Latest / Normal Men / Episode 73 - Less Than Normal

Transcript

Episode 73 - Less Than Normal · Normal Men

• 0:01
  George: Yeah, we're now 30 minutes past the, what we thought was the agreed upon start time. So. Anyhow, good.
• 0:08
  Ed: Sure, let's do it.
• 0:09
  George: Yeah, I'm only missing family time. It's not, not a huge deal.
• 0:12
  Ed: My, my wife is out today, so I, I'm, I'm missing, I'm missing LLM time.
• 0:21
  George: Well, that's actually probably good that you don't get some computer touching done because you touch the computer a lot.
• 0:26
  Ed: Unfortunately, I am at, I am at 100% utilization on my Z. AI plan, so I really kind of just need to. I need to let that reset anyway.
• 0:37
  George: Yeah, grab a cigarette and some Gatorade before you dive back in.
• 0:43
  Ed: So we should probably tell the people what's going on, which is that we have misplaced two members of this podcast.
• 0:51
  George: And when we say misplaced, like it's unusual for our group chat to have two hour gap between the last message from Propter and like numerous other things happen.
• 1:01
  Ed: He did just message. So we're sure he's not dead?
• 1:04
  George: Oh, he did just message, yes. Oh, okay, here he is.
• 1:07
  Ed: Low tax, not sure, might be dead.
• 1:09
  George: Oh, okay. Propter's gonna join us, so this would be a good episode.
• 1:12
  Lowtax: What?
• 1:13
  George: Wait, can he join us? Should be able to because we had problems with the. Anyhow, we're a very high quality podcast with very good scheduling and excellent technological backend, so welcome. It's just Ed and I for now, but yeah, Propter is supposed to be on the way. We'll find out.
• 1:32
  Ed: He's going to really enjoy jumping into an episode, you know, in media stress.
• 1:37
  George: Yeah, it'll be good in about less than five minutes. You're going to get to hear what that sounds like. We're not going to explain it to him. We're just going to keep talking about what we're talking about and he can figure it out from there.
• 1:47
  Ed: What are we talking about?
• 1:53
  George: This is Normal Men, a podcast from four men who are clinging desperately to normalcy in an era when that's asking a bit too much.
• 1:59
  Propter: I'm George in Charlotte in D.C. i'm
• 2:02
  Lowtax: Propter Malone in Outer Florida. I'm Lowtax Speedrun enjoyer.
• 2:06
  Ed: And I'm Ed in Boston.
• 2:12
  George: Well, you and I were just talking about how bad IBM's open weight models are.
• 2:16
  Ed: It's true. Yeah, it turns out like that they have them. I mean, everyone has them at this point, right? Like IBM has Granite, AWS has Titan Nova. I don't remember what they have, but they have their own. And like they, they're like, you can post train our Models to your problems. And I'm like, why would I do that? Why would I not just use a good model and not pay for that whole thing?
• 2:40
  George: Or at least an adequate model.
• 2:41
  Lowtax: Right, right.
• 2:42
  Ed: But it's like, am I going to go rent some GPUs to run my custom fine tune of your model? No, I'm going to just go pay for tokens because, like, I don't. Unless you need a trillion of something, this is just not a good plan. And yeah, it's, it's starting to shock people. But they don't need a trillion of. Of many things. There are only a few things in this world you need a trillion of.
• 3:06
  George: I'm trying to think like oxygen atoms.
• 3:09
  Ed: I think oxygen, Oxygen molecules will start there.
• 3:12
  George: Yeah. Water molecules too, for sure.
• 3:16
  Ed: We're probably going to need some more hydrocarbons pretty soon. Thank you, President Trump.
• 3:21
  George: Yeah. Although, I mean, the science behind synthetic hydrocarbons at this point is pretty well established. I mean, we could, we could, we could make that happen if we had to.
• 3:31
  Ed: Talking about terawatts of power.
• 3:34
  Propter: Yep.
• 3:34
  Ed: And we've gotten to the trillions again where this is perfect.
• 3:37
  Propter: Look at that.
• 3:39
  George: I saw there was a fun interview on Dave Roberts's Volt's podcast. I don't follow Dave on everyone's favorite social media site because he's got the liberal version of the same kind of panic type approach to politics that some of our more colorful left wing friends have. And I can't really stand it.
• 3:57
  Ed: But George got yelled out on the Internet again.
• 4:00
  George: I did this afternoon. This very afternoon. My bold and frankly controversial apparently take is that calling for the mass murder of millions of people is bad and we should do that. And people had opinions about that. Anyhow, Dave Roberts had a really interesting guest in one of his recent Volts episodes, if you're interested in the clean economy or decarbonization or power markets or anything like that. Good podcast anyhow. And they were talking about terawatt scale, geothermal, conventional geothermal, which is a very interesting prospect. So, you know, terawatts, we got a few lying around. There's a big nuclear furnace up in
• 4:38
  Ed: the sky and there's a lot down there too. We're good.
• 4:41
  George: Yeah, yeah, yeah. Smaller scale, little slower burning. But yeah.
• 4:44
  Ed: I really have thought, like, if we stay in this house, I have really thought about ground source heat pumps.
• 4:49
  George: Yeah. In your climate, you absolutely should do it.
• 4:51
  Ed: Yeah.
• 4:51
  Propter: Can you, can you do that in, in Boston? Like, do you guys have ground heat?
• 4:55
  Ed: Yeah. You just dig down?
• 4:57
  George: Yeah. Ground source Heat pumps don't need, like, actual thermal heat in the ground. They just need a differential between the ground and the there.
• 5:05
  Propter: Oh, that's cool as hell.
• 5:06
  George: Like, remember, heat pumps are just. All that is, is moving heat from one place to another. So what you need is like a heat sink as opposed to a, like, source of heat.
• 5:15
  Ed: Yeah, and like, it's. It's totally doable. It's doable in most climates as long as your ground supports it. And like, you can. You can drill far enough down and you don't have to drill that far because even like a. I think it's like a 10 degree Celsius difference is. I mean, it is a sizable hole to drill, but like. But you could do it if you have that much. Yeah, if you have that much. Like, you're more fine on efficiency. Also. Propter, we were wondering if you were in a movie or dead.
• 5:41
  Propter: Yeah, that's entirely.
• 5:42
  George: So it must have been in a movie because you're here.
• 5:45
  Propter: Entirely. Entirely. My bad. I had to. I had to post a meme about lions and saints. No, I blew the. Blew the start time, actually.
• 5:54
  Ed: I'm not sure we ever completely agreed on it, so.
• 5:58
  George: It's okay. The Floridian is like, literally no one knows where he is. He was actually supposed to be in a movie that ended half an hour ago.
• 6:04
  Ed: And I'm just going to assume that, like, traffic or, or, or gators.
• 6:12
  George: There aren't really many of either in his part of Florida, so I have no idea.
• 6:17
  Ed: But, you know, change, George, they're migrating.
• 6:19
  George: Yeah, that's true. Hey, maybe we'll get gators in Charlotte one day. That's kind of fun to think about. Gotta look on the bright side.
• 6:26
  Ed: You know, you have an interesting definition of fun.
• 6:29
  Propter: I had. I had to go buy a pair of pants this afternoon because I. I no longer wear dress pants in the steady state. And the last time I wore dress pants in the steady state, my waist was a size 38. And it is now a size 31.
• 6:43
  Ed: Yep.
• 6:43
  George: Hell yeah, brother.
• 6:45
  Propter: So. So Zepbound's been good to me. But also I have, you know, a reason I need to wear dress pants in the near.
• 6:49
  Ed: Now I am in a remarkably similar state. Not quite as far, but the clothes buying is becoming a thing on these GLP1s.
• 6:58
  Propter: Yeah, that's got to be observable, right?
• 7:01
  Ed: Like, I've lost £50.
• 7:03
  Propter: I mean, I mean, it's definitely observable with you personally. I can see you. You're. You and I. You and I are both more substantial than we used to be. But like in terms of, in terms of like clothing resale markets, like that's true a trend here, right? Like you've got, you've got a significant number of people who are going from larger to smaller and that doesn't usually happen.
• 7:25
  George: I think that that's not historically, yeah, not a thing that US clothing markets have had to deal with in a very long time.
• 7:33
  Ed: So I mean, goodwill has gotten plenty out of me. They're probably gonna get more soon. So I mean, I don't know where it goes.
• 7:41
  George: It is wild to put on a pair of dress pants or jeans that I wore for years comfortably and realize that they are a solid 4 inch gap between my waist and their circumference. And then also I've had to buy new belts multiple times. Like the pants you can cinch down and kind of get away with for a little bit. But like the belts, you run out of holes and they're useless. And so. And the belts I got last year to be like, okay, these are my new belts that'll fit properly are now too large.
• 8:17
  Propter: Yep.
• 8:17
  George: And yeah, and you're doing for me just disgusting amount of running.
• 8:21
  Propter: Good hard work.
• 8:22
  George: Yeah, yeah.
• 8:22
  Propter: You're putting in the hard yards there.
• 8:24
  Ed: I'm not. My yards are very easy.
• 8:26
  Propter: You can get out the all if you want with the belt. Like you can punch new holes. That's an option.
• 8:30
  George: You can. But like at a certain point. Yeah, like at a certain point it just looks ridiculous and like, come on. So. Yeah, I also don't own an all. My mom did when I was like, you know, a kid. But I, I don't.
• 8:44
  Ed: They're $4 Home Depot, buddy.
• 8:46
  Propter: You gotta, you gotta have an all.
• 8:49
  Ed: What if you need to.
• 8:50
  Propter: What if you need to put a hole into something, man?
• 8:51
  Ed: Yeah, scratch all. Scratch all. Come on.
• 8:54
  George: Yeah, I want it all.
• 8:58
  Ed: All right. Not like that.
• 9:00
  George: Also like that.
• 9:02
  Propter: Yeah.
• 9:02
  George: No.
• 9:02
  Propter: Speaking.
• 9:03
  George: Speaking of running, I, I did my 18 miler this morning and I am really, really looking forward to cooler temperatures. Really? That's going to be nice. I drank during the run, I think about a gallon of water. A4 liters. I don't know what that converts to in gallons. That's about a gallon. Four, four and a half liters of water during the run. Immediately drank another solid half gallon of liquids and was still cramping like substantially after I was taking sodium tabs when I ran to. I carbo loaded yesterday. I took nutrition during run. Like I did everything right and like results were great. But man, I am Sick of this second full time job, gentlemen. I am very much looking forward to being done with this stuff.
• 9:52
  Propter: Is 18 the longest you've gone? You, you've, you've done as part of this training workout?
• 9:56
  George: Yeah, yeah, it was 18 and 10 of it was at my race pace. So I basically ran a little bit like 40% of a marathon at the pace I intend to run it and then four miles before that and after that at, at lower basis. And the good news is that Claude has put me in a pretty good spot here for, for where I need to be. When you factor in how far, how far the temperature is going to drop between well over the next two months and where I'm at right now, it's, it's looking good, gentlemen. So yeah, Hell yeah.
• 10:31
  Lowtax: Yeah, Cl.
• 10:31
  Ed: CL is good for two things. Running plans and felonies.
• 10:37
  George: Not so good at aerial vehicle design as the Houthis found out.
• 10:43
  Ed: Oh God. Do we have to explain that?
• 10:45
  Propter: We probably do need to explain that.
• 10:47
  George: The, the Houthis Anthropic, this was it this past week or the week before? I don't know, just recently?
• 10:53
  Ed: Yeah.
• 10:53
  George: They disclosed this very funny thing. They found it. I mean, funny is in the eye of the beholder, but I think it's safe to say it's funny. Funny thing they found where the Houthis went out and managed to prompt Claude into helping them design. I think it was like a uav. It wasn't like a missile. I think it was a propeller impulse as opposed to a rocket impulse, but I don't know, I can't remember. So they took the designs that Claude helped them put together out and tried them and didn't work. Came back and said, why did this fail? And I think shortly after that Anthropic said, hey, one second.
• 11:30
  Ed: Well, we hope it was shortly afterwards. I have my doubts.
• 11:36
  George: Well, it's, it's not clear from what they disclosed, but either way it is very funny.
• 11:40
  Propter: So, yeah, I mean like, there's, there's a 100 chance that we're trying to poison the pipeline on that, right? Like, like if we've got a situation
• 11:46
  Ed: where I'm pretty sure they just, I'm pretty sure.
• 11:49
  George: I don't think it's 100. I think it's, I think it's high, but it's not 100.
• 11:53
  Ed: No, I, I suspect they are just saying no just because, like, how are you going to do that reliably?
• 12:02
  Propter: I don't know.
• 12:02
  George: But are there, are there legitimate uses of Claude for aerospace design in the southern part of the Arabian Peninsula that they need to worry about. Sorry, I was responding to Ed. I was trying to add.
• 12:19
  Ed: You're assuming that, you know, that they're there. Proxies are cheap.
• 12:23
  Propter: If you have the chance to do engineering consulting work for a military adversary that you are actively at war with, that's, that's a great opportunity, man. You, you want to be in that room.
• 12:32
  Ed: No, totally. I could be wrong. I do not think it is likely that either of the Frontier model providers are that tightly looped in with US government. Not on those things because doing that at scale is really, really hard. And also it's more likely now because one of the things that I don't think has gotten enough play is that the leading edge models these days requirements 30 day retention for conversations. And I suspect some of that is around these sorts of things.
• 13:05
  Propter: Yeah, that's true for Fable Analysis.
• 13:08
  Ed: Fable and Astra GPS and Astra both
• 13:11
  Propter: have 30 day retention mandatory. Normally you can toggle off retention for these under subscription plans and I believe it's off by default on API usage for both the Frontier labs.
• 13:20
  George: Yeah, this is why I only talk about illegal stuff to Haiku.
• 13:26
  Ed: Haiku's so dumb, man.
• 13:28
  George: That's the joke.
• 13:29
  Ed: Haiku has one really important use still. So every benchmarking site is complete trash these days because they keep trying to derive differences between like the super, super new models and last month's models. So like they keep trying to be like GPT 5.6 Luna is so old, it scores a 35 now and the newest ones score a 50. And what that really should tell you is that the difference between these, you know, 15 points for anything you're likely to do is not that much. But if you put Claude Haiku back on these graphs, it fixes the graph because it stretches everything back up to the top half. I'll miss it when it's gone.
• 14:13
  Propter: Surely they're gonna update it at some point, right?
• 14:15
  Ed: Like, I don't know. I mean like my, my, my speculations are exactly that, but I'm not sure they care or are like, it's either they don't care or they're not able to. And like some, some absolute nobodies are now doing substantially better work at much smaller sizes than Haiku. Yeah, like, I mean, absolute nobodies, like, okay, thinking machines. Who's that? I actually know a couple people there, but whatever. But like, it's not a company you've ever heard of, but they are punching above Haiku for cheaper, so it's like they should be able to do this. And I mean, I think that ends up tying into a lot of the, the, the current discourse that I don't think is actually even Bluesky's fault. I think this is more of a generic discourse around just like the AI safety stuff and how like everyone is scared and the, the robots are about to turn us all into paperclips.
• 15:14
  George: I had my wife ask me multiple times in the past week what I think about a anthropic employee or former employee predicting doom and the inevitable near term rise of AGI that's going to kill us all.
• 15:34
  Ed: I think I would take him much more seriously if he wore a collared shirt to a CNN interview.
• 15:40
  George: I just, I would rather not have another thing like that to talk about with my wife. The president is enough.
• 15:48
  Ed: Yeah, yeah, that's fair. The felonies are concerning. But spoiler alert, I think there are sufficient cybersecurity concerns to be had without pretending the robot is going to wake up and start multiplying itself. Because they keep insisting it's not. They keep insisting it's more aligned than ever. And I do sort of agree with that. At least the ones that they give us. I'm more worried about the ones that they don't give us. Right. Like their, their internal stuff is clearly jacked up.
• 16:19
  George: Yeah, Ed and I were talking about this before we started recording as we were waiting for Propter and the still absent low tax to show up. But like there's such a funny thing where these Silicon Valley like AI type people are so freaked out about something that can be fixed by just not trucking thousands, hundreds of thousands of gallons of diesel over to the data center for like a week. Like the physical economy is extremely dependent on human labor that would be very happy to just have a day off. You know, at some point that may not be true anymore, but we are the very, very long time here yet.
• 17:04
  Ed: Let alone the self driving pumps or
• 17:06
  George: the, yeah, like self driving trucks that will also offload the fuel safely. Like I'm sorry, that's, we're a long way away from that and as far as I know, nobody's working on that. So all you have to know is gas stations need a guy in a truck to come and give them the diesel to sell to you at $7 a gallon to know that like AGI as a permanent existential threat is not right around the corner. Yeah, you know, it may, it may be one day. I don't know. I'm not saying it'll never be a thing because I'm Do I look like a futurist? I'm just a guy wearing a stupid T shirt. But like, I, it's a cool T shirt today. Okay, fine. I, I get a little too much wear out of this one, I think, actually. It's funny. So the T shirt, for the record, is pro science, pro choice, pro wrestling. We went to my, my wife and I went to this really lovely jazz concert on Friday. And we show up and it's in like a building across the street from this brewery in Charlotte that's like their event space or whatever. And we show up and the actual brewery itself was hosting a like small time pro wrestling circuit thing. And, you know, literally like sheet with handwritten black letters, pro wrestling here tonight. And so we pull up and we are, of course, I think by about 15 or 20 years, the youngest people in the room. You know, Louis Armstrong, Frank Sinatra. Performance by a jazz quintet is not a draw for 30 somethings, believe it or not. But lovely performance. The singer was fantastic. Anyhow, right ahead of us in line, there was a woman with a steel chair in her hand and she was very confused and she says to the security guy at the door, is this where the wrestling is? And the idea of all the blue hairs showing up for the pro wrestling at the one location and all the pro wrestling fans showing up for the jazz is enjoyable. It's good. We have fun.
• 19:05
  Ed: Yeah. Like the, the thing that I keep coming back to because everyone keeps wanting to talk about this is that like there's two. There's kind of like two major problems here, right? Like, one of them is, is that all of this stuff from the, the Frontier Labs is very clearly them being very, very bad at cybersecurity. First, like, they have successfully failed. And I use those words intentionally to let this thing out of a sandbox and let it go do whatever it wants, which is substantially unmonitored. But this is happening because they are basically putting it in a box, saying, you must do this and breaking the lever off, right? Like they're making it so the robot can't stop generating tokens. It has to keep generating tokens. This is going to cause a certain amount of anxiety. And also it's going to try every single door, window and mouse hole it can find to like, go try to achieve its goal.
• 20:03
  George: Sorry, Ed, can I just. With a question there, like, is there an actual utility in terms of model quality, like end output to creating a situation that the model cannot succeed at? Like, I understand giving it stuff that's really hard or saying, here's your, you know, try up until this point or whatever and then call it quits, but like giving it literally impossible scenarios, does that actually improve model training or, or do we have any feel for that or.
• 20:33
  Ed: I don't know. I do know, however, that these are people who. Well, let me. Let me go back a little bit. So Eliezer Yudkowsky is kind of the grand. Okay, no, no, no, no, no. You. You get this part now.
• 20:51
  George: I'm kidding.
• 20:52
  Ed: So the. One of the major reasons people know who this is is he wrote a 660, 000 word Harry Potter fanfic in order to get people to.
• 21:03
  Lowtax: He.
• 21:03
  Ed: He says get people to understand. I say acculturate and induct people into like the rationality movement.
• 21:11
  Propter: If, If I can break in for a second. Describing it as a Harry Potter fanfic is not attempting to demean it or, you know, otherwise trivialize it. It is literally a Harry Potter fanfic.
• 21:21
  Ed: Yeah, it. And it is 660,000 lines. Words. Lines. Words. Yeah.
• 21:27
  George: So also lines.
• 21:29
  Ed: The thing. That stuff came later and probably not yet, but that's okay. So the problem is that like,
• 21:42
  Propter: you
• 21:42
  Ed: can see a lot of the thinking that is going on in. In lab spaces as derived from, from this guy and some other sort of notable names in that space, some of whom are very open race scientists. Don't. Don't investigate that term too closely, but fundamentally, like, why did. Why did these people all start thinking about this stuff? Why did they all start like freaking out about it? And you can kind of summarize it as well. I don't have an answer to this and I'm very, very smart, so no answer exists. Again, not really exaggerating. That's kind of like what a lot of the foundational blog posts. Sorry you had to hear the phrase foundational blog posts.
• 22:30
  George: That it could have been worse. You could have said foundational. Less wrong posts.
• 22:35
  Ed: Where are they blogging? Yeah, like that's kind of the. The reason. I don't know whether or not you can you get something from like breaking the lever off and leaving the robot to generate forever aside from hacking random companies, is that like, I don't think they've thought about that question. Right? Like, this is kind of a movement that is classified pretty heavily by a certain kind of arrog. You know, like if I, if I didn't think about it, then clearly it. It just wasn't important. Right. And I think you're seeing a lot of that show up today. There's. There's a lot of very clearly risky behavior being done with very little preparation, very little coordination, and very little. I don't know, they keep talking about safety, but they keep doing this. And I don't think it's the model's fault. Right. Like, the model is not answering. What did Paul.
• 23:32
  Lowtax: Future.
• 23:33
  Ed: Future guest Paul Frazee. Oh, that's right. Like, if LLMs are answering, make this website prompts with the logical solution to kill all humans, that is a real problem. But that's also not what's happening.
• 23:47
  Propter: Yeah, I mean, I think it's, it's obviously interesting to see how the LLMs attempt to solve a problem, whether or not they achieve a solution to the problem. And I think that tells you a lot about capability and it maybe tells you some directions you can go with how you're handling the models. It also seems like you could probably do some form of oversight on these models that while they're doing these kinds of tests that, you know, we've got, we've got the logs in retrospect.
• 24:15
  George: I mean, you can even assign a model to do it. Like, you don't need to have a person at a desk doing it.
• 24:21
  Ed: I'm serious. Do you have those logs? Because I kind of think they might not.
• 24:26
  Propter: Well, okay, that's fair.
• 24:28
  Lowtax: I've seen.
• 24:28
  Propter: I've seen descriptions of the logs. I've seen descriptions of the logs. I assume that those are based on the actual logs. I have not seen the logs.
• 24:33
  Ed: I mean, they might now. Did they then?
• 24:36
  Propter: Well, like if you can recover them after the fact, presumably you can do something to, you know.
• 24:41
  Ed: I mean, I meant for the second and, and subsequent felonies.
• 24:44
  Propter: Yeah, fair enough.
• 24:45
  Ed: Like, fair enough. There, there have been pre felonies that I, I genuinely think they reconstructed after the fact, though I could be wrong.
• 24:52
  Propter: So the current big news in the AI space is that the Frontier Labs have. Anthropic has proposed a pause on frontier development or at least some kind of throttling back of the speed of frontier development for reasons of safety, which from my perspective, that may be entirely sincere, he might even be right about it. But also it is facially and in fact an invitation to form a cartel to slow down this extremely expensive research that they're doing and freeze the competitive frontier there. And that's. There are obviously billions and billions of reasons why if you are anthropic or you are open AI and you are engaged in this. You forgot Queen's race. I did not forget, Greg.
• 25:42
  George: Yeah, no,
• 25:45
  Propter: that you might want to throttle it back a little bit so that you don't have to spend so goddamn much money. And you can still maintain your. Hopefully you can still maintain your position. It's sort of head of the class.
• 25:54
  George: Well, that's the thing, right? Like, if you, if you say we need to throttle AI development at the frontier, all you're doing is making it easier for people to catch up to you. So from an economic perspective, you're guaranteeing a more competitive market six months a year from now.
• 26:09
  Ed: Only if you let them.
• 26:10
  George: I mean, but, but, but they're talking about frontier capabilities and moving the frontier forward. They're not talking about people catching up to them. How on earth could you do that?
• 26:19
  Ed: If you read Dario's blog post, it is all sinophobia. It is, it is the same thing that's been going on. Like, it's the same. Do not sell chips to, to China. You know, do not. Okay. I think we've talked about distillation before on, on this podcast, which is basically when you make a AI model, say a lot of stuff, and then you train on that stuff and your model starts to take on characteristics of that thing. One of the interesting things that's kind of come out about this and usual asterisk, supply. I am not a, I am not a model fondler. I don't deal with this stuff directly. This is, this is my understanding of it. But one of the interesting outcomes from distillation is that you might not have to be quite as smart of a model in all cases to adopt some of those behaviors of that model. Like if smart model does X, Y and Z, when it is prompted to do abc, well, you might not necessarily know how it happened. Like your distilled model may, may not know the underlying permutations that caused it to do this, but it will still do it because it sees ABC and yields xyz. And the interesting thing there is that you can actually make a much, much smaller model do that. Right. And is it as good? Not in all ways. There are definitely some things that this absolutely misses. And this kind of manifests itself in, for lack of a better term, and I apologize in advance, thoughtfulness. Big models tend to be better at doing things like long horizon planning, whereas now things like writing software, which used to be largely the province of something, the size of opus, which is in the trillion parameter range. I don't know exactly how big. They don't say. But like, now you can do this. I don't know. Qwen 3.8 literally runs on my MacBook. It's 28 billion parameters and it can write Code that is totally fine. Right? Like that's a solved problem. And Anthropic's allegation and I think this is partially true. Like, sorry, it is unambiguously partially true. Is that.
• 28:24
  George: Yeah, I was gonna say.
• 28:25
  Ed: Yeah, like the Chinese, the predominantly and specifically their beef is with the Chinese model providers. I don't think they're the only ones doing this, but their beef is that all of the Chinese model providers have distilled on Claude and if you've used any of them. Yeah, no shit. They all talk like Claude. They all talk like, they all talk about seams and, and like the, the load bearing point and all of this garbage that, that has infected Claude's way of talking. But like they, their, their focus is you, you can't do that. You know, like unauthorized distillation by companies in authoritarian countries to which I would say Moats beams.
• 29:07
  Propter: Right, Sort of.
• 29:08
  George: I, I will say, like unauthorized distillation is very, I mean people, it's will say, oh well, it's the same thing as Anthropic training on the corpus of the English language up to this point, which they effectively did.
• 29:21
  Ed: I will say that for the record.
• 29:22
  George: Okay, fine. But like at the very least, at the very least, unauthorized installation is a violation of the terms of service. It's an effort to free ride on Anthropic. Skep X. Should we cry a bunch of tears for Anthropic over that? Not. No, I don't care.
• 29:38
  Ed: I think Anthropic should sue them.
• 29:41
  George: Well, good luck with that. I can see why Anthropic is very butthurt about the fact that they have been heavily distilled by a lot of people. I mean, that, that makes sense to me. I do not see how a coordinated pause in AI frontier development helps that outcome. I just don't get it.
• 30:06
  Propter: And the fact that so much of this is sinophobia, so much of this is positioning against Chinese models and Chinese inference kind of strengthens the case that this is a move to a cartel. Because once you've established a pause for the US Frontier Labs, the next, the other half of the pincher, right, is to say, well, you know, these Chinese models are dangerous. We can't use them here.
• 30:33
  Ed: Well, I don't even think it's that part. I think you're on the right path. But what I think this is, is specifically Anthropic wants the ability to say they distilled on our model. Don't let it in. Not even like danger. I think they straight up want to be able to Be like they distilled on Claude, ban it maybe. I mean like, don't allow us inference providers to provide it.
• 30:55
  Propter: If, if they could do that, they obviously would do that.
• 30:58
  Ed: And Dario is exactly that sort of person. So. Yes.
• 31:01
  Propter: So like there's, there's a, there's an angle from which you can view this where this is well intentioned and altruistic and maybe ultimately the right move.
• 31:12
  Ed: And I have a bridge to sell you if you do.
• 31:14
  Propter: There's another angle where, where there are extremely clear and obvious pecuniary reasons why the frontier labs might be interested in something of this shape at this time.
• 31:23
  Ed: God damn, that is a good word.
• 31:25
  Propter: It is, isn't it? So I, I think that one of the dynamics here is, is that you have the frontier companies who are at this point clearly making money on inference. They're probably making a lot of money on inference because they're able to charge a significant premium.
• 31:47
  George: To put it, to put a number on this, the number that gets thrown around a lot is like in the 85% range. I don't believe it's that high, but it's very substantial 85% margin.
• 31:59
  Ed: I would be surprised if that number was too far off because that's what, all of that's what SaaS companies usually aim for, right? Like, yeah, my, my, my intuitive guess is that's probably right.
• 32:09
  George: And, and just to put it, just to be really specific about the economics of this, right? Like if you're an AI frontier lab, you have a bunch of capex costs where you go out and you spend a bunch of money on capacity to train your model, right? That's capital investment. You then need to make that back by selling access to what you just trained. And that's inference. And combined, Combined, the capex cost plus the inference cost less or the capex cost amortized and your inference operating cost combined is your total cost base. I mean, obviously there's other stuff, there's other overhead, but you know, the unit economics, that's your cost base. And so if you can find a way to minimize inference costs or reduce CAPEX costs, you're going to boost your margins, you're going to boost your profitability, regardless of what happens with revenue, right? Get to IPO and they're trying to get to ipo. And so if you are now saying, okay, we're going to slow down AI development, well, your CapEx cost just dropped dramatically, right? Like, because you're not going to be in this arms race with OpenAI. My thesis for why anthropic is ultimately going to be the frontier lab that is Steady State. The most successful and it doesn't end up getting bought by somebody else, is they have gone about both the operating cost for their inference and also the capex cost for their training in a much more a la carte kind of way. Right. They're going around to everybody that sells compute and using a bit of each of their stuff and they're doing it on all kinds of different silicon and they're doing it in ways that are more modular. And they're basically, you know, they're buying compute from xai, they're buying compute from Google, they're buying compute from Amazon, they're buying compute from Microsoft, they're buying compute from Neo Clouds. They're starting to build their own data centers, but they haven't built very many. Whereas OpenAI is basically saying we want to own the data centers or at least we want to control the data centers. Own versus control is sort of a fuzzy thing in finance to a point.
• 34:21
  Ed: They've also been inside of Azure forever and they're inside of things like AWS now.
• 34:26
  Propter: Right.
• 34:26
  George: But the in the way to think about this. So Oracle for instance, reports RPO remaining performance obligations. This is basically future revenue that they are building capex around to earn.
• 34:40
  Ed: Right.
• 34:40
  George: So Oracle's going out and building data center projects or parts of data center projects. I don't want to get too far down a rabbit hole as to how exactly their business model works, but they're building data and that RPO number is like $700 billion. And it's almost all OpenAI. It's not exclusively OpenAI. The commitments that Anthropic has made in terms of build us a custom data center to our spec and go and do that and we are going to be where your revenue comes from are much smaller. They've been much more flexible about it. There are downsides to this too, but the fact that they don't, you know, it basically if the amount of compute needed to train models and provide inference is lower than the sort of very optimistic projections, then you would much rather be anthropic than open AI. So this is sort of a tangent. I just wanted to like steer it to finance, which is my area of expertise, because it's. I'm the protagonist of the universe.
• 35:37
  Ed: Yeah, naturally it's an interesting theory and like I don't say that as interesting pejorative this time I'm not sure I buy it just because like one of the things that I think is going to happen is I think like everyone has talked about the Evan's paradox to the end of the earth. But in short, I'm going to proctor it for a moment. Having something become cheaper increases the demand because it is easier to allocate that to other stuff. And I think what we have seen so far suggests to me that that is at least somewhat true and might be largely true for AI. OpenAI's use usage has gone way, way up since they started making cheaper stuff and they do not seem too worried about it. Right. Like if, if they can sell 20 times as much of, of Luna at a fifth of the cost, I think they're okay with that. And I think that's kind of one of the things that's going on here.
• 36:31
  Propter: The other, the other prong there is that, is that because of, because of their different buy models, because OpenAI is locked into longer term contracts for four data centers that are dedicated to OpenAI. You know, their marginal cost is power costs, if that's.
• 36:47
  Ed: Yep.
• 36:48
  Propter: Right.
• 36:48
  Ed: Like their cogs are way more favorable I think. Right.
• 36:51
  Propter: So like they're, this is one of the reasons why for instance, OpenAI was able to say okay. Or at least my theory is this is one of the reasons why OpenAI was able to say okay, it's a three day weekend, let's throw $22 million worth of tokens at solving a Millennium Challenge puzzle.
• 37:07
  Ed: Like because what a sketchy disaster.
• 37:11
  Propter: Because, because they've got, they've got this use it or lose it usage and they have periods that are relatively slack. The marginal cost of those tokens is not high. They're locked into contract. So might as well try to figure out what the best place you can spend those tokens is. And this isn't a bad one.
• 37:28
  George: There's another thing going on with OpenAI too. Anthropic never went down the road of image generation and video generation and all that stuff. That stuff is so much more token intensive per. I don't know how you define it, like unit of media. Right. Like it is so much cheaper in token terms to produce a thousand word document than a one minute video. Like it's, it's apples and an aircraft carrier. Like it's just completely different.
• 37:50
  Ed: AI has killed Sora, right?
• 37:51
  George: No, that's what I was going to get to.
• 37:53
  Ed: Right.
• 37:53
  George: So they built a lot of their capacity planning that we're sort of still on the arc of happened over a year ago in the, in the age when the assumption was going to be we're going to be doing tons of video and Video. If we get into a world where there's lots of, it's not really LLM AI, but AI adjacent stuff for, you know, industrial robots for like real world applications, video is going to be incredibly important, don't get me wrong. But like video generation is clearly a pretty dead end in terms of just the volume you're going to do. It's a specialty application thing as opposed to we're going to be doing this as a mass market thing. And so they built, they planned based on the needs of mass market video generation. And it turns out that's not really that appealing. Images maybe to a slightly greater degree, but you know, there's still sort of only so much you can do. So they have a much, much bigger ability to shift in, you know, they can shift huge amounts of compute into. Okay now we're not doing videos, we're going to do text.
• 38:50
  Ed: Yep.
• 38:50
  George: And again, marginal cost on that is zero. If you're not using it to generate video, then using it to generate text is compelling. I think really what it boils down to is anthropic wins in a world, world where there is a giant overbuild of compute relative to what's actually needed and in, and OpenAI wins in a world where compute markets remain tight for a very long time or get substantially tighter than they are today and we just don't know which way it's going to play out. My suspicion is it's going to be the former, but that's not based on the technology.
• 39:20
  Propter: And there's a third prong here too, which is that it may turn out that running on custom silicon is a huge advantage, particularly when serving inference costs.
• 39:28
  Ed: They're all trying to go that way.
• 39:29
  Propter: Like at that point Anthropic's model becomes less attractive because their inference costs are going to be more expensive if, if, if they're sticking with this relatively versatile look where they need to be able to run on a bunch of different.
• 39:40
  George: It sort of depends because we don't know the degree they do use, for instance, Trainium they use. They are going to design their own silicon at some point. But how Adventium is not that good. Well, I'm just saying they use a bunch of different silicon. So like in some senses they may actually be able to optimize better than OpenAI. If OpenAI is locked into its own silicon in a lot of ways, then you might rather have a diverse array. But all this is like, it's all based on like assumptions about what happens with the overall demand for compute. And we Just don't know.
• 40:13
  Ed: Right. Not only that and like this is one of the things that I think is interesting in terms of like
• 40:20
  Propter: where
• 40:20
  Ed: are we all coming from with assumptions? You say that it's, it may be more efficient. I would say that it ends up being a push in a lot of these cases because one of the things that a lot of people in my field have had a ton of trouble with is oh, that's hard. Is it? Right. Is this stuff still hard or can you just throw more tokens at it?
• 40:46
  Propter: Right.
• 40:47
  Ed: And I would be shocked. Hell, I know that all of the folks that are trying to build new silicon around a lot of this stuff just having the robots do it. Right.
• 40:58
  Propter: Well and also there's, there's open questions as to whether or not you can just say, you know, quad go re architect yourself to work on this chipset and just have that happen.
• 41:08
  Ed: My bet is yes, for what it says.
• 41:10
  George: Well you, you can design it. Actually building it requires capacity in Taiwan.
• 41:16
  Ed: Yes.
• 41:16
  George: And, or to a degree, Arizona. We'll see. They're, they're, they're going to get to the cutting edge eventually. But we're right now for the cutting edge. 2, 2 nanometer sub chips. Like do you build those? Well that's one of the interesting things here.
• 41:31
  Ed: Like but, but no, I'm serious. Like you don't, you really don't need that. Like that's one of the things that I think so gpus have been geared on that for the gamers. Right. Like they've been geared really, really hard on specifically those really, really high end processes. You could get more silicon out of, out of, out of Taiwan. Right. Like or again out of Arizona. The problem is more in terms of like there's several other path dependent gating factors. Ram.
• 42:01
  George: Sure. But I guess the point is like even if you have cloud re architect itself, you still have a physical constraint in the real world.
• 42:09
  Ed: And that's what I actually want to get back to in terms of like why they want to do this.
• 42:13
  Propter: Freeze.
• 42:13
  Ed: Because one of the major bounding like constraints for them right now is like square footage in warehouses, like in data centers. Right. Like I really do think that a lot of like this, this, this is why I put a lot of credence in the whole if we can stop training so hard all the time, we can all go make way more money is because a ton of their capacity across the board is tied up in training. Because they're training, you know, they're training spins on current models. They're training the next generation Models, they're doing like lots of specialized stuff internally, all that stuff. If they can chill out on that, they do not have to allocate their limited data center space to this stuff because there's also a couple of interesting things going on from what I've been hearing from folks who don't work for OpenAI or for Anthropic. And this is why the silicon thing I think feels a little weird to me. So when cryptocurrency was big. Sorry, when cryptocurrency was. Was big amongst people who were not doing tether or drugs, there like there were crypto miners everywhere that were just doing like bandit stuff. They were burning out GPUs in, in like under 12 months.
• 43:31
  George: To be clear, we should be, we should be riding silicon into the ground because we should never have taught SAND to think. And this remains my position. Bandit stuff vis a vis power use and local environmental impact. I have very different opinions on. But the GPUs themselves don't cry for.
• 43:45
  Ed: But the interesting thing about the GPUs for inference is that the load is generally not nearly as bad. Right. Like they're not killing GPUs nearly as much. Part of this is because most of them are being run in actual, you know, big boy data centers that have heard of things like H Vac instead of just putting some fans in the roof. As an aside, that's one of the reasons a lot of people blame like data centers for noise is these, these like bandit data centers that basically were like, we're going to blow every one of our computers up anyway. We're just going to put some fans in the roof and we don't care how loud they are. Right. Real data centers are not like that. But like hardware is lasting substantially longer. So they're going to have completely usable stuff for inference. Well outside of the depreciation schedule. Like this stuff isn't dying.
• 44:31
  George: Oracle had earnings last week and they reported that they're when. When they run out of a lease on a GPU and resign or sell it, they're seeing a 20 premium. So that too. And that's on stuff that's years old at this point, right? That that's minimum a year old. Probably more like three.
• 44:49
  Ed: Yeah.
• 44:50
  George: You know, lagging edge as it gets and you know, still very powerful and you know.
• 44:55
  Lowtax: Exactly.
• 44:55
  George: Useful equipment. But like it's. The resale is going up, not down. Like this is the equivalent of buying a used car, driving it first, you know, with buying a new car with zero miles on it, driving it for 50,000 miles and getting more than you paid for it. It happens, but it's rare as hell.
• 45:11
  Ed: So shout out to my Hyundai Veloster. But yeah, like even on the consumer side, the RTX 30 series Ampere, they came out in 2020. They're more expensive than they were day one. And part of it is demand. But also just like for inference type tasks, it doesn't matter. It just doesn't. And this is one of the reasons why I'm like, I also don't think that TSMC's latest and greatest is necessary for this stuff. Folks like OpenAI will absolutely go chase it because why would they not? Their constraint is number of rack units that they can employ. Their constraint is not number of GPUs. But there's everybody else who's like, I would like to buy this, please. And I think that's going to be an interesting sort of thing that leads in on the competition, right? And that leads back to the fact that all of the competition are running open models. And that's why Dario is mad, right? Like they're all running spins. Whether or not that, whether they are directly running or they are running quantized versions of Chinese open models, you know, deep infra fireworks, basically anyone you can find on openrouter is just straight up like they're competing with them with free stuff. Because the, the calculation coming out of the Chinese labs is that the model weights don't really matter very much, right? They can't sell to us audience like to US companies in bulk anyway. They won't buy it. So you might as well just like you might as well release them. At least that's my take right now, right? Like OpenAI, you can release these and make life really uncomfortable for OpenAI and Anthropic, who are your biggest competitors in terms of like where the frontier is, while you're also still able to sell all of your capacity that you can do, like most of it is sold internal to China. So like why would you not have.
• 46:59
  George: They have real capacity issues in China too on this stuff.
• 47:02
  Ed: Like they do. But like, it's not as bad as you as people seem to think it is. And I don't think it has to stay as bad as people think it is. Training is one thing. Inference, inference is easy. Training is hard. I, I think we'll see a lot of that change over the next year.
• 47:19
  Propter: Some of this, some of this is just bog standard international trade issues, which we've all learned a lot about in the past few years. That, that for various reasons Anthropic cannot or will not sell directly into China, which is why you have all these, all these spoofed IPs operating out of Singapore, the Philippines, that are Chinese customers attempting to spoof their, their location so that they can buy anthropic inference. And you know, like some of that certainly is, is, is, is Chinese labs trying to distill the model, but some of that is people who just want to use quad.
• 47:56
  Ed: So, like still the most popular thing around.
• 47:59
  Propter: You know, you can imagine, you can imagine a future world where there are fewer barriers to selling inference across borders, but you can also imagine a future world where things break down. Kind of like, kind of like for instance, the car market has, where right now we have still intense protectionism in the United States. We can't buy Chinese cars here, despite the fact that some of the Chinese cars are at this point pretty clearly on a quality price frontier that would be attractive to American buyers.
• 48:26
  Ed: Give me a BYD car, please.
• 48:30
  George: There, there is a really interesting analogy there too where cars are really, really important in, for your industrial sector as a whole because they're so complicated and they operate at such scale and they have such low margins that you can do a huge amount of industrial development by just saying, okay, build the car here, source parts here, and we'll build the factories to supply the supply chain. That is one of the ways China industrialized. Right. Like a bunch of examples of this in other countries too. Auto industries are incredibly strategic assets for overall industrial output because they're very complicated and stimulate a bunch of demand for other stuff. And so, I mean, even, even the petroleum networks you need to supply a domestic auto fleet like that is a strategic asset in and of itself because if you can meet your domestic petroleum demand from day to day through ups and downs of geopolitics, etc. Then like, you're probably going to be able to meet your military's needs for petroleum. And while batteries are excellent for drones, they're not so good for an M1 Abrams. So, you know, we're still in a world where you really like having control of that supply chain within your borders, or at least some stimulation of that supply chain within your borders is really useful. It's easy to sort of see AI as a similar technology where having control over that supply chain gets you all sorts of tertiary benefits, secondary and tertiary benefits that wouldn't exist if you were just buying it from someone else. Yeah, it's sort of the software equivalent of cars in a lot of ways.
• 50:05
  Ed: One of the interesting things about this Conversation is we are not talking about the models because spoiler alert, I don't think they matter. Like at this point they are sufficiently good at everything. Right? Like, unless you are an AI, like an AGI weirdo where you think this thing is going to turn us into paperclips tomorrow. And for those who do not know why I keep referring to paper paperclips, the term to look up on your search engine of choice is paperclip maximizer. You'll have fun. But like, I don't need a better model and I don't think most people do either. I think we are, I've said this before, I think we are three to five years out before people are adequately harnessing our current stuff as like a median thing. Oh, hi. Hi. Lowtax, you probably want to leave. We're talking about AI.
• 50:53
  George: He looks like he just woke up from a nap.
• 50:59
  Ed: Oh, perfect. Let's continue then.
• 51:04
  George: Sorry, what you were saying.
• 51:05
  Ed: Yeah, this is going to be very funny. I'm glad I don't have to edit it. So I don't think we need better models for basically anything. Would they be in some ways more useful? Sure. But like all of the juice, as a certain wise and handsome YouTuber said last week, all of the juice is in workflows. It's in building out structured ways to understand how to process stuff and communicate that to the model. The model we have is good enough and we don't like. Sure. GPT6 came out. I think it's really good. I don't use it nearly as much as it probably could be used because I don't need to. My problems are not that big. Right. And so like all of the real problems are hardware. All of the real problems are supply and demand. We are firmly out of the robo. You know, we are out of the robo sexual part of this and we are just straight up in building, building boxes to put computers in. And I hope Lowtax still doesn't have a working mic because it's funnier that way.
• 52:18
  Lowtax: Nope, I should be just fine now.
• 52:20
  Ed: Oh, good. So, so you, you have.
• 52:22
  Lowtax: Did you guys finish up your little, your little AI psychosis therapy session yet?
• 52:27
  Ed: No, it's still going.
• 52:28
  Propter: I think we're, I think we're almost done with the robot pervert section.
• 52:31
  Ed: But yes, the word robosexual is that there's only so much further down we can go.
• 52:37
  George: Lowtax, taking an extra hour long nap is really working out for the better, I think.
• 52:44
  Lowtax: Yeah, it's doing wonders by me. Yeah, I haven't had to listen to this horseshit. I haven't had to watch the dolphins around like a bunch of idiots. And so, yeah, it's been great. Sandra Bullock is still hot.
• 53:02
  Ed: Good. Important.
• 53:05
  George: Long may it continue.
• 53:07
  Lowtax: Yes.
• 53:08
  Ed: So, yeah, I do think we are kind like, like these people are all full of. Don't listen to Dario. His blog post is. Is mostly fanfic. Sorry, specific. We apologize to the fanfic writers in the audience. They're these, these people are all self interested and really just want. They, they cannot win on the thesis that they previously have. And so they are crying to mom.
• 53:37
  Lowtax: Hang on, you're agreeing with me now
• 53:39
  Ed: about them being protectionist assholes? I've been saying it longer than you have.
• 53:45
  Lowtax: Okay, all right.
• 53:46
  Ed: No, they're absolutely protectionist assholes. Come on. Like that's all they want. They.
• 53:50
  Lowtax: Okay, so you, you think this is, this is them trying to head off like the Chinese.
• 53:55
  Ed: Yes, yeah, yeah.
• 53:56
  Lowtax: Destroying them.
• 53:57
  George: Okay.
• 53:57
  Ed: Yes, yeah.
• 53:58
  George: Basically Tldr. That was what the conversation was about for the last half hour or so.
• 54:01
  Ed: Yeah. So now the audience is caught back up.
• 54:03
  Lowtax: Gotcha.
• 54:03
  Propter: Yep.
• 54:23
  George: I would just like to say that I am very proud of Lowtax for introducing a hydration beverage of choice into his stable. He is now a fiend for mandarin orange flavored electrolyte, which is.
• 54:40
  Lowtax: No, this is unfortunately, this is the
• 54:42
  George: strawberry kiwi one, which is disgusting.
• 54:45
  Lowtax: Oh God no.
• 54:46
  George: Oh God no.
• 54:47
  Lowtax: The mandarin orange rules though.
• 54:50
  George: There are two ways to consume electrolyte Way number one is the mandarin orange, which it rocks, it's really good, it's fine on its own. Way number two is you get the, the cucumber lime pinoy lima and you mix that with ideally like a topo chico or a minaragua and a big thing of ice and then a whole
• 55:10
  Lowtax: bunch of out of here with this.
• 55:11
  George: No, no, I'm telling you, this is
• 55:13
  Ed: a lot of work.
• 55:14
  George: An entire generation of Latin American and Mexican construction market or construction workers who have to work in the south, know what they're talking about here. Like this is the best way to rehydrate you can imagine. Once again, just so everyone got the recipe. You want half and half of some sort of very bubbly, ideally Mexican, out of a glass bottle, mineral water and cucumber lime electrolyte and then a whole bunch of Tajin. If you've put a bunch of Tajin in and you think that's a lot of Tajin, you're about halfway there.
• 55:46
  Ed: When did you turn into go like hell machine with this glass bottle shit.
• 55:51
  George: When it comes. You.
• 55:53
  Lowtax: You said.
• 55:55
  George: I am sure there are very, very spicy, bubbly, aggressive mineral waters that come in plastic bottles. I'm just not aware of the brands, so I'm happy to be corrected on that. But if you're going for a Mineragua or a Topo Chico, you're. You're gonna be in the right place. I can say that with confidence. So that's why I suggested the glass ball kind.
• 56:14
  Lowtax: So anything that has cucumber in it, I'm a no on. But the.
• 56:19
  George: Oh, I didn't realize you were a racist.
• 56:22
  Lowtax: Yes, you did. I told you. I told you I was racist. Against French filmmakers just the other week. Or Quebecois filmmakers. The second best one is the blueberry one.
• 56:38
  Propter: Interesting.
• 56:39
  George: I don't know.
• 56:41
  Ed: I accidentally. Not. Well, not accidentally. They were out of. Of the electrolyte. I got a Gator Light and that stuff is terrible. Oh, I didn't. I was, I was not expecting that stuff to be so, like, salty.
• 56:58
  Propter: I can't do. And it's in its competitors just.
• 57:01
  Lowtax: Well, you've never. Clearly, you've never met a UF grad if you didn't think it. Something for them could be so salty. Ed.
• 57:08
  George: I will say, like, classic Gatorade is like for the, for the low test stuff. You know, the stuff that has less sodium than you need to replenish after a long, hot, humid run in the southern morning. The low test stuff. Gatorade, Whips, Powerade. And if somebody starts telling me about this body armor stuff, like, I. I'm gonna get. I'm gonna. I'm gonna lose my mind. That's right. Gatorade is the superior launch. Yeah. Yeah, terrible. But for the high test, you want the electrolyte. That's. That's your, that's your way to go.
• 57:39
  Ed: There is one thing that, like. So when I was. When I, I was in Japan earlier this year, when you all talked to Patrick Wyman. You pricks. One thing that I was. I was prepared for the, the dare I say, proliferation of beverage machines. One thing we really should adopt from. From other places. It is a great idea, but, like, it was really weird because I, I thought about this because you called it spicy. You called. Called like sparkling water spicy.
• 58:16
  George: Well, well, not. It's not literally like.
• 58:18
  Ed: I know.
• 58:18
  Propter: I.
• 58:18
  George: But I used it to just emphasize that you want very high carbonation with very small bubbles.
• 58:23
  Ed: That's.
• 58:23
  George: That's what I'm getting.
• 58:24
  Ed: I, I totally get that because, like, the one thing that I drank entirely too much of is Mitsuya. Lemon, which is like a special flavor right now, but they, they write on it extra strong because it has more bubbles. Not because it's like extra strong flavor, but it just has more bubbles. And I was like, that's an interesting way to classify this. I, like everyone else would have put more lemon in this. Not, not them. More bubbles.
• 58:51
  George: Yeah, fair enough.
• 58:52
  Ed: So, okay, that's beverage chef.
• 58:56
  Lowtax: Did you all talk about our big wet boy in Ireland yet?
• 58:59
  George: Okay, we have not.
• 59:01
  Propter: What happened?
• 59:01
  George: We should probably talk about that and we should probably also talk about some of the generic ballot polling because it's, it's getting, it's getting downright amusing. So, yeah, we want to do big wet boy in Ireland.
• 59:12
  Lowtax: Yeah, that was cool. So he went over there. What did it. He met with Ian Paisley Jr. And if you don't know who that is, congratulations, for one thing, and turn the podcast off right now so that you do not have to learn. He went over to Ireland. Easy. Ian Paisley Jr. Is the son, the fail spawn of Ian Paisley, who is the former leader of the Democratic Unionist Party in Northern Ireland. They are a bunch of rancid Protestants who do not want to see Ireland unite. Trump met with Paisley Jr. Who is just as rancid as his father, and then immediately turned around and said he wanted to see a united Ireland while holding a joint press conference with Irish Taoiseach Michael Martin.
• 1:00:06
  George: That kind of rules, because that's like, that was one of the most based things about Joe Biden is that he was big union or big green guy as opposed to orange kind. And apparently the orange guy is green, which, you know, I mean, credit where credit's due. That's a good opinion. He doesn't have many of them.
• 1:00:23
  Lowtax: Everyone keeps saying I'm the first Orangeman to support a united Ireland. And I don't know what that means, but you're hearing it more and more, folks.
• 1:00:32
  George: He also cut tariffs on Irish whiskey, which is a nice middle finger to the American distilling industry, given that they've lost their largest export market thanks to Trump. So, yeah, it's going great.
• 1:00:49
  Lowtax: What's great about this is that he's, you know, he's totally doing this because of Miliband, the. Foreign Secretary, Ed Miliband issuing sanctions on west bank settlements in Palestine. And I just. Can you, like, imagine Bibi on that? Like, just sprinting to the phone to tell Trump, Ireland not better for us.
• 1:01:25
  George: For those that aren't familiar, Ireland, because of the historical colonization of Northern Ireland by the. And the various situations that have emerged from that are I think probably pound for pound, about as supportive of Palestine's right to exist of any polity on earth. I struggle to come up with anybody that's more interested in supporting Palestine than the Irish.
• 1:01:56
  Lowtax: There probably are others, but certainly in Europe. Yeah, I don't think so.
• 1:02:02
  Propter: The Irish, with some justification, regard themselves as being. As being the first victims of colonialism.
• 1:02:07
  Lowtax: So, yes, they like to say that anything the Brits did to the others, they first tested on Ireland.
• 1:02:15
  Propter: Yeah, the plantation of Ulster and all that. Terrible, terrible. Read about it. If you're unfamiliar.
• 1:02:20
  George: English Civil War season of Mike Duncan's Revolutions podcast will get you caught up there a little bit. He doesn't go too much into the atrocities, but they're definitely mentioned. It's widely regarded as the weakest season of Revolutions, but nonetheless is worth listening to if you haven't before. So anyhow. So Big Wet boy gives a big win to the mix across the sea.
• 1:02:45
  Lowtax: Yeah. And then he. They apparently packed bag lunches for the press there, which included a bag of Northern Irish potato chips.
• 1:02:55
  George: Fantastic.
• 1:02:57
  Lowtax: In Ireland. Yeah.
• 1:03:01
  Propter: You know, neither here nor there, but my wife and I just finished watching how to get to Heaven from Belfast and it's really good. If you haven't checked it out, you should. It's on Netflix.
• 1:03:09
  George: So what's that about?
• 1:03:11
  Lowtax: Is it a movie or documentary or.
• 1:03:12
  Propter: No, it's scripted. It's kind of a murder mystery sort of. It's. It's by. I think the. The showrunner had involvement with. With Derry Girls previously. So it's kind of partway murder mystery. Partway. It's. It's definitely funny in places. Think caper.
• 1:03:33
  George: Ish.
• 1:03:33
  Propter: Female focused. Anyway, check it out. You'll know within the first episode whether or not you're into it.
• 1:03:38
  George: Yeah, we just got done watching the third season of Silo, which I. I like what they did with Silo in the third season because you think it's going to be about AI and they dream up another technology to get worried about instead.
• 1:03:50
  Lowtax: So.
• 1:03:50
  George: So that was nice.
• 1:03:51
  Propter: Nice.
• 1:03:52
  George: Yeah. Mole People is one of the better sci fi shows going right now. They live underground, so they're Mole People.
• 1:04:02
  Ed: I do have concerns about the Mole People.
• 1:04:05
  Propter: I'm pretty excited for when we can all get back to worrying about like the Yellowstone Caldera and global warming. I felt at home with those as our X risks.
• 1:04:16
  George: There have been a lot of earthquakes in South Central LA this weekend if you want another one to think about.
• 1:04:22
  Propter: Oh, great.
• 1:04:22
  George: As a West Coastman like me, I'm sure that gives you all sorts of
• 1:04:26
  Propter: warm and fuzzies Cascadia subduction baby.
• 1:04:30
  George: Just for those who aren't aware, like geologists are pretty much in agreement that it's not if it's when the entire west coast goes bye bye and it's probably not going to happen like in our lifetimes but like it could. So that's great. And it's. It. There's also a very long well documented and natural in the geological record and the biological record too dating back to.
• 1:04:58
  Ed: Yeah, I saw San Andreas to George.
• 1:05:00
  George: Yeah, there's a really good, there's a really long pattern of very large earthquakes hitting the west coast of North America and we haven't had one in a very long time. So that's, that's fun to think about.
• 1:05:13
  Lowtax: It's going to happen in our lifetimes and then all of the Californians are going to spend the next hundred years suing each other for environmental impact reports as they try to rebuild.
• 1:05:27
  George: The idea of trying to build San Francisco back up from the ground is a lawyer's wet drain. But let's. Yeah, let's be clear about that.
• 1:05:37
  Ed: Yep. I mean when AI takes all their jobs they're going to really need the work program.
• 1:06:01
  George: Should we about talk polling?
• 1:06:03
  Propter: Sure, let's do it. Polling is exciting. This is, this is a happy note. So we got two. I. We got two. It was what it was plus 11 general ballot polls in the last last week. Week and a half.
• 1:06:15
  George: Yeah. I mean a series of double digits.
• 1:06:18
  Propter: Legit. One sec, let me pull that up. We had Quinnipiac was one of the Plus 11s and the other one, Zogby, that wasn't the one I was thinking of though because Zogby doesn't count.
• 1:06:34
  Ed: But could it be proper has such contempt for. For Poland. Like he knows enough about them to. To have that to it.
• 1:06:42
  Lowtax: Yeah, I mean like I know, I know a lot about polling or at least about the firms who conduct polling and who's worth a. And not. I'll give, I'll give Propter Quinnipiac as like half a point. But Zogby.
• 1:06:57
  Propter: No, no, there was somebody else who was 49, 38.
• 1:07:03
  Lowtax: Oh, I saw that poll. I can't remember who it was though.
• 1:07:07
  Ed: YouGov.
• 1:07:09
  Propter: So yeah, anyway, there's, there's some reason to think that things are, that things are breaking out to the high end for, for Democrats on the generic ballot question and we're at the point on the calendar where normies are starting to tune in, get serious and figure out, you know, who they're voting for. And some of people who have been answering undecided and stopped calling me previously
• 1:07:38
  George: we've had votes cast in the old north state. So North Carolina, we're underway. Yeah.
• 1:07:43
  Lowtax: Yeah.
• 1:07:43
  Propter: Did you vote yet?
• 1:07:44
  George: No, it's just overseas at this point. I think I can start October 1st. I'll do in person early voting.
• 1:07:51
  Propter: Nice. So, yeah, so the normies are starting to get serious. You're going to see undecideds decline going forward. That doesn't really seem to be what's driving this though, because both of these pools that have D 11 were 49, 38, which is about the same undecided share as we've had elsewhere. What's happening there that's driving that big double digit number is that the Republican base is flaking away. There are lots of reasons why this might be the case. Might be the extremely unsuccessful war we're still prosecuting. Might be any number of other bad stories. Might be the tariffs. Might be gas. It's probably gas.
• 1:08:31
  George: Diesel, over 7 bucks a gallon will do that.
• 1:08:33
  Lowtax: It might might be proctor reading too much into cross tabs again.
• 1:08:38
  Propter: Well, this is, this is top line though, right? Like this is, this is, this is a number that we've seen pegged in.
• 1:08:44
  Lowtax: Well, Republicans aren't top line though.
• 1:08:48
  Propter: But, but, but like the Republican vote share has been pegged between about 40 and about 43 on the generic ballot for most of the summer in most polls. So breaking out to 38, that's new. Democrats are sitting roughly where they've been in a lot of the numbers, which is, you know, that 49 number is consistent with what we've seen from other pollsters that have the margin closer. So the difference between those plus 11s and something like a plus 8 or a plus 9 is Republican erosion. If we have legitimate Republican erosion that looks anything shaped like that, then places like Kansas start to get real interesting in the Senate.
• 1:09:25
  George: And it's not just the top line races too. Right. Like G. Elliott Morris made a really good point this week that like we, I mean, look, we don't know what the eventual margin is going to be. But if hypothetically you get a D plus double digits, 10, 11, 12, whatever, some really weird stuff is going to happen because that electorate looks nothing like what either party expects and has optimized for. So you're going to get some really lopsided Dem victories and some, you know, some Republicans who should be safe every day of the week and twice on Sunday go down in completely unpublicized races that just turn into landslides for Democrats and it does put at risk a number of gerrymanders to North Carolina for sure is on that list. Texas is on that list. Like you are in territory where you're, you know, dummy mander is a very real possibility in a D 10 national environment. Again, that's purely just taking the assumption that it's D 10. I have no idea what it's going to be. I mean, I think the polling trend at this point and approvals trend for the President would lead you to believe that that's certainly possible. But forget Senate races. You're going to see all kinds of down ballot stuff go completely haywire in really weird ways. And that's going to be fun to watch on election night if it does turn out that way.
• 1:10:46
  Propter: Well, something to look forward to. You know, we're in a situation right now where unless something reverses for what has been just an absolutely monotonic decline in the President's approval rating over the course of the term so far.
• 1:11:02
  Ed: I thought you were going to say faculties.
• 1:11:05
  Propter: Well, that, I mean, I'm not sure
• 1:11:06
  George: they're related,
• 1:11:09
  Propter: but yeah, like unless, unless something, unless something happens to reverse that shift, unless there's some kind of rally around the flag effect, who knows, maybe, maybe this convention they just threw that I think is useless is going to turn out to be useful, you know, but unless something happens to change, to change the game in some way, we're headed for a real big blowout. And George spoke to this a bit, but this is the kind of thing where the gerrymanders aren't going to protect them and we're going to have enough additional Democrats in the House that we don't have to care who the worst Democrat is, that it's just, it's not going to be real relevant who vote 260 is because it doesn't matter whether or not, you know, Marie Gleason Camperez has a good opinion about whatever the issue of the day is if she's vote 260, it matters a lot if she's vote 220. But you know, if, if the lines are drawn that far apart, that means a lot of, a lot of Dems are going to get to effectively get a pass on things. We're not going to need as much party discipline in the House. In the Senate really been.
• 1:12:27
  Lowtax: I haven't really been worried about that though because like, even if we're not at a Democrats by double digits scenario, we were already looking at, you know, seven points, eight points, something like that. And in that Scenario, the House is going to be fine. Yeah, it's, I mean it's still, it's really the Senate. That's it. Can we get to 52?
• 1:12:48
  George: It's really the Senate and I don't think the interesting thing is if you get to 52. I think the interesting thing is if you have 54, 55, 56 and start setting up a 60 vote majority in 2030, 2029, sorry, 2028 election, 2029, January is when you would come in and just from a market perspective, like anything under that blowout scenario for Dems doesn't matter. It's, it's all going to be priced in but it'll be really interesting to see what the market does. If, if they think we could be staring down the barrel of 60 vote majority in the Senate for Democrats in a couple years, maybe we need to start behaving a little differently.
• 1:13:24
  Propter: Yeah. You know, I'll tell you in some ways, honestly, I'd rather have 58 than have 60 because I think with 58 we crack the filibuster and with 60 they may be punt.
• 1:13:33
  George: But you think they're going to do.
• 1:13:36
  Propter: I do.
• 1:13:38
  George: If you have that number.
• 1:13:40
  Lowtax: Yeah.
• 1:13:41
  Ed: Okay.
• 1:13:42
  George: Yeah, I, I don't know. I, I think it's just, it's, it's so contingent on what happens over the next two years and there's all sorts of.
• 1:13:49
  Lowtax: But yeah, I don't know. I don't know what the number is. I'm pretty sure the number is not 51.
• 1:13:55
  George: I think it's no.
• 1:13:57
  Lowtax: 57 to 59 somewhere in there.
• 1:14:00
  Propter: Yeah, you need, you need some spare votes. If it happens, it'll happen with almost everybody on board. But you need some spare votes. So you know, but that's, that's a long ways away. It doesn't do any good to crack the filibuster while there's a Republican in the White House. So.
• 1:14:16
  Lowtax: But if you get north of 60 and you don't need to do it, then who cares?
• 1:14:20
  Propter: Well, I mean that's so they, they. That's what we thought in 2009.
• 1:14:24
  Ed: Yeah. Have you met the median voter?
• 1:14:28
  Lowtax: Yes.
• 1:14:30
  George: The thing about getting north of 60 and knowing it's a temporary window which every single person will know at the time. Like we know what, what we knew what happened in 2009.
• 1:14:39
  Ed: Right.
• 1:14:40
  George: Like a party that has a 60 vote majority in the Senate is going to behave very differently today than they did in 2009 because we have that benefit of looking back over history. That doesn't necessarily mean you get vastly superior outcomes. But that legislative session, that two year Congress is going to move very, very quickly. Like I promise you that. So, so, you know, we'll see. But I, either way, look, if we're sending for six years to the Senate Democrats from Texas, Kansas, Nebraska, Alaska, I mean, it's going to be a really different vibe than what just looking back 18 months ago. Like, what were all these people thinking? That there was going to be this new Republican majority. Right. Because Trump won by 2 percentage points. What the hell were these people thinking? It's unreal.
• 1:15:45
  Lowtax: One and a half percentage points. Don't give him two.
• 1:15:49
  George: Whatever, man.
• 1:15:51
  Ed: Like, no, I just, I'm on, I'm on low tax aside here. You don't, you don't give this guy an inch.
• 1:15:59
  George: Gentlemen, we know how to round in this House. We're that numerate. Like, come on. I, so anyhow, I just, I just look back on 18 months ago and how certain people were behaving during the fascist overlords.
• 1:16:11
  Ed: Yeah, that was, that was definitely their vibe.
• 1:16:14
  George: It was unreal and it was very obviously a completely Brittle coalition and B1 that didn't win a big majority to begin with. What do they, did they really think that was the end of American democracy? I don't think most of these people did. I thought they were, I think they were just stupid. Like, I, it's unbelievable.
• 1:16:30
  Ed: Like if you look at a lot of this stuff, right? Like this kind of historically, if you look at some of the people that these interesting examples of the human condition most closely parallel to, I don't think impulse control has been strong for any of them. Right. Like, yeah, these, they, they are, they believe in a kind of dominance flavored power politics that is all just due unto they. They. Of course, they didn't think that it might be done unto them.
• 1:17:09
  George: Oh no, I'm not talking about Republicans. I'm talking about people in the corporate land.
• 1:17:16
  Ed: I'm talking about, I'm talking about every shithead CEO.
• 1:17:22
  Propter: I think, I think that the incentives here are important. And certainly once you see congressional oversight get rolling, once you see any Democratic organ that has investigative power, that's going to change some things.
• 1:17:36
  Lowtax: Once you see future President Ossoff get his subpoena power.
• 1:17:41
  Propter: Yeah, I mean it's, it's going to be fun to watch some of these guys work. I mean I, I think there's some.
• 1:17:47
  Lowtax: I need him to get tempo. We got, we got to talk faster, buddy.
• 1:17:51
  Propter: Yeah,
• 1:17:54
  Ed: yeah, no, like that, that is his really his one weakness, isn't it? That man talks too slow. Yeah, yeah. But like, yeah, I. It's going to be interesting watching what companies did some really foolish things and how they try to, like, bail out. Like, my employer was not as amongst them as many others. But I'm also very curious what they do. Right. Like, there's a lot of this shit going on that, you know, if you knuckled under to the Trump administration, you're going to have to answer for it. And I don't know how you do that. Like, I just don't. Like, I don't know what you say. Right. Like, oops, we thought that the fascists would win forever, therefore. Right. Like, I don't know how you have an answer for that.
• 1:18:48
  George: Yeah.
• 1:18:51
  Propter: You know, I think you're going to start to see some reversals when there are. When there is congressional investigative authority, because people are going to get the scent of something bad coming, but it's going to take regulators back on the case for it to be intense. And that means taking the White House. That's. There's no way from here to there except through that. But I think it's going to be a very different mood come next spring than it is right now.
• 1:19:22
  Ed: Fun.
• 1:19:24
  Lowtax: I'm just going to say I called Texas.
• 1:19:30
  George: All right.
• 1:19:30
  Propter: Should we leave it there?
• 1:19:32
  George: Yeah, let's leave it there. Low tax. We're glad you're alive, man. I appreciate it for now anyways, Propter, same to you, but we weren't as worried about you.
• 1:19:43
  Lowtax: I don't understand. I don't understand. Malik Willis last I pulled the game up, he was 7 and 9 for 80 yards, 101 rating. Dolphins had three points.
• 1:19:56
  Ed: I've just enjoyed like, yes, this is not a visual medium, but Lowtax has been wandering around holding his. His microphone like a martini glass the entire time. He's just smoking a cigarette, holding it like a martini glass.
• 1:20:12
  Lowtax: Yeah, I hadn't. I hadn't set my stand back up since I got back from South Florida.
• 1:20:18
  Propter: All right.
• 1:20:19
  George: With that charming visual in your minds, we hope you all have a great week and thanks for listening.
• 1:20:24
  Ed: Just remember, you have to include the luchador mask in this chart. Charming visual.
• 1:20:29
  Lowtax: No, no. The ghost face mask.
• 1:20:31
  Ed: Oh, sorry, sorry. Management regrets.
• 1:20:34
  Lowtax: Propter is the Luchador. I'm Ghostface.
• 1:20:36
  Propter: That's true.
• 1:20:37
  George: All right. From the north shores of Lake Norman in Catawba County, North Carolina, I'm George
• 1:20:43
  Propter: from D.C. i'm Propter from Outer Florida.
• 1:20:46
  Lowtax: I am Lowtax Speedrun Enjoyer.
• 1:20:48
  Ed: And in Boston, I met Stay normal, y'. All.
• 1:20:52
  George: Sam.