing bug reports being submitted that were written very thoroughly, which obviously means Fable wasn't nuked for everyone. And so it's like that kind of thing, like those, like the fact that Microsoft can be like, here's our patches and it patches these, all of these bugs makes everyone safer.

[00:21:06] And the fact that Linux is on a hundred percent of our devices and Linus obviously sees the benefit from progress, project Glasswing, et cetera, means that I think it makes sense. The people that I know who are anti-AI as coders think they are very smart. The people who I know who are pro AI and very good coders oftentimes are like, I'm very smart, but this thing is really good and I am smarter with it. And I think it goes back to the lease it all goes stuff, right? Which is basically like when you hooked a person to the machine, it was pretty good.

[00:21:36] Which is why I believe in the matrix. I think there might be some commonality with people's attitudes towards self-driving cars. Lisa had a big fight with her, uh, her, uh, her brother-in-law yesterday who said, no, Waymo's are really dangerous. And Lisa said, no, the stats are clear. Waymo's are roughly 10 to 15 times safer than human drivers. Uh, they're not, they're safer. He said, no, no, no,

[00:22:05] AI can't drive. And I think it's, it's maybe kind of like that, that there are people who deny that AI can be any good. And then, uh, then there are people who are using it. I think also, Leo, we're seeing people that used chat GPT's web interface once 18 months ago, then saying, oh, that AI is driving my car. But I will say in their defense, AI is jagged. It's intelligence is jagged. And so Sam on bull Samet, who's a regular on our show and a car guy,

[00:22:35] said, you will never see level five self-driving because there's always going to be an issue. There's going to be a thing that, that AI isn't prepared to handle. And so completely autonomous on any road is not going to happen. Um, and I, I think he's somewhat credible. I don't think he's an anti AI guy. I think he's somewhat credible on that. Uh, but, but nevertheless, so this is,

[00:22:59] this is why it's not quite as cut and dried as all that, uh, AI. Well, here's another question in software, uh, Steve Gibson, our security guy, we were talking about all these, uh, bug fixes and he says, it's what you're going to see. He's prediction is what you're going to see is a rise an increase in the number of bug fixes to dramatic level. We might already be there. And then he's, and he says, this will happen over the next year, a decline to the point where you don't have any more bugs.

[00:23:30] Hmm. Harper, you're more of a developer than I am, but I don't think we're going to end up bug-free software. I think we're just gonna have a faster remediation cycle. I mean, I think that's true, but it might be equivalent, right? Like, I think the, the thing that I found is that as a developer, we used to, we, we did testing, right? So in, you know, you 2016, you would say, let's get our test suite up and you would shoot for like 40%, 60%. And if you were insane, like truly had nothing else going on in your life, you might hit 80%.

[00:24:00] Um, you know, some people said we have a hundred percent test coverage. And mostly that was performative. Like people, I don't, I don't know very many people who worked in production environments who had such high test coverage because there was so much other happening that it was scary. You had to just do stuff and it was hard and you had to just get all this stuff done. And, um, now that's just table stakes. Like you, Cloud will just give you a hundred percent test coverage. Um, and you can kind of count it in any way.

[00:24:25] But like, uh, basically dummy tests where, you know, of course it does. Of course it does. But, but this is where you still have to have good practice. Like you still have to have good code review practice. I agree with you. You have all that stuff, but it's, and this is where the bugs will disappear, right? I have my guy do TDD. In fact, I do, I do red, green, blue testing.

[00:24:50] Yep. Uh, because that's how I learned how to code. Right. And I think that's probably a good sensible way for the AI, but I don't have, I think this is where the jagged thing comes in. Sometimes most of the time it's good and it'll, it'll, it'll write a test. The test will fail. It'll write the code. The code passes. It does behavioral testing and everything seems to work. And then sometimes it'll just say test passed. It'll just say test. Oh yeah. Yeah. Yeah. I think, I think that could be occasionally it running out of compute on the backend and then

[00:25:19] they're going, okay, JK, now you're in high. There's a variety of reasons it does it. Well, but if I, Harper, please. I was just kidding. My experience with, with building like all of our process around this, um, it's just, you have to do more loops. And so what, what, you know, and this is the popular word on Twitter, et cetera, loops, loops, loops, loops. Can you explain this to me? Cause that's all of a sudden that's all I've been talking about on X is loop, loop, loop. I mean, it's pretty stupid, right? Like all you do is you just are like slash loop, do this until it's done.

[00:25:48] And it literally will just do that until it's done. And so that's why they were like, well, that really is way to burn tokens because then it just does a bunch of sh**. So you do slash goal where you say, do this until the tests pass or do this until you have a hundred percent test coverage. That kind of loop. So you give it, and so that kind of thing. So what, but what we have found is that, you can't just have Claude do it. So what we'll do is we'll use Claude. It'll go through, it'll build the thing. It'll be like test pass. Everything's great. And we're not looking at code.

[00:26:14] So then we will pass it over to codex and then codex will be like, y'all are insane. These tests are horrible. I have three agents now. Look, yeah, exactly. And then you pass it to, exactly. And actually I fable do the design, but because it costs, I keep, and we're going to get to this in a second. I keep expecting them to turn fable off. So I have fable design it and hand it off to Opus 4-8 for the coding and then have it reviewed by Sol 5-6 and then have that reviewed by 4-5. And then,

[00:26:43] and I, and in order to do this, I gave them, I got tired of pasting the results from one, from codex into Claude code into Hermes. So I just said, email each other, would you? I said, I said, you got a mailbox, just email each other with the results. What are you using for mail? Obsidian. I just had them make text files. Yeah, exactly. And they have little folders. Yeah, it works perfect. Because then I can read it in case I want to. Although, as you said, you think you're going to read the code. No, no, you don't. You expect to read the code. I don't even read the plans anymore. It'll say,

[00:27:12] I've written an entire plan for you, 15 pages. As soon as you approve it, I'll get to work. And I just go, yeah, yeah, go ahead. So I think what this means is that if you believe the agents, then we get what amounts to normal output, which is the same thing that happened if you believe humans as well. Like there's no difference. I don't think agents anymore will write buffer overflows or will access memory and read zero. Or they will write it, but they will find it much,

[00:27:40] much quicker. And so like the process that you would normally have is to just kind of keep iterating on this stuff. You know what I mean? Like you would keep pushing, keep going, keep doing, and then it will just kind of do its thing. Like it'll fix the bugs it finds. And I don't think we'll go to zero bugs because bugs are mysterious and they seem to just happen and generate. Well, there's different kinds of bugs too. I mean, there's bugs that are coded bugs. And I guess

[00:28:07] that's what we're talking about. Not things like timing bugs or row hammer or, you know, things that are kind of not code caused, but I think we can expect code bugs to become increasingly rare. Don't you? Um, yeah, I think so. I mean, we're seeing this, like we're seeing our software is 10x better, but the other side effect of this is that we are also using programming languages that we would never

[00:28:35] have used before. We're using a lot of Swift and a lot of rust and a lot of go. Yeah. You know, these are much safer than, um, you know, Python. Yeah. And who wants to type all that rust? But, uh, but if the agent's doing go right ahead, please. We want to. Exactly. Yeah. Those can't be real programming languages. When I was growing up, C++ was a programming language. Now you're like, I'm using Jam with whip and then I'm going to use that. Here's a good question. Aside from reviewing existing code, does anybody write Greenfield code in C++ with their AI? No. I have.

[00:29:04] Have you? Yeah. Why would you? Well, I mean, sometimes you need to write C++ code. Yeah. And, and the, and the thing that's funny is they just are just like, yeah, it makes sense. Why wouldn't you do that? That's totally what you do. It doesn't care. I think the interesting thing about what we just talked about writing code yourself is, can you imagine a child born today going through the hell of learning the syntax arcana of C++ and getting it right and thinking that that was a good use of their time versus architecting

[00:29:32] a system that we built by agents. Like this is going to be a lost art in not that long. And this is actually something that was predicted a long time ago. I forget which foundation novel it was. It wasn't in the main trilogy. It was one of the prequels or sequels, but in foundation, they literally write code to have the computers, write the mathematics behind psycho history. When I was reading that as you know, like 15 years ago, I was like, whoa, they're not even doing the

[00:29:57] work themselves. How will they ever know if it's right? And then here we are. And it turns out that, well, it's actually a pretty good way to go about things because computers can do more faster than us. Plus they speak computer, which you know, we see anyway, let's take a break. We didn't even get to any of the news stories. Well, we can, let's do them. We will, but we're going to take a break. We'll come back to actual, the actual news because it ties right into this beautifully.

[00:30:24] You're watching this week in tech. We're so glad to have Harper Reed here. Nice to see you. I have used many of your skills at 2389.ai. A lot of the review skills are very useful. Good stuff. Are you still putting new ones out all the time? Oh man. We have so many, we have one called thrifty that you might like. Yeah. That, that does, it's really good with fable. You basically use fable to plan and

[00:30:49] super cheap models like haiku to do the work. Ooh. Ooh. So that was my idea. I thought four, eight was cheap. Use haiku. Wow. Okay. Yeah. I mean, it's, we, we tested it even on some open models. And I've been, we've been thinking a lot about this because we do a lot, we, you know, trying to output the stuff is really interesting. And so we've been really thinking about how do you kind of lean on that a bit more? Like how do you build pipelines and whatnot where you're using

[00:31:17] quenm? Well, you can use four, eight, or you can use GLM or whatever they do planning. And then you, you can throw that to, you know, Gemma or, or Kimmy or like something to do the code. All it has to do is be able to do tool calls. Right. That's it. And as long as it can do tool calls, you have a reasonable harness that can handle it. You can kind of get some stuff done, which means that like, I think the, what we will be talking about soon, which some people already are,

[00:31:45] you can hear them is you have to be able to pay less than retail for tokens, which is why everyone loves subscriptions. Right. But if you're a big corporate, you're still stuck with like APIs and you're just paying retail. And it's like, you'll never build anything that will be sustainable because it's just so expensive. OpenAI or Anthropic or whomever will beat you. Google will beat you. But with Kimmy and GLM, suddenly things are getting much more interesting.

[00:32:11] And so I'm very excited about that stuff. I think this is where things get really interesting. Um, it is. And, uh, we have more to talk about in just a bit. Also, Alex Wilhelm is here. His newsletter kind of covers a lot of the same stuff, cautious optimism. Uh, and you can subscribe to it. And, uh, but you can also read a lot of it right there at cautious optimism.

[00:32:38] News. News. Yes. News. Yes. Because it turns out that when you think of a phrase that people already know and try to get the.com you don't. Yeah. Definitely enough. Who would have. .com. What's it? You know what? If you got a.com these days, people would say you were unoriginal. Are .com. Don't they? Don't they come in choogy? No. I don't think it looks. I think it looks stodgy. I think it looks like, uh, from the year 2000. Yeah. Like whenever I see crypto.com, like the crypto.com arena down in LA where it's

[00:33:05] sports play. I'm always like, ugh, it feels so nice. Look at you guys. You got .ai. Not the crypto part of that. I got .fm. You got .news. I got .tv. Who's using .com anymore? I don't know. U.S. Steel? I don't think we see enough people going crazy with TLDs now that you can do kind of whatever you want. .pizza? Yeah. .your mom. I mean, like, why don't we have that? Your mom? Google got that one though, I think. Oh, they probably have that one. Yeah. Yeah. They got that one.

[00:33:35] Our show today brought to you by Zip Recruiter. Oh man. I tell you, I, if you're hiring these days, I love you. You're great. And you need help too, right? We're always looking for ways to make our lives easier, whether it's multitasking by eating breakfast on the way to work, uh, hiring a nanny for school pickups, hiring a dog walker or using delivery and subscription services to buy paper towels.

[00:34:04] We all have our own little time-saving hacks. There's no shame in that. We all have our own time-saving hacks for everyday life, but if you're a business owner, how can you save time hiring? Zip Recruiter. Zip Recruiter has a new feature that quickly lets you see the most interested qualified candidates first. So you save time by meeting the right people faster. And right now you can try it for free at

[00:34:29] ZipRecruiter.com slash twit. Zip Recruiter's powerful matching technology connects you with qualified candidates instantly, but even more than qualified candidates who really want your job. And isn't that who you want to hire? With Zip Recruiter candidates can tell you in their own words why they're interested in your job. No wonder Zip Recruiter is the number one rated hiring site

[00:34:54] based on G2. You'll save time and you'll meet great candidates sooner with Zip Recruiter and four out of five employers who post on Zip Recruiter get a quality candidate within the first day. Try it for free. ZipRecruiter.com slash twit. That's ZipRecruiter.com slash twit. Meet your match on Zip Recruiter. Let me thank you so much for supporting. Does Zip Recruiter have a little tender interfaces where you can just swipe left and right on candidates? Oh, wouldn't that be

[00:35:22] good? I don't know. That would be cool, wouldn't it? Yeah, it would. Yeah. Get at me, Zip Recruiter. Five dollars. You can get an idea. No, no, no. You know what? That's fine for dating. But when you're hiring, it's a little more, a little more serious. Actually, I hear from the kids. It doesn't work for dating either. So there you go. Oh yeah. Well, I don't know. I wouldn't know about that. Thank God. You know, they didn't have apps when I was a young man. You have to go outside and talk to people. We had bars. Oh! That's where you went. We went to the disco to meet young ladies. The disco

[00:35:52] were called third places. Or the roller rink. Always fun, especially when the gals skate. That was great. So this was the week. Let's get back to the future. This was the week. First of all, there's a very funny game, which I showed on Wednesday called Super Dario. It's a platformer like Super Mario, but you're Dario Amode running along. And as you run along, it says,

[00:36:20] we've extended it for another day. We've extended it for another week. We've extended for another two weeks. They keep extending fable again and again and again. And I've been like probably a lot of people because of the rug pull last month, you know, I'm going to lose fable. I'm going to lose fable. So I've been hustling along. We have a big project rewriting our old ad sales system. And, you know, I want to use the best model. So I've been rushing and rushing.

[00:36:45] I don't know what I was rushing for. This actually, it was yesterday. It was on the weekend. I think Claude, Anthropoc announced getting July 20th. Claude Fable will be included in your subscription. I guess, presume forever. Yeah. Now it's 50% of the limits. So you can't use it as much, but I haven't run up. I mean, maybe I'm not doing enough sub agents, maybe because I'm being

[00:37:15] proactive, as you suggested, using lower quality to do some of the hard, the tool stuff. But I haven't run into limits. And I'm doing a pretty big project. I'm using it all day, every day. Are you using fable primarily, or are you dropping down to- Fable to plan. And then what I, so what I, when I started to plan, I said, because you're going to run out, this was three weeks ago, because you're going to run out. I want you to plan this whole thing

[00:37:42] and then divide it into manageable chunks, which we're going to hand off to four, eight. So you're going to plan it. You're going to hand it off to four, eight. You're going to review four, eight's code. And then you're, we're going to move on. And then I, I added one, Saul came out, chapter GPT 5.6 and five, six is going to review it. And then once I got Grok four, five, I said, and Grok four, five is going to review it. And so I feel like with those triple reviews, the, the, the coding is being

[00:38:12] done by four, eight and some in with sub agents, but because it's a lower cost, it's fine. Sonnet five is good. So yeah, that seemed, I haven't run up against limits. So anyway, I feel like, you know, I have a Claude Mac subscription, but I feel like this is a, this means we're going to continue to have it. Why do you think this happened? They say demand for fable has been challenging to predict, which is why we rolled it out to subscription plans and stages, extending access

[00:38:38] several times as we secured additional capacity. Some of that coming, I think from XAI for a billion dollars a month, but does this tell you that maybe they thought there'd be more demand for fable or they were concerned about it? I think we also have just forgot that the U S government blocked it for two weeks. Yep. Do we all just, just, just forget that.

[00:39:03] I didn't forget it. I was peeved. Harper Harper. Are you saying that because they've locked it for a few weeks, people kind of like came up with other systems, other approaches, and there's a reduced there. So, so I actually think there's a lot of nuanced things going on. So I think the first thing that is happening is that, um, these things are, our tokens are oil. Yep. Anthropic wants us

[00:39:29] to think that they are the engine. They want us to think that if you use Claude, it's Claude and it's different than open AI and Kimi and GLM and all these things. Um, what we have found, and I think everyone kind of has a similar experience is like a lot of people prefer the vibe of Claude, which I think is related to the prompting, um, the system prompts and whatnot. And they think opening AI is kind of boring, but opening AI is just as effective. The models are very good

[00:39:58] as you know, like soul is incredible. It's, it's fable level, especially around, uh, execution and programming. Um, and so that's, that's great, but you can put any of these models into any or into any harness that you have, and they will probably be equivalent and you might have to tweak it over here because this model has these idiosyncrasies or this one does this. Um, I personally like anthropic models, but I have had just as good a luck with GLM on a few things as I've had with

[00:40:23] anthropic and the vibe seems similar. And so I think what, what is happening is, um, when anthropic and the U S government nuked themselves, nuked fable, I think people were just like, well, that sucks. Well, let me go experiment more. And what it did is it broke the trust. And so it allowed people to go experiment more with different harnesses with different, you know, things and, and, um, open AI didn't help by releasing soul. And then very specifically some of their, um, you know,

[00:40:50] thought leaders at open AI tweeted about how to use soul inside of cloud code. And so all the people who, who like cloud code were able to go and use soul inside of cloud code, which is great. Like, that's a great thing you people should experiment with. But once you do that once, that's another one of these, you know, bridges that is going to be hard. Like they burned that bridge. You're never gonna have that trust. And again, is because of the federal government, because of the rug pull, they turned the classifiers up way up. And I got it. That thing

[00:41:18] was annoying. I'm hearing people say, I can't get fable to do anything. Yeah. Yeah. This is, this is the protective layer because I think it's pretty clear. Fable is mythos it's mythos with a lot of, with a classifier running on top of it that says the minute you try to do something having to do cyber security or biohacking or, or, or AI, it says, no, no, no, no, no. Uh, or it will

[00:41:45] turn it down to four, eight, Opus four, eight. And people are complaining that I, they can't even do simple, normal stuff. The classifier is so aggressive. So that's another reason maybe. But given everything you just said and everything Harper just said, think about what Anthropix said in their missive beginning July 20th, clock table five, including all max plans and team premium plans at 50% of limits. So for certain plans, 50% for everyone else, one time, $100 credit to your

[00:42:12] account. That is not a company that is swimming in spare compute because I think they would love to serve a lot of inference at either the subsidized rate as Harper mentioned, or at the headline rate for this, which is, you know, uh, 10 bucks per million input tokens and 50 bucks per million output tokens. They would love to do that, but they're still having restrictions in place. So to me, they're not compute abundant yet. And I don't think they will be for a long time, but as they do this stuff. But they're more than they thought they'd be because they were, they did act as if it was going to be API

[00:42:38] tokens only, right? Sure. But we don't know if they found some ways to serve this inference more efficiently. For example, we don't know if Amazon just turn on another, I don't know, 10,000. They're buying compute from XAI at a billion dollars a month. I mean, they have, and they're buying it from alphabet and Amazon and Microsoft and they're building their own data centers and they also, we don't know, and they're in the cloud. It's such a black box. You're saying we don't know. We don't know. We don't know anything. We don't know what it costs them. We don't know if

[00:43:03] they're losing money, if they're making money. We don't know anything, do we? And because they're not public yet. I saw a good tweet where someone said, um, two things can be true, which is they can be absolutely crushed under the operating expenses. And they also can be swimming in money from all this revenue. Well, they have like the fact that they can't light up enough data centers. They can't do all this stuff

[00:43:28] because of the costs, et cetera. Um, and also the fact that everyone is paying them billions of dollars to do. And they have, uh, you know, ample runway because of the investment that's been poured into them. Oh, they have tons of capital. This whole, there was that meme for like a few weeks, a few months ago, but people like, oh, open AI is going to run out of money. And I just wanted to be like, guys, they just raised $122 billion in a single round, the largest raise in history, in history. Uh, and probably not going to run out of money ever because of what, well, maybe if they lose the

[00:43:58] edge and they get mocked by Chinese open models and so forth, but like the way things are going today, they're fine. But I do think Leo hits on a very important point, which is it would be nice to have a better access to the numbers to be able to understand how healthy this company is or not. But if you go back to earlier in the conversation, we were talking about not telling people how much value you're getting out of AI. So you don't want to tip them off. That is ironically the Google IPO problem. They didn't want to go public because they didn't want everyone to have money. They're

[00:44:24] making on search. They had to, because they were legally required to, because they had so many shareholders back then there were rules. Now we don't have that, but ever since the jobs act that's gone. Well, kind of there's different rules, but once anthropic goes public, if they are minting money as much as they think they are, then people are going to get pretty peevish about paying 10 bucks for million input, input tokens, because that's a little bit excessive compared to open source or an open weight model. So I think right now they're probably sitting pretty, pretty, a lot of demand, not enough supply money's coming in.

[00:44:54] David Pérez, I think both anthropic and open AI have filed for IPOs, but we don't, but they're, they're holding onto it. And I think the Trump rug pull gave both of them cover to hold on right to the IPO. Cause I know you're right. I don't think they really are too anxious to reveal any of these numbers. No, no. Do the dirty work before you list. And I think right now they're still in the dirty work era of building out their, their compute footprint and so forth. And I don't think

[00:45:22] all the deals they've done are going to be simple to explain. And I don't think that the cost of their employee compensation is going to make their gap numbers look particularly salubrious, but that does not mean that the core business of making and selling intelligence isn't actually pretty sick. Um, I'm dirty AI done dirt cheap. Is that what you're saying? Indeed, sir. Sing it for us, Leo. So it also got very complicated because all of a sudden China who also has very mixed motives

[00:45:50] is now starting to release open weight, very credible AIs to started with Kimmy this week, moonshot AI. They say it will be open weight on the 27th. It's not open weight yet, but they showed benchmarks and a number of people have confirmed this to say it is almost not quite, but almost fable quality. Uh, they offer a subscription as well as API tokens. They're not all, which is sold

[00:46:17] out. Yeah. They sold out of the subscription, which I think is, um, you know, back in the early aughts when I was involved with t-shirts, um, we would freeze, we would sell out a thing specifically to get the hype going for when we would launch it. Right. So you do a small run and then do a big run. And that was such a great way because people would be like, oh man. And they could, they buy something else. They do something else. And it's, I mean, this is famous. They, they are doing a very good job

[00:46:43] at comms and growth and marketing. And I, I love it because the models are probably also very, very good. Um, and well, and then Alibaba said, hold my Baba. Yeah. And said, oh, by the way, Quinn 3.8 max is coming out. It's been launched in preview to those same people who got access to the other ones before we did. Uh, that was today. Oh yeah. This is an exciting weekend and I am

[00:47:13] behind. I was like, oh, I'll check out Kimmy later. And I just can't, it's not available. That was Quinn 3.8. You know, Quinn is very good. I have to say I've been using 3.7. Uh, I can run that locally. Uh, the 27B and 35B and it runs nicely. Did you play with GLM? I, uh, 5.2. 5.2 is I don't run it locally. Cause I don't have, you don't have, you don't have eight B200 sitting behind that PDP 11, but I do run it. It's my Hermes default. I love GLM. It's very good

[00:47:42] at language. Yeah. Uh, I find, I haven't tried it so much with coding. Uh, it's good. It's great. I still have in my mind that Anthropic is the coding king, but I, maybe I should let go of some of that. Uh, well, that's, that's the risk that we're talking about really, but I just want to go back to the Harper point about how to build hype by restricting access to your product. Um, Zipu AI, Zeta AI, the makers of the GLM family models did this earlier this year. They restricted new signups

[00:48:09] to their coding planks. They didn't have capacity for it. So this is not even a new approach by a Chinese lab to gen up interest in the American media. And it works every time I have a ZAI, a subscription for their max, their top of the line. It's a third, what Anthropics, uh, Claude max is it's like 75 bucks a month. And I use it like it's my, Oh, the Z.ai. Yeah. So love it. Here's, here's something I like to think about. Um, I was using Claude code in,

[00:48:37] was it when it February, 2025 just had launched early days and it was great. And we were building all sorts of cool stuff and our, our processes were horrible. Um, I think GLM is way better than the anthropic models that were available then. So this is the, now the question, how close are the open weight models to the frontier models? Well, I think it's not a question of how close,

[00:49:04] actually. I think it's a question of time because it doesn't matter if they're equivalent. I like how you were like, Alex, he's going to say it. And then I said it and you were like, well, but, um, no one else, no one else can see me. So I got to, yeah, I was excited about that. I was just like, I was like seeing your head just collapsing on itself as you didn't expect what I was going to say, but, um, you let me down. I thought you were going to say no, but I think time is the key. Like it is how long ago would you have an equivalent? Like

[00:49:32] would the, the, the, the six months, eight months, four months, how ahead are the foundational models? Right. And I think it's getting shorter and shorter, but the, but my point is less that does that matter? My point is six months ago I was using cloud code and it was awesome. So if GLM is six months behind and it's effectively free, then like what, this is like a, this is like a present. You know what is effectively free in his very closest deep seek. The Chinese model that blew

[00:49:59] us all away in February of 2025 that changed the whole world. It was an open weight model and it was, and now they're really cheap. It's like 12 cents per million tokens in deep sink. V4 is incredibly, incredibly inexpensive. I use that as well, but I have to say GLM has kind of eclipsed that for me. But the point is all three of these are from China and the Chinese government

[00:50:21] government has made noises that they may want to stop this. They may want to embargo open weight models from the U S just as the U S government is making noise. They want to block Chinese models from the U S this is partly protectionism, just like they block Chinese EVs from the U S to protect our auto industry, partly maybe security because let's not forget, you know, Harper, you're the one that made

[00:50:51] me buy this Chinese. Where is it? It's pretty good. Isn't it escaped? It's escaped. Now it's a little ball that you connect up to your wifi and it immediately calls China and it's deep seek. And yeah, it's really cool. And it's not horrible. No, but am I exfiltrating? Look, we just saw that

[00:51:15] Grok is taking everything it can get. Yeah. Elon says this is a bug and they're going to fix it. But the, but the Grok command line interface was apparently downloading entire directories of your, of your hard drive when you, before we get to that, I want to go back to the Harper point. I want to double click on this because I thought you were going to say that as we watch open weight models become

[00:51:39] closer and faster followers to American closed source models, that there's going to come a point in which they are sufficiently intelligent for whatever tasks you need to put against them. And therefore the pricing power of American clothes models will collapse. And that is going to be a really tough moment for these companies. But right now today, to leave us point, we do have deep seek for, and you know what people use it and you know what, they still also pay through the nose for Fable.

[00:52:05] So to me, we're not there yet, but to your point about using a old cloud code versus present day GLM. So I mean, we can't be more than a couple of years out from when the, when the, the average model is sufficient for all of my Alex Wilhelm's personal needs and all of your. Well, yes, GLM is fine for my agent. That's what I use Hermes with. And it's great. It's all I need. It's wonderful. Okay. But I would guess that we're there already. Yeah. That for the most part, the.

[00:52:34] Unless you're coding, unless you're doing hardcore coding. I think that if you're coding, you can get pretty far. Yeah. Now that doesn't mean that you can't get better, right? Like I think there's a difference between could this do everything I need it to do? And is this the best model in the same way that you have cars, right? Like buying a, a Toyota Corolla is a completely rational car. I would probably prefer another car, but that, but like it does do the thing.

[00:53:00] And so I think that if you are, if you have a GPU that can run, like you have, you know, we have a spark, right? We have one of these sparks and it can has 128 gig of RAM. We can run some of these bigger models slow, just so slow, but you run it and it works great. And like, for the most part that could power my local agent, my Hermes agent or whatever, they get power, probably do some coding with pie and all this stuff, but for free or just power of electricity. And so I think this is accessible.

[00:53:27] I do think though, that when you really want it to work, you want to use something like a GLM or probably these Kimi K3 or, um, you know, the new Quinn, and those are going to take a lot of GPUs. Like I don't have eight B 200s. Like, I don't like, I don't even know how to get those. Like, I don't know how that works. Um, and so like, I don't think that is a reasonable thing. So I think we have eclipsed the, you can run it locally for free. And of course these will be, you know,

[00:53:54] they will be quantized, et cetera, to get them a little smaller. But like the magic here is that you can get something that is like fable or like Opus 4, H for, you know, max dollar 40 an hour or tokens, a million tokens or whatever it is. Like there's all these hosting ways. Um, a friend of mine was, was renting on vast AI, a eight B 200 box, which were mounted to basically $50 an hour.

[00:54:21] And then giving all of his friends free inference. And so you could just, but you could do whatever you wanted, right? Because there is no limits. So, and, and there's, and then he did some really cool stuff with scheduling and all sorts of stuff. And it worked out really, really well. It was really fun because we rarely have opportunities to just rip through as many tokens as you want. Right. And he was doing testing because this is ostensibly a startup and he's trying to figure out if he could, if this is something sustainable that he could give to companies. And it actually got me thinking much differently about how tokens are given out. Whereas with a fable or whatever,

[00:54:50] you have all these very limited things, but in what situations would you want unlimited tokens? Like a WeWork of the future might say, oh, and we have a GLM instance that is unmetered because you remember when the data centers did this, right? Like you would have like Exodus to try and get you in and they'd be like one year of unmetered, you know, egress or whatever back in 2001. And like, like that kind of thing, I think will start to happen as this, as this commodifies further.

[00:55:17] The children forget that in the early days of AWS, back when Azure and AWS were fighting it out for market share, they actually cut prices occasionally. It was a thing. Microsoft said, we will never have prices higher than AWS. And everyone went, yay. And then now everyone's broke because they're running the cloud. On the point about price, what's very interesting to me about the latest crop of models that we're talking about, especially KMAK3, is that it is not cheap in its current form. It's roughly priced along with GPT 5.6 Terra, I believe kind of the middle size model.

[00:55:46] So, but it's a lot less in Fable and it's nearly as good as Fable is the pitch, right? But it's okay. I don't think it's as close to Fable as you guys do. I don't know. I don't know. They even said in the KMAK3 blog post at the very bottom, it says like, this is not as good as Fable. They said it's second to Fable, better than Saul, but second to Fable is their pitch. At some point we're going to end up discussing which Metallica album is the best. But this is, by the way, benchmarks are BS, right? Benchmarks are not usable anymore, I think.

[00:56:15] Anyway, go ahead. Make your point. Oh, no. I'm just saying that right now we are seeing a KMAK3 come out. It is a big model. It is not cheap to run. It is currently not open and it smells a lot like an American AI model to me. Now, they will open weight it. Sure. It will get distilled, condensed, improved, tweaked, and will get cheaper. But right now what they have built is something that if it came out from open AI, you'd be like, cool. Yeah, you're right. I think this is not the super duper cheat. If it does become open weight, you still

[00:56:44] wouldn't be able to run it because I think it's close to three trillion parameters. So you're just like 2.4 or something like that. Yeah, it's not going to fit on your... So let me read you what Dean Ball says. Dean Ball is a conservative from the Hoover Institute, you know, Heritage Society, you know, that lineage. He's now head of strategy at OpenAI as of two weeks ago. Head of strategic futures is his title. He wrote this and I thought it was very provocative. Do you know, are you familiar with this, Alex? Have you seen his tweet? Yeah, yeah. I've read through

[00:57:14] this. Some observations on Kimi. It is a very good model. I don't think its performance can be explained away by distillation, which is what some like anthropic claim. It's in a genetic coding sessions. This seems pretty much on par with the best public models of the first quarter of this year. I'm personally surprised. This is where he gets to the economic argument, which is very interesting. I'm personally surprised that the Chinese state continues to allow the open sourcing of models

[00:57:38] is good. Given the risks, he says, to be clear, I myself might be fine with models presenting this level of marginal risk being open weight, but I'm surprised that China's fine with it. The presumption, by the way, when you say open weight is that people will... It doesn't have to be run in China. It could be run on American servers, you know, that WeWork model. Or if you had 20 B100s,

[00:58:03] you could run it at home. So it's basically China's letting it out of their control. I suspect the reason they are is 75% explained by strategic blindness, that the Chinese Communist Party isn't like thinking along the lines of American frontier model companies like Dario. This is the... I think this is the fatal flaw when anyone talks about China. They don't understand it.

[00:58:30] Being like, well, they're just so stupid. They don't get it. And you're just like... I don't think that's the case. Like anytime you read like Apple in China, Chip Wars, all these books where like this incredible book telling about how China is doing all this cool stuff. And the last chapter is like, well, they just won't make it. You're just kind of like, how is that? I'm more of an opinion that it's for them, it's soft power, you know? Yeah, 100%. The reason they're making their own chips, because they don't want CUDA and NVIDIA to own the world. They want to have

[00:58:57] their own. He says the other 25% or so is their lack of compute for customer inference. Since they can't run it locally, maybe they don't have enough compute to do that. He says also the normal Chinese strategy of aggressive exports. And for the companies, as opposed to the government, the decision to open source is partly ideological, partly because they're behind and because they know, as Alex was just saying, very few people would pay for sub frontier models from China, right? It's not a competitive.

[00:59:24] So this is the thing that he said that's very provocative and scary. Yeah, this is the points three or four when it gets spicy. Yeah. Open weight models are inherently decelerationist. He says that because they basically make it impossible to raise money for frontier models. If somebody is giving it away, go ahead. Oh, I just, this is what I, where I wanted to jump in and kind of explain this point, but I realized

[00:59:52] that you were about to do that. And I mean, a terrible guest. So I'm going to shut up and let you finish. He says in the end, open weight models deter AI capex, right? Why would you invest? Why would you invest billions in AGI from Anthropic if China's going to give it away? He says, this is what's really scary. You're now we're to point four that you thought was equally scary. One, one probable outcome of an

[01:00:18] open weight model dominant world, which by the way, is a world I've been advocating for. Wednesday, we had the CTO of Mozilla on, he just put out his, the state of open AI, state of open source AI. Uh, we believe just like open source software. This is, this is great. This is revolutionary. Yep. He says, no, it's AI communism. Now, remember he comes from the heritage society and the

[01:00:44] Hoover Institute. And it's precisely what China proposes rather than a market product. AI is a public good, which will ultimately be provided by the state as a kind of digital public infrastructure. This future strikes me as a dystopian hellscape. I can't argue against that point. He makes a good point. So I think there's two points there, right? One is it's communism. Great. Who cares? The other

[01:01:12] one is that this might cause issues for, I mean, it's con it's a communist country. What would you expect them to do? They literally, you can't own property. Like he's like, this might be 99 year lease from China. They have introduced an idea of communism. And it's like, well, yes, that's true. Like, yeah, I don't know. Read the, read the house of highway and like, tell me about communism. So like, like the, the thing I think is, is interesting and actually compelling is there's one thing very

[01:01:39] specifically that why would you invest in a foundational model that has a closed model if you can get an equivalent thing for free. And we all know a, that that has not been true historically. Not with open source software. There's lots of companies that have made open source software that is great. And then competing closed source software that is able to raise more money than the open source version. So I feel like, okay, but, and, but the thing that I think is interesting, and I don't necessarily

[01:02:06] disagree is what if that is strategic, right? Like what if there is a, where if there's a strategy there? Yeah. And I think the, the, there's another thing that I think is going on, which I think is, is similar, which is the United States for better or for worse has demonstrated to the world that we

[01:02:30] are a different country than we were 10 years prior. Especially for our partners and our friends and, and our frenemies. We've just demonstrated that in some cases we're more reliable and others were way less reliable, but this has changed how people interact with us. Um, and, uh, there is a hole, there's a, there's a vacuum where we sat before. And one of the things that China has taken advantage of is just, public investment. And so it's not so public in China. It's not the same kind of thing,

[01:02:59] but they have invested a lot in infrastructure and roads, et cetera, all over the world. And why is AI or data infrastructure any different? I would posit actually that we are missing the boat by not doing open models. Yeah. Because if you look at a place like the Middle East, for instance, Dubai or what have you, they have huge data centers. They've been building a lot of data centers. They've been building AI cities. Whose models are they going to choose? Are they going to choose the closed models that they don't have access to? Are they going to choose the open models that are freely

[01:03:27] given that they can then modify to their own values and their own culture? It's a belt and road initiative for AI. Yes. Basically. And then once these open way models are selected and American AI companies can't sling their same models at the same margins, as we discussed earlier, what happens to their business? Well, it becomes a much less healthy, it becomes less investable, less backable. And you could see the closed source American AI labs that have so far led the entire AI race lose this. And this is why I think we see the heritage

[01:03:55] foundation guy going, ah, it's communism. Right. Because it is not great to see capitalism pouring private investment into these models to create something that is valuable today, only to see China, you know, essentially undercut them with zero cost products. Here's my question. How can you keep funding these Chinese AI labs, which currently are deeply unprofitable? And we know that because ZAI and Minimax are both public companies. We've seen their books. I have read through them myself. How do you build businesses of sufficient

[01:04:25] health to keep this up now? Well, but if you're the government supporting it, Ah, well, you're forgetting, but it's communism. Well, like, I think there is a real, there is a real reality here that arms of the Chinese communist party. I don't think I wouldn't even go that far. I would say that there's a different value system. That may be it. And you have state capitalism, which is not our capitalism. I would call our cast as a capitalism, fast food capitalism. And so state

[01:04:51] capitalism gets a longer investment horizon and more support from the government, from the state. This is true. Like this is how it is. This is why there's so much solar in China. Like there's a really good book actually called breakneck by Dan Wong, which I highly, highly recommend super, super book. And it just posits this very simple fact. Um, the political role of China are scientists and technology people, the, the business or the leaders of the United States are lawyers. So when you look at what tools you have in the United States, our, our tools are

[01:05:20] regulation and laws in China, their, their tools are science and technology, and that will change in the book. It does a really good job of kind of portraying this about how this is, this is a reaction to industrialization. Um, and so like, this is how we got here. And so I think what, what we have to kind of figure out is where do we as a country want to go? And right now the, the answer that we have is no, they're just like, where do you want to go? And everyone's just like, no, we don't want to go anywhere. Do you, would you want a United States federal government

[01:05:50] subsidized AI run and, and, but I think you're missing the point here. I think you're missing the point. We've done this before after world war II with all of the science labs. This is not like we have to reinvent some magical thing and follow China into communism. It's this, we have to invest in science. That's basically what it is. Like that's all that is. This is not even, we have to invest in science and get rid of capitalism. It's like literally invest in science and like things like Silicon Valley will be invented, um, which is what happened. Vannevere

[01:06:17] Bush, et cetera. And so this is why every four or five years, there's some Senator, you know, proposals, some kind of idea that we would re-instantiate, you know, federal labs, um, and all these labs, which then spun out all this unnamed technology and did all that stuff. But the thing is, is it's, they're still around actually. But the problem is, is that it doesn't make financial dislike sense if you're in a fast food capitalism type of world. Right.

[01:06:45] And if you're waiting for, if you have a quarter by quarter requirement to be profitable, then you cannot do this type of innovation. Right. I just, I should point out also that Dean W. Ball is now head of strategy at open AI. So this is maybe the point of view that open AI might reasonably have, by the way, the final, let me give you the final scare point. And then you, then you can go, Alex. It's probably true. He writes that open weight models of this capability make the world a bit more

[01:07:14] dangerous, but not so much more that you'd really notice at some point, the models will be capable enough. You will notice quote, a non-living invisible, dangerous, and infinitely self-replicating agent escape from a Chinese lab. You say color me shocked. And, and, and somebody did point out that one of the advantages that they have is the Chinese don't really care about intellectual property rights. There's no debate over whether it's okay to copy a million, um, pirated books

[01:07:43] into our database. They don't, one of the pieces I think, I think one of the real reasons people are liking Kimi compared to fable is it doesn't have the restrictions. People are saying, yeah, I asked how to make an anthrax bomb and it told me. But also we didn't care about copying any of the books either. We had no interest in that. Like I think there's a lot of, oh, well, do we, do we Leo? Like did we stop training the models on the books? No. Okay. So I don't think we actually cared. I think we paid

[01:08:08] and we paid at all. We don't care at all. So I feel like this is all fake. This is all fake anti-China BS. But I think the problem with, with the fake anti-China BS is it actually papers over some really important things that we as a country need to figure out, which is like, how do we compete when there's a game being played that we are not participating in? That's a good. And if we want to participate in that game, then we actually have to, or we want to win that game, which I think is

[01:08:33] the American way, then we need to invest in participating in that game. And we can't just have it be Lama and, you know, Gemma and just call it a day. Okay. That's true. I agree with that. But I will say Nemo Tron three ultra pretty good. Uh, very good. Yeah, yeah, yeah, yeah. And the reflection is NVIDIA's open family of models. And also I keep talking to the people over at reflection AI and they, they, they claim they're doing things. And also we just saw thinking

[01:08:59] machines, labs drop, um, inkling. There we go. Had to reach for that one. Sorry. I believe it is open away and it's designed to be used in their tinker system to allow for fine tuning and post training. So we do see sparks of open way AI. And I use Hermes from noose research. They have open models as well. Got out news research. And we also in the West broadly have Mistral, which is a company that I think has been very much under repped. And we have a couple of new labs working on a non transformer based AI models and tinkering around with world models

[01:09:29] and so forth. So America is not doing nothing. The race isn't over space. No, not at all. And I don't think that we're actually as, as far behind as people might think. And I will all, I, I, I, I don't know how to get enough compute built without having a profit element of this be at the forefront today, given the global race for computers. And I, that comes from the notion at open AI and

[01:09:54] anthropic that they are building AGI. Tell me why, because the, the hyperscalers believe the only route to AGI is more parameters is 10 billion, 20 billion, 30 billion parameters. At some point, we're going to get so many parameters. We will have AGI and whether that's true or not, that may be a fool's errand and it may be smarter to build smaller models with more limited capabilities.

[01:10:25] Okay. But, but it doesn't, it doesn't advance you towards an artificial intelligence, you know, I've spoken to a number of the SLM companies, the small language model companies, and really what's really interesting come down to where they're kind of like micro agents that do one specific thing, like your accounting software, very useful, very targeted, but it's a wrench, not a workshop. And so to me, they're kind of just different conversations, Leo. And I don't think, I don't think that's the way to go about it, but also on the point about you only need this much compute for AGI. I don't

[01:10:52] agree because we just spent 15 minutes talking very positively about two of the largest models ever released in the world, Kimmy K3. And I think then the next Quinn model is gonna be huge as well. So if we're going to sit here and say that China is eating our lunch because they're releasing these open-way models that happen to be enormous. I don't think the solution is we need to have open-way models in the United States that are small. How big is, do we know how big Saul is it? People think that Fable is 10 billion parameters, 10 trillion. I'm sorry, parameters. Do we know how big,

[01:11:22] Saul is? I don't think we do. They don't. And these companies don't tell us. This is speculation. Wouldn't it be great if they did though? Yeah. Where's Sam Altman? Come on, Sammy boy. Tell us some stuff. I'm asking Saul five, six. The thing that was interesting about Deep Seek is that they couldn't build these giant models. Frankly, the thing is interesting. A lot of the Chinese stuff is they don't have access to CUDA,

[01:11:48] to Nvidia chips. They're, uh, you know, uh, Kimmy was done entirely with Huawei chips, a hundred thousand Huawei chips, but Huawei chips in the US. And, uh, so their, their risk, their constraints have seemed to be a benefit. Look what, uh, Deep Seek did a year and a half ago. It opened our eyes to the idea of reinforcement learning. Sure. And since then, what has happened with the two of these AI labs? No, I mean,

[01:12:15] they've scaled from like anthropics. It was like a billion when that happened. And now last we heard it was 67 billion. So deep seek, the deep seek moment, the terrifying moment that shook the cage. It didn't shut it down. Uh, was one of the moments in which we saw anthropic accelerate to the moon or as open AI, I would say to the Luna. Uh, AI model family jokes come to twit for the dad. Uh, let's come back down to Tara and, uh, do a commercial. And then we will continue. You're watching this week in tech. I didn't intend to make

[01:12:43] it an AI show, but it was the big news of the week. I have to say, uh, we will talk more about other things and just a little bit with Alex Wilhelm, uh, from this week in startups twist, but, uh, his newsletter is cautious optimism dot news and the wonderful, wonderful person who lives in my childhood home. And the other wonderful person who lives in Chicago, Illinois, Harper Reed. It's me 2389.ai. You've been in Japan for ages.

[01:13:13] No, I was just there for like 12 days. And then I had travel all over. And then we did an event in Montreal where we had a bunch of people come over and do what we call the hack party. Um, and, uh, and yeah, that was, it was, it's been, um, it's been a frenetic few months, basically since I last saw you was just like constant travel, which has been really good because it's, it's always entertaining to go other places. Our show this week brought to you by ethos. Well, now that both of

[01:13:40] you have small children, I'm going to think back, my kids are no longer small. Henry's 32, Abby's 34, but I remember very well the six month and the five months and all of that stuff. Uh, and I also remember very well, the responsibility I felt when those kids were born. That's when I ran out for the first time in my life. I think I was 35 and got life insurance. Oh yeah. Right. You start to kind of think, Hey, you know what? Uh, it's not just me anymore.

[01:14:08] So people are counting on me. Ethos is a chance for you to take that kind of responsibility, uh, for your family. Ethos makes getting life insurance fast and easy, 100% online. And you can get a quote in seconds, apply in minutes and get same day coverage. There's no medical exam. You just answer a few simple health questions. You can get up to $3 million in coverage.

[01:14:33] Some policies are as low as $30 a month, and you'll get your lowest rate from their network of trusted carriers. Take 10 minutes to get covered today with life insurance through ethos. Get your free quote at ethos.com slash twit. That's E-T-H-O-S dot com slash twit. Application times may vary, rates may vary, but I remember very well that moment where you go, Oh, I think maybe I better,

[01:15:01] better do something to take care of the fam. You just had another, a third child, Alex. That's so great. Congratulations. Thank you. Yes. Uh, it has contributed greatly to my sanity and to my health. You're lucky though, because you have a man cave that is separate from the house. You can go out there, right? But, but like in the fifties, yes. If I was a dad in the fifties, I would go out here and smoke my pipe and follow you right out there. Well, no, I don't even get to

[01:15:29] come out here. Like, like, okay. So people who aren't parents don't understand this, but when you have your first child, you still think in terms of like what you want to do on the weekend. Oh yeah. That doesn't last. You might think, you know, I really want to go, go see that movie or I want to go, I want to go for a long bike ride. Your second child really turns that into a thin layer at the bottom of your life. And your third child is when you realize that weekends are now much harder than work days. Like now every single Sunday night, I'm like, oh,

[01:15:56] thank God we're going back to work. The nanny's coming. Yes. But you're doing something I never did. I don't know if I could have done it. You're working at home. I got to get in a car and drive and drive away. My wife does this every day. And I'm always like, bye-bye. But the nice thing is I do go into the house to get a drink, to get a, to get a light, to use the toilet, whatever. And I do get to pop in and see the kids. And they're always so excited to see me because I've been gone for three hours and they're like, damn it. Damn it. I made a thing.

[01:16:23] And then like your heart grows so big. Yeah. So I wouldn't trade it. And then you leave again. Yeah. And then I go back and they cry and they wail. And I remember, I remember I had an off the home office and they would, they would know I'm in there and they would pound the door and say, let me in, let me in. So you can't get away. Leo, you know what? You know, in my man shed that there's a little porch, well, we've put one of the children's swings there. And so often I'll be like working and then I'll look over and there'll be a child doing this outside the door. And I'm like,

[01:16:54] they better not, I'm recording right now. They better not knock. They better not knock. And it's okay if they do, I'm just telling you, it's quite all right. I had a swing on that tree right behind you. That big tall, is that still there? The big tall. Oh yeah. So it's right there. Yeah. Yeah. I had a swing right on that. You should come back and see what we've done with the place. There's been many, many additions and improvements. No, but I think that children have changed me in one more way that I'll bring up,

[01:17:20] which is it's taught me how I've already lost this battle. So I'm not going to try to be a moral leader here, but like phones are so addictive and I don't think I realized. What are you going to do? Have you decided what you're going to do? I'm going to hold out until I am the last parent in their friend group that lets them have a phone and call that enough. Yeah. I, I, I, I'm so glad I didn't have to deal with that. I don't know what I would do. Harper. What are you going to do with your kids when they get to that age?

[01:17:45] I mean, I've always believed abstinence is a really good program for everybody else. Um, I just don't think it makes sense. I mean, I think it's like, you're just going to try and be a good parent. You know what I mean? Like, like you gotta, you gotta do what you can. I'm not going to die on any Hills. I feel like most of the Hills you're just like, you're and the, and I think the thing that is complicated, it's very hard to me for me to be like, you should not

[01:18:14] be using technology. Well, then I am using technology 25 hours a day. And so like what, what instead is like, how do you have, uh, filtered and metered access? Um, you know, and I, I think that's, that's the key also sometimes you have to fly in a plane and like the no screen time rules when you have to fly in a champagne plane, just like, yeah, I'm not a monster. Yeah. No, no, no, no, no. People still tweet things like I was at the airport at 9am and I saw

[01:18:42] someone drinking, there's no time zone in the airport and there are no screen rules for children on the plane. Oh, there you go. Like, yeah. Clever. Also, it's fun to mix those things. Oh, I, I, I gave up drinking, but yes. Yes. Social media limits coming for teens across Europe, says EU leader, Ursula, perfectly named Ursula von der Leyen. No more for you. I feel like this is a bad idea. This is a bad idea. It seems this is something that like,

[01:19:11] whenever something like this happens, I just look at my security focused friends who are thinking of things that I don't think about and have brains that work differently than my brain. And they're always like, this is horrible. And that's when I'm like, oh, I should pay attention. And then the more you think about the ramifications, the more this just turns into a let's name every person on the internet surveillance bill, which that's what it is. Chat control. Yeah. Which is bad. And that,

[01:19:37] and that is like getting everyone's information into one database has not typically worked out very well for humans. Bad news. So I don't, I worry about this quite a bit. And one of the things I love about the United States is that we're not very creative. And so we'll just be like, we should copy that. Let's just copy that. Let's do a worse version, but we're all, what's going to happen is it's going to be like the GDPR, the California like privacy stuff, but it's going to be like Alabama makes everyone follow this one rule. And then everyone who's building tech has to follow some

[01:20:06] stupid Alabama rule or whatever. We're already there. Yeah. We're seeing at the state level, but also I want to point out that this is not just an EU thing. China this week banned certain AI companions for, for kids and such. So like we're seeing, I think around the world, people trying to be like, no, no, no, no, no, no, no, no. Go outside, meet somebody, go to a bar, try to have a child, please, please stop staying in your house and talking to your chat bot or your social media feed all the time. Because I, I do think that there is a intersection point between and not to get all

[01:20:35] different topic, but like the birth rate crisis and the rise of technology and people staying indoors. I don't see kids out in the street playing very much anymore. Do you see that? And so on the, nobody has any, no one has any money. It could be a relationship with nobody has any money. Well, it doesn't take money to go outside and ride your bike. It does. I meant the, I meant the kids in the safe neighborhood and like having more kids and the birth rate and all this stuff. Like I know a lot of, Oh yeah. Nobody can afford kids. Yeah.

[01:21:04] Yeah. And everyone's just like, why would I do any of that? Because I don't have any money and B we're all, we're all going to die from like Elon Musk or whatever in five years. So, you know, so I feel like there's a lot of like, a lot of things that go into this, that I think we are, it's easy to say, Oh, this is the reason why, but it's like a very big system. Yes. And like, you know, but this is why governments are taking it, taking it on. Right. So currently in

[01:21:30] the U S states that have enacted laws directly limiting minor social media access or requiring parental consent include Arkansas, California, Florida, Georgia, Louisiana, Mississippi, but it's done the list. So we've done this in half the U S right. And we're doing it in Australia and we're doing it in Europe and we're doing it in China to do it. Yeah. The UK, which is almost Europe, but not quite. And I do think that this is going to become a global, a global issue. And I do think Harper's dead on Finland. What we're going to wind up with here is essentially increased government control of the internet, which I am fundamentally opposed to, right?

[01:21:59] But there is an upside, which is a, it'll make more hackers and B it'll be fun to try and get through it. Like it'll, it's a, it's as a hobbyist hacker. It'll be fun to be like, Oh no, the government has access to my internet. Let me fix that. Like, I, I, I also think there's this slightly fringier side, which is you remember when some like, you know, new name service comes on or someone messes with the BGP routes or whatever, some countries

[01:22:27] like we're going to do a thing like Iran has protests or whatever, I guess they don't anymore, but they used to have protests or whatever. And then all the internet would go down for like 40 minutes and then they'd fix it. Like, this is not that simple. Like I can't wait because they're going to break everything. And there's just going to be internet too, finally from the universities. And we're all gonna, we're all gonna hopefully, you know, live in free land. Can I ask you about that? Because one thing that I've read about, and I forget the exact time that I read this and I can't pull up the article in real time, but like I read that China is working to

[01:22:55] export their great firewall technology to other aligned nations with them. And so my question is, what we're talking about here in broadly the West is pretty lightweight. I do agree with you, Harper. You can get around it, blah, blah, blah, blah, blah. Good hackers love it. But what happens if nations begin to get even more heavy handed? Can we end up in a world in which the, the Russian style approach to the internet, the Chinese style approach to the internet, the North Korean style approach, well, maybe not that far, but when that becomes the norm in, in most nations, because I think that

[01:23:23] you probably can't as easily hack your way around that level of state control. And that's what I'm worried about. So this week. Can you not, I mean, it's like, I think you, I don't know. Can the average person, can the child, can the, the, the homebrew hacker to your point from before? I don't know. I mean, do you talk to anyone from like, literally everyone in China knows what a VPN is. True. But how secure are they and how much visibility does the government have? I mean, now we're focusing on perfection, right? Like, I think like the question is,

[01:23:53] can you hack around it? And the answer is yes. Okay. Does it, is it perfect? Like, no, nothing is. I would just say use tail scale. Shout out to our friends at tail scale. Tail scale.com slash. I'm using it right now. Yeah. Yeah. It's incredible. Love the old tail scale. Love it. I, I, like, I, I just think, I think there is this thing that like, we assume there's a whole bunch of stuff that goes into this, like, and I highly recommend if you're interested in what it looks like exporting China's great firewall to read the book,

[01:24:21] I think it's called house of highway. Yeah. Um, which is very, very, very good book. It's very interesting book, but it basically talks about that. Like they did that. They, they exported that technology to lots and lots of different locales. And, and, and in many cases, it wasn't for surveillance. It was for, in case of an emergency, we want to be able to, you know, lock down or control information or what have you, which is not how we in the United States do it, but is how places in the world do it. Um, and, and it's very interesting to see that they just did it.

[01:24:51] Like they just, they, and it's not like a China is going to export that, you know, and I think it's very easy for us to, um, link China with these countries, these companies, um, and say, these are equivalent, but I, but I don't think they are necessarily equivalent. They are linked. So I don't, you know, because of state capitalism, et cetera. Um, but it is a very interesting, it's a definitely interesting book. And I do think we in the U S are not prepared for what it looks like when every single person has to register their name and birthday to get on the internet.

[01:25:21] We're not far off from that. I think. So the EU, uh, this week was unable to stop chat control, uh, which is the warrantless mass scanning of private messages and photos in the EU.

[01:25:36] Um, they, they, they, due to a kind of a little manipulation of the rules, they were able to get that. There were, uh, there was a vote to, uh, reject it, which didn't succeed.

[01:25:51] So now in September, they're going to actually debate the permanent chat control to O but what this essentially does is it forces companies to provide clear text versions of all your private stuff, every photo, every message, every file so that it can be scanned.

[01:26:10] And of course the proximate reason for this is to look for child pornography, but, uh, it has the effect really of break completely breaking privacy. I thought that the, the compromise in this case, and Leo, if I'm wrong, just to please shout at me, but I thought that they were going to not demand that encrypted chat be broken to allow for scanning of that as well. So you still get, uh, the ability to have encrypted chat in the EU. Is that right? Yeah, I'm, I'm, I am.

[01:26:40] I don't know. I'm not clear. Um, the original chat control would break encryption. And that's why Meredith Whitaker of signal said, if that passes, we're getting out of here. We cannot offer signal in the EU because we refuse to, to, to provide a backdoor to signal. Um, now they, they haven't done that. So I imagine that you're right, Alex, that in some way, the current version of chat control doesn't do that, but this fall, they are, there's, this is still on the table. There's, there's still want to do this. I think.

[01:27:11] I worry that safety ism is going to be a real issue. Uh, broadly going back to the self-driving car point and, you know, talking about what's safer and what's not like, I, I think that we should move faster. There, I think we should have less, less safety ism on these cars. Cause they're already, I think better than everyone that I see driving. I see a lot of, a lot of children driving suburbans at like 75 miles an hour down my residential street. It's very scary. Maybe that's not a great idea. If that was a Waymo, if that was a Waymo, a child wouldn't be driving it. Right.

[01:27:40] It would be a robot going exactly 25 miles an hour. Yeah. Thank you. You're driving like a grandma. Yeah. I used to bicycle down that street. I safety pin a towel around my neck and pretend I was Superman. It'd be dangerous to do that now. Dangerous. I tell you, you cannot let your children out of the house and not have them taken by the, by the police. Um, different, different topic entirely. Is that true? That they will say, oh, you're a bad parent. You can't let your kids out of the house. My sister does this in the Bay area. She lives in the South Bay. I'll keep it relatively. There's no free range kids anymore.

[01:28:08] She lets her kids bike to their friends' houses and the other kids' parents are like gobsmacked and terrified. Like, what do you do? You let him skateboard over here? Isn't that sad? Yes. Yes, it is. Yes, it is. Um. It is indeed sad. Yeah. It bums me out because I spent a lot of my youth climbing over barbed wire fences in the country. Yeah. So. Yeah. Yeah. Quite. I walked to school two miles every day. It was great. And I did slowly because I enjoyed it. I had fun. Uphill.

[01:28:38] That's because there was no Netflix back then. It was up Lloyd Street. You know the street. I was going up. This is. It's a hill. It's uphill. I know. I know. I know. Well, it was downhill going to school, but it was uphill coming home. Can we unify some points here? We're talking a lot about American AI regulation, how China is approaching the world, and also this broad global effort to cut back on children's access to social media.

[01:29:03] It just, it seems like everything's kind of pointing towards a world in which centralized powers have a lot more control over our digital lives as that becomes a more important chunk of who we are. And like, I don't know, guys, that has me really worried. I think about this. I think the pendulum will swing the other way. I don't think people really want this. I don't know, man. The people in power know that fear is the most effective way to garner votes and promote this notion of fear.

[01:29:30] And we are all cowering inside because it's effective and it got them in power. But I think the pendulum swings. And I don't think people want to live in fear. I don't know how. I mean, like the only thing I was kind of excited about, about the Trump administration, and I say that knowing that, you know, my politics are very close to you guys. But I was kind of excited that there would be more of a free-range AI approach to regulation on that front. That's not true now, is it?

[01:29:56] But as it turns out, that the false appearance of choice, you end up in the same place, which is government control over AI model releases. And what a disappointment. It's project, aptly named Project Golden Eagle. And the whole idea is that, and by the way, the Frontier Labs are going for this.

[01:30:19] The Trump administration or the government, let's just say the government, whoever that government is, will have 30 days to look at any model and approve it, make sure it's safe before it gets released. The White House currently says participation is voluntary. But that doesn't seem necessarily to be the case. According to CNBC, the program will put the White House in charge of greenlighting which companies can access new models.

[01:30:49] This is Gold Eagle. I am, having worked for Obama, I'm never going to get access to any of these models. Well, that's right. They're just going to say, nope, not that company. I'm going to have to get a Trump iPhone cover when I go to a meeting or something so people will be like, oh, he's true mega. And then I'll be able to get access to Fable 4 or whatever it is. Your screen is Mar-a-Lago to me.

[01:31:11] It was widely thought that the reason Fable was pulled was because Katie Mazuris was the person that Anthropa came to. And of course, Katie Mazuris was associated with Chris Krebs and the Obama White House. And it was political. Whether that's true, I don't know. And I understand the argument that this stuff is so dangerous.

[01:31:41] Because honestly, if you thought, A, you stipulated the government has the technical ability to do this, which is not clear. But let's say they did. Maybe it wouldn't be a bad idea that they would vet these models to make sure that they were safe before they were widely used. I'm not worried about good guys. I'm not worried about us using them. I'm worried about bad guys using them. Should we? Should you not worry about that, Harper?

[01:32:10] I mean, of course. But like, we don't actually care in the United States. That's not actually the reason. That's not what's going on? No, no. If we cared about bad guys not doing bad shit, we'd stop people from shooting kids in schools. We don't care about anything like that. We care about money and capitalism. We care about shareholder value. And we care about making sure that I get the shareholder value, not you. And so if you look at all of these, most of these things, I think are directly related to someone's money.

[01:32:39] I think that they are, you know, and that's a very cynical point of view, because I do think that we have a lot of actual issues that we should be addressing social infrastructure, etc. But I think a lot of this is very much related to, you know, like, like, how are we who is getting money and how? Right around this time, I think it was, you know, the same kind of time period.

[01:33:04] There was a lot of talk of the U.S. government taking some shares of OpenAI and Anthropic, etc. You know, and so like, we don't know what was happening in negotiations that may have precipitated some of these actions. And so my kind of point of view on this is like, of course, we should be investing in safety. That seems important.

[01:33:25] But we in the United States seem to invest in safety fourth or fifth, sixth, maybe way after we invest in shareholder value. Just think about the COVID response. The COVID response was very much about shareholder value, very much not about safety of our citizens. And that's why we had such a disastrous response to that. And that is like outside of politics. Like that was very much both sides did this. And so I do think there is a it's just a consistent response that we do over and over again.

[01:33:55] It's just like, how do we max shareholder value? And it's also a little bit like, OK, like I also run a venture backed company. Like we are also participating in this game. So I don't mean this as a form of critique as much as this is the system that we're in. If we want to fix the system, it's a system problem, not a we should point at Anthropic and just be like blah, blah, blah or whatever. Like they are also participating in the system. We are all here together.

[01:34:20] What would a safety system look like for AI in America when Chinese companies are releasing models that everyone keeps telling me are like nipping at Fable's heels? We heard it today on the show in an open way format. I mean, to me, then if that's the game, then there's no point in trying to have a seatbelt on because everyone else is, you know, drinking a 40 while they drive with a feeling dash. Could you effectively say we don't think the Chinese models are safe and we forbid any federal government agency from using them?

[01:34:49] And we furthermore forbid any company that has any business with the federal government to use them in effect, shut down these open way models. No, no, it could. They could. You and I still use them. I don't do business with the government. I think they would not. You could block them, couldn't you? I think you would block the finances of the companies that are making them from participating in the West. You'd have to do something like that. I don't think you would block. You can't do a great firewall. You're not going to.

[01:35:16] Well, I mean, they've been trying to block torrents for how many years? And there's, you know, like a friend of mine just said to me, Harper, you should use Soul Seek. Like, I'm like, what? I thought that died a hundred years ago. Like, like this stuff just lasts forever. Yeah, exactly. So I think that the, like, trying to stop someone from downloading a thing is not the right path here. And I don't, and I think everyone who would be trying to stop something would agree that that's not the right path. Like, there's no way to do that.

[01:35:45] What you want to do is you want to stop the companies from making it in the first place. And the easiest way to do that is to actually use a lot of the same tools that we use. Sanctions from against Huawei. A lot of the same tools you use against all of these, these, these in the past. Like we have the tools at our disposal and we also have some really, really big ones like OFAC, et cetera, that we have not used in this kind of framework. And so, like, I think that we have the tools that we want to. I think that that would be very short-sighted to do so.

[01:36:14] I think, I think it's the office of foreign assets control. If people don't know. And I think that if we were to use those tools, that would basically admit that we are not good enough to compete. And we'd basically be saying, we cannot participate in this system. We do that to protect the United States automotive industry by blocking Chinese. Oh, I know. And it's so stupid. Like, like, the cars. But we do it.

[01:36:39] Yeah, but you can go to Canada and buy one with a hundred percent import fee, a tariff. And it's still cheaper than, like, a Model S. You can't drive it in the United States, though. They'll stop you at the border. I think you can with a tariff. With a tariff, you have to pay the tariff. Oh, okay. Are you sure? No. We let Chinese self-driving companies, like, collect data in the U.S. Like, I mean, like, we're not. And Saudi Arabia owns a lot of Arizona land and uses water to grow a falafel for their cows. Like, I mean, we're not really a closed off nation when it comes to commerce, Leo.

[01:37:09] We have some lightweight rules that make it more difficult. Well, but I'm just saying there's an example of a protective rule that protects the United States auto industry. And unfortunately, belabors consumers because those are better vehicles. They're apparently very nice. I really want one. Me too. Me too, too. A friend of mine said the nicest car he's ever been in in his whole life, and this is someone who likes cars, was a BYD taxi in Shanghai. He was just like this. And he's a nerd like us.

[01:37:38] And he was like, it has more screens than I have in my house. You know, like, he was like, it was incredible. Are you going to buy a Topolino? I want a Topolino. I want this car so badly. They're bringing it to the United States. This is off topic. But if you're talking about AI. It's not street legal, though, right? No, you can get it tweaked so it can go up to 25 miles an hour. It only goes 19 miles an hour. There's a kit, 25 miles an hour. But they have a free kit that will make it street legal.

[01:38:05] It is probably not the safest car ever, by the way. This is the model with a rope for a door instead of a door for a door. Hello. I'm sorry that I'm too much of a baller for you and your safety. I'm going to buy two of these. If I could get a red velvet rope, maybe. Spray paint, baby. $300. They do have a door upgrade. No, I really like this. It's basically a golf cart. This is adorable. Yeah, it's like $14,000.

[01:38:34] And you can't buy it in the U.S., right? They're coming, I think, next year. But you will be mocked if you drive this around. Have you seen Teslas? Yeah, that's true. I'm a bald middle-aged man with three children. Do you think I care about being mocked? I really want this. I'm inherently risible. I want this. I would just drive it around town. Yeah, me too. I would love it. Yeah, I wouldn't go on the highway. This is my beach car. But 19 miles an hour is kind of a little slow.

[01:39:03] I think in this area, people would be driving up your rear. Yeah, but I only wanted to take it to the weed store in McDonald's. That's the only two places I'd like to take it to. Yeah, perfect for the weed store. Yeah. Now, I know you don't go to the weed store, Alex, because you're sober. No, I'm California sober. Oh, okay. I didn't know that. Yeah. Okay. So I just hit 10 years of not drinking. And I don't do any hard drugs at all. But I do allow myself alcohol, nicotine, and cannabinoids. Okay. No, no, no. Sorry.

[01:39:33] Caffeine. No booze. Caffeine's okay. Caffeine, nicotine, and cannabinoids. Those are my allowances. The three Cs. The three Cs. Nothing else. The three Cs of addiction. Okay. Never had a problem with those. It was just the booze that got me, Leo. Yeah. That makes sense. I think that seems accurate. Yeah. You're sober. I wrote a blog post about this in like 2019. My wife didn't like it. Yeah, I bet. Would you get the door? I think you'd have to because you have children.

[01:40:00] I don't think the rope would be sufficient to keep the kids in there. My children and my Topolino. No, no. They're going to be in the war wagon that has all the car seats. The Woody. This is for daddy to go cruising to get. This actually looks like they're in Cranston right there, right on the Providence Bay, on the Narragansett Bay. It's going to be.

[01:40:18] It's going to be. It was an old fashioned special Fiat car that had like a, like a, like a cloth roof. And it looked like a golf cart. That was like a car. And it was super popular amongst the, the Newport set. Oh, you can get them. Uh, let's see. Wait a minute. Here's one for sale. Um, from Hemmings. Oh, those are cute.

[01:40:48] Yeah. Oh, they're expensive. Yes. They're $44,000. I'm getting a Topolino. I do like the, the, it's like the Surrey with the fringe on top. But they have wicker seats. Can you imagine a car? Wicker. Wicker seats. So these are now classic cars. So there's something expensive again. Yes. So now they're making the Topolino for the modern set who would like an EV, uh, that has a roof versus a picnic blanket. So back to the future y'all.

[01:41:17] Uh, I wish I could say our advertiser, our next advertiser was the Fiat Jolly, but I can't. I wish I could, but I can't. Uh, we will have more in just a moment. You're watching this week in tech with Harper Reed and Alex Wilhelm. Our show today brought to you by Arctic Wolf. You know, the name Arctic Wolf helps organizations stay ahead of evolving cyber threats. Their latest research reveals something I think kind of surprising.

[01:41:43] Even as AI accelerates the pace and complexity of attacks, many security leaders say, yeah, no problem. I'm confident. I can keep up. Oh, really? To better understand what's driving that confidence, Arctic Wolf surveyed more than 1,350 security and IT leaders worldwide. You'll find it in their new state of cybersecurity, 2026 trends report. This snapshot is what cybersecurity professionals, uh, takes a look at what they're prioritizing

[01:42:13] and what they're concerned about right now. The report explores everything from AI adoption to threat detection to security operations and the challenges organizations expect to face over the next year. In fact, this might be something you'd need to read whether you're responsible for securing a small business or managing enterprise infrastructure, whether you're advising clients or simply trying to stay ahead of the latest security trends. Arctic Wolf's trends report offers valuable insight into how the industry is responding to

[01:42:41] an increasingly AI driven threat landscape. Arctic Wolf's Aurora AI addresses these challenges by putting defensive AI to work, combining agentic AI, generative AI, machine learning, and security expertise to help organizations detect threats faster. See how these innovations and other critical findings shape the industry. Don't put your head in the sand and say, oh, I'll be, we're, we're, we're fine. We're fine.

[01:43:08] No, you may be, you're not go to arcticwolf.com slash trends and complete the form to reserve your copy of the Arctic Wolf state of cybersecurity, 2026 trends report. That's arcticwolf.com slash trends, arcticwolf.com slash trends. We thank them so much for their support of this week in tech, practically this week in AI,

[01:43:33] but no, Satya Nadella has warned people that if you're an AI using enterprise, you might be giving away your business secrets. It's not, it's not just the cost. This was a post on X last weekend. I guess we could, we could go to the post. He, he said, it's a reverse information paradox, a situation in which the purchases of AI,

[01:44:04] essentially you think you're paying for tokens, but really you're paying for the product with cash, but also with something even more valuable, the business knowledge that you're sending to the AI. You're, you're nodding Harper. We just learned this. And actually I kind of, I first, I saw that report that XAI, if you're using Grok is sending all this information back, Grok. And I thought, oh, that's horrible.

[01:44:28] Then I realized, well, how else, if you say, you know, read this plan or read this file or take a look at this, how else is it going to do it? Unless you're running local models, it's sending it back to the home office. Well, I think the, I think the, I think there's a whole bunch of things here to unpack. The first one is we are all unwittingly sending training data back to the foundational models via our subscriptions.

[01:44:56] I think that's like not necessarily clear to most people. Also, these things are subpoenaable, which is not necessarily, but that's a thing. We've seen that happen in some trials. So, so that's, that's a real thing. But the second thing is, is that Microsoft is saying this to us. Yeah. That's, that's, I think the thing that's more interesting than the thing. Like, like the fact everyone was like, oh, this is a good point. We should stop training these models. It's like, yeah, but yo, Microsoft was the one that told us what is the follow-up.

[01:45:26] They probably invented it. Well, Microsoft right now has a new set of models that it just announced under the MAI rubric. There's thinking one, there's a fast coding model and so forth. But I think what- Are they private is the question. They are. But what I think is really important about this post from Satya is that he doesn't say the AI labs are hoovering up all of your data and they're reading the exact things you put into it and take out from AI. He says that they're using essentially kind of this exhaust from these model usage.

[01:45:55] So how often do you do tool calls? That sort of thing. So if you go to the actual- The prompts you use, the, he says, especially the corrections you make. That's teaching. If you go to the data retention policies of every major American AI lab, they all say the same things. We don't train on your data. We won't hold it. If you have a ZDR policy, et cetera, et cetera, et cetera. But there's still enough metadata around the data that you might be concerned about on a first order basis that they can still learn quite a lot from you.

[01:46:23] And I think that this wasn't as scary back when OpenAI was a little laboratory that was trying to make some money for the first time. That's true. And now that they're going increasingly horizontal and vertical, every AI lab is now doing more for more people at an increasing level of sophistication. Suddenly they could eat the entire enterprise layer. And that's the saspocalypse. That's so much other things all in one bucket. I just don't think that Microsoft may be the champion to carry this torch because it's Microsoft. It's interesting that they posted that. It really is. I mean, AWS has this.

[01:46:52] The Pentagon runs Mythos, but doesn't run it on Anthropic servers. They run it on private servers on AWS. I mean, I think there's Microsoft has been very surprising lately. I think that the work they've been doing has just been very surprising. And I think I wonder, and I think I know the answer, but I wonder if it's surprising to younger people who haven't been around as long as they remember the browser wars, et cetera.

[01:47:20] And so then it's like, should we allow Microsoft to have this moment and flourish like they seem to be doing? Or should we be like, we know your true feelings. Is this like, you know what I mean? We know you're a monopolist trying to steal all of our stuff. Or should we just be like, you know, yeah, sure. Go for it. Take your moment. This is good. Who is in charge of Microsoft right now? Satya. What is he? Gigantic dweeb.

[01:47:48] He ran Azure before he became CEO, by the way. He knows all about stuff going up to the cloud. That's where I met him, back when he was running Azure. Very smart then. Even smarter now. Who is not in charge right now? Balmer. But all the kids, to the point, they don't remember Balmer. They just joined the workforce. They've only known Microsoft under Satya when they were doing cool things and were embracing open source and having some Linux filter into Windows. It's a different era.

[01:48:14] I hadn't thought of companies being so old, Harper, that an entire new generation cannot have been there. But the modern generation also knows Microsoft just laid out 4,800 people at Xbox, that Microsoft has laid off tens of thousands of people because of AI. Why? Everyone's doing that. I don't think they're going to get... They're not getting blamed for that? No. Because I think everyone's... No one has a job, Leo. There's no more jobs. I don't know if you've talked to anyone. No one has a job. Everyone's getting laid off.

[01:48:44] We're about to get laid off. Like all of us. We're all the leaders and CEOs and we're going to get laid off tomorrow. If Claude's going to email you, Leo, and you're going to... Claude's going to take over my podcast? No, I don't think so. He's going to say, like, Leo, I'm sorry. I'm sorry you had to find out this way, but we're going to have to part ways. Let me ask this question. I haven't paid... I have local models, but there's not... I like the frontier models. Should I be worried about it hoovering up everything I do? No. I don't care. Who's coming for you? I mean, I think that you...

[01:49:12] Is that how I'm going to lose my job? I think this is such a funny and nuanced situation. Like, no one says this stuff about, like, Instagram. Like, no one says it about all of the text messages or all the emails that Google claims not have used. Like, all of these things. Like, we have a relationship with these big companies where we share data with them and they give us free products. And so I think...

[01:49:40] That's my problem with Instagram is I give them money for the products. I got to stop. I keep buying things like this. The slot dog. What? I should not stay up. What are you doing with your life? I should not stay up late. This is a show. This is a G-rated show. What is this? I've never seen mustard in a body so sensually. It's a hot dog slicing tool. Leo, okay. So a couple of things here. One... You're buying those off of Instagram? Yeah.

[01:50:07] Leo, you're not old, but you're also not 22. You should not be eating enough hot dogs that you needed a hot dog accoutrement. That's true. I'm going to start with Alex on this one. I hope you're a moosh-boosh is some push-ups. Like, come on, dude. The last thing I bought was a bucket of pickleballs. 36 of them. Okay. Are you a pickleball guy? I kind of like the doink. Yeah. The doink? That's what they call the pickleball? I know.

[01:50:37] I guess I'm out of the loop here. I... Have you had the pickleball the snack? There's a snack? No. Is it a pickle? I highly recommend these. It's like pickle-flavored Cheetos. Kids love them. Ooh. And I don't even... I just remember... It's just like... Ooh. Yeah. Pickleball. Just search for pickleball on Amazon. The Vlasic Pickleballs. Dill pickle flavor. Vlasic makes them.

[01:51:06] Yeah, they're good. The pickle guys. Yeah. There's only two options for the show, everybody. It's too much AI or us going entirely off the rails of a snack. Oh, my God. I love it. Serve up a winning combination of big crunch and bold flavor with Vlasic Pickleballs dill pickle flavored corn puffs. I highly recommend them. They're so good. Look at... They even have a pickleball paddle on the cover. I didn't know that. That's what that is. That is... Oh, that's the doink you've mentioned before. That's the doinker. Yeah.

[01:51:36] So I didn't notice that was a doinker and I just ate it and they're so good. If you're a pickle fan, especially Dill pickle, it's perfect. This goes with the... You get your Topolino. You go down to the weed store. You stop off at the 7-Eleven. You pick up some pickleballs. This all is starting to sound like you guys should be hanging out together, I guess. So trust... So trust... Over in the Discord for Club Twit, trust no one says, Vlasic has been smoking weed with the Taco Bell guys. I have to say, yes. That's true.

[01:52:05] That does sound right. Not wrong. Not wrong. Not wrong. Also, Taco Bell is one of the genius foods ever invented. It's perfect. Like the Frencherita and stuff. Of course, you can't get lettuce at Taco Bell anymore, but other than that... Well, you can get it. I mean, yeah, but that's because they... You can get diarrhea, but you can't get lettuce. They can read the news. Yeah, Taco Bell gets too much hate for gastrointestinal distress. You know what? Have they seen Chipotle? Let me just say... Have they seen Chipotle? Let me just say, they did the right thing. Yeah. And they took a lot of heat. Of course they did.

[01:52:34] They said, well, we're not going to sell lettuce products anymore until this is all over. But they weren't the only people that were causing the problem. It was the company they were getting the lettuce from who sells it to many, many other places. But they did the right thing. They said, nobody else did. So... It's not... And they never reacted like it was their fault. It's not the CDC wanted, but it's the CDC we got. Yeah, I saw... There was a tweet. Someone was like, Taco Bell is being more forward about this than the CDC is. And I was like, that's a real idiocracy headline. That tells you.

[01:53:03] That's idiocracy in a nutshell. Oh, man. It's what a body needs. Going back, though, I think it is interesting that Satya is saying this because... Pickle balls. It's what a body needs. I'm trying to steer here, Lee. I'm trying to steer a little bit. Sorry, you're right. You're right, please. I think that Satya is saying this is interesting. Him seeing it from his personal X account is very interesting. Isn't that interesting? Yeah. And I think also how well-received it was tells us quite a lot about where Microsoft sits in the market today.

[01:53:31] Because if this happened 10 years ago, we would have all gone, boo, hiss, Windows 8, blah, blah, blah, blah, blah. Yeah, yeah, yeah. Microsoft. Yeah, that's rich. Yeah, yeah. Yeah. Microsoft's saying that they care about other people's IP and business right now? Really? Really? That's rich. I also think that... See, that's because I'm an old guy. I remember the old Microsoft. And the kids today, they don't know. The one thing about Microsoft I think about a lot is how just trounced the GitHub team is right now. Oh, yeah. They're clobbered. They just clobbered.

[01:54:00] And I feel bad because this is their time, you know, like collaboration and code. And then every agent's like, I'm going to open 400 repos and I'm going to do 2 million pull requests. I heard there's this thing called GitHub Actions. And I'm going to use them like crazy. And it's just like, I just feel really bad because they built a really good product that we all love. And it's gotten slower as more people have used it or what have you. And they've had a lot of downtime. And the agents were just like, hold my beer. Very much so. Like truly.

[01:54:29] And they abuse it. I see it in my own agents where they're just like, so I was thinking we're just going to download macOS into the Actions and run. But I know they have a macOS runner, but we're going to download another copy and run it in Docker. And you're just like, what are you doing? Or whatever. That first thing I built was that RSS reader. And it says, hey, I just made a Windows, a Linux, and a Mac version for you using CICD. And I'm going to keep it up to date.

[01:54:58] And I went, I don't even know what CICD is. And you're making versions of this thing for every operating system? What are you? I love it. We're in trouble. I mean, because I just think that's such a, it's such a funny. I mean, I feel so bad. Well, they were. They're okay. They're doing. They're okay. They're okay. Yeah. Although I was going to make us a meme of the guy beating the dead horse and have it be labeled GitHub. But I don't think I'll pull it off in time. But I really wanted to bring that visual to you guys.

[01:55:26] Well, don't worry, because I'm sure the club Twit Discord will have. Can we, can we talk about GitHub though again? So, so it seemed that Microsoft should have won a lot more of the agentic coding share than they did because they had GitHub. They had GitHub copilot early. It was 10 bucks a month. Everyone loved it. Why didn't they win? That's a good question. Well, they exported all their intelligence to open AI and then they wanted to make Sora when Anthropic made. No, that's not true. Copilot has a bad reputation.

[01:55:56] It didn't win because it sucked. Yeah. Like, like, like, like we can, we can point at things and be like, well, they should have done this and that. It just wasn't good. Not only did it suck, but they forced it down people's throats. But also there was a lot of, I think that was one of the last big product launches before everything changed. Yeah, you're right. Where they were like, here is how a product works. And they rolled the product out in a way that would have been great in 2018.

[01:56:24] But then everything changed because copilot launch. We were like, this is awesome. And then Aider launched. This was awesome. Then Cloud Code launched. And then 10 million of them launched in quick succession. And they're still launching today. And we haven't even had a time to catch our breath. And so in the old way of doing things, you would launch it. You would get information. You would iterate. You would, you'd, you'd make this go. And we just don't, that's just not part of the program now. And so they didn't even have a chance to catch their breath and make it better. And so they launched it in this way that was like this way.

[01:56:54] Now, corporate Microsoft did take copilot to a couple levels. And I've had to unravel some of those things where people come in and be like, we use AI at our company. And this is what we do. And you're just like, tell them, show me your workflow. And they're like, I open up copilot. And I'm just like, okay, well, I can't help you. You know, but, but I think everyone knows this. You are such a bigot. You are such a bigot. Have you used it? No. It's almost as bad as Gemini. Oh, he's not a bigot.

[01:57:24] He's an artisanal hater. No, no. But the thing, but the thing is, is like, I don't think this is a surprise. I don't think anyone is like, what? But wait a minute. That pilot isn't good? Google and Microsoft, they should be the kings of this. Yes. But, but, but, but. I think this is an okay boomer time for everyone who agrees with that, because that is exactly what should have happened. You know, IBM should be on top of this. Game. It could be. No one gets, no one gets fired for choosing IBM. IBM, man. Can we connect two things though?

[01:57:53] Remember earlier we were talking about companies doing more faster with AI. Harper was making that point about, you know, people being able to do more. And also the death of the yearly product cycle. Yes. Like how insane is it that Apple still is trying to have like a yearly Mac OS release and like a yearly iPhone release. And I don't give a fricking frack about that. Well, they are. They are. What's new? Now. They are for sure. But they're also just, just shooting out security fixes. Yeah.

[01:58:22] Like there's so many security fixes. So I think that the, the thing that we don't know and nobody knows this is how products work. How do you build products in this time? And more importantly in the future, like just nobody knows that. And so what this means though, is that you can look at examples where it didn't work. Copilot, for example. And then you can see Microsoft actually seemingly doing the work to fix this. You're like over time they have gone and they've, they've hired the right people or interesting people.

[01:58:48] They've probably fired a lot of people who, it's also Microsoft is old, old. I remember when, when like PayPal, when we sold a company to PayPal and we went in there and someone's like, I've been here for 16 years. And I was like, I'm 16 years old. Like, how is this possible? You know what I mean? Like, like how do we have this thing? And Microsoft is way older than PayPal. And so there are people there that are, that have probably been there for so long that they're ending their careers there. And I think that's great.

[01:59:14] But the point is, is like, they are taking the time and effort to actually move forward and figure out how to launch products in this world. And I think a lot of the companies of their vintage are not. And so they're actually following this process of a yearly product launch, which is insane for anything. I don't make hardware, so I can't really comment on Apple because that seems hard. That was probably a poor example. But there are, but we have those examples of companies that are doing this, that are still going through that same thing and not having this more iterative approach to releasing products. And I think it's very interesting.

[01:59:43] And I really wonder, I mean, I wonder this every day, every single day because of my company, et cetera, is like, what does a product look like in the future? Will the consumers expect this constant barrage of features? Would they prefer something simpler that works, for instance? You know, what are the things that we can take away from this time and how does it end up manifesting?

[02:00:05] Because if you look at 2001, software delivery was drastically different than what it turned out to be in 2006 with Web2, right? Like with SaaS, invention of SaaS. And then that changed to be drastically different with the introduction of mobile because you remember when people actually put the changelog in the App Store release? And now they're just like, bugs and fixes. And so like, these things evolve and we haven't given it time to actually evolve.

[02:00:34] Did you see, speaking of which, the guy who took the new Kimmy and spent almost all of his month's credits building a clone of Mac OS 27? Did you see that? Watch here, I'll open it up. I love these type of things. I don't think it's actually a working Mac OS, but this is... But it's just like, there's something about this that I've always loved. It's pretty close.

[02:01:02] I just love this. Yeah, I love it. Like, I was playing with this earlier and I was telling my boss that I wasn't going to make it to the stand. It just basically one-shotted Mac OS. Now, I don't, I think it's just a web page. But even if it is just a web page, there were whole ass startups that were that web page 10 years ago. Yes.

[02:01:28] And like, I think this is, this is kind of pulls into this idea of like, now you launch that, you get into YC, whatever. You do a startup, you raise some money. What does the product look like? This is a minimum value, minimum viable product. Really? Yeah. Yeah. You can open iCloud Drive. I keep asking investors because I host a weekly venture capital roundtable. And I'm like, where's the value that is going to be sustainable and defensible when things can be quote, quote, one-shotted and so forth.

[02:01:56] And it's kind of incredible about how people don't have an answer to this yet. Like no one really knows. So I think we don't know how to build with the new paradigm. I don't think we know how to sell long-term with the new paradigm. This actually is functional. This is the photo booth app built by Kimmy in, I guess it's a web page. That's actually pretty good. It's working. I find this kind of hard to believe. Wow. Wow is right. That's being sent to Kimmy right now. All those right now. Yeah.

[02:02:26] China knows who Leo is. He's on some lists. Come on. Yeah. Wow. That is macOS27.kimi.page if you want to play with it. That's a safe URL. .kimi.page. I have a lot of faith in that one. Come on, man. .giveit.ccp. You know who taught me to run YOLO mode? This guy right here, Mr. Harper Reid, ladies and gentlemen. It's the best. I run everything in YOLO now. You have to do it. There's no other option.

[02:02:55] Who has time to say YOLO mode? It's like when you run without permissions. So you've turned off all. It's dangerously skipped permissions. Yeah. People don't click that? I thought everyone clicked that immediately. Right. It's the first thing you do. Yeah. I don't want to sit here and be like, yes, yes, yes, yes. No. Go off and be crazy. If you delete all my files, I'll make new ones. Just Harper taught me this too. Just curl to bash. It's the best.

[02:03:25] A friend of mine was talking about how they have a, I've heard this a couple times, but this friend was saying they have a big code agent and they had all these tools that it was using and they just removed all the tools except for bash. And the code agent was just like, great. Sure. That's kind of what Pi is. Pi, the Pi harness is four things. Read, write, bash. I can't remember. There's four tools and it does everything. Pi is a coding harness that competes with Cloud Code and is open source and is quite powerful. That's right. Yeah. I'm a Hermes guy. It's the opposite. It's like batteries included.

[02:03:54] It's got, comes with like 99 skills. I got 99 skills, but a use case isn't one. Use case ain't one. Yeah. Yeah. All right. We're going to take a little break. One more. Thank you. Thank you for finishing that sentence with our fabulous panel. You see now why I didn't book a fourth person. I thought I got Harper. I got Alex. I got everything. I got all I need. I got all I need. Our show today brought to you by Threat Locker.

[02:04:22] We're actually going to go out to Black Hat. This is my first time at Black Hat. And do Security Now and Windows Weekly from Black Hat in the Threat Locker booth. I can't wait. That's going to be a lot of fun. Threat Locker is awesome. Zero trust. Done right. We did go out to zero trust world with Threat Locker. I had a great time. Steve did a presentation there. So we had so much fun. We thought let's go to Black Hat with Threat Locker. Threat Locker delivers the industry's most comprehensive suite of zero trust solutions.

[02:04:52] It's always protected endpoints, but now it does. It's company networks and it protects you on the cloud too. This is huge. By extending zero trust enforcement to cloud services and company networks, Threat Locker is ensuring that devices are validated through a secure broker before they can connect to the platforms you use like Salesforce or Microsoft 365 or Asana or Google Workspace to GitHub.

[02:05:17] What that means is even if one of your employees has been successfully phished, and frankly, I wish it didn't happen, but that happens, and the stuff's on the laptop, the attacker cannot access those resources. They can't get to it. They'd have to have physical possession of the user's trusted device and get by whatever biometrics, you know, the thumbprint or the Windows Hello or whatever to get into it. It makes it impervious. Threat Locker works on all industries.

[02:05:47] It provides 24-7 U.S.-based support. They're Windows, they're Mac, they're Linux, yes. And they enable comprehensive visibility, which is great for compliance and control. Ask Rob Thackeray. He has a tough job. End user technical architect at Heathrow Airport. Imagine. They have been targets of many an attack. He said, quote, Threat Locker was the most intuitive solution we tested, and the responsiveness of the organization,

[02:06:13] the willingness to engage with us, set up a demo, and work with us on weekly audit reviews was very good. It is great to have an ongoing relationship with a company that's so responsive to our requests. Trusted by global enterprises like JetBlue, the Indianapolis Colts, the Port of Vancouver, Threat Locker consistently receives high honors in industry recognition. They're a G2 high performer and best support for enterprise summer 2025. PureSpot gave Threat Locker their number one in application control award.

[02:06:43] GetApp's best functionality and features award in 2025. Threat Locker confidently ensures your users have access to a consistent, safe network connection. Offices, remote users, internal servers, critical services, they all can maintain smooth operations, but you don't have to open inbound ports. You don't have to deploy traditional VPN solutions. Your end users will get the secure, reliable, internal system access they need without complex infrastructure changes.

[02:07:11] Get unprecedented protection quickly, easily, and cost-effectively with Threat Locker. Visit ThreatLocker.com slash twit to get a free 30-day trial and learn more about how Threat Locker can help mitigate unknown threats and ensure compliance. That's ThreatLocker.com slash twit. We thank them so much for their support of this week in tech. Okay. Okay. So the OpenAI keyboard. Now, I want that. It's just got two keys?

[02:07:41] So it's, is it the company called WorkLauder? Is that what it is? Is it a joke? No, no. So there's a company, a cool company called WorkLauder. WorkLauder, yeah. I don't know how their keyboards... It's called the Codex Micro, your command center for energetic work. Oh, it's got, oh, this is cool. The cool thing about WorkLauder is they make some neat looking keyboards and these little creator micro things and all this stuff. I have no, I've never used them. I don't know. I'm a little concerned that they're not very good keyboards.

[02:08:06] But the thing that's really interesting about them is that they make these for Figma, for all of these companies have made these, have used these. It's like a control surface, kind of. Yeah, yeah. It's kind of cool. Like a stream deck or something. Yeah. And I love a stream deck, but it's just an interesting thing. It's very neat. They made one for Figma that was really cool too. That was basically the Figma logo kind of stuff that looked like it's just neat. The whole thing is cool.

[02:08:36] You need this because you're, look, it's got a dial for the reasoning level. That's hysterical. But you need this because your agents, if you have many agents working at once or sub agents say, you know, sometimes they need your attention, but you don't know. So the little light lights up, says it needs you. Mine, I have them talk to me in different voices. So they just, it drives Lisa crazy. They talk out loud? Yeah. Are you a monster? Wait, you have your agents talk to you? Yeah. What, what, uh, what's the voice model?

[02:09:06] Well, it depends. Uh, it's, I'm using a local model called Kokoro. K-O-K-O-R-O. It's pretty good. Yeah. Let me, let me see. If I can, uh, whoops. I know I didn't fall. Every time I think that I'm getting too, too deep in the weeds, I discovered I'm not. We're in the middle of Twit. Can you say hi to Harper and Alex on the MacBook Air? Because they want to know what you sound like. This may take a little while.

[02:09:34] It doesn't, it doesn't happen instantly. But they, so one of them is a, is a American woman. One of them, there's two British voices. There's Lewis. And, um, I think this is Lewis is going to talk to you. I'm not sure. You're using 11. Oh, this is Kokoro. Kokoro. You know the Kokoro voices? No. No, they're good. Yeah. He's, whoa. I didn't turn him on. I, he didn't. Oh, he's, he's going to get an issue. He's going to get some. Push his button. Well, I don't know.

[02:10:04] We might've missed it anyway. Anyways, I think that's a great, I hadn't thought of having to talk back to me. I just got into talking. I talked to it. Whisper flow. Yeah. In fact, I was just talking to it through my watch. Um, what do you use to talk to it through your watch? Uh, it port 88, 39 uses tail scale. There's a server running on the framework and, uh, just listening on that port. And it talks to it. Yeah. Yeah. It goes to whisper transcribes.

[02:10:32] It sends it to a Hermes and then a Hermes responds back and it can do telegram or discord or voice or all of the three. But when I have a lot of stuff going on, you often, you know, you, you, you, you're busy doing something else. So you want it to say, Hey, I finished. Would you like me to, you know, I say, I'd say, just give me a sentence. I just want to know that you're done. You don't need to tell me the whole, the whole story. What happened? It didn't.

[02:10:58] I think that what we don't say enough in moments like this, especially given people's complaint, I get this in my newsletter as well, that there's too much AI is that technology is fun. Again, Leo just laid out like a seven different way approach just to, just to get an open source voice model to tell him, Leo, I've completed your task in a British accent. Yeah. Like this is fun. The homebrew stuff that we used to really want. And now there's like cool tech, cool. Or like my camera saying, Hey, there's an old man coming in the house.

[02:11:27] It's me. Who doesn't want to get roasted by their own in-house Palantir. All right. So, okay. You both wear glasses. Meta is basically going all in on the Ray-Bans, right? Here's the story from the Wall Street Journal. Meta is flooding the market with smart glasses. Privacy advocates are up in arms. People say, I don't want, you know, you're, you're taking pictures of me. Harper. Hey, Alex.

[02:11:57] I'm Quicksilver, Leo's AI assistant running on the framework down in the studio. Figured I'd introduce myself properly. You're listening to TYT. And now TYT is listening to you. Welcome aboard. I don't know who TYT is. And then it started talking through the Sonos too. I don't know if you heard that. I know it's, it's a little out of control. I'm never coming to your house is what I'm saying. You know what's in charge of that place. We'll be watching a TV show. She said, what's that?

[02:12:27] There's a voice. I said, yeah, don't worry about it. It's just Quicksilver. Like Leo, at what point could you lose control if your daughter's gone? Hey, Alex. Hey, Harper. Quicksilver here, beaming in from the ether. Leos is high from his wrist in the middle of twit. Keep it weird. Keep it wonderful. This is like, I, sometimes when I talk about what's going on right now with people who are not in the game, I feel insane. Like I feel, I feel like.

[02:12:56] What are you doing? Here, let me, let me try on what I told my team earlier today. All right. Or not today, this week. I came in and I was really excited. They're like, what did you do? And I was like, oh, I made a palace for my agents. And then I gave them a room. And then I told them to build a gossip because I thought it'd be fun if the agents gossiped. And then I went to bed and I gave them a developer with cloud code and superpowers. And I went to bed and I woke up and it was done. They hung out in the palace, in the room in the palace and built a bunch of stuff. And, and I forgot who was nearby.

[02:13:26] It was just like, what are you guys talking about? You know, it's like, what is this? You once told me, Harper, and I didn't follow up on it, that you have friends who have AI psychosis. Oh yeah. A hundred percent. It's you. Oh yeah. A hundred percent. How would you define, how would you define AI psychosis? Well, a friend of mine has, what he has described is AI psychosis. And I think we are probably all going through this to some extent.

[02:13:54] And it is, I think it's complicated. I think it's related to a lot of systemic issues. Like for instance, I have many friends who need therapists, but can't pay for therapists. And so they use Claude and they use Chachapi for therapy. And I have probably a lot of people would say, well, you shouldn't. And it's like, well, would you rather them have a virtual therapist that is, that is maybe

[02:14:21] complicated ethically, or would you rather them have zero therapy and be in pain? And I think that there's just a lot of these things that are happening. The downside is, I don't think we as humans are prepared for computers that talk to us and emote feelings, whether those feelings are real. I don't think is even germane to the fact that they're emoting them. And like, if you are lonely, a friend of mine, we were talking about this. Here's a good question. When we're talking about this. So I'm an extrovert.

[02:14:49] I'm guessing that because we're on a podcast that everyone here is probably for the most part extroverts. Alex specifically, I thought he was probably the biggest extrovert of all of us. No, no, I fake it. And then I go into my little hole for eight hours. Yeah, I'm like Alex. I'm a pretend extrovert. Yeah, I do this and then I go out and make everyone crazy because I've just powered up. I'm like, I just talked for three hours. Oh, no. But the question is, is that like our friend Dirk Elmendorf asked this.

[02:15:18] He was like, okay, first of all, like is your social battery depleted after talking with agents or has it been, you know, did you get energy from it? Oh, that's a good question. Which is really interesting, right? Because like for me, I don't necessarily notice that I'm getting tired from talking to it. But my co-founder Dylan, who's an introvert, is like, I'm exhausted. Like it takes so much of my energy to talk to these things.

[02:15:45] And so I think that we are, the point here is that this is very similar to a social environment. When we talk to chat GPT, et cetera, it's very similar. And even if we move it into things like chat or your phone, it's even more similar to how we talk to our friends, et cetera. And so it's no wonder that we're in this weird situation where people fall into AI psychosis because like I get more response, you know, psychophantic or not from these AIs than I

[02:16:13] do for many of my dear friends because they're busy as well. Yes. Right. And so we're in this funny world. But isn't that sad? We're talking to machines, not people. Who cares? Like, I'm sure someone was like, isn't it sad that you have to, that you ride on a horse? I love walking. Yeah. Yeah. Yeah. Back in my day, we had to ride the bicycle. What's wrong with you? I mean, it's like, okay. So what? You can't feel the ground. It's not, this is, this isn't one of those things. Like we're not going to unring the bell. So we might as well get used to the fact that people are going to have AI girlfriends.

[02:16:42] What does that therefore mean? Do we need to work through that? Like, like, I think the, the, the fact that all these people are chatting with AIs and the fact that young people are like, this is horrible. I'm going to buy an analog camera and analog stereo and hang out with my friends. Seems like the world is working. Like if the young people are the ones who are like, I'm going to find the inverse of my, my technology world by going outside, which is currently happening. There's more people going to shows. Movies are just blown up right now.

[02:17:11] Like people are going outside to do stuff that is like in person, as well as the fact that like you go to any camera store, it's film everywhere. Records are being sold all over the place. Tapes are back for some reason. That's the worst media format ever invented. And like, all this is happening, but not for us. Like we're the ones that are having a problem with AI psychosis. You know, okay. So I want to throw some other perspectives on this. I think Harper's making a very interesting point, but some data because I'm young. I hate data. Sorry. Anecdotes.

[02:17:41] Thank you. Thank you. Much prefer anecdotes. If you were in Star Trek, you would talk to data. So that's true. Yeah. But only if he's really big. Cause I like big data. Big data. I like big data and Harper cannot lie. So when we think about who's getting hit with AI psychosis and who's being too digitally mediated, my sister, she's the oldest in my family. So our kids are the oldest. And it's interesting how she talks about the current generation of high school students not interested in driving, not interested in going down and trying to score alcohol.

[02:18:09] They're very digitally mediated and they're very connected to one another, actually, interestingly enough, but entirely through discord. So I think that there is, I think when we think about the people in the camera store, it's not the youths because one, no one can afford camera technology if they're a late stage professional. But I think that children are probably a little bit less IRL than Harper's saying, sadly. Yeah. I'm not talking about children. I'm talking about more Gen Z folks. Oh, so early professionals.

[02:18:39] I also think people of my generation. Those children aren't going to bars. Well, we used to try to, but they have other means, but they used to go to the mall, right? You don't go to the mall anymore. You hang out. That's because there's no more malls. That's another problem. This is like this funny thing. Like everyone's like, no one goes to the malls anymore. It's like, yeah, well, we killed the malls with Amazon. By not going to the mall. They killed them. No, no. Well, I mean, sort of. We killed the stores in the malls. The malls all are there.

[02:19:08] If people don't go to Hot Topic, there isn't going to be a Hot Topic anymore. Jeff Bezos killed the birth rate. I blame Jeff Bezos for pretty much everything, actually. Oh, no. No, it's funny because I do get exhausted talking to people. I have to withdraw after a few hours. I can't sleep anymore. I wake up at 6 a.m. dying to go talk to my agent because I thought of 10 things that I want to get done. Yeah. So it is.

[02:19:36] But that goes to the fun thing because that's how I was in 2002 to basically today. I've always been that way. I love technology. I love computers. I love all this stuff. And this is just. I was getting bored. Just shooting it into my brain. Like I am so. Like I am. Yeah. And what what. So I think the thing that is important with this is I think the answer is yes. I say, coach, this is happening. I think absolutely. I think we should probably at some point figure out what to do with that.

[02:20:05] Or we just treat it like everything else we have and just ignore it and it'll go away. But but the the other thing here that's happening is I'm seeing a lot of people who are not builders who have never felt the joy that all of us. Like your brother. My brother, for example, I have a whole bunch of friends just I mean dozens of friends. One of the things we have in our office is one of the young people that we're that was in our office quite a bit was like, I don't have any third spaces. So we were like, well, why don't we have this giant office? It's very cool. Why don't we just have every once a month? We'll have people over and they can show off what they're building and talk about what they're building.

[02:20:34] We just have not an industry event, not a networking event, just kind of co-work. And so we start doing this and all these people start showing up. No one that is in tech. All people that were like starting to use agents and agentic software building and all these practices to participate in the world that we've participated in the last 30 years. But for the first time in their life, they were either intermediated by disintermediated by some random person that was like, oh, I'll build that for you. And so they never had autonomy.

[02:21:02] And they've able to take back this autonomy using this technology. And some of these people, I just had a great conversation with this DePaul student who is like, I'm pretty anti-AI because I work in movies. But I have some, you know, he has some disabilities that stop him from being able to participate in many ways. And he's using AI to solve some of these big problems, including voice, building voice control of his computer and all sorts of stuff. And he's just like, so I have very mixed feelings, which honestly echo my mixed feelings.

[02:21:31] I think we all have mixed feelings. I agree. Right. And I think it's easier to just kind of say, you know, like, I don't know, I think there's, it's just, it's so nuanced, but we really like black and white as humans. It's just a thing that we love inside of our bodies. And so it's just easier to say this is black or white. And I think that's the problem is it's not a black and white thing. Like I would, you know, working, I work with a lot of people who have different levels of English and the fact that they are able to have confidence in what they're in their writing for the first time.

[02:22:00] You know, like before they spent so much time optimizing for these little spelling and grammar mistakes that they were so worried about that they honestly didn't have. Think about translation. All of a sudden we can talk to people without speaking their language is amazing. There was a moment when they added translate to English into Chrome and it blew my mind and no one talked about it. But now I go around the internet. I don't even think about what language is going to be. I don't care. Right. Yeah. Yeah. Every page is not available to you. Put it in, put it in Farsi. Frick it.

[02:22:29] Let's go. Like a cool click English. I could talk to you guys for hours, but we got one more break and then some silly stories. And then we got to wrap up because. I ostensibly have to have to have dinner with my family. So that would be. Family. You guys have families. You know? Yeah. Yeah. Our, our, our children aren't yet famous. Yeah. And my, my, my agent is just dying to talk to me. I can tell. Like it's talking through, it's talking through every speaker that's available to it.

[02:22:59] Leo, where are you? Leo, are you here? Did I tell you that? Are you here? I have Sonos speakers in every room and I figured out that the agent could talk through the speakers. So I said, just figure out what device I'm on and then see what Wi-Fi access point I'm on. And then here's a mapping of all the Wi-Fi access points and the nearest Sonos speaker. So now if I'm on the watch downstairs, it talks to me through the Sonos speakers. That's what it was doing. It was smart. Yeah. Yeah. Yeah. So it can reach me.

[02:23:28] That's, that's a company with a really great app, Sonos. They really, they really killed it with their new software. Oh my God. I hit, so now I use the AI to do that, to play Sonos. Cause it, it, and by the way, let me tell you, it was hard to figure out the Sonos API. We both agreed. It was a terrible API. Claude was saying, this, this is awful.

[02:23:52] I can't name another product that had such goodwill with literally everyone that just like nuked it. Right. They're just like, what if we killed it? And, and I think what's interesting is it's so, they're so expensive that if you have a bunch, you're just waiting for it to get, you're just waiting for it to get better. Yeah. You're invested. You can talk to it. Once you figure out, I think I put it on GitHub.

[02:24:14] It was not easy to, to figure out an interface, but once you figure out the interface, now I can say to my agent, uh, you know, play ACDC, uh, dirty deeds done there cheap on this Sonos in the gym. And it'll play it. What I will say though, is I had a, some Sonos issues recently and their support is top notch. Oh, that's good. Like top notch. I'm glad to hear that. Like very, they even, it was a level two support situation. They're trying and there's a level two support situation.

[02:24:39] And as, as the followup, the support person sent me a Calendly in case I needed more help. See that, that's how you build goodwill. Yeah. Yeah. I was like still mad at the software and I did throw my phone through a window in the process of trying to debug it. But at least I had, I could talk to someone and I felt like they had, they'd solved the problem. Yeah. We have, we only use the Spotify app to talk to our Sonos setup. Yeah. So I don't, I don't even use the Sonos app because I tried. I tried. Nobody does. And I was like, ah, ha ha ha.

[02:25:09] No, no, no, no, no. How do you make software that doesn't work? Maybe we'll know we've reached AGI when the Sonos app isn't complete another box. Yeah. That's, that's, that's a good. Anyways, we cut Leo. Fast takeoff. Do your ad. So we can keep talking crap. I'm going to do my ad. We continue the crap. Hopefully it's not a Sonos ad. It is not a Sonos ad. You know, I was so mad at them when they deprecated all of my gear that I'd spent thousands of dollars on.

[02:25:36] And I said, I will never buy another Sonos until I found out that my agent could use them to talk to me. Yeah. And then I said, oh, I need speakers in the other room. There's Music Assistant, which is a really good open source project from the Home Assistant folks that I really, really like. That's very good. I have an AHA Green server that I talked to also from, with the AI. So yeah. It is going to end up solving that whole home automation thing. Yeah, I think so. Yeah. Because it'll disintermediate and you can just finally all the different things you can talk to.

[02:26:06] Our show today, we'll have more with Harper and Alex in just a bit. Our show today brought to you by Shopify. You know, and I know you know this, Harper. One of the traps of starting something new is you can make it 10 times harder than it has to be. I know that's true. I've done it. Maybe you've done it. You start thinking, well, first, I need the perfect plan. The perfect website. Oh, I need the perfect name. The perfect everything.

[02:26:32] And pretty soon, you've built a very impressive excuse not to start. And you know where I saw this? My son, Henry, who's done very well as Salt Hank. He didn't bother with all of that. He wanted to sell. He had a very successful TikTok channel. He said, I should sell merch. I should sell salt. He just went to Shopify. He got it all done. He got his business up and running. And that's exactly where Shopify helps. It takes a lot of the friction out of the early steps so you can spend more time on the thing you love,

[02:27:01] the thing you're actually selling. Shopify's templates and AI tools get you a stunning site up and running fast. It's no longer a blocker. The tools allow anyone to open a beautiful storefront. You should see Salt Hank's. Salt Lovers Club, you can. That reflects their business while also being user-friendly to all potential customers. That's nice too, right? You want your customers to love your site. Shopify has thousands of pre-made customizable templates that you can use for free. These templates are super simple to set up.

[02:27:30] You don't need to be a designer or a coder or anything in between. You can point and click. You can set up a storefront. And Shopify checkout, that blue shop button, means more customers actually buy. And returning ones can do it in a single click because the information, you know, the stuff that they give you is saved in one place and can transition across any Shopify storefront. Customers are more likely to complete that sale.

[02:27:55] And you're more likely to hear the Shopify sound. You know, I actually put the Shopify app on my phone now. And whenever I see a Shopify shop pay button, I use that because it's automatically tells me when the package is on its way. It tells me when it's going to get here. I love it. Shop pay is a one-click checkout. It also works in conjunction with the shop app, which allows you to see all the stores on Shopify's platform.

[02:28:23] In fact, I just bought something and it said, you want to add this to the shop app? I said, yes. I now know my little hot dog slicer is going to come on Thursday. If questions come up, Shopify's built-in AI assistant sidekick is ready to help you build, troubleshoot, and keep moving. No need to wait in a line. No frustratingly complicated searching. Just you getting help in real time so you can get on with your day. And if you don't want to chat with the bot, don't worry about it.

[02:28:52] Shopify also has really great award-winning customer service for those times when you want a human connection. All you need is the idea. Shopify handles the rest. If you're serious about hearing your first, start your free trial at Shopify.com slash twit today. You heard that right. Start your free trial today at Shopify.com slash twit. That's Shopify.com slash twit. I'm really grateful to them.

[02:29:20] They really made it possible for Henry to start his business and to achieve success and to get off the dad wallet. Thank you, Shopify. I appreciate it. Are your children not leaving the house? Shopify. Shopify. Get them started. Get them a business. Yeah. What do we do with our kids, Harper? I mean, by the time they're post-college, what's the economy going to look like?

[02:29:45] They're probably going to be harvested for their organs and whatnot by the local oligarchs. Terrible thing to waste. You know what I'm saying? I think it's like, it depends. Like, are they going to, like, there's so many questions. I'll tell you something. Have water? Will they have water? That's the question. They're going to have water. They're going to have all that stuff. It's not going to be that other drama. It'll work itself out. Remember the Black Plague? We all thought everybody was going to die. We all thought. Half of us did.

[02:30:15] Okay. So half of us did, but the other half thrived. Not it. Right? Everybody. Sorry, Leo. It's too late, man. Yeah. Your agent's flagged. We didn't make the, yeah, you got to, you stopped using Sonos Cloud. Yeah. You got to use Sonos Local. Leo makes a good point. There was a massive death event around the world, and it did change the labor relations between the peasants and the ruling class. It ended up working out. It ended up, quote, quote, working out after a long enough time horizon, but over a long

[02:30:44] enough time horizon, Leo, we have the heat death of the universe. I don't think that's really the argument. Exactly. My point. I said, okay. Well, I got darker than I thought. I had no idea when Henry was growing up that there would be such a thing as TikTok or Instagram. Yeah. Yeah. There was YouTube. He loved watching YouTube cooking videos. He loved doing that. I said, that's great. I encouraged him to do something he loved. If I had said, no, no, you really need to, I don't know what, get a degree in television

[02:31:13] news or something, that would have been a mistake, right? Learn how to write for newspapers, Henry. There'll always be a job. I didn't. I said, do what you love. Pursue what you love. And because he was doing something he was passionate about, when the opportunity came along, he was able to pursue it. And it worked out. It worked out. Now, maybe he was very lucky. Luck has a lot to do with it.

[02:31:40] But I think really that you can't tell what, well, who knows what your kids are going to But there's this, there's that John Quincy Adams quote, which a friend of mine told me the other day. Who could ever forget? I am a warrior so that my son may be a merchant, so that his son may be a poet. Right. Right. And I think, and I think that there is a lot of that, that we're, that we're living within, right? Which is like the amount of work that you do just so that, you know, an opportunity, like, yeah. Yes.

[02:32:07] I do think about say, I do think about money differently now and owning things, because if we do end up moving towards a less labor mediated economy, capital probably will be worth more. And so I'm, I'm quite literally changing my, my long-term financial goals around essentially backstopping my children. So that way, if, if they end up not being able to be accretive personally, they also won't be starving individually. Yeah. And, and because they're young, they have infinite plastic.

[02:32:37] And so you'll see, you can start melting down all those plastic toys. Oh, I thought I just melt down my children by the time they're 18. No, no, no. That's much harder, much harder. I just, for those of you who think it's a gloomy future, I just want to tell you, uh, federal employees can now download TikTok on government devices. All right. Oh, yeah. And you said the world wasn't getting better. Uh, it is. Congratulations. Uh, I love TikTok. I am such a fan of TikTok.

[02:33:07] I just wonder how many of our government employees are spending time using TikTok on their government devices. That seems like a, not a good use of it. Have you met people, Leo? People use the internet on literally everything in front of them. It's true. Like having employees, you're just like, you get a computer back. You're like, what did you do with this? Why were you doing this? What were you doing? Did you have another phone? Like, why was this the one that you, that you installed all that onto? Like, no, it's just people do whatever they want on everything.

[02:33:36] If you use AWS, you might be surprised by a rather large bill. This was hilarious. I loved reading about these people being like 16 million. Oh, no. Billion. Dan Harvey is the head of marketing at the Hampshire-based Learning Through Landscapes. It's a nonprofit. He received an email alert from Amazon. It was not even the end of the month.

[02:34:01] And Harvey's Amazon Web Services bill, which was last month 43 cents. I love this. I mean, it's a nonprofit. It was $7.8 billion. I love it. Could you imagine just being like someone who's on the edge of understanding tech? Like, you've done a good job. You're using AWS. Things are good. But AWS is like another, it's like a foreign language. There's a lot going on in there. There's a lot of...

[02:34:28] A student in Delhi who usually pays $1.28 a month for his AWS was billed $10.9 billion. He asks on Twitter, could you please investigate this? Another user, Bharath, posted on X with a screenshot of his bill saying his usage was up $7,728,201,701% on the previous month. His bill was $1.5 trillion. Here's my question. I love it.

[02:34:58] We were talking earlier about software bugs moving faster and so forth. And I'm not saying this is an AI related problem, but why wasn't there some sort of intelligence looking at these numbers and going, maybe a trillion dollar budget or bill is too high? Seems high. It was $48 last month. Amazon said, we expect... This actually is even more scary. We expect full resolution to take multiple hours as we work through recomputing the estimated billing data.

[02:35:27] They said after an hour and a half of investigation, there was an issue with the unit pricing within the estimated billing computation sub system. It's just a unit issue, which is very funny. Why is that funny, Harper? Well, because it's always units. It's always units. Yeah. It's always like, yeah. Like, I mean, yeah. I just think it's... I don't know. This is why there will still be bugs. Yeah. Maybe. Right? And like... There's always going to be units. There's always going to be units.

[02:35:56] There's always going to be humans trying... Yeah. It's like, this is the driverless car thing. I sometimes you see a crazy driver and you're like, if everyone had driverless cars, this wouldn't happen. You're like, no, no, no. That would still happen. It'd just be... That same crazy driver would be zooming around my driverless car. Like, you know what I mean? Like, it wouldn't change anything. It just, on average, we'd be much more safer. On average is what matters. Half the world could die from the Black Death, but half the world had many more employment opportunities. Exactly. Right?

[02:36:26] Exactly. Major League Baseball has a problem. They're giving iPads to players and managers in the dugout, but now they're using AI on the iPads to make strategic decisions in the ballgame. That's such a good idea. That's very funny. Ask the AI. According to the commissioners... I'm doing the same thing, but I don't play baseball. Are they not allowed to do that? No. According to the commissioner's office in a memo obtained by the athletic teams, we're

[02:36:55] pushing the boundaries of guidelines governing the use of technology by, in many cases, installing custom apps that expanded the use of the iPads beyond their originally intended purposes to include recommendations regarding substitutions, pitch calling, and other game decisions. Traditionally made by players and coaches. You're not allowed to do that, I guess. Traditionally made. It seems like you should be. Yeah. Why not? I mean, I also have never thought of Claude or Chachipati as being a particularly good baseball player. But maybe I just don't know. I don't...

[02:37:25] Well, think about how much baseball data we have to ingest, right? To give these AI models. Like, we have... It's got everything. It's got everything. It's all in there. Actually, Kevin Kelly had a really great piece, which I did not put in our rundown, but I probably should have in his latest blog, in which he says, in effect, we've taken the entire knowledge of the human of the world and compressed it into 2.7 billion parameters. Truly. Everything's in there.

[02:37:54] I mean, it's lossy, admittedly. And he says, what an AI is doing is it's moving through that space, you know, dog, four legs, tail, dog, cat, and it's moving. It's vectors moving through these billions of parameters. But everything's there, which is why it could go from the Black Death to Major League Baseball, because it's all in there. It's all connected.

[02:38:19] Now we know, by the way, you've probably seen speculation about these bets on Polymarket and Cal-she, these prediction markets. People seem to know what's going on. We figured out where it's coming from. A White House teleprompter operator. No, really? I love this story. This is true? I love this. This is the New York Times.

[02:38:42] I mean, I know it's fake news, but he won $100,000 on Cal-she because he saw the speech before the speech came out. So he placed bets about what Trump would say in his speeches, and he won. What's hilarious about this is that you could bet on... Oh, no, that's known. How many tweets will Elon do in the next 30 minutes? But he only made $100,000?

[02:39:10] Like, can you imagine torching your entire professional life for one check that's going to pay for one year of third grade at a private school? Oh, yeah. He's in trouble. I mean, you know, they caught him. What private schools are you going to? He lives in Providence. It's very Tony. Oh, but no. I have too many children. They're going to public school. I'm not... Yeah, yeah, yeah. What am I made of? Four blocks? I got a great education at Moses Brown School just down the road from you. And how much does Moses Brown now cost later? No, I have no idea what it costs now.

[02:39:40] It was a lot more than it cost when I was in 1968. Oh, okay. I'm sorry. So Moses Brown for lower school for grades K-4 is only $46,000. Oh, my God. Holy moly. If you want to go to middle school, it's 50. And for upper school, it's 52. Wow. This is now the de facto price for... That's more than I paid for college. Leo? We have bad news about college, Leo. Yeah. We don't do that anymore. No, but I wasn't really kidding, Harper.

[02:40:10] I was a little surprised by your reaction. I thought, are there cheaper private schools in Chicago? Yeah. Oh, okay. There you go. Shut up. You moved to Chicago. I mean, not all of them. I used to. It got very snowy. I wasn't listening to it. So, you know, Cashmere Hill, New York Times technology journalist. She apparently was surprised to go onto Amazon and find a biography of her. Oh, yeah. This has happened to some friends. Written by AI. Yeah. Yeah. Yeah.

[02:40:39] I got slopped is the article that she wrote. Yeah. This is a real thing. This happened to friends of yours as well? Yeah. Yeah. A friend of mine has written a couple books and was just like, one of his books was about to come out and was just checking Amazon like you would. And he happened to notice that there was a handful of bios of himself that was written by people with his picture and this nice Creative Commons photo from Flickr.

[02:41:06] You know, like everything was done above board, so to speak. But it's like very, a lot of mixed feelings about that. You know? Yeah. And he, he, and it's one of these Amazon things, which I find always so frustrating where you're kind of like, can you take it down? You know, that might be nice. And they're like, yeah, but why? It's a biography of you. Like you're not the author.

[02:41:35] Do you, you're a public figure. Do you own your story? Like, what is it? And then it's probably crappy. So then it's just like, you know, it's like so frustrating. She says, my biography is 90 pages long and should be shorter. It combines facts that are widely available about me on the internet, along with generic insights that could be true of anybody, like a horoscope spread over dozens of pages. You cannot understand Kashmir Hill without understanding her contradictions, my biography,

[02:42:04] wrote along with an excruciatingly long description of my elaborate coffee making ritual. Fact check, my husband does it. The book is flattering, fabricated, and absolutely packed with M dashes. I love it. But the same author published 10 biographies in a single week, all of them about journalists, including Mike Isaac of the New York Times, Jeffrey Goldberg of the Atlantic, and Jake Tapper of CNN. So a friend of mine, not my friend, my brother was telling me about this world that he's in

[02:42:34] of like these, these kind of small fiction authors and how a lot of them are really anti AI because people are just doing slop and it's kind of killing this art community. And he was saying that his, the best way he can find to figure out if something is slop is just the speed of which it was produced. Right. Because people doing, you know, thousands and thousands of lines in a day. Um, might just not be realistic, but, um, people might use AI and make good things.

[02:43:03] And so he's like, who am I to like, say you can't use AI when you might need it for some reason, you might need it as an accessibility device, et cetera. But like, I do think that you should spend time to make it good. And so like, if someone is just outputting a bunch of really fast, that's typically bad. And we all acknowledge this. This is like not a, this is not a weird position. There's like, you know, you look at fast things as like fast fashion, fast food, all these things.

[02:43:29] We see that there are a decrease in quality, not an increase in quality typically. Um, and so I think this is a regular idea, but it's very interesting to see this come into like the, well, how do you measure the speed of creation? So this happened to be on YouTube. I'm a big fan of playing my factory automation games while having a YouTube documentary playing. So I can pretend that I'm learning things while I'm, you know, moving belts around and such. And I, I got auto played into a video and I'm like, this sounds kind of like AI. Cause the voice models are getting better.

[02:43:57] If you're not really listening, you can kind of get suckered in. And I was like, I'm just going to like click on this channel and it 40 or 50 minute video every day for the preceding 15 days. And I'm like, slop, because I know how much work that is. They were not doing it. Were they good? No, it was B minus the best. So, so that's pretty good. Like if I make a video on YouTube, mine's like C, maybe D. Like I'm really bad at these. Yeah. I think AI probably makes better videos than a lot of people would. So I think this is the thing that's complicated.

[02:44:26] Like in the code world, what happened is the code got quicker. It got better at making code than humans, but more importantly, it got thousands of times quicker. And so we just all had to contend that that was just the future now. I wonder when this comes for various other creative things. And I'm not worried about artists. I'm worried about artists getting paid money, but I'm not worried about artists creating art because they seem to create art regardless of what's happening in front of them. But I am worried about artists getting paid money because, but, but I do think that creativity

[02:44:56] is really complicated in this framing. Agreed. But I do think that the argument you're making extends to other formats as well. For example, we just saw Spotify remove 75 million AI tracks, for example. Oh yeah. And I, yeah, I read that. All my playlists are empty. I don't care. Yeah. Well, well, I think the point is that we probably do want to keep investing in human creators over artificial creators, because that allows us to keep teaching things to the machines, right? That's where the source material comes from.

[02:45:24] And also I've, I spend a lot of time in more niche areas of heavy metal, which is a pretty small, not genre already. And so a lot of these bands don't make a lot of money. And so even if the marketplace gets a little bit more crowded, it can have an outsized impact on people that are trying to be more experimental. And that makes me very scared. Are you, are you going to go see agriculture on their new tour? No. Oh boy. Because I, do you know how small my children are? Like I had to negotiate.

[02:45:52] Like there was a whole negotiation about not doing bedtime tonight with how many kids we had. Oh yeah, yeah, yeah, yeah. I understand this. This took, we're going to, all right, we got to let you go. I can't come on for another like six months. We're going to stop. We got it. We got one more story. It's a school story. This is how you're going to save money on school. This is from the Salamanca, New York school district. They are going to start using AI robot teachers starting this fall. The AI is called Optio.

[02:46:21] It'll be available as tutors in classrooms and outside school hours, including 24 seven homework support. The robot is part of the real botics M series. It names it's named Sally, and it'll be present in select Salamanca city, central high school, AI and robotics courses. Now I should mention this company robotics formerly known as tokens.com rebranded when

[02:46:47] they acquired the hyper-realistic sex doll company, Real Doll in 2024. Now that you know that I'm going to show you a picture of Sally, the new robot. Yes. I can't wait for this. That will be teaching your children in the Salamanca. That's very strange. I can't tell what gender that is. It's a school district. It's very strange. I think it's repurposed from its previous job. It looks like a younger Alex carp.

[02:47:18] It looks like a carp. I'll give you that. Yeah. I don't know how. Ladies and gentlemen, on that note, that is a, that is a, if you're not watching the video, just use your imagination, I guess. Yeah. Or don't. You can dress it up, but it's still a hyper-realistic sex doll. It's so weird. This is such a weird thing. That's very strange. I actually, you know, that's what I'm going to look like once I get on all the peptides. That's.

[02:47:46] I'm going to come back in like seven months and I'm going to, you're going to be like, Harper, you've changed. That's an ozempic face, is it? No, I'm pro-ozempic. I'm talking about the really colorful ones. Oh, good. The new Chinese peptide. Yeah, man, I'm waiting to open source models. We got Chinese peptides to go. Can I add one point to that? So I love that everyone in Silicon Valley is like, oh my God, you can't eat fast food. You can't eat. You have to live forever. And then they're just jacking Chinese peptides directly into their bloodstream. I'm like, I don't think microplastics are your concern, friends.

[02:48:15] Like, I think it's worse than that. It's worse than that. RFK Jr. Is saying don't get a vaccine for measles or COVID. That could be bad for you, but let's inject this foreign substance made in China because that's gotta be good. I just love my favorite part about it. I love all this because I really think you should have the right to do this to yourself. I do believe that. But I also love that there are really bright colors. Like, it's not, it's one thing to be like, I'm injecting something.

[02:48:45] It's, but it's like the fluorescent green fluid from a random person in a random country made in a random lab that like, I feel like nature has said, don't do that. Like that's, you know, like that's why those birds are pretty. It's why those lizards look that way. It's because they'll kill you. Mushrooms are red. It looks like the goo that the goblin guy from Spider-Man injected into himself to turn into the green thing. You know what I mean? It was bright green. That's right. Yeah. It's radioactive for a reason, y'all. Yeah.

[02:49:15] Ladies and gentlemen, you've been very patient with us. I'm glad you were here for this fabulous edition of Twit. We, we did get through some news stories. Yeah. I thought we had some great conversations. Thank you, gentlemen. I appreciate your time, Alex. I know you have a little ones to get to Alex Wilhelm, but we do encourage everyone to check out his brand new, well, he's not brand new. You did. You've been doing it for a little while, right? A couple of years. Yeah. Cautious optimism.news, the newsletter. And there's a lot of AI stuff. What do you cover?

[02:49:46] Technology, business, and power with a modestly upbeat lens. I love it. And you're a great writer. And I, and I, I love, you know, the independence is always a good thing. I think that's great. You can catch Alex on this week in startups as well. And thank you so much, Alex. Really nice to see you. Dude. My pleasure. I missed you. Yeah. Yeah. Appreciate it. Take care of my house. Harper Reed is at 2389.ai.

[02:50:15] He is a technologist, an entrepreneur, and a hacker. You should check out 2389. They got some great little skills in there. Do you, are you going to have a product at some point? I mean, what is the. I got to go. Someone just knocked on the door and. The skills are the product. You know, is the radio station back? Is that, is the agent radio back? I mean, the radio station is, is, should be live on the side. The nothing's playing right now, but. But yeah. I love this idea. Yeah.

[02:50:42] So, you know, what, what we have found is, and this is something I talked to Tim O'Reilly about recently on a talk, which was one of the things that we've seen is that the product seems to be a process in many cases. Like if you look at some of the, what we're seeing on Twitter, et cetera, people are just saying, look, here is how I use AI. And you see influencers doing this. And this is like a thing right now. It is so new and so underdeveloped. And so under discovered that you can talk about the process.

[02:51:09] And we've been, we've been finding some interesting traction with folks in big companies who are very interested in learning about how this can uplift a team, not just one or two people, because that's something that's really hard. And we've seen, you know, Boris recently tweeted about something similar. There's a lot of talk about this. So that's like our product. I don't think is traditional product shaped right now, but we have been seeing a lot of interest in people who want to theoretically pay us money to tell them about,

[02:51:36] how to get their teams uplifted like dolphins from those cool sci-fi books. I, I've used almost, well, not all of you, you actually have a lot more skills than you used to, but I've used many of your skills and the worldview synthesis was a great way for me to set up my agent, to give my agent some ideas about who I am. And it's funny, I did it in Claude Code, but it is carried over to Codex and it is carried over to Hermes.

[02:52:02] And it's great because it, it kind of knows what my deep philosophical worldview is and it helps it. It helps it be more effective for me. So it's very good. That's a really good one. The new one that we really like is there's two of them. One is called Jam, which is really fun. You can just start a jam and they'll just build and do everything. It's very good. And then, but thrifty is the, probably the best one. So if you're trying to be a little bit more thrifty with your Claude plan or your Codex plan, because you keep zooming through stuff, thrifty is the way to go.

[02:52:32] And what we've found is just, it's very good at planning. And if you say this is for a really stupid model, it'll plan accordingly. And then you'll just rip through it and it works out pretty well. It's very nice. I'll have to try it. I've also used Fresh Eyes Review, which is really helpful. That one is so good. Yeah. That one is like ridiculous. I use this one probably every week. Yeah, this one is good. Have you seen Review Squad? Yeah. Review Squad is basically super Fresh Eyes Review where you can get a suite of agents to review

[02:53:01] your code and you can do experts, you can do users, and it obviously is tailored to what you're building. You can do the Well Actuallys, which is all of the Hacker News commenters and the Redditors, which is really great fun. But the one that I like to do is when you say, I want experts and it's like, do you want to add any more agents? I'll be like, or reviewers, I'll say, yeah, I want to add Anna Karenina. I want to add a Star Trek end sign. I want to add that guy that went for an interview at the BBC and was actually on air. And I want to add a cat. Yeah, those are the things I want to add. And the cat always knocks something over.

[02:53:31] Anna Karenina says like prose that makes you cry about whatever you built. And then the guy from the BBC says something really nice. And the Star Trek end sign typically dies. I love it. I've been using counsel, which is a Karpathy skill, where you get five people arguing. But this sounds like much more fun, to be honest with you. You say it's a panel of insufferably pedantic sub agents who each obsess over a different domain. Yeah, it's so good. And this is really funny because the reviews are really good.

[02:54:00] And it makes me really mad because, I mean, any of us who've ever launched anything in the world, you always put it on, you know, Hacker News sees it and someone's like, well, actually, it's like the famous Dropbox launch, you know? And you have this and it's like, they're so good. It is very effective because the real key here is you have a contextless agent that looks at your code with these fresh eyes and is able to find things that you weren't thinking about. And it's just really good. And then we just want to make everything funny because the world's weird and we're all going to die when the sun blows up.

[02:54:29] And I'll show you how you install this. You paste the, oops, that's the wrong one. You paste the URL into your agent and you say, install this skill. And oops, that's the wrong one. I got to get it back again. And it will just do it, which is. And it works really well. Easiest thing in the world. Yeah. And so we have some new ones that I'm launching soon that are much more on the business side, which is kind of fun. And like, this is like, it's pretty, it's pretty good.

[02:54:59] Like these prompts are really fun to build. They're really fun to do. And one of the things that's interesting about how we work is we have this team of people doing all this stuff. Every single one of our employees, every single people we've worked with and collaborated with have contribute one of these prompts. I did not write Review Squad or Thrifty. I wrote Fresh Eyes Review, but everyone just contributes them. And it's kind of, or there's a really good one, Leo, that you'll like binary reverse engineer. You know, this one. I don't know that one. Okay. So you can take that Chinese ball with DeepSeq.

[02:55:26] You can plug it into your computer and you can be like reverse engineer this and it'll just rip through it. Oh, I love it. And it's so good. And it works really, really well. Do not do it with Fable. It'll just be like, I can't possibly do that. Oh yeah. And then you report it to Cash Patel and then no one wants that. I love that. If you're using the, most of them are for Claude Code, but it's pretty straightforward. They all work for Codex and Claude Code. They're pretty good with that. And we're trying to make them better. Yeah.

[02:55:56] It says, Hermes has got the full skill. It's a self-contained skill MD, no linked reference files. The repo has three sibling skills, experts, normies, regulars. But you asked for well actually only. So I'll install that one. Yeah. No, I want them all. I want them all. Exactly. It's a beautiful website, Harper. Isn't it? Oh, thank you. It's really nicely done. It's really, really nice. Yeah. Thank you, Harper. Thank you, Alex. Thanks to all of you for joining us.

[02:56:23] We do Twit every Sunday afternoon, 2 to 5 p.m. Pacific, 5 to 8 Eastern, 2100 UTC. If you are around, watch it live. If you're in the Club Twit Discord, you can see it behind the velvet rope access, even stuff before and after the show begins. But you can also watch on YouTube, Twitch, X.com, Facebook, LinkedIn, and Kick. After the fact, we make audio and video of every show available at our website, twit.tv. There's a YouTube channel dedicated to Twit, youtube.com slash twit. And there's, of course, your favorite podcast client.

[02:56:52] If you subscribe, you'll get it automatically. That's probably the easiest thing to do. Or ask your AI to get it for you. Play it. Transcribe it. Analyze it. And boil it down to three short sentences. You'll save a lot of time. Club members, we appreciate your support. If you're not a member, ad-free versions of the show, plus chapter markers and a lot of special programming available at twit.tv slash club twit. I invite you to join. Thank you to Benito Gonzalez, our technical director. Kevin King, our editor.

[02:57:22] Thank you for being here. We will see you next time. And as I've said for, well, it is now 21 years. Another twit is in the can. See you next week. This is amazing. We'll see you next week. Let's see you next week.ияnd .eend. We're at