00:00:15
Speaker 1: Pushkin. Welcome back to Risky Business, a show about making better decisions. I'm Maria Kanikova and I'm Nate Silver. So today on the show, note, we've got we've got a pretty interesting Risky Business e news week. We've got GPT five being released, which is maybe less of a deal than the launch hype announcements were, but still still a big deal. And then we have some interesting trade stuff going on with AI chips right with Nvidio.

00:00:52
Speaker 2: Interesting is an interesting reader?

00:00:55
Speaker 1: Yes, well, well we've got gp.

00:00:57
Speaker 2: Now there have been five GPT since the last GTA Grand Theft Donald. Sorry to throw that in there.

00:01:03
Speaker 1: Fascinating to well, before we before we get into it, Nate, I just wanted to say, you know, we're taping this on the twelfth of August Tuesday, so a few days before listeners will hear it. But congratulations. Today is the launch of the paperback of On the Edge.

00:01:32
Speaker 2: On the Edge the Guard We're seeing Everything, the bestseller by groundbreaking author Nate Silver. Yeah, so the paperback is out. There's a new forward, or I think it's called a preface technically. I was just at Barnes and Noble signing some copies. A little sweaty. I walked here from there. It's hot in the middle of August in New York, breaking news. But yeah, no, I think the book holds up really well. It covers a lot of really risky business esque topics. And you know, it's a big book, cheaper now with paback. It fits better on a shelf, not quite as thick, and there's new content. So I would strongly recommend, of course. I mean, you know, a little biased here, but like, thank you.

00:02:09
Speaker 1: Of course. Well, I'm excited to read the new preface, and yeah, I definitely recommend the book to everyone. I will try to repost the review I did of it on my substack so that people can get reintroduced to it one more time. Anyway, it's a fantastic book. Congrats Nate. And let's get into some rivarian topics, like the release of GPT five from open Ai. Nate, have you had a chance to use GPT five yet we don't.

00:02:43
Speaker 2: Have much choice, you know, they kind of steer you into GPT five if you And at first I'm like, oh, okay, I guess I pay for the pro plan and so I'm like, oh wow, I'm one of the privileged few and now I don't think there's an easy way to get back into all the all the old family of GPT products that we knew in love before, Like people did get it attached to the different models for different reasons, and so now it kind of is implicitly trying to figure out what you want, which is kind of part of I mean, it's interesting. Right. So on the one hand, you know, I kind of made the joke before about kind of comparing it to like a video game release. But anything new, if we release a new presidential model or something, right, anything new might have some kinks and bugs. I mean, as much as you might say, okay, we have to have it perfect before it's ships, I don't think it's practical because like most people are gonna only learn thing. I mean, you know, I think they did discover that's not doing what like groc did and calling itself Hitler for example.

00:03:37
Speaker 1: Right, Yeah, I mean I don't know if we could really call that a win. Like that seems to be like a baseline.

00:03:43
Speaker 2: Or like Google Gemini drawing multicultural Nazis, you know, so that the bar is pretty low. But no, I think it has a little bit of new car, new model kind of smell a little bit. The first thing I asked it to do was, my partner are planning a trip to Bestly, Scandinavia or the Nordics, technically, right, I want to visit these cities. Give us in our itinerary, right, and like thinks and things and has something that seemed plausible to me, very detailed, right, And then I'm like, emails use a PBF and it like freaks out right like shows some complicated like you know code, Python code, and it's like, I don't know how I know this. I don't I know this, So you know it's it's up and down. Well, hey, at.

00:04:21
Speaker 1: Least it gave you a plausible itinerary.

00:04:23
Speaker 2: I so.

00:04:26
Speaker 1: We've been warned. We were given an explanation by Sam Altman because I tried to test it, you know, when it was just released, and apparently the auto switching tool was down, so it seemed like it was a lot dumber than it was that it was supposed to be. So one of the one of the things that they're kind of really touting on this model is that it automatically knows what you want right, whether it should think deeply or not, which does not actually seem to be the case. Even now when the swings back, it's like.

00:04:53
Speaker 2: Hey, who got a cheek here? I don't think we need too much deep thinking. She probably wants a quick answer, get back to the cooking.

00:04:59
Speaker 1: That's that's exactly honestly, that's that's been my experience. And one of the first things I actually tested it on because it says that, you know, one of the things that it's much better on is on hallucination. So I actually gave it a psyche question and asked for some sources and like papers, and it unfortunately still hallucinates when you go down that route. I was because so I'm writing a piece this week about kindness contagion. You know, when someone is nice, you know, how does that spread? And so I asked it some questions about the research in that field. And I didn't really need its help in the sense that I know the field pretty well, but like, I was just curious to see what it would come up with. And it gave me some good stuff, but it also gave me some stuff that just simply does not exist. And I know, you know, as you say, and as a lot of people say, like you need to know what to ask it but the reason I tested this specifically was because one of their big claims was no more hallucination basically, which is not true.

00:05:58
Speaker 2: Yeah, I've used it for you know, I almost always put the articles I'm writing for the newsletter through a copy of it in fact check, in the GPT models or cloud. Sometimes I think this is one of the most useful things that AIS are good with. It took seven and a half minutes for what, from Nate is a relatively short article seventy birds or something, Right, I am using, like the thinking version, right.

00:06:21
Speaker 1: Did you tell it to use the thinking version or did it?

00:06:24
Speaker 2: I think I was on thinking by default, right, And usually I say have a high threshold, and I didn't say that. But it's really nitpicky. It's like, you know, I made some line about like Elon Musk is like tweeting out like anime smut was a term I used, right, I already tone that down from porn. And it's like you should be more careful, you should say NSFW images smut implies an editorial stance, you know. So like it's just like it was kind of nitty as a poker term, right, and like I I had it actually a friend sent me a poker hand because i'd we wrote before or I wrote before about how Chick's bad at poker, and it played the first hand really well, and then I'm like, okay, simulate more hands, and then it was still kind of not not great. Maybe a little better, right, Yeah, No, it's a little weird because I don't look to me, you kind of had four point five and you had like oh three and oh one. Right, Like I noticed in the spring, I thought some of the reasoning quote unquote models or the type of stuff I'm doing, which is not using GPT as a chat bot, right, it's like a workaid, you know. I thought there was an improvement then, and I kind of feel like there's a bit less this time. You know. Also, the open air models or some of them are kind of slow, right, I've found sometimes like Clode or Gemini or even groc will like spit out things faster, and so you.

00:07:42
Speaker 1: Know, and sometimes that matters, by the way, like sometimes you wanted to take time, but sometimes like you're like, okay, come on, like let's let's get let's get moving, and uh yeah that if.

00:07:53
Speaker 2: You can get me a simple you know, and often like in a fact check, I mean today I had to go to the pricking Barnes and Noble, right, and so like the time pressure is often relevant, especially, you know, and oftentimes I also use the models a lot for like little programming task right, Like I forget the programming using language US is called STATA. Some people would say it's kind of a fake, but you know it's a it's a real language, right, and I'm like, I forget how I do this instat or full disclosure, even Excel. I'll be like, oh guy, what's the complicated formula for this? Right? And like it seems to me to be seventy to eighty percent reliable the AI models in general, for like you know, where I want a snippet of code that's like, you know, three to ten lines long, right, you can kind of vibe code, right, You're just like I want to do this right, and it's like, you know, it's pretty smart about it. I mean, every now and then they won't work. And you know, I always tell people like these things screw up when they're trying to chain together different steps and they don't really quite I mean they're trying to train themselves, right, but they don't quite know how to stop right, and so it's like, okay, when I build a model, we're going to NFL model. You know, I stop at every point and ask, okay, does this output make sense? Do you form a bunch of complicated discoperations sort from the bottom to the top. Right, you know, hopefully Tom brady ease list is one of the best quarterbacks, right and Ryan Leaf is one of the worst. But like, yeah, I'm keeping myself in the loop. And for that kind of thing, it's like, okay, I'll just say fifteen minutes trying to figure how to program this or you can debug, right, You're like, why isn't this working? As somebody's clawed and it was like, because you made a typo. You misspell the word nate. That's why you've been tearing your hair out for forty minutes. Is like you misspell the word yeah. So yeah, no, look, I think it's kind of halfway a branding it did it did your size or some people seem to vouch for this a lot. I don't know.

00:09:41
Speaker 1: Well, you know, obviously, like I don't code right, and people have said that it seems to be a lot better at coding certain things, which great, you know, if it is great, but when you're what you're saying actually like seventy to eighty percent to someone like me, that makes me actually much less likely to use it because I'm not you. I don't have that background, so I can't do I can't always do like a check to figure out, you know, does this make sense? Right, And in the sense that I don't have that technical base, I can't review it in any real way, and so I need it to be accurate, right, because I don't trust myself to spot any potentially would.

00:10:18
Speaker 2: Say the old way is to look at a manual or like stock overflow or whatever, and there you're starting through a lot of crap too. Right, it might not be pertinent to your particular KSE, or it might be an old version of software or instead of there are lots of little fussy language things with local variables and scalers and one of those things mean and what the different rules are and stuff like that. Right, it slightly fussy language, and it's good for a handle that kind of thing. And again, to me, I'm working in ways where it's fairly failure proof. Right, You're doing one thing, you have an expectation for what that will do to transform the data set, right, and if it doesn't happen, then it won't work anyway. Right, But like the notion of like I'm just gonna sit back here and trust it to do all these things, I mean, I think it's probably, you know, to code an entire NFL model, which involves a lot of original research collection and involves a lot of like knowledge about the sport, knowledge how to build models right, a lot of trial and error, Like you know, I don't think the AIS are particularly close to doing that kind of work.

00:11:21
Speaker 1: So I think that you just made a really important point, which is to just not expect the world from it and to like to know what it can and can't do, which, by the way, already takes a certain user intelligence and like knowledge to say, Okay, you know what, I don't trust it to do this, but I do trust it to do that. So there is you know, even though GPT five was kind of hype, does you know you don't have to think anymore. You still actually kind of do right in order to get the outputs that you want, and to realize, Okay, I can trust this task, but not this task. I can I can get it to do this, but not that and I think that that, you know, that human in the loop is still very much a thing and still very much needs to be a thing.

00:12:04
Speaker 2: Yeah. There, I just kind of keep like a little running mental tracker of like here my expectations for AI models, LLM SLARGE language models, and are they exceeding those or were falling short? Right? I mean? And then do weird thing, you know, I like I had a situation where, like I had a bunch of latitudes and longitudes of NFL stadiums that we'd code up quickly, right, and we're like, I want to reverse look these up and tell me what city they're near, right as a double check, and like put Atlanta as like Chattanooga and just a little it's you know, and so like for things like that, because for data, I want all my data to be perfect, right, I don't want to misattribute one city that doesn't really matter if they have the you know, Falcons playing in Chattanooga, it doesn't matter that much, right, and like and so for that kind of thing, you know, I I still would rather have my, yeah, my research assistant do it or me do it myself. Right, things that require like a lot of person But you know, it's very frustating, use them enough where I have like particular rules where I think they're likely to be helpful and not and how how safely can you fail and things like that, But like, yeah, I mean my general view is that getting kind of savant like and that they're not very bright about some things and they're freaking genius is about others as opposed to this notion of like general intelligence, where I mean, but you know, look, even the things it's worse at, it's like as good as like a you know, high school sophomore or so, you know, I mean, it's not terrible and there aren't too many things where it's terrible, right, But like, yeah.

00:13:32
Speaker 1: Yeah, well, I think I think it really really depends on what you're asking it.

00:13:36
Speaker 2: You know.

00:13:36
Speaker 1: One of the main things that I've read about people kind of responding to which highlights an issue that you know, Sam All was like, Oh, we didn't realize how big of an issue this was, which I think is very interesting because people have been trying to say it's an issue. Is the change in voice, right, the fact that past models were very psychophantic. You and I were talking before taping today and I was like, we should really be pronouncing the psychophantic because there's been some there's been some real psycho behavior here, and when they introduced GPT five, the default tone of voice was very different. Right, They did try to address this, and then they got within not even twenty it didn't even take twenty four hours. Just like immediately they got all of this pushback with people saying, no, you know, I've lost my boyfriend, I've lost my best friend, I've lost the person who told me I was a genius. And it makes you realize how many people were using this really not for what it's intended, and something that can be incredibly bad for a lot of things, right, mental health, just social connections, all of these things that you know people people were like, whoa, whoa, whoa? What happened to my significant other? And they brought it back right, so now you can actually select that voice again. We have the default voice, but we also have you know, the listener of voice, and there are a few other voices. None of their descriptions actually map onto what that actually is. I was reading there was like a cynic and I don't even remember what they said the cynic voice was, but I was like, that's not what a cynic is like it was, it was they're really they're really weird. They're really weird descriptors. But basically you can get you can get GPT to interact with you in different voices, and I you know, my reaction to that is like you shouldn't always give the people what they want in a lot of ways, like this was bad and like you've fixed it, like don't go don't go unfixing it, because even though you fixed it, they didn't fully fix it, right, they just made it less overt, which you know, subtle psychopancy can also be bad. But they've tried, they at least initially tried, and now they've really gone back on that immediately caving to pressure. And if you always give the public what it wants, like we've talked about p doom and like a lot of people.

00:15:59
Speaker 2: Want things you really should not be getting these, it's pretty hard kind of in an equilibrium, like not to optimize for what drives engagement in the show. I mean, you know, on the one, thecept something that they don't need revenue like right away, but like, yeah, no, I think sometimes these models like you give them an inch and they take a mile with it, right, like if you look at what happened with ROC when it had its moment, the prompts that system prompts that elon or ate an unnamed engineer XAI was using. We're like not that wild, right, but like if you go down a rabbit hole, we keep kind of getting like reinforcement feedback. This is good, this is good. And you know, you know open aies models used to have this thing where did you like answer A or answer beast. They're now outsourcing it to some of their users, right, and like yeah, I mean, look, you know what are you kind of optimizing for objectively? Right? It's kind of easier when you are trying to train it on like a math problem where there's an objectively correct answer, right. You know, the NFL model turning to build the end of the day, how accurate is it to predict NFL games? Right? That's that's the bottom line.

00:17:15
Speaker 1: You actually have a metric, right.

00:17:17
Speaker 2: And for you know, for feedback or goodness of an answer that's more subjective, then it's a lot trickier, right. You know, I think we've seen with some of the you know, some of the reason that like Grock hasn't had as much reinforcement learning training right or or or you see that right? You know, the works are pretty rough. I'm mixing metaphors here and yeah, I mean there it's a weird technology, and I think people understand more about how these models work than before. And like, by the way, like one reason to be optimistic unless you're a doomer. I guess it's just like the amount of like human another capital being poured into AI research, right is like, you know quite something, right, It wouldn't surprise me at these companies start saying, oh, we have a lot of smart people's to kind of spin off technologies and energy or quantum computing or whatever else.

00:18:08
Speaker 1: Right, yet, know, I think that there's so much promise, but I think that with this particular thing, the incentives are misaligned at least for now, which you were kind of hinting at, right in the sense that sure, like they don't necessarily need it. But if people if this is one of the things that fuels revenue growth, right, that people want to feel not lonely. They want you know, someone who's kind of reinforcing their ideas that they interact more, which is good, right if you're actually paying more in order to be able to kind of spend more time with the model, and they're interacting more when they feel like, this is my girlfriend, this is my boyfriend, this is you know, my best friend, This is my counselor my psychiatrist, you know whatever, it is my teacher. There was a guy who was who Cash Hill just wrote about in the New York Times.

00:18:56
Speaker 2: Who.

00:18:58
Speaker 1: Believed that he'd created a new mathematical theory right that solved everything, and like and tried to actually he tried to fact check the delusion, which was crazy. He's like, I feel like I'm sounding crazy and the a I was like, no, You're absolutely not crazy. You're the most everyone else is crazy, right like you're saying so. So was one of these very strange kind of things where he went in. By the way, his first question was he wanted it to explain how you got the value of pie, because this was someone who never finished high school and his son needed to help with homework, and so he just asked chat japt about pie. That was the start of this insane rabbit hole. And as and I'm not I mean, we're talking about chat JEPT because of GPT five release. But this doesn't just apply here. We're just talking about. You know, that's Lum's in general probably will be susceptible to this. I'm not sure, right because all of these examples are from CHATJAPT But you know, you have this innocuous question that then leads someone to a very detrimental spiral. And it's not a standalone case, right, And now we know how many cases that were below the radar because nothing bad, nothing quote unquote bad happened, yet there were given the outcry when the model shifted. And I don't think honestly, like, do we really trust a company that's clearly profit driven? We know, I mean Sam Milton's companies profit driven. I know, I know it's it's crazy.

00:20:26
Speaker 2: Gambling a casino, But do we.

00:20:30
Speaker 1: Really trust a company that's profit driven, right, that's bottom line driven to sure, they will fix all these other problems if it's good for them, But do we trust them to fix these things that are really undermining mental health? We've seen that meta, you know, Facebook, all these none of them have right for years, they've known that there are issues that have existed and they haven't addressed that.

00:20:51
Speaker 2: I'm not as convinced about this particular problem. I mean because first of all, what it's substituting for. Is it subsitting for Twitter or Reddit forums or some other dark corner of the Internet.

00:21:02
Speaker 1: Potentially No, because from a psychological standpoint, there's immediate reinforcement and a conversation that goes back and forth, which is very very different for the brain than like saying something and then.

00:21:12
Speaker 2: On Twitter your media feva.

00:21:15
Speaker 1: But it's not quite the same thing. Just from a psychological standpoint, it can be much more pernicious when you feel like you're talking to an actual person and personalities do start developing. I mean, what do you like? Obviously this is this is not ideal. What do you think is ideal?

00:21:31
Speaker 2: Right?

00:21:32
Speaker 1: If you're interacting with chat GPT? What personality do you want me personally? I want no personality whatsoever. I just wanted to give me the damn facts.

00:21:40
Speaker 2: You can give it custom instructions, right, which is like appendittus. So I tell it like be honest and straightforward site sources. It's fine to speculate, but if you are speculating, label it speculation that kind of thing, right, And.

00:21:53
Speaker 1: It sometimes lies about that too, Like when I say give me sources and it lies about the sources and you say hey, the source doesn't exist, and it says something like you caught me. And I'm like, okay, well you're you're ostensibly following instructions, but you're not really And I actually, I don't know, I didn't try that you caught me thing with GPT five. I don't know if it's still going to do kind of a version of that. But Nate, I mean, if we're looking at the reliability of these models, right, you had mentioned the Chattanooga example right where it thought that Chattanooga and Atlanta were interchanging.

00:22:28
Speaker 2: Basically, well, Atlanta wasn't in my gazzeteer file. I forgot about it. Man. It ha'd a litle arch just sitting in Georgia. But I fixed it now, right.

00:22:35
Speaker 1: But if it's getting things like Chattanooga wrong, like you start questioning how how good its output was on other things.

00:22:41
Speaker 2: Yeah, and if there are you know whatever, four hundred rows of data and it screws up five, it's a lot. And the analogy here is something like, Okay, if you take the subway in New York, you don't have to like look up the timetables because it's rarely more than like a five or seven minute wait for a train, right, you just go on.

00:23:00
Speaker 1: Station unless it's the B, in which case you'll be waiting forever.

00:23:03
Speaker 2: Sorry, B fans, I'm on the L now. It's a real it's a real adult train. Rent there.

00:23:11
Speaker 1: The B stands for bullshit.

00:23:14
Speaker 2: Bullshit Okay, And that's when it's part of your continuous workflow, right, chechup is more like you go to like the amcrak station and you have to plan around it a little bit. Right, It's not it's you know, it's not kind of the bonus productivity from from kind of turning your brain off and and just having to handle the work. You know. With that said, you know, I've come home uh knights when I'm tired or busier and had a couple of glasses of wine, right, and it sure is nights then to be like, oh, I forget the stupid fucking command and static, Can you just tell me what it is? Right? But I'm you know, still using it in like a piecemeal way.

00:23:53
Speaker 1: And we'll be back right after this.

00:24:07
Speaker 2: All this might be kind of not bad for AI safety, right, Like I kind of think.

00:24:12
Speaker 1: Like, yeah, well I think it depends. So I don't know the nitty gritty of what the improvements were, but some of the some of the reviews that I read, some have said that the lack of transparency into kind of what's being used and how it's being used is actually potentially not great. That before you could, you could query much better to actually figure out, you know, what processes were being used, what models were being used, et cetera. Now you can't, and that if some quoting gets to a certain point that it can do like some self replicating things. And this is kind of what we talked about briefly when we were talking about the AI twenty twenty seven report, that some of those things are potentially becoming closer to being a.

00:24:58
Speaker 2: Real Yeah, I mean this is almost another show, right, but like you know, humans currently have an important role, a couple of important roles in the process, right, one of which is to provide the corpus all the time that the models are trained on, the second of which is with reinforcement learning. And like again, math problems are a weird exception in that you kind of you kind of know the answers, right, there's objectively correct answer, it's just really hard to figure out, right, and a human can say, oh, that's correct, right with other things. It's it's trickier if you're if you're, you know, wanting to patent some novel protein that could be used in drug discovery, right, then you got to test that and make sure it actually works. Right, And like so if you don't have a human reinforcement learning, if you if you're trying to train corpses that are beyond them, I mean there again, there are a case where you can extrapolate and get fifty percent than the best human or or two hundred percent better. Right, the notion of like an explosion of superintelligence, I think you know, I mean, these things still can't book good fucking delta flight to Chicago, right, I'm probably the year they can.

00:26:01
Speaker 1: But by the way, by the way that talking about like security risks, this isn't p doom, but like personal security risk as it becomes better at doing that, the chance of someone hacking and like actually being able to insert some malicious code without your knowledge so that you end up, you know, so that they can steal credit card information, et cetera, et cetera. Don't underestimate. I'm not talking about you, but I think we as a society underestimating.

00:26:27
Speaker 2: Maybe even more than to Google, you know what I mean.

00:26:30
Speaker 1: It's crazy and that is available somewhere, right They're storing all of that data, and so that means that it can be hacked, and if it can be hacked, it will be hacked at some point. I think that that's kind of the rule of the Internet right now. I've spent enough time with you know, con artists and bad actors that I know that, like, there's always someone, if there's a new technology, there's someone one step ahead figuring out, Okay, how do we exploit this? Does it call me Maria? Now?

00:26:54
Speaker 2: Okaymate?

00:26:55
Speaker 1: Sometimes no, it hasn't called me Maria. But I've tried not. I mean, obviously you have to sign into it, but I try to uh minimize uh minimize my sharing of any personal information. But that only takes you so far. But yeah, I think that that's kind of a personal security risk that people are probably underestimating. I'm guessing there's some people who are very well aware of it, but I would hesitate before giving it access to you know, my travel plans, access to any itineraries, credit cards, et cetera, et cetera, because those are things that seem like there could be vulnerabilities and we see you know that as you said, you know, this is a this launch has the new car smell, so like at the beginning, like there are going to be bugs. There are going to be issues, and sure eventually like some of them are going to get sorted out, but there's always going to be like the initial data breach that prompts it to get sorted out. And I don't want to be part of that initial data breach.

00:28:00
Speaker 2: Yeah. I don't know that these giant tech companies had behaved in particularly trustworthy they have, right.

00:28:06
Speaker 1: Right, Yeah, I don't think they have. So I don't know how much we want to we want to give them and what you know, Nate, there might be some there might be someone who's like, uh oh, does Nate upload all of the specifics of his models to these Maybe we should hack Nate's uh chat GPTs that we can steal his model and improve on it. I mean that seems silly, but like it actually makes corporate espionage those types of things much easier too. It's what I've always said with you know, con artists, that it's become so much easier, and the barrier of entry to conning people has become so much lower simply because we share so much information online unthinkingly, and so it becomes the case where before it would take someone a lot of kind of research to try to figure out, you know, oh, what are the things you know? Where does Nate like to go? What does he? You know? What are the pressure points that I can How can I approach him? Et cetera, et cetera.

00:28:57
Speaker 2: And maybe you can trade out on poker tells, right, yeh, watch five hours of Who's a Poker Player? But yeah, Adam Hendrix and figure out like what are his tells? Right?

00:29:06
Speaker 1: But but now I'm you know, con artists can use all that information like very quickly because you've shared it. And with chat GPT, people are sharing so much right on such a personal level, not thinking that this will become public, and I don't know why they think that it won't. So it's a very you know, it's a very interesting conundrum. And I think there are so many amazing things that are kind of come out of this, and then some very dystopian things and some things that will potentially really hurt individual Yeah.

00:29:32
Speaker 2: I mean, look, I'm not even talking about the audio and video caare I mean, look, it remains the case that if you beamed to twenty twenty five from twenty twenty right, you would be amazed by what these models can do. Absolut would be considered a freak if you had predicted five years ago that you have this machine that for many things can like pass the touring tests. Some areaster don't use Turing tests, but like it's you know, it's basically giving you plausible human level performance across a variety of cognitive tasks, deficient in some, excellent in others. Right, Like that still is quite amazing, And like part of what I'm reacting to is like, you know, where is the hype relative to the reality. And it felt like a year ago it was like, Okay, people outside of Filekan Valley are just not seeing at all the power of this, and now they kind of do. And I still think that kind of like the political types are like significantly behind the curve and calling them chatbots or whatever. But like also like you know, you read these really smart researchers saying that, oh, we think there's gonna be a singularity in two years, right, and I'm like, you know, look, you can drive. There's a lot of trucks you can drive in between. Oh, it's just a chatbot and Singularity by twenty twenty seven, right, right, Yeah, it feels like pretty safe bounds.

00:30:46
Speaker 1: I totally agree with that.

00:30:48
Speaker 2: You know, just to come back to the question of your ideal AI chat bot personality, this sounds like a weirdly like the Howard's Stern someone or something. But Maria, what's your what's your what's floats your boat?

00:31:00
Speaker 1: Well, Howard? You know, earlier I had said that I want an AI that just gives me the facts, right, Like, I do not want the damn thing to have a personality. This is an AI, it's a computer. Like, this is not my friend, and I don't want its opinions. I just want it to kind of give me factual answers. Now I know that that's not actually possible because, as we've talked about many many times, like I'm probably not asking it about math problems because I don't have any use for that. I'm probably asking it for things that will inevitably be opinion tinged because you know, the inputs were made by humans. But yeah, I want it to kind of be as as neutral as possible, And like.

00:31:45
Speaker 2: Do you not get were you familiar with FIVEY Fox? Does that mean anything to you?

00:31:48
Speaker 1: No?

00:31:49
Speaker 2: Five E Fox was like the mascot of five thirty eight models, right, it was a cartoon fox.

00:31:54
Speaker 1: Oh, I've seen the picture of the cartoon fox.

00:31:57
Speaker 2: I'm presenting my modelty like, maybe i'd be a good personality, right, Yeah, little, you know, a little furry animal.

00:32:02
Speaker 1: I am so so Nate. Are you familiar with Microsoft Office's paper Cliff Gliffe? Oh my god, I think we all have stories about clipping.

00:32:14
Speaker 2: I mean, there are things about paper clips and AI. You do you probably don't want it to.

00:32:17
Speaker 1: No, we do not want paper clips anywhere near our AI models. That the paper clip problem has given us enough headaches. By the way, for those of you who aren't familiar with the paper clip problem, you know, you might know Microsoft clipping, but not the paper clip problem. It's an AI philosophy problem first proposed by Nick Bostrom about basically how paper clips can cause the end of the world. Well, we'll talk more about it in today's Pushkin plus Day. What about you, what's your ideal personality?

00:32:46
Speaker 2: Yeah? Like, I mean, you know, my custom instructions are to be straight forward, to provide a lot of detail. I'm not looking for the AI for like, you.

00:32:55
Speaker 1: Don't want emotional, you don't want fiery foxy or fighty foxy.

00:33:00
Speaker 2: I take you fivey fox. No, I want five fox.

00:33:04
Speaker 1: Yeah, all right, five e fox. So you want five e fox. I want to be like a cartoon, except not a paper club.

00:33:12
Speaker 2: Not a paper club.

00:33:15
Speaker 1: Let's take a little break, Nate, and then talk about n video and another element of AI, the chips that make it happen, Nate, this has been such an AI E AI E. That's a weird word, but you know what I mean.

00:33:39
Speaker 2: Week.

00:33:40
Speaker 1: The other kind of big news has been in video and the fact that you know, we've gone through quite the cycle on in video where at first there was a ban on in Nvidia selling its chips to China. Then within the last month the ban was softened and Trump announced that they had kind of reached a deal where in video was going to be able to sell some of its H two O chips to China. And then all of a sudden, there was an announcement that now Nvidia can sell these chips as long as the US government gets fifteen percent of the profits. So this is starting to seem a lot like we're now in the world of the Godfather or the Sopranos and less like we're in the world of the US government. Give me a taste, Give me a taste, Nate, and then you can do whatever you want. But Papa wants his taste of the action. Yeah.

00:34:37
Speaker 2: Look, I mean Trump had his, or the White House had. It's like ai action Plan, which you talked about a couple weeks ago, and like, you know, people I trust thought it wasn't that bad. But like these are people who, you know, one thing you might think is good for Trump are good for I don't know. Right, what I call the river in the book having more influence over the White House is like they're all really competitive. They want to be China. They want to be China, right, And here the US now has like an incentive for the best chips in the world to be sold to China. Right, I haven't tried. I assume it's going to like the Treasury and it is not like Trump's personal stash. But like that seems that seems a bit weird. And like granted, okay, you manufacture a chip and it's kind of hard for China not to get it. Eventually, I'm sure, there are black markets and gray markets, although you know, people have used the parallel like nuclear facile material. We track that pretty carefully potentially, But yeah, no, I imagine that. I mean, let's go with that analogy. Right, It's like, oh, okay, you can sell radioactive material to Iran.

00:35:33
Speaker 1: Right as long as as long as hes goverment get.

00:35:36
Speaker 2: Think China's Iran or exactly right, but they are the right now only country that's competitive with us on AI, right, yeah, the third place, right, maybe maybe the least you know, let's give them some two.

00:35:49
Speaker 1: Yeah, no, it's kind of crazy. And and Trump just says that the more advanced black Well chips, He's like, oh yeah, I'm open to us selling those as well if we can also get a percentage. So, all of a sudden, the national security concerns and you and I did a whole segment about the AI Action Plan, and the one thing that was kind of a concern in it was China, right, And all of a sudden that seems to have gone out the window. If we can you know, grease the palm a little bit and get get the fifteen percent kick back, and so from you know, all of a sudden, you realize, well, it was just a talking point, right, like, it really didn't matter. It wasn't national security doesn't matter. If we can if we can get a percentage of this. By the way, after these announcements, China has itself said, hey, companies, as in like Ali Baba, you know, Chinese companies, we don't want you buying these US chips because we're worried that they're going to kind of insert location tracking and back doors and all these things into them. We want you to be buying local That would be smart. And Video said, no, no, absolutely, we would never do that. But so they say, you know, we want you to buy local Huawei chips and you.

00:36:58
Speaker 2: Want your local pitches and exactly.

00:37:03
Speaker 1: So China is actually like a little bit skeptical of this, but it's not a law, and you know, it's not actually clear yet how it's going to be enforced because the agency that issued this directive doesn't actually have enforcement power. And by the way, there's also already been a pre order of I think it was like seven hundred thousand something like that. There's been a pre order of a shit ton in other words, of chips that are presumably going to start getting shipped now and so the now you know, we talk a lot about incentives, but they're just for the US. Like now, it's all out of whack. The other part of this, by the way, is that there is a ban on the in the Constitution on export taxes, and so I can see there being a legal challenge here. I mean, this kind of is an export tax Actually, think about it A fifteen Like they can probably.

00:37:54
Speaker 2: Ban in the Constitution of export taxes. Yeah, I didn't know.

00:37:57
Speaker 1: Yeah, there's a ban in the Constitution on export taxes, so we cannot lovey export taxes on companies. Yeah. So if you can argue that this is.

00:38:07
Speaker 2: We're going to change constitution, then it seems.

00:38:10
Speaker 1: I mean, I think that we have seen that this particular administration has no problems changing the constitution or trying to change the facts when the facts don't agree.

00:38:20
Speaker 2: If you were allowed one free constitutional amendment, what would you do?

00:38:27
Speaker 1: One free constitutional amendment?

00:38:28
Speaker 2: One free I guess you're by FIAT allowed to enact a constitutionalism.

00:38:33
Speaker 1: Oh, I know, it's a really good question. I don't have a ready made answer for it to you. Is it something you've thought about?

00:38:39
Speaker 2: It's harder to produce some practice than in theory, but like to ban gerrymandering is.

00:38:44
Speaker 1: One that would be amazing.

00:38:46
Speaker 2: Yeah, yeah, pretty good one that would be a really good one. I mean the Senate, I mean, don't mean fucking you know, I think we should sell North Dakota to Canada.

00:38:57
Speaker 1: All right, all right, I need a little bit more explanation here.

00:39:02
Speaker 2: Too many fucking Dakota's. It's their nice states. It's such a there's surprising beauty in South Dakota in particular, I'll tell you that much. But I don't think that Dakotas need four centers between them.

00:39:12
Speaker 1: This is true. I mean, I do think that the senatorial representative kind of model of government is broken.

00:39:20
Speaker 2: You trade North Dakota for Greenland.

00:39:27
Speaker 1: This podcast is devolving.

00:39:30
Speaker 2: Like, oh, I've just came home from Denmark. Oh you mean Fargo, Denmark. Never mind, we're typing in the studio. If you're listening to the auto version, we have our producer turning around and.

00:39:43
Speaker 1: Take it back to the video, Yes, take it back to n video. So, yeah, we we have this incredibly perverse situation right where the incentives are just now completely fucked. Now, let's assume the reason we got on our tangent is because of the export taxes. Let's assume that legally this is upheld, that the fifteen percent is allowed to stand. It also sets an incredibly dangerous precedent for you know, for everything, for like, it's a really it's a really scary proposition to think that, Now, well, as long as you kind of give a kickback to the US government, you're fine. So we do see this kind of quid pro quo mentality where like, you know, you help me, I help you, And that is something that norm that should not be happening in a healthy democracy.

00:40:31
Speaker 2: Yeah, I mean I think I think that train flew right past the station. Marie, I don't, yah, you know, you look, it's it's part of It's the same thing with the tariff policy. Right. Why are we doing these tariffs that try to encourage the Eastern American made products. Is it's trying to do industrial policy, is trying to do foreign policy right, or just trying to make a bunch of money for the government, right. And sometimes Republicans are kind of caught or tariff defenders are kind of caught in between saying, you know, actually the ideal amount of money is terifts make is zero, because then we on shore everything and people saying, hey, that's going to make up for like a loss of tax revenue elsewhere. Right, And it's kind of the same thing with this with this China thing and again, you know and videos like okay, sure, yeah, by the way, I own a video stoic, right, so yeah, yeah.

00:41:16
Speaker 1: All right, yeah, well no, but you can actually envision a future, right where for companies they're like when they're calculating profit, they're like, and this is the cut, Like just like before you had to like this is the cut we give to the mob boss, like, this is the cut that goes to Trump for it's the cost of doing business. And so we're willing, we willing to do that.

00:41:35
Speaker 2: The Italians, you know, Italians, French, we're acting like the fucking Europeans, you know some yeah, yeah, the southern less efficient Southern Europeans. Yeah.

00:41:48
Speaker 1: So so it's a obviously, I mean, it's an the understatement of the day to say it's not a good look. But it also doesn't just it doesn't bode well for a lot of things. Yeah, so I think, you know, bottom line, like this really is I think bad for our economic prospects and for the way that other countries see US as well. Right, like that matters, especially when our currency matters, and kind of our reliability as a trading partner and as a lender, and all these things matter. So reputation, reputation matters, and foreign reputation matters, and in other stupid things. By the way, we're we're going to have be hosting Putin on US soil, even though he has a warrant out for his arrest. Really nice to see you.

00:42:39
Speaker 2: He's a warrant who has an unrest warrant in the US for him.

00:42:42
Speaker 1: Well, the ic C has one, Okay, so he technically cannot leave Russia because any other country would have to be.

00:42:48
Speaker 2: Kind of fun a Vladimir, that would be funny.

00:42:53
Speaker 1: Funny that is not happening. Yeah, for war crimes against EU crime is it in Alaska or something? It's in Alaska. Yeah, yeah, so we'll see what happens there. But anyway, all of this not great news. So we have yeah, we have a mixed bag for you today on a risky business and export taxes. Let's see, let's see what happens on the legal side of things, and if this is in fact deemed an export tax every time will be challenged.

00:43:23
Speaker 2: GPT five hallucinates they have to pay the US government fifteen cents.

00:43:26
Speaker 1: Some of that that would be That would be something, Nate, that would be something. Let us know what you think of the show. Reach out to us at Risky Business at pushkin dot fm. And by the way, if you're a Pushkin Plus subscriber, we have some bonus content for you that's coming up right after the credits.

00:43:48
Speaker 2: And if you're not subscribing any come on really, but consider consider signing up for just six ninety nine a month. You get access to all that premium content and ad for listening across Pushkin's entire network of shows.

00:43:59
Speaker 1: Risky Business is hosted by me Maria Tanakova.

00:44:02
Speaker 2: And by me Nate Silver. The show is a co production of Pushian Industries and iHeartMedia. This episode was produced by Isabella Carter. Our associate producer is Sonya Gerwick. Sally Hilm is our editor, and our executive producer is Jacob Goldstein. Mixing by Sarah Briguer.

00:44:18
Speaker 1: If you like the show, please rate and review us so other people can find us too. But please only rate and review if you like the show, because you know we like good reviews. Thanks for tuning in.