Why ChatGPT Sucks at Poker

brain power and the US competitive edge. So for people who don't subscribe to Silver Bulletin, which you absolutely should, Nate and I both had poker posts this last week. Mine was about cheating, was about chat GPT, but his post this week was one of the funniest things I've read in a while about his attempts to get chat GPT to simulate a hand of cash game poker. And one of the reasons I mean, I found this amusing because I mean, first of all, it fused cards together

for an image. From the beginning to the end, it was pretty spot on in terms of being spot on wrong. But Nay, You're also someone who is constantly writing about how good AI is at so many things, and so it was funny to me to see it fall so short of something that actually tests intelligence as opposed to being able to kind of pull together a lot of things and spit out an answer.

There's also making adjustments for how other players play. There's the conversation you're having, the physical reads that you're getting their stuff, like knowing what the rules are right, which might seem trivial, but we've yeah, chatchpt also.

you're tired, you make mistakes like that. But and you also have to have like a lot of short term memory, which sounds trivial, but you you know, you want to remember how you arrived at this current spot in the hand, right, you have to keep crack of all these stack sizes, which again seems trivial, but like it's a good test of general intelligence of a certain type, especially a live poker hand, like I'm asking.

there are a couple of issues with that, right. Like one is that, like, clearly, if you train a transformer model, a machine learning model on a particular type of problem, like it can do fairly well or very well often, right, Like poker is to a first approximation, and it's an important approximation. Right, But if you have poker dedicated tools, I wouldn't consider a solver and AI that's a technical distinction I think might not be that important for our listeners.

But like, you know, if you want to use computers to play very good poker, then they can play very very good poker. Right. The question is like, can you take a text based model and have this organic property where intelligence emerges and converts from the data set towards superintelligence without training it on poker specifically. And the answer, at least as of last week when I did this, is no, it fails miserably.

broader scale, that would mean something in terms of kind of broader intelligence capabilities. And before LMS were developed, that had not happened. Right, There had been computer programs, kind of AI algorithms that have been able to beat heads

up one on one opponents in poker. But when it came to full ring games, So for people who don't know poker, that means basically, when you put it in a rich environment with you know, six players, eight players and you have a computer there, it was still not quite there. And one of the you know, one of the reasons why poker so interesting is it's not just math, right,

especially when you have multiple players. There's so many dynamics and there's actually a very thorny problems for AI, which is that if you program something to follow an algorithm right to be GTO game theory optimal, then it can be exploited by a human if the human figures out the GTO strategy. Now, if it were another human, then you'd adjust immediately right the moment the human adjusts, the other human would adjust and kind of figure it out.

But if you're an a if you're a computer, and you build in that adjustment parameter, what ends up happening is that the model adjusts too much and too quickly, and so that lacks kind of some of the nuance and flexibility that marks that's kind of the hallmark of the great poker players, right, kind of that ability to

into it when it's time to slightly deviate. And I mean, it would be terrifying but amazing if a model could spontaneously, like an LLM, could actually spontaneously figure that out and so far, I mean, instead instead of doing that, Nate, you know, your model decided that the person who won the hand actually lost the hand, right, It just got basic things completely wrong, and you just you know, obviously we've talked about pe doom a lot on the show. I think you and I are both worried about that.

But when I see this poker stuff and like, oh man.

It's late at night, you forget the Stata command for something, You're like, right, give me the right command and write five lines of code, and like usually it works, and it kind of now shows you like the chain of reasoning and like what it's thinking. Right, And but you know, people are saying that we're going to have like AI revolutionizing all programming or you know, well, first of all, the big claim is going to have AGI, meaning that a machine. It's a little ambiguous when people say, does

this mean like chat GPT itself is AGI? Right, Well, clearly not. It's very far from human being in some things, right, Does that mean that a combination of transformer or machine learning based technologies is or zooming out the definition of AI. It's really AI still, right, so that you have chechiptiza seventy percent of things well, and then you specialized applications for twenty percent, and then there are ten percent where you're not doing it very well. Like that would still

be a big achievement, very disruptive achievement. I'm not sure it would kind of quite count as AGI. Like one thing you could do is, like some of these new models will say, okay, I can detect that you ask for something that requires mathematical precisions. So let's take a simpler example. Right, Let's say I want to take the

distance between pick two cities, Boise, Idaho and Tampa, Florida. Right, that's pretty easy to calculate mathematically if you look up the latlong coordinates for Boise and Tampa, and then there are various formulas that.

of mine. Where now I'm going to go and I know to search the internet for the coordinates or maybe it's stored somewhere right of Boise and Tampa. And then the formula. I know the formula and I can apply it, right, so like you have the scaffolding. It's sometimes called a different processes on top of one another. And I guess that kind of I guess that counts as I mean, it's kind of how humans solve problems, right, They're like, oh, and now I have to go look something up, right,

It's still I think subtly different. So one implication I think this has is artificial general intelligence versus super intelligence ASI, sometimes called where there's like an explosion of intelligence just from kind of mining text or other data sources. I have actually become a little bit more skeptical of that, or I think it's presumptive to assume it can find that.

I mean, like, let me give you another example, right, like one I had an idea for either a silver bulletin post or maybe even a risky business segment, right where like I asked, chatchipt deep research, design a meal with foods that have no clear precedent in other countries, right, huh? And I thought, and the idea was that, like, you know, at some point somebody invented pizza or spaghetti or whatever else, right, or or I.

It was like their little chefy, meaning like things that like have one too many ingredients, right, and like and I don't, yeah, look so bad chefe.

for a certain task, it's very helpful. For certain task, it's somewhat helpful for certain tasks, not helpful at all. I can get you in trouble, right, which again contrast with the experience of people at the AI l apps to say, yeah, it's going to like displace all programming within a year or two.

it was, but one of the big ones. And someone had asked him earlier, what was something that was surprising about the way that this AI functioned, and he said, you know, one of the mistakes that people is kind of they trust it, And what we know is that it's remarkably accurate ninety percent of the time, right, and then ten percent of the time it isn't. But the ten percent is getting harder and harder for humans to spot because it doesn't make errors in a way that's

intuitive for people, right. It makes errors in different ways. It's like computer errors AI errs, that's not the way that the human mind makes errors. He's like, but like, it's fine because we have really smart people who can figure out where it's making errors and can work with

it and can kind of calibrate it and help. And so my question, my follow up question, was, Okay, you know what happens when the new generation that's being brought up with this AI comes up and they've been educated with it and they've been using it the whole time, and they didn't necessarily get the same education that we got, right, because they aren't the incentives are different, and they're not

getting the deep knowledge. They're not actually able to figure out what are the programming errors because this person was talking that, you know, some of the biggest potential of AIS and like programming and biology and kind of those those types of things, and I said, well, what if that person doesn't have the neuroscience and the biology background

or the programming background. And what the CEO said was, this is a major problem, and what we're hoping is that we can develop even more advanced AI to help and to help fix the problems that AI is causing. Because he was like, yeah, this is a big problem and we're not quite sure, like we're hoping for the best, but we don't know. And the fact that this was our conversation didn't exactly inspire me to heights of thinking this is going to replace intelligence or become AGI or

what did you call, Nate the other ASI artificial super intelligent? Yeah, artificial superintelligence. So it didn't inspire me that that was going to happen, and instead it was like an uh oh moment where like, what happened when the humans who are kind of helping push it along and make it better?

When that generation, like when they age out and they're new people out there who who lack that sort of skill and expertise and who grew up with AI that is ninety percent accurate or even ninety five percent accurate, but they can't. It's harder and harder to spot that five and ten percent. And that to me was actually, like, it's a worrisome thing, and it's clearly worrisome to the people who are developing these technologies as well. And so that I think is something that dovetails with what you

were talking about here. What happens Nate if all of a sudden, all poker players are training with chatchipt right, and that's that's actually great for us, but not great for the future of poker.

saving device. Often they say substantial labor, right, Like I use them to copy edit articles I posted the newsletter, and I you know, I'm a fairly proficient user of the English language and like for programming, right Like you know, if you're like using AI to like design a website and some of the functionality breaks, right, that might be okay.

application of heuristics. Like one of the things like the AI did so at first, I had it simulate just one hand, and it had like a backstore for each player. It fucked that up, right, then like Oka, I'm gonna use deep research now and have it simulate like an orbit of eight hands.

various excuses. One thing it said is like, well, the quality of text based content on poker on the internet really sucks, Right, it's aware of that is it says it's worse than go or chess, although when I talk to people who know how AI's played chess, it's also a disaster. Even something like wordle I heard from a user.

You know you think wordle is a word game and it does very well, but like it can't quite figure out the strike sure of the problem, right, it kind of was learning a little bit more by by a rote And to be clear, like this is a very good strategy for like many types of problems. And also if you were to say, okay, here's a bunch of we bought data from PIO solver or gto wizard. These are solvers if you don't know listeners or a database

of high stakes hands. Like if you trained it on that and then had GPTs say okay, poker question, I'm going to call them especially trained database, then it would do well. And maybe if people are criticizing its for on poker, then the labs will start to do that. Right, but it's still kind of like slightly cheating from the

standpoint of super intelligence. Also, it says like when you asked me to do a whole bunch of things at once, right, So I told it, give me characters, give me eight characters who have like they were all kind of like ethnic stereotypes, right, it was.

of stacks. I have to store a bunch of stuff in memory, and then when you ask me to do a whole bunch of things at once, right, and like you know, and even including like so in the hand I show a sylver bulletin it like it misreads the board, doesn't recognize that a pair of nines has a higher

two pair. If you ask it that out right, then it thinks about it more and gets the question right, right, But like it loses track of things in the context of these scaffolding situations where it has to keep track of a bunch of things at once and it doesn't know it doesn't know what to prioritize right. Like you know, you can make a lot of mistakes in poker if you don't know which hand be twitch hand. That's a more elementary mistake than anything else.

when we keep saying thinking. But one of the first things, so when I was just learning poker, one of the first lessons I had with Phil Galfon, who is one of the people who kind of I worked with and who taught me a lot. I remember very early on he said, you know, I can give you a bunch of charts and a bunch of outputs, and you can memorize them and you'll be a very decent player very quickly. Like I know, I can give you this shit, you can memorize it, you can spit it back at me.

You'll be fine, He's like, but I don't want you to do that, because that's going to make you a fine poker player, and you'll do fine in the short run, but you'll never be a great poker player because you have no idea why you're doing it. You're not actually thinking,

you don't understand. You've just done a rote memorization, which is might make you more money in the short term, but in the long term is actually going to be detrimental to you because you're going to do stuff unthinkingly because you've memorized it.

two lllms, they'll be doing some rote of memorization. You know, at this point they're not understanding the why, and so that will lead them still to make big mistakes, which happens when you just learn solver outputs, no matter how sophisticated those outputs might be and how correct they might be in one specific hand, in one specific spot, if you don't understand the why, you're either going to overgeneralize it right or misapply it like you're going to screw

it up because you don't understand the underlying reasoning. And there's the difference between looking like you're thinking and actually thinking. And yeah, you know, as a human you have other pitfalls, and even if you're you know, thinking through things, you can mess up, So you mess up in different ways. But I think that's a really important distinction and probably one of the reasons why AGI is not as close

like poker. Actually illustrates in a very practical way, why the notion that you can just have a lot of data and all of a sudden have this you know, flurry of insight might not be as I guess, as intuitive as one might think.

because it's adversarial. You know, understand that there's some game theory there, but it's adversarial, and like there's no one I mean, there is like I guess some superstructure of like a solution for all poker hands if you get very zoomed out, right, but like but like subtle things when you have to calculate like an adversarial equilibrium on the fly basically, and it's that's very difficult.

Plus I mean, I think that anyone who has used solvers, and anyone who's talked about this understands anyone who doesn't even play poker but has worked with algorithms, the common saying garbage in, garbage out is absolutely accurate.

all of a sudden, your solution means nothing. And as a human you can figure that out, and you can kind of make adjustments from baseline, right, if you understand what the baseline strategy as you can adjust as an LM, as an AI who is learning that but doesn't kind of have that nuanced experience at least at this point.

I don't pretend to know what AI is going to be capable of in you know, in five years and ten years, but at least at this point they're not capable of making those sorts of nuanced extrapolations and figuring out, Okay, what were my inputs accurate, right, or were my inputs not quite accurate.

Like there are some applications where it's pretty good. But like, you know, the other day, I was getting getting a drink with my partner and a guy who looked like he was homeless comes in and like hands like the hosts slash barked like a note saying call nine to one one or something like that, right, and you kind of have to make this judgment call about like is this like a paranoid schizophrenic or is someone actually in danger?

And like I think was pretty clear. I mean they're probably you know, two bartenders and eight customers clear to everybody. Like this guy I think is not like acutely in danger or anything. But I'd never quite experienced situation like that before. And like the fact that we can use like our general intelligence about like human behavior Yeah, Like I think everyone made the right decision that we're not going to call the police either to rat on him

or to say he was in danger. And like you know, and that kind of thing is is is hard for language yells to do, you know.

than a lot of kind of super intelligences. Right, Like it took forever for a robot to be able to pick up an egg, right, It's something that we don't even think about, but this was like a robotics problem that was absolutely unsolvable, Like how do you get a robot to pick up an egg without without breaking it? And we do it without just without thinking about it.

And we make judgments all the time, just silly judgments that we don't think twice about that are so easy for the human mind, but that are that we don't even realize we're making right things about safety, things about just perceptions of the world. So I think that you know, humans are much smarter than than we think in like in dumb ways, if that makes sense, Like in waste that seemed like they're not hard problems, but those are actually some of the hardest problems to solve.

And like, you know, my heuristic is that like within a few years we might have roughly an even divide between those four and I'm counting, by the way, tasks that involve manipulting the physical environment, which I think AI will mostly be pretty bad at. Right, I don't know how much I should extrapolate from the poker example. It was just so incongruent with these predictions of like imminent AGI, right, where a year ago would have said, okay, yeah, of

course suck at poker. Right, it's not really trained on this and it's a hard problem, and ha ha ha, Right, But like if you want to have these complicated, like structured, nested tasks, like I'm much less worried now in periods of oh, let's say five years of AI like being able to build like atiscal model to like forecast elections or the NFL or whatever else, because that like just requires like a superstructure of lots of little tasks, all of which are fuck upable, And if you get the

superstructure wrong too, then you're kind of just drawing dead. To use a poker term, right if any if you have to chain together twenty steps and any step, there's a ninety percent chance you get it right. Well, point nine to the twentieth power means you're almost.

Like again, I I, you know we are talking about we don't other segments where we're like amazed by AI progress. Like it's very good at verbal reasoning, or at least faking that. But in ways I if it's faking, it's doing a pretty good job. Right, Like it's very good

at verbal stuff for the most part. Right, And AI is traded on poker, are very good at mimicking Selver solutions, right and like and so like and machine learning can get you a lot of ways when you have good data, but like when you don't have the data in the training set and it's not seeming to extrapolate very well, and then these complex tasks that require more and more compute like, I have not been impressed by deep research, which is where it goes away and says I'm going

to perform several tasks for you that require deeper thought, right, and then it takes fifteen minutes you come back and like, are you fuck this up? Right? And like, anyway, so I think it has affected my priors a little bit.

of demands. Harvard said fuck you and sued and yeah, and tried to go that route, and the funding was frozen. And then the administration last week, so we're taping this on Wednesday, the twenty eighth of May. So last week, the administration decided that it wanted to punish Harvard even more because you know, they didn't like the fact that Harvard wasn't complying, that it was being defiant, that it was taking them to court, and so said, hey, Harvard,

you can no longer have foreign students. Anyone who has visa's effective immediately. Those are going to be revoked. So Harvard's class of twenty twenty five that's graduating next week is going to be your last basically foreign student class, and we're not going to be granting visas anymore to those students. Harvard has sued again, so now we have a second set of lawsuits. The initial court cases have

frozen that. But then after that happened in the last few days, the administration said Marco Rubio actually said that we are going to freeze interviews for all foreign student visas, not just Harvard. So they not only you know, went after Harvard, but doubled down and said, you know what, you can sue us and you can freeze this. But if we don't grant the interviews, last laugh is with us.

of framing what's going on. And he is someone who has attacked Harvard. He's a tenured professor at Harvard, but he's attacked Harvard many, many times for not standing up for free speech, for not protecting conservative viewpoints, for being a little bit too woke. Right. He has been one of the main people who said, hey, like, we have an issue here. But this his op ed was, this ain't the way, right, this is not the way to

solve this problem. And you're actually attacking the students who are some of my most critical, thinking, least woke people who are on campus.

is Harvard trying to achieve? Like I support maybe some listeners will get pissed off, right, you know, I think clearly some of these colleges are not complying with like the spirit and maybe the letter of the Supreme Court's affirmative Action decisions, right, and so like, I wouldn't mind if like the Trump administration's like aggressively enforcing the law on those or things like diversity statements, which are enforced

political speech. Although Harvard's a private organization, so it's a little bit more complicated, right, But like I you know, I don't it's such a blunt tool to like target

foreign students. And even if you hate Harvard, you know, as a patriotic American, I want the best and brightest from all around the world to come here and contribute to our economy and pay our taxes and everything else, right, And like it's just like purely destroying the value of these very bright students from helping our economy.

And that's super smart, right, Like, universities all over the world should be doing stuff like that right now, because what the administration is doing is kind of bankrupting at the future of kind of intellectual development in the country by saying, oh, foreign students, you know, we've we've just frozen all these applications. So to me, like that's super smart, right, Like that is the we've we talked about this several months ago, Nate, Right, what can China do to capitalize

on the moment it's doing it. It's being like, come here, like you don't even have to apply, you don't have to do anything to show us the letter that showed you were accepted and welcome. Come.

to be any strategic value to this. There's not any any value in terms of trying to get at the types of problems that the Trump administration says it wants to get at, because it's not even addressing that right, Like at the originally anti Semitism was the pretext for this, and this is just like so far divorced from it. Everyone is having their visas revoked or not granted, or their interview is not even granted. So it's just, you know,

the pretense is now crumbling. We always knew it was just a pretense, but at this point, it's just not accomplishing anything other than to further could of create this absolute chasm in higher education and future brain power, future research ability, future creativity of the United States basically mortgaging its future right saying that we were going to give up one of the biggest, if not the biggest edge that the United States has had historically, which is kind

of this brain power creativity place where people can can develop and find support for for for their ideas, and now that is not possible and if this actually stands, right, if foreign students are not allowed into US universities, because right now, like I said, this interview process being paused for everyone, not just Harvard. I mean, that is going to be just detrimental in so many respects to what happens here.

had it tweet. I don't know where he was like, he's oddly transparent sometimes about his thinking, and he's like, well look just like, look, these colleges aren't serious about complying with the law citing like the students for fair admissions. I think it's called which was the affirmative action decision, which I do think they've kind of not been serious about complying with in some cases case by case basis, and so like, unless we kind of show real muscle,

then what can we do. We have to fight them because they're in the wrong, and so we're overreaching. It is a subtext of that, right, which is like maybe not crazy and they have a real leader because Harvard isn't that sympathetic perceptions of higher education that's gone way down, so it's like harder to marshall public opinion on the side of this, right, and like and like, you know,

I saw some Harvard professors on Twitter. Ever we can be like this is the greatest threat to American I'm like, please don't say that, because it's it's it's maybe more of a threat to American competitoris and not sure it's a threat to democracy per se.

all of these things in different ways. And I have no idea, you know, how successful that will end up being, but it is. It's a it's a tough you know, he's been put in a really tough spot because obviously

he hasn't been president for long. The last president departed on lessons tellar circumstances, as did a number of presidents of Ivy League institutions, and you know, you're picking up the presidency at a very difficult point in time where, yeah, you have to acknowledge that Harvard has made mistakes, right, not just Harvard, a lot of universities I think were trying to be I don't even want to use the

word woke, because I don't think that's accurate. Here but we're becoming, you know, places where people were afraid of

voicing their minds. And I wrote, I think almost a decade ago, I wrote a piece for The New Yorker about basically, you know, the problems in academia that come from not having enough conservatives right that it actually has trickle down effects in research and the types of things that are being researched in the findings that are allowed to be voiced, and that it can be incredibly problematic. And so I think you need to acknowledge that, and

those are things that need to be rectified. At the same time, you are currently fighting this battle against you know, Donald Trump, who's trying to destroy higher education in the United States, and so that is a you know, overriding priority, which doesn't dismiss all of the things that have gone wrong and all of the things that are wrong with academia, but it's just a it's a kind of this fine balance where you're like, yeah, I know I'm not sympathetic

because I've fucked up and I've done all of these things. And yet right now I'm one of the people best positioned to fight this, and so let's just try to figure out how to prioritize this and how to not spread ourselves tooth in so that we can actually protect higher education and then start to remedy the problems that we have internally, because I do think those are problems that need to be remedied, right, Like, you can't just be like, well, now we're all united against Trump, so

let's forget we ever did anything wrong. I think that all of these things you remember, Nate, like less than you know. Like six months ago, I was railing against Harvard because I was really pissed at the way they were reacting to a lot of things. Now I think

now I'm supporting it because they're being attacked. It's the classic psychology thing, right where you can actually unite a lot of people who who criticize you when you attack, right, all of a sudden, these various factions, when they feel these life threatening attacks from the outside, all of a sudden, these factions start uniting. And I think that's what's happening

at Harvard right now, as it should be. But we still have to realize that, you know, there are all of these different issues and it's a really I mean, it's a shit show. That's a technical term.

might like and that people would find tolerable. On the Harvard side, Right, we'll collect data on the political affiliation of our students and faculty. We will make concerted efforts to I'm not going to try to frame this right conservati efforts like higher conservative scholars. Right, maybe you have like you know, if you have an ethnic studies department, you have a conservative stay.

like you have blind admissions. You are willing. You just don't you don't ask any of that, and you don't disqualify people because they have conservative leanings, right, and that it's not even it's not a thing because you don't really care, Like I don't care if my uh, you know, neuroscience professor, Well maybe not neuroscience as insofar as I study psychology, but if my astrophysics professor is, you know, what their political views are, as long as they correctly,

you know, as long as they're one of the best astrophysicists in the world. So so I don't think that that's you know, that what you're proposing is affirmative action for politics.

hold up their quid pro quote. This is kind of very transactional and explicit, right, So they're not even like presenting clear enough information to make themselves easy to bargain with, even if it were some theoretical optimal solution or win win.

You and I are not going to be out there for another It's another basically two weeks, I think for us week and a half. So let's wish all of our Risky Business listeners who are already in Vegas for the World Series of Poker good luck. We hope that you all crush it at the tables and we're looking forward to joining you soon. 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.