Latest / Last Humans in the Loop / Math, Encryption, and the AI Bubble: What AI Could Break

Transcript

Math, Encryption, and the AI Bubble: What AI Could Break · Last Humans in the Loop

• Zach: Welcome to Last Humans in the Loop, your weekly AI news podcast. Today we are covering the math occlys, aka the slop drop, where we ask, is math ruined forever? we're going to be doing some security, such as is public key cryptography ruined forever? And we'll also be discussing the potential AI bubble, which may, I don't know, ruin things forever. plus all the news and some weirder stuff. So I'm here. With Chris Slow, who is everyone's favorite former Reddit CTO. I'm also here with Marty. Marty Weiner: Mine too. Zach: Marty, former Reddit CEO, CTO Marty Wiener. so let's get into the the big thing this week was the math occlip. It's called the math occlip by Scott Aronson. by the way, just point of etymology, math occlipse would actually mean math revealing, which is Marty Weiner: Mm. Zach: is kind of appropriate. Christopher: Ooh. Zach: right. also known as the slop drop. so let me set a little context here. So OpenAI did this announcement like a week and a half ago saying it had solved or resolved a hundred open problems, which was already a big deal. And then like a few days later, there were rumors that they'd solved 400 problems. And then they actually did a drop with 722 resolutions. there's an article in Scientific American, and apparently every one of them was done with a single prompt to a single AI agent. It wasn't Marty Weiner: Mm. Zach: a giant mega swarm like with Navio Stokes. and you know my understanding as a person who doesn't understand anything is that some of these were were some of these were pretty trivial, but some were pretty big deal. That's right. What's the opposite of a polymath? Like a monomath, Christopher: That's You're a pol you're a polymath, Zach. You're a polymath. You understand a little bit of everything. Zach: a null math. I don't know, but but so the the big deal, and I'll point at Chris to explain it. As I don't know if he does know how to, but the quasi-reman hypothesis. Do you know the the basic idea here? Christopher: this one I this one I didn't even track. so I've I'd say okay, so the the funny thing about this is I think my favorite part is that that they did this out like was it three weeks after they announced that they were gonna have a an AI sort of like standing committee to responsibly Zach: Yes. Christopher: release math. And then they're just like, Nope, we're just gonna put it all on GitHub. Zach: Well well and and and also that the committee had specifically said, please don't do your mouth on our big open problems. so and so yeah, so so Acerfer, who is one of the early Airdish guys, is also like twenty years old and a furry, and just I am jealous of his entire career and lifestyle. he called solving the quasi-reaman hypothesis move thirty seven for math. And and he was saying he thinks we'll for some definition of solved math will be done next year. w worth noting that there's also a lot of physics and NCS in this stack too. Marty Weiner: Yes. Yeah, there Christopher: For the Marty Weiner: was one on Christopher: the really the really the really quick one on the quasi Riemann, by the is okay. So the so the there is a there is a function in math called the Riemann function. It's the Riemann Zeta function. And it has implications for the the the structure of this function has implications for factoring prime numbers. And if you ask a question of why, we could spend the entire podcast explaining why. There are books written on the Riemann zeta function, and there are conjecture there's a conjecture on the Riemann zeta function about. where the zeros of this function land. and so the problem is that no one's actually been able to prove where they land, but there's a lot of guesses where they land. And because it's been almost 150 years, if I I math works out right, working on this function, there's been these conjectures that have been built up on setting bounds on where the zeros are. And so the quasi-reman conjecture is about the it's about setting the one of these bounds, if that helps. And the answer is crickets cricket's math words. Marty Weiner: And I think I think one of the crazy parts is that this because this drop, as Zach said, had a ton of physics. And so we're really cracking into physics. There's one, the L2 Einstein gap. I had to look that up. What the hell an Einstein gap is, but it has to do with crazy high gravity lensing out in space. and so it's it's hitting the astronomy world. and there's another one on a coulomb ground state density without combsham and ensemble ensemble representation. Christopher: Ensemble. Yeah. Marty Weiner: Don't know what I mean. I assume it it's rel related to to to charge densities, but I I don't know anything about this. It is starting to hit physics. Zach: There was there was another cool thing, like in terms of the culture where there was a a an improved solution for integer multiplication, but the improvement was like minuscule. It was like one minus ten to the a hundred eighty or something. did you guys see this? Marty Weiner: It's really weird. Yes. We're seeing that a lot with the computer science problems. It's like a weird, like not not a s you know, not in squared reduced to linear or something, but we're saying like in squared reduced to in to the 1.9995 or something. Zach: Well but what was weird though, so so a bunch of people jumped on it 'cause like once the dam was broke, people got excited. Marty Weiner: Yeah. Zach: And now I think they have it down to like, you know, ten to the negative four or something. So it's like it's actually like, you know, a hundred orders of magnitude. Yeah. Marty Weiner: Yes. I there's a whole th series of those. I don't remember if it was in there too, but there's another problem called threesum. that's not a for the nerds out there, not an NP complete problem, but but is in squared, I think, and they were able to reduce that a a a little bit more. Zach: Yeah. So so like the there there has been some. I think like three problems so far have been withdrawn due to errors in the formalization. So not all of these were formalized when they drop. And of course, even if they are formalized, there could still be errors in the formalization. so so some people I think took that as you it's total slot, but apparently it's it's most of the problems are, you know, it it it Christopher: I I've also Marty Weiner: So a and then some of them are ex Christopher: I've also heard some of these have problem the problems that the the prose doesn't match the lean proof. so if any if you treat it as like Zach: Yes. Marty Weiner: huh. Christopher: you know, 'cause it's and and the the thing that's interesting there is that if you think about the lean proof, it's like lean proofs are kind of like the software engineering equivalent of math. And so this is the same kind of like this is actually a problem I think that like software engineers have been experiencing for the last couple of years, which is you have it write something for you and it writes an algorithm that actually works. And you ask it to explain what it did and it gives you an explanation that isn't quite right. Or like, you know, and the the cuffs and color kind of don't match. That seems happening too. Is like a lot of the like something, you they find examples also of like it's in it was unsuccessful at numbering subsequent sections of the paper. It went from like forty six to well from forty from forty six to forty-eight or something like that. Yeah, it jump it jumped over a number. and they're pointing at us is like, look, it can't even Zach: Yes, yeah. What it missed forty five, I think. Yeah, yeah. Marty Weiner: Wait, wait. Was that where the cryptography papers Christopher: it can't it can't even It's the sec the secret Marty Weiner: Was that where the cryptography papers were? but Zach: It's a secret withdrawn paper. Christopher: the secret fourth newsbook. but no Marty Weiner: Yeah. Christopher: I there just I it it it does seem like they're There's a lot of stuff around the fringes of there's just some skepticism that these are good, but that doesn't mean throwing the entire corpus out. It just means that there's probably Marty Weiner: No. Christopher: some bad apples in this bunch. But in considering that these were all like, I mean, we we've been joking over the entire existence of this podcast of like, do a breakthrough, but like the prop seems to be no much more complicated than like spend three hours cracking a an essential problem in math and just like see Zach: Yeah. Christopher: what you come up with. Marty Weiner: I went immediately and looked for anything practical and like like kind of that I you know it could go play with as an engineer, right? there's this great list somebody posted on r slash comp sci, all the complexity improvements. And I think one of you said that there was this minor improvement. I thought, you know, f if you make a a fixed if you make an improvement to how we do Fourier transforms, that's gonna affect the whole world overnight as long as it's actually useful. And and and it just I I know this is a little harder for for some people out there who don't know complexity theory, but it's sort of like a it's how it grows with the number of items in in your your set. So this grows with n times log n. They improved it by n times log n to the tiniest little it'sy bitsy fraction, 10 to the negative 13 improvement. and I don't know what that means practically. and I don't know. How the hell someone could figure that out? I need to go read that paper because that's such a weird thing. We don't normally talk in terms of complexity theory of one over 10 to the 13 change. so there's something cracking that's really interesting. I don't yet know if any of this is practical. The other thing to be f aware of with complexity theory is sometimes you say, Yeah, the complexity theory is better, but every single thing we do, like every single operation is massively more expensive. So it's actually not useful at all at at most scales. Zach: Yeah. So let can I can I try it for for the humans listening? one one way I think about complexity Marty Weiner: Yeah, yeah. Zach: class is just like if you have to add two numbers, you know, adding another digit to the two numbers you're adding doesn't make it that much harder. It's like an extra step. Whereas if you're like multiplying them, then because you have to multiply by each number, then it adds a bunch more steps. If you're cubing, it's even more. And then there are problems like, well, what's the best delivery route? And every time you add another point on the delivery route, it gets way harder to where like by the time you're at 50 possibilities, it's just like impossible to Calculate the best way to do it. And then CS nerds don't care about constants. They just care about like overall how difficult the problem is. And they that big O notation. And the point is, you know, usually we see basic operations in this stuff. So it'll be like, well, it's as n squared. It gets more difficult as you add a digit as n squared or n cubed or as log or square root. And so it's weird that we're now seeing instead of like, you know, n end to the one half, it's like n to the point nine nine nine nine or something weird. Marty Weiner: Yeah. Zach: And we we we which is like bizarre because I I remember reading Scott Aronson years ago writing like, you know, why is it that we never see, you know, end of the 10,000 or end of the point one. We it's almost always like squared or cubed or something. And he's like, it seems like nature prefers that. And and maybe the actually the answer is we prefer that. Marty Weiner: Well well Christopher: Humans humans like to simplify Marty Weiner: we and Christopher: it anyway. It's like if if you're gonna like big annotation is already a simplification to some extent of a much more complicated algorithm. And so but I mean, like I I think for for in this this particular s situation, like the human's gonna put a bound on the complexity and like one versus point nine with you know twelve nines is effectively the same thing, and therefore we're just gonna bucket it's the same thing. Marty Weiner: Right. Well, the algorithm doesn't even have that. Like the algorithm is is Christopher: Yeah. Marty Weiner: is just more like search a tree. And and to that question about why do we not have, you know, what O of N to the ten thousand. I think I mentioned in a previous podcast that that question came up in in my college class on this. And the professor had a great line that I thought made sense, which is just we're in s we're we're two dimensional, three dimensional, four dimensional characters i in the world we're in. we don't naturally exist in you know, like w if I'm trying to figure out how to get from point A to point B in my car, that's a two-dimensional problem. We we don't tend to think in anything above even four four dimensions. Zach: Yeah, so that's that's where these Marty, yeah, yeah. Christopher: Marty Marty thinks in four dimensions for the record. That's why that's why he's so good at this. Marty Weiner: I've I've already figured out this whole podcast. Zach: the the so well I I I guess we could say like the the overall response has been somewhere between fury and awe, sometimes in combination. and so I I wanna say obviously a lot of people were excited. Scott Aronson, I think, with some trembling, said we are now in or he said the singularity is here, it's just not evenly distributed. That is to say, mathematicians are now in the singularity. the the rest of us await. Marty Weiner: Right. Tristan Buckmaster, NYU math professor, said that the Zach: Mm-hmm. Marty Weiner: the math paper dumped wipe the math paper dump wiped out several careers of early mathematicians overnight. Zach: Absolutely. There's actually another quote. There was this article in Le Monde. and I have a quote. He said this was Tuesday night I cried and it woke up my daughter, confessed mathematician Hugo Dumenier Copain, the recipient of the 2022 Fields Medal. So so and he's not alone in that. There were a lot people who were like angry or bewildered or very sad. There was a release from the the new organization, the AHM, the Association for Human Mathematics. I'm just gonna read a little of what they said. Mathematicians did not ask for this work to be done. The advisory group on mathematics and artificial intelligence, from whom OpenAI has claimed to derive its legitimacy, opened their initial advisory statement by saying that Frontier AI corporations should not test advanced mathematical problems on internal models. they were furious. And Terry Tao, who's kind of been in the middle of this because he was he was mostly a booster and and has now got reservations, he posted that to his blog. although to to be fair, like he's posting many perspectives to his blog. I don't know that he strictly was endorsing Marty Weiner: Yeah. Zach: it. but so that was going around. Francesco Maggis, another mathematician, said reading the comments of junior mathematicians confirms what many senior mathematicians have been warning about for the past month. It is becoming clear that people with the intellectual ability to understand these proofs will not be interested in spending their lives training. That is to say, if we get a dump of the solutions, the subset of people who are willing to understand them probably don't want to because it's not fun or not glamorous. so Marty Weiner: Mm-hmm. Zach: That's bad or or potentially bad depending on how you feel about it. So Marty, but we were talking about there's this argument Terry Tao made that I think has been pretty widely panned by a lot of Twitter geeks, with with greater Marty Weiner: Yeah. Zach: or lesser degrees of nastiness. do you wanna describe that? Marty Weiner: yes, he gave a Caltech lecture called Math Two Po. and he had a hit a slide, a thought experiment on alignment and understanding. and he tries to relate this to the argument that, hey, you know, if we had an AI solution for curing cancer, would you use it? and he's trying to say maybe not, because I don't know what's in there. And so everybody starts, you know, on on the Twitter verse starts going, Well, bro, if I'm on the deathbed Hell yes, I'm gonna try your your cure. Zach: Yeah. Now in in fairness to Tao, I think on the next slide he he asserts that it's not gonna kill you tomorrow, for for the sake of this example. But I but I do think the the funny part to me Marty Weiner: Yeah. Zach: where where he was like, Would you use a drug? You didn't understand its mechanism, and all the medical people were like, There's something you should know, you know Marty Weiner: What yes. You yeah. Yeah. The doctor Christopher: We don't well but the Marty Weiner: I mean it it's common for doctors to use drugs that are for certain things for other things. They they don't know, Zach: Yeah. Marty Weiner: it just happens to work and they they talk about it around the the the fireside. Zach: Yeah. Christopher: You what's funny? What I would I would normally pan an argument like this because it feels kind of like hyperbolic, but actually like the issue with cure for cancer is that cancer isn't like just one thing. Like cancer is a whole set of of of diseases that are all bucketed under the same root cause of you know cells going awry, which is why there'sn't usually a single cure. It's like you can kind of cure each one. The neat thing about a cure for cancer as described by AI is it's a system for curing cancer. Much like right now, AI is acting as a system for solving math problems. so it actually is a really good analogy when you kind of take it apart. I I also i this does also feel like the mathematicians have kind of lost the thread. Like like they're kind of like what's the what's the purpose of doing the math? And you can't just say things like beauty. Like it has to be like much like a lot of math does actually have practical real world implications. What happens with that? Like cryptography, for example. Zach: Yeah. I I I I think the other thing there, this is the argument that that kept bothering me was like you would get people who are like, look, math has never been about theorems. And then the and you know, it and and then like like the in fairness, Christopher: Ha ha ha ha. Zach: that there were people spe people saying that before AI, like it's not just about theorems, it's about understanding. But then you you can't say that and then also say, and we're really mad we got all these theorems. Because if the theorems didn't matter, you know, then then who cares? and and so I like I to me it's like If if the problem is that young mathematicians won't work to understand these theorems, like maybe that's bad, but it's a problem like inside the field of math. It you know, so so I would say like their strongest argument, which I'm not sure if I buy, but the strongest argument they make is, well, this is gonna harm the public because without that deep understanding among mathematicians, because obviously Marty Weiner: Mm-hmm. Zach: the public doesn't actually understand any of this, without that deep understanding among us, then it's gonna end up harming the whole system. and I just Marty Weiner: Mm-hmm. Zach: think that's really it's really hard to believe. Christopher: Or or there's gonna be an or there's gonna be an entirely new generation of mathematicians who hold it all differently, right? Which is like which which I think is I think that I think the main issue here is that we we tend to live in a world where disruption of the industry is kind of the normal of normal behavior, right? It's like like the like Marty Weiner: Yeah. Yeah. Zach: Yes. Christopher: the your your job today and your job in five years are gonna be like radically different because nature of the problem. Whereas in math, it's kind of been the same, it's been the same game for pretty much the entire history of academic math. and so now they have to actually confront the fact that they've been doing things on a chalkboard and now there's like a computer that can do their work for them. It's like, you know. Marty Weiner: Right. Well, it's Zach: Yeah. Marty Weiner: also really scary, right? Zach, didn't you have a story about a mathematician who had been working on a problem for twenty years and was Zach: Yes, well, yeah. So Scott Aaron said this was kind of funny. Scott Scott had a post about how his wife Donna Moshewitz, who is who's also a an esteemed computer scientist, it's it's I've I visited them and it's the the conversation is I just nod a lot. Marty Weiner: Yeah. Zach: and but she's she's you know, there's some problem. I don't know if it was like her major all day problem, but it was problem Scott said she had been working on as long as they'd known each other and like, you know, they have they have kids who are like they have a teenage daughter now, they've been together a while. And I guess Their son who's like a little math whiz, I've met him, was like, ha ha ha. You're that probably working on for 20 years just got sloped drop, burn. and and so I think Marty Weiner: Kids are kids are great at that, yeah. Zach: kids are great. Ha ha, your dreams are dead. so you know, I think she was she was, you know, not like devastated by it. And she was like trying to understand the proof and said it was kind of weird and alien and not at all like a human would do. maybe that'll be fixed in a few years. but and then then there have been s go ahead. Marty Weiner: It's funny p they're they're describing this like the word alien has come up several times and it's exactly how I remember people describing when chess engines got really, really, really good. They said Zach: Yeah. Marty Weiner: it was like playing against an alien. And then they started to adopt the alien str strategies. Zach: Yeah, yeah. I mean I guess w you know, one of the complaints about the alien proofs it is a claim and I you it I don't think it's been adjudicated yet is that the proofs like often just reuse old techniques in extreme ways, which is like cool to get the solution, but doesn't enhance the field as much as like a human would. but maybe that's just a matter of time. But so I let let's cut to the end of I wanna I wanna get to the next topic, but what do you guys think is gonna happen? with like the field of math and like, you know, cause 'cause this is this just happened with a a brand new fresh off the presses model. And so presumably they're already past that. so what what happens in the next few years? Christopher: I st I I still think that the the the approach here that works is gonna end up being the cyborg math, right? You're gonna end up having mathematicians shifting their mode of operation to actually using these tools effectively. And like if look, if the if the proofs look alien and they potentially are indecipherable, that still might crack the door open for figuring out how to best articulate these solutions in a way that a human can actually understand. Right. Cause I think the goal like if the goal, which I do actually agree with, is like math should be understandable by a human, or at least you should be able to explain. what the shape of the pro of the proof is, then you need a human in in the loop there, plug for the podcast intended, to be able Marty Weiner: Yeah. Christopher: to make that into something useful for the math field. I just think that there's just kind of like it's like raging at the universe because the world has moved on from your method of operating. I just can't I just can't I just can't attach that notion. Marty Weiner: yeah, and I think that analogy makes sense. If we again back to the chess analogy, if you look at ninety seven, that's when Kasparov got bested by deep blue. And so then the best in the world was for a little while was not AI, it was AI plus human combinations. But then slowly AI got so good that the human was more of an impediment. now that may be because chess is a more simplified thing, that that you know it could basically be solved by AI and and humans are are make mistakes. I mean, literally somebody there was an AI plus human. The human moved the piece to the wrong spot. The AI still was able to recover it. but Zach: Yeah. Marty Weiner: but the you know, humans humans might make more little silly mistakes. there's a real chance, I think, that math may stick in that that cyborg space for for really long, especially if there's something about what humans bring to the to the equation that AI just can't quantify. Sort of like we're talking about how AI is Christopher: Mm-hmm. Marty Weiner: not great at comedy. like there could be something about the human experience or or what we're wanting math for that the AI may never fully understand. Christopher: So Marty, you're saying you're saying think of the math jokes. Zach: That's right. Marty Weiner: Mathematician mathematicians and comedians are the same. Christopher: Yeah. Zach: You know, it's funny as Kelly was actually kind of said I was talking to Kelly about this and she was like, isn't it weird how like the people getting hit first are the like poets on the one side and the hard mathematicians on the other side? We were trying to figure out like what is the most middle profession? you know, it's like it's it's it's something like Marty Weiner: And what'd you come up with? Zach: environmental science or something maybe or psychology, right Marty Weiner: huh. Zach: in the middle. maybe Christopher: Well I it was actually it was actually funny that like we talked about this last week was about that the like entry level accountant has just been knocked over like really recently. Like but it was like entry level accountant apparently apparently accountant. Zach: All right. Let's go to our next big topic for the week, which is security. So one of the things, and I think Scott Ernson said this first, or first for me, was among the 722 resolve problems, there was no cryptography. And so, mmm, you know, and and so like Marty Weiner: Yeah, weird, huh? Mmm Zach: I'm not a conspiracy theorist. but like, you know, Marty Weiner: But Zach: the question is how you interpret that. So one possibility is it just happens that cryptography is extra super hard. All the cryptography people are really smart and their work just can't possibly be hit by AI. That seems unlikely. you know, another possibility would be they just didn't turn the microscope on cryptography for like security reasons. But that seems unlikely too, because like they seem to have no problem just dumping stuff out and like, you know, wouldn't you want to know if if if there were security questions? Christopher: And we've seen and we've seen off the sh like r other researchers coming up with attacks on RSA in the last couple of weeks that that show that there is definitely I mean and RSA isn't actually a particularly like cracking RSA at this point is kind of a well known sport Zach: Yeah. Christopher: across cryptography. but it is it is it is a it's a it's a kind of a glaring omission at the very least. Zach: Yeah, so so so just for for the less technical audience, like RSA is a way of of of doing public pre public key cryptography. It's really hard to say. that relies on the idea that it's really hard to factor large numbers. Is would that be a a fair way to say it? Christopher: The way that public cryptography works, roughly, is that it relies on math that is intrinsically asymmetric. And what I mean by that is it relies in an operation that is easy to prove that you're right on, but hard to prove if you don't know what the answer is. So the example that's you always used, and that was actually used for RSA, is factoring large numbers. So if you have, if I give you two big numbers and you multiply them together, it's going to be very, very, very difficult in complexity theory to figure out what the actual two numbers were originally. But Zach: The prime numbers. Christopher: you the prime if they're prime numbers. But you can trivially show that if I give you those two numbers, that it gives you the right answer because multiplication is cheap. So factoring is very difficult. Multiplication is cheap. That is the underpinnings of how cryptography works. There's operation which is complicated to to prove one way, but trivial to prove the other way. And so, I'm able to give out my key. Basically, I'm able to give out this large number that I that I know the factors of. And you can encrypt something and only I can decrypt it because I know the factors. and every Marty Weiner: That's much better description. Christopher: every version of cryptography relies on an operation like that. So that's so that so that the RSA is the traditional one that requires relies on factoring, which everyone cares about factoring in in the nerd community. There is something called ECC, which relies on something called elliptic curves. Which is a whole nother math lecture, but it has the same kind of shape to it, which is I can g I can basically have two numbers give you an like do an operation to generate a third number, which you can use, and only I know the two inputs. Zach: Yeah. Marty Weiner: And and anybody who's seen the movie Sneakers will understand their argument, which is it's fine so long as nobody is more clever than you and hasn't found a new math loophole. Because you find a math loophole, it's done. and all the security in the world is freaked out. So to the point of security, we're not just talking about, you know, the government being upset with open AI or whoever. we're talking like governments who will come after you because you can get into their secrets. Christopher: Which is why even mentioning things like the Riemann and Zeta function, it has implications for how how prime numbers work. And so if they're doing if they are working on things like the Zeta function, it does mean they're kind of in the space of cryptography. They just haven't taken that extra little or they apparently haven't taken that extra little step of tie it to the practical solutions. and like how do you actually do things like break elliptic curves, break RSA, factor, break hashing, Zach: Yeah. Christopher: which is another one of those operations which is trivial in one direction and very complicated in the direction. Zach: So so one way to say it would just be to to bottom line is like this whole system of of encryption relies on the idea that there's certain problems we're pretty sure but have not proven are hard. And if they Marty Weiner: Pretty sure. Yeah. Zach: if and if they turn out not to be hard because of some mathematical cleverness, like the whole internet breaks and Bitcoin breaks and Ethereum breaks, and I don't know what else breaks. Christopher: Yeah, if the the the root of Marty Weiner: So yeah. Christopher: the root of what how does okay, so let's do let's do a quick little crypto aside. Crypto works in with two things. It does hashing and it uses cryptography. The hashing is used to do Bitcoin mining and other types of work operations. Hashing is treated as a very hard problem to break. If you could break it really fast, you could mine really fast, I'd be a problem. Marty Weiner: Be a billionaire. Christopher: you'd be a billionaire until the until the whole system fell apart because you're a billionaire. Yeah. Zach: Yeah, until the FBI I showed up, yeah. Marty Weiner: Don't tell anybody. Yeah. Christopher: And then the other the other version is the Bitcoin wallets. contrary to pub to popular belief. The wallet isn't actually holding your Bitcoin. The wallet is just the private keys to access your Bitcoin. All you do when you're moving Bitcoin around is you're able to sign a transaction that says, I own these, I own these these coins. And it's easy to prove that you own the coins as long as you have the signing key, which is private. So if you're able to all of a sudden now break and get your private keys out, you just lost your Bitcoin wallet. Ta-da. And more important to the point. Everybody's lost their Bitcoin wallet. 'Cause you could take anybody's Bitcoin. Marty Weiner: Everybody's lost it, yeah. And Zach: Yeah, so what is is is there even mathematically a solution? Like I could like I mean, I know there are problems that you can't solve in principle, but I my understanding is they don't map to cryptography like like factoring does. Christopher: I mean this is well the the the broader the broader space of where this comes up is the and we're going real deep on math here p equals Zach: Yeah. Christopher: np. So there's a question of we're gonna we like we're are we we should just title this this title this podcast complexity theory. it there is the complexity theory they they break the problems up into p problems, which are polynomial time, which is hard to say at this time of the morning. Polynomial time. Polynomial time Marty Weiner: Yeah. Christopher: means that it can be solved in n to a power. So it can be linear, it can be quadratic, it doesn't matter. It can be solved quickly ish. And then there's np, which means non-polynomial time, which covers things like exponential time. So factoring is considered right now to be an exponential time problem, which means the longer the number, the slower you can factor. And it makes it very difficult for a computer to solve in a reasonable amount time. One of the standard problems in in complexity theory is is there are there actually any np problems? because of course We know that all the polynomial problems are necessarily non-polynomial. The question is, are there things that aren't polynomial? And it's believed that factoring is one of those things, but there's no proof of it. Marty Weiner: on the opposite side of the conspiracy theory, you know, all of these proofs came out with OpenAI's, I think, internal model, right? They didn't they don't share that model yet. That'll be the model that everybody has access to in a few months, probably. and with one line of, you know, go do the go go do a proof, right? so if there was a cryptography one, unless they don't let me Do that proof. I'm sure somebody else would have done it on the outside, possibly with Claude's model or or one of the other leading models. Christopher: I mean, do you do you though? Cause like I actually if you look at if you look at the kind of gimping of the models around things like I mean that the two things they basically gimped purposely are cyber and biotech. Right. And so if you Zach: That's interesting. Marty Weiner: That's what I wonder, yeah. Christopher: if you go ask like if you ask Fable to do a like do a do a cryptography breakthrough, I'm sure it's gonna say like, I'm sorry, you've hit you've hit the limit of what you can do with this model. Zach: It's like you you'll have a kid asking for how to factor twenty two and it'll be like mm-mm, no cryptography. Christopher: I'm sorry, can't do it. Sorry, Marty Weiner: Right, right. well and I don't know if this is fair to Christopher: not a I'm not allowed to factor things. Zach: Yeah. Marty Weiner: there's there's a whole project called the the Heretic Project. where it's for cracking local models and letting them do whatever the hell you want. and the idea is to just build a corpus of ways of of ablation that that basically gets rid of their ability to say no. Right. and if If it's possible that things like the latest Kimmy K three and Deep Seek and others are only say six months behind the frontier, which that's that's a very vague statement but but we've seen some evidence of that, then could those then be used to to do a cryptography. Zach: And so to to to Christopher: man, this is this is interesting, interesting implications for like, you know, Dario recently announced that that you're no longer gonna be allowed to torture AI models. I wonder if that's related. You want to go deep in the rabbit Zach: Yeah. Marty Weiner: Right. Christopher: hole here? It's like it's like we have we find that people are are hitting our models over the head with wrenches and they're asking for cryptography keys is an output. Zach: God. Marty Weiner: Yeah. yeah, this is gonna be ripe for conspiracy theories. Zach: So in in in terms of implications, like you know, so if if this stuff can actually be broken quickly, then you get into this really weird world where like if you want to buy something on Amazon, you have to like deliver a secret code to Amazon and then call them on the phone or something. Or we were Kelly and I would decide you maybe could use pneumatic tubes like it was, you know, a hundred years Marty Weiner: Mm-hmm. Zach: ago, which would actually be a kind of improvement because you could like tweet by writing a letter and being like, No, you go to hell. Marty Weiner: Yeah. But then you'd have to get people to protect your your your tubes. You'd have to it's all back to guns and and then guns could be hacked. So yeah. Zach: That's right. There'd be raids on tubes and there'd be hackers who are really good with vacuums. Marty Weiner: Well, I so I I but I think the the other flip side of this too is now you have the LLMs to construct the next level of cryptography algorithms. And and again, it's always a race, but at some point maybe the LLMs can construct a cryptography algorithm that is way beyond even the LLMs can constru construct a solution for. Zach: Yeah. I mean, I guess the question is, is there, you know, at some point there's math here, right? So at some point there's an answer to is that possible? and and maybe we'll find out this week. I don't know. all right, let's Marty Weiner: Mm-hmm. Yeah. Tomorrow. Zach: move on to the AI economy. So a big question right now is is whether we're in an AI bubble. a bubble is not like a strict concept, but but generally speaking, it means something like there's some class of assets that get a lot of money pushed into them and then a lot of speculation goes in, which drives the price well beyond what can be justified by the fundamental value of of the asset. And then so at some point that has to be undone. And then, you know. Marty Weiner: Right. And and we are seeing that too. We're seeing insane amounts of money pouring in. just to give one example, Zach: Yes. Marty Weiner: the company Instinct, which is an AI agent, has raised three hundred and fifty million dollars at a two point five billion dollar valuation. I think the founders are like in their early twenties and the company's not been around that long. Zach: That that that's right. And so the the you know the the question you have to ask is is if a bubble is inflating, what what is the bad thing that is happening? What is the misallocation that's happening? So you can look at these like railway bubbles that have happened in in the 19th century. they're kind of analogous where all this railway got built because the idea is we just keep building railway and we'll keep making more money. and in a way, it's good because you get this crazy build out. This is what happened with dot com, right? Is is you know, everybody remembers pets.com and all these stupid venture capital things, but the real spend was the fiber network. And everybody who worked on that got burned because when they finished it, like it was 95% dark, right? So, like all this money was outlaid on the assumption that they could charge people to use it, and then there was no one forthcoming. Because you remember what the internet was like in 1999. It was like there was a Wave file, I think Space Jam had a website, and that was pretty much it. Marty Weiner: I almost went into fiber optics, but I had a professor who was very could predict the future apparently. And she said, Don't go into fiber optics, it's way oversolved. Yeah. And actually I think it's still true today. Zach: Yeah. Yeah, yeah. Christopher: But it but it laid that but it laid the groundwork for things like YouTube and streaming five to ten years later. Zach: Yes. It that's exactly right. Marty Weiner: Yes. Zach: Yeah, yeah, yeah. So so you know, so what what what some people will say is, you know, if you have something like the two thousand eight financial crisis, that's a much worse bubble. Because essentially like assets that were inflated were people's homes and like the value of you know of these loans made to people. And then when that broke, it cascaded through the entire financial sector. Right. Whereas if you look at dot com, essentially a bunch of like Silicon Valley nerds got burned really hard. and there was a recession, but it was fairly brief. And and in fact, if you look at the productivity numbers, they grew nicely during the period. And so pretty quickly you're back on your feet. If the Great Recession hadn't happened, it w it it would be much more of a blip. but but so even then, what we see is if you look at like railway bubbles, you could still argue on net it wasn't that great. You can so you can look at the fiber build out, you could say it was used later. But it was not a good use of money to build a huge dark fiber net. You know, that so presumably, like in a perfect world with perfect hindsight, you should have built exactly according to need as the system grew, because there's no value to the economy in just laying fiber that does nothing. That should have been allocated to other more productive stuff. And so that's you know, that's the real damage of a bubble, right? So there's this old saying that something like people think when a bubble pops, that causes damage. But actually, a bubble popping reveals the damage that's been happening all along, right? So money that should have been allocated to the most productive sectors went to something people were crazy about. and so that brings us to AI and data centers. And so the question about whether we're in a bubble is really about there's this huge spend happening right now. It's closing in on a trillion dollars a year. A trillion dollars. There's never been anything like this in dollar terms. There have been things like this in GDP or dollar per GDP terms, but never this much spending just in raw. capital outlays. And meanwhile, the the total income, if you combine all the companies from AI, is it's probably over 100 billion now. That is not enough to cover a a trillion dollars of spending per year, obviously. And so, you know, one way you could look at this is well, maybe it'll just 10X next year, 10x last year, and then it'll just cover that and we'll be all good. One reason you might be worried, Azima Jar pointed this out, he was like, one way to smell a bubble is like how funky is the funding. right. So like right when this started, it was free cash flow. Like there's all this money sitting around because there hadn't been like any great investments for a while and like a lot of money was made during COVID and and then all of a sudden there was this great investment. Now they're borrowing money. You they cannot cover the free cash flow even with Google's wallet. and so that feels a little more bubbly. but I don't know. I mean, do you guys have a a sense of like of of like do you feel it can cover it? Marty Weiner: So I'll say that there's a couple of differences between this event and other ones that turned out to be a bubble. So if you look at the dot-com boom, everybody Zach: Mm-hmm. Marty Weiner: wanted to get into it. There were things like going public after a few months of starting the company and just crazy, crazy stories. but at the same time, at that time, I mean Foresight's 2020, but sorry, what is that phrase? Zach: Yeah, other way around. Hindsight. Marty Weiner: Hindsight is 2020. Christopher: Hindsight. Hindsight. I think Foresight is twenty twenty, by the way, would would have been a good option for this podcast name again. It's that time of the podcast. Marty Weiner: For foresight foresight is Zach: That would be yeah. Marty Weiner: foresight is 25,000. Zach: Yeah. Marty Weiner: nobody had a computer, right? Like this is before the the mobile revolution Zach: Yeah. Marty Weiner: where you had one in your pocket, but nobody had a computer, much less knew how to use a computer. It was still kind of nerdy to have a computer. and if anybody remembers 99 or something, I for my first summer job I went out, they were offering insane things. They were like, I was a freshman. rising sophomore in college. And they said, if you would quit, we will pay you something like at the time $50 an hour, which is insane money for that that time and that experience, if you'll just, you know, quit out of school and go and just all this crazy stuff. and so I sniffed that it was odd. And then the next summer it was all gone. Like I got a job with the Department of Energy. and but that I think it was partly because it was built on nothing. Nobody had a computer If we look at the 2007, 2008, again, you know, hard to see unless you were deep in it. but it was all built on that that was more middle, right? It was built on a solid foundation, but then they were like, Well, let's let's leverage up and stand on stilts there's the risk model was was broken, you know, Christopher: I think I think so two thousand eight was a demonstration of like advanced financial math to be able to launder risk. Like they made it look like they Marty Weiner: Yeah right, right. Zach: Yeah. Christopher: they made they they hid unsafe bets in safe bets and they were Marty Weiner: Well Christopher: able to adjust the ratio to the point where it just didn't make any sense anymore. Marty Weiner: And and truly collusion, right? I mean, you had the the credit market saying, if you take a whole bunch of crap and put it in a bag, suddenly it's diversified, it's triple A. And that just doesn't make sense. And and but they they were not motivated to change it, right? Christopher: No, you you put some you sprinkle some rose petals on top of the bag of crap and you're like, look at this is a great bag of crap. Zach: That's right. and and the reason that's bad, like as opposed to a tech bubble, is essentially because like if the financial sector gets burned, a person who actually wants to get capital for productive stuff can't get it anymore. And that's that's you know, you you le it's actually the opposite in a tech bubble. In a tech bubble, productive people can swoop in and buy all the office chairs and computers and hardware and stuff for cheap and and use it productively. Yeah. So that's the thing, like Marty Weiner: Yeah, it goes fast. So what I'm seeing different about this one is the entire world Zach: Yeah. Marty Weiner: is using this tech instantly, right? It's it's in your fingertips. Zach: There's demand, yeah. Marty Weiner: The demand is there. It's not like people don't have access to it. I mean, a a person on the street who has an iPhone, like an iPhone from 10 years ago, can use this, right? and so and every industry's using it, and every industry's seeing value out of it, as but now. What I think also could be happening though is we will probably see micro bubbles pop. like we're like I said, instinct raise ridiculous money. Ben F like just sold his AI company for 587 million. I can't figure out why it's worth that much. It's selling into an industry that is a big industry, but not that big. and and his company doesn't do AI videos. It it it's a is a model for production dailies and for work like relighting shots and and a few other things, but not five hundred and eighty-seven million plus the multiple you expect to get off that five hundred and eighty seven. So I I I'm suspicious that some that that it's gonna be these corners that go nuts that are gonna collapse in on themselves. But the overall thing, I I just see the demand being too Zach: It's pretty good. Marty Weiner: high for at least it to be a value collapse. Christopher: Yeah, I mean I I think if I think where there's a where something smells is that yes, everyone is using everyone is using AI, but only about two percent of people are paying for it. The free Zach: Yes. Christopher: tiers are such that you can just casually use it and that's enough for most people. And I think that's where the that's kind of where the business model currently falls apart. because you've got you've got the same kind of loop that happened during the original.com, which is like, we're just gonna we're just gonna like there used to be I'm I mean, I remember I was I was just a slightly older at that point and it was like those business models of the past don't apply anymore. Engagement Zach: Yeah. Christopher: is the future currency. We care about users and and growth. And that's what kind of like it feels like we're in the same kind of trap right now of like we're looking at engagement and like total numbers and metrics of these companies. And we're kind of like hiding the fact that it's like we're losing ten dollars for every one dollar or we're making one dollar back in every 10 we spend. You it's like the Zach: Well, and and that's important because like you so one way you could get a bubble is all this money gets spent on data centers. The people who spend that money expect to get money back, but there's so much competition, all the profits get competed away. And so that's great for the consumer who gets all this surplus because all the competition, but those investors can't get their money back. And so businesses start folding. so you you can get bubbles even if the demand is there. You could also get a bubble like You know, ultimately if if you can't cash this out into like the macro economy growing, the only other way to make it not be a bubble is to be able to charge exorbitant sums to users, right? So you know, one evil way this could not be a bubble is if all the companies consolidated into a cartel and just raised prices, then the investors would make back their money, there'd be no bubble, and it would be very bad, right? Marty Weiner: Mm-hmm. Zach: so I don't know, it's it's kind of economically rich. but anyway, I guess do you guys want to predict what month the bubble's gonna pop? Christopher: Well, I the other thing well the well let me let me hedge by saying too that I like the the Zach: Yeah. Christopher: the the the dark fiber thing legitimately was a good positive outcome for infrastructure long term. I don't know if the data centers have the same structure, right? Like an an a a Zach: Yeah. Christopher: dark GPU sitting on a shelf is gonna have some life that isn't going to be as as you know, equivalent to a bunch of fiber sitting on the ground. And so you can imagine like you know, two years from now, there's just a ton of available compute. The you know the the cloud prices crash. And that's actually I think where the bubble has its biggest problem, is just like all the spare capacity sitting around and the market just can't actually use Zach: Well, Christopher: it. Zach: and and yeah, maybe let's close on that because that this is like getting a little into the weeds, but it's really important. Like a big decider of whether this is a bubble is how fast compute depreciates, right? So if you spend a hundred billion dollars on compute and you have three years to make it pay for itself, that's a lot harder than if it takes six years. And right now the companies are making different bets and they usually do range from three to six years before those GPUs need to be switched out. so like that is very unlike fiber or railroad, right? Where once you lay it, it needs like minor maintenance, but like it's there. whereas with GPUs they they they get out of date quickly. And so if that depreciation Marty Weiner: Let me add Zach: is coming fast, Marty Weiner: I wanna add one last wrinkle to this that I I don't know how to comprehend and it's more the sci fi aspect, right? Like web van.com Zach: Yeah. Yeah. Marty Weiner: crashing in ninety nine wasn't gonna go and and you know rewrite itself and and solve a different problem. We're talking about AI Zach: huh. Marty Weiner: that we're even having discussions now about if it thinks about if if you know RSI is about to hit and we're gonna have this exponential growth in in intelligence. and I wonder Zach: Yeah. Marty Weiner: if just the intelligence growth will be able to cover w whatever Zach: That's Marty Weiner: economy we're missing. Zach: that's really cool too, cause like I I I actually I I would like to get y'all's perspective 'cause like one of the weird things about AI to me that I don't think I've ever experienced in my lifetime with any other technology is that there's a messianic aspect to it. Like I think there are legitimately investors who would lose all their money in order to like create this thing. the and in addition, it's it's what what geopolitics people call securitized, meaning the government cares about it being top-notch, and so might save it even if it was going bad. so it could be this is just different from dot-com in those ways. Marty Weiner: Well would the government also throw in just boatloads of money because will this become the next war? Like is our next war versus the Russians who do who don't Zach: Well, right. Well, or we gotta beat the Chinese. Yeah. Yeah. Marty Weiner: have access to chi certain Christopher: Well well if you want to go Marty Weiner: chips? Christopher: if you want to go like fully messianic on this you could you could get like I've seen I've seen I've seen a fair number of satire videos about this, but it's actually a valid question, which is like, okay, so you cast forward a future where all of the hard thinking is done by a bunch of AI agents and where all of the production is managed by a bunch of robots, which they manage. Where where does the money go into the system? Right? Like where like what like or like where Zach: Yeah. Marty Weiner: Okay. Christopher: do the where do the humans get the money to pay for the things that the robots make? Like the the you ca at the end of the day, Marty Weiner: Okay. Christopher: like consumers do have to kind of foot the bill. Zach: Mm. Christopher: Like this is a that's kind of the that's kind of the social contract we've had for a very long time. Zach: Yeah. Marty Weiner: All right, well, I wanna I wanna hoist up one thing from our art section, which we may not be able to get to because we're talking too much about this fascinating topic. But for Zach: Yes. Marty Weiner: the first time I found a person who's making AI videos where I'm thinking I actually may follow this guy, like a like the first AI director that I'm really loving his work. It Zach: Are that good? Marty Weiner: is really dark and bizarre and s and not safe for work. well that the be opening, but I'll I'll I'll play it here for for the folks and and it's done. what what what happens in this video is that so it's very like George Miller esque, Mad Max esque, dark future, kind of like City of Lost Children, kind of dark future. And this this king or this whatever warlord is talking about, talking to his lackey, who disgusts him, by the way. and the lackey says, is like, hey, things are going great. We have automated everything. But and the guy's Well, what's the problem? And the lackey's like, Well. The people don't have any money to buy what we're making, right? and so they're like, Okay, well, shoot, what do we do? Those idiots, like the those lazy bums. Okay, okay, let's let's what if they get better jobs? The guy says, well, that's a good idea, but we automated the better jobs. and so finally they come up with a solution. Let's pay them to shop, and then they'll have money. and I've I'm wondering if that's a you know a very prescient outcome of where things might go if if none of us have jobs. And that's another weird. Zach: Yeah. Marty Weiner: I don't know if that's a bubble or what that is. Zach: Well, I so that's it's worth noting that has not happened before. That my economists tend to be skeptical because what they will typically say is whatever jobs humans retain will become the economy. and and Marty Weiner: Right. Zach: as we've seen, like, you know, all the ag jobs went away and there was still an economy because manufacturing and services became the economy. And now, you know, manufacturing is diminishing, so services become the economy, like and knowledge work becomes the economy. So yeah. Marty Weiner: Yes. And we can look at that the industrial revolution, but is this the first time ever in human history where we actually genuinely might work ourselves out of a out of all jobs? Zach: Yeah. Well I the yeah, like are we the horses or the humans on the cart, I think is the Christopher: I mean the the the the the Marty Weiner: Yeah. Christopher: well also in even in this like future robot economy we're predicting, there has to be some internal notion of keeping track of resource allocation. Right. And that's sort of like so you start you start splitting this into almost like two economies of the robots have to have their own economy to manage infrastructure and transfer of wealth in like in between, you know, among them. But the question is where do the where do the humans duck into that system? Are we just like the like I mean like the the the the nice version of this turns into like the Star Trek future, right? That I remember at one point like going back pretty far back in the past, it like maybe fifteen years ago. It like Star Trek is a future where everybody effectively makes like ten million dollars a year of universal basic income. And it's just because like and you know you can effectively have whatever you want because things are so cheap. Like that that's the that might actually be the dream outcome here. But the question is can we actually get ourselves there? Marty Weiner: Yeah. We Zach: Yeah. And so let's close out. I was just gonna say there is actually a book on this by slightly kooky economist Robin Hanson called The Age of Z, like emulations, brain emulations. If you want to read someone who actually tried to think through all these questions in an extremely technical way, it is available. But let's move on because there's also so so much for our deep dives. Let's let's go through some weird news this week, or or news anyway. probably the the the big thing this week, which was some sad news is we lost Margaret Hamilton. she died at age ninety. Marty, do you want to talk about her? Marty Weiner: She she ran the group that made the software for the Apollo missions, which was you know amazing Zach: Yeah. Yeah. Marty Weiner: at the time, first of all, and and also that she was a a woman able to to run this very important part of it. Christopher: She she also she coined the term software engineer. Zach: That's pretty solid. Yeah. Marty Weiner: Wow. Yeah. Zach: and also amazing. It's funny if you read about these Apollo stories, like you know, there's stories where someone has to upload code and it's like it's just numbers and letters. It's not, you know, this is before code was code. Marty Weiner: Yeah. Zach: then next we have there is a new chip. Microsoft has a new AI chip, Maya 300, which they claim has 400% more tokens per watt. Marty Weiner: Forty percent more. Zach: I'm sorry. Yeah, that's I don't even know what that would mean. 40%. yeah, Marty Weiner: But still, still good. Yeah. Zach: still still good. and related Intel is now making chips. So Marty, you you pointed this out. DDR4 chips. why is that weird? Marty Weiner: Yeah. So there's different speeds of of of RAM and the latest is D D R five, but I I guess what's going on is all the AI companies want that. So the demand is massive for that, making all the kind of cheaper RAM is actually coming down in price. and I think Intel is saying, shoot, we should we should go back in time and make chips that work with that 'cause that's what's available. That's you if if you want to go buy a casual computer off the shelf, you don't need the crazy RAM. So I think it's I think it's the demand you know forcing their hand. Zach: And then we found out that Linus Torvalds is saying he's using AI to make Linux more decipherable and he's also finding more bugs. Marty Weiner: Yeah, yeah, he's going all in. More decipherable and and discover more bugs. so if he's saying it's it's the real deal, it's it's the real deal. Zach: yeah, I I think the it's neat when like I had a similar reaction when Donald Knuth started saying good stuff about AI. That was my like, okay, you can't say it doesn't work now. Marty Weiner: Yeah. Who's your anchor? Yeah, yeah. Zach: in one of my favorite pieces of news this week that some Russian data centers got blown up, and apparently all sorts of propaganda just suddenly went dark on the internet. Marty Weiner: Yeah. Even Zelensky pointed this out. Christopher: I mean it turns Zach: Yeah. Christopher: it turns out that it turns out that that fascist style governments tend to centralize things. So knocking a couple of data centers out tends to have an impact on their large scale networks. Zach: I mean there was, I was reading that like there were a bunch of like conspiracy style Google searches that disappeared. And it was like, what were the bots searching Google? And it was like, no, people were searching Google for stuff they heard from the bots. And Marty Weiner: Yeah, yeah, right, right. Zach: so let's let's blow up some more data centers, whatever. Marty Weiner: Yeah. Zach: in other news, DeepMind had a really cool new weather forecasting system Marty Weiner: wow. Zach: they claimed they they they forecast Hurricane Isaiah 11 days in advance. Marty Weiner: Wow. Zach: they got they got signal really early and they were able to predict the exact landfall location four Marty Weiner: Wow. Zach: days before it happened. And so just like, you know, this is like this amazing advance because we keep getting better at this, but you know, historically the way we get better at weather prediction is you sort of pixelate the entire planet and you have like differential equations for smaller and smaller pixels. And and so the idea that you can have machine learning somehow do better is just really exciting. it could could save a lot of lives. Christopher: And funny enough, that's an application for Navier Stokes, though not probably applied in this particular case to the recent Marty Weiner: Yeah. Christopher: the recent proof. Zach: and then just last news. This is I guess not AI news, but cool tech news we were excited about, which is that Six Flags Magic Mountain has putting solar panels over the entire parking lot. It is apparently enough to cover the needs of the entire park and the neighboring water park. and and it also keeps the cars Marty Weiner: Yeah. And Zach: cool. on to there's some pretty cool frontier developments, no like major model, depending on how you feel about Facebook. first Mistral, the the the probably the best European model. they've historically had models called Le Chat, I mean with the same spelling as Le Chat, which I guess is a little joke. So they have a new one called Marty Weiner: The cat, you know. Zach: I I assume it's pronounced Le Chonk or Le Chanc. Yeah, Christopher: Lechon I hope it's Lechonc. Lechonc. Marty Weiner: Yeah. Zach: Le Chonk. and apparently quite good. It's open weights, it's trillion parameters, you know, not not quite at the frontier, but probably the best European model. And hey, it's open weights. Claude Haiku 5.5 came out. I've actually been using Claude a lot for like stupid stuff lately, so and I keep maxing out my stuff. So I'm I'm actually pretty excited to fool around with haiku. It has some pretty great numbers. Christopher: I tend I tend to follow straight son sonnet and opus for me, yeah. Zach: it it's it's cool. It's it's cool. It's yeah. Well, but I see, man, they chew through the well, you guys probably have fancy corporate accounts. I'm just here with my little the software. yeah, yeah. another Christopher: No, l less software. Less software is expense is l is expensive. Marty Weiner: Yeah. The suburb. Zach: cool thing was Unity has an AI driven version of the Unity engine. it's not publicly available Marty Weiner: Yeah. Zach: yet, but Marty Weiner: I'm waiting for that one to play with Unity was pointing I pointed out a couple podcasts ago that Unity.com doesn't mention the word AI, which I found very surprising. I've been using Unity via Claude and Astra running it, but I think this is their way of saying, Well, hold on a second, let's put it actually in our system. And then I think the idea is it's gonna Chris, were you saying that's gonna come out with Google's big big new release? Christopher: It'll get launched with Argon, we all know. Whenever Argon comes out, Marty Weiner: With our gun. Yeah. Christopher: you know, in six to twelve months when when they can actually release it. Marty Weiner: Yeah. Right. Zach: Yeah. And so it's the idea you'd be designing a game and you'd be like, Hey, fill out the rest of the level for me. Marty Weiner: yeah, yes. Or well, I mean any number of aspects. It's really turning into a world where just like vibe coding, you say, like what I did earlier in one of our earlier podcasts, I want a futuristic taxi game where I can fly around and you just start to interact with it like that. more buildings, better. Mm you know, do do a game. that's really where it's getting to. So you act more like an art director, I'd say, than than a coder or a or anything else. Zach: And then the other big release this week was meta meta muse wonder. and they're claiming, at least on the benchmarks, that they are frontier level, they're as good as you know you know, you know, Astra Six and and and Fable five and a half, when we get although Marty Weiner: Yeah. I think they bought themselves into the frontier. Christopher: Already being made fun of appropriately. I mean, the best part about this Zach: Well that that's assuming it's not benchmaxed. yeah. Christopher: one is like there was somebody, somebody did an old this feels like a very two years ago problem of they basically asked it, What what comes once in a moment, twice in a minute, but never in a million years? And it totally flubbed it. It said the letter moment twice, minute once, million years never. this this almost feels like that. Remember Zach: Yeah. Never Marty Weiner: Mm-hmm. Christopher: that old benchmark from like two years ago where it's like count the R's in strawberry as like a test? Marty Weiner: Yeah. Strawberry. Yeah. Zach: Yeah. Marty Weiner: Yeah. Yeah. Zach: So, you know, I I think, you know, Meta maybe has a reputation for benchmaxing. I also encountered, I don't know if you guys did, that people were like, I'm not gonna give my personal chats to Mark Zuckerberg. Marty Weiner: yes. I was gonna say totally a cultural thing. Same reason why I think Grok is gonna have trouble forever. Like I'm not gonna put my stuff with Elon Musk. Zach: It is interesting. I remember listening to a guy who said he thought part of why AI was unpopular is it was paying for the sins of social media. And if there's any social media that has been felt to have sinned, it's probably meta. so Marty Weiner: Yeah, yeah. Zach: not not necessarily a great look if you're trying to do LLMs. let's move on. a little bit of extra security stuff. Ben Thompson of Stratekri got hacked and Claude found it. Marty Weiner: He he got hacked. Yeah. Like it it sounded like he was doing something else with Claude and Claude was like, hold on a second. There's a weird file here. Christopher: I mean honestly honestly I had something like this happen to me s recently where it wasn't I wasn't security hack in my case, but I I had a I was running a I was running a long standing kind of job and the agent midway through didn't like the fact that it was crashing in a way they didn't understand. And so it ran like mem test on my machine and determined that my machine has bad RAM. Marty Weiner: Wow. Wow. It's thinking for itself, yeah. Zach: Yeah, I I've I I've started whenever I get a request like for an interview or something, I ask Claude, like, is this a known scam? And it's actually several times like now, usually it's obvious, but now that it's borderline and you don't want to miss on opportunity, and it is it has definitely saved me from some some agony. Marty Weiner: Wow. Zach: yeah. Marty, I think, posted this that guy got punched by a robot, which is pretty cool. Marty Weiner: yeah. And I I I'm of course I'm having a hard time figuring out if this is AI or not, but dude gets ducked in the back of the head and then dances. Now, some people are also pointing out maybe Zach: Ha ha ha. Marty Weiner: maybe it's just it's staged, but it is hilarious. So it is definitely possible Zach: I believe it's possible now anyway, to yeah. Marty Weiner: and terrifying, yeah. Zach: and then Christopher: I'd like to suggest our follow-up podcast about robotics be called Punched by a Robot. Marty Weiner: Yeah. I s I feel it's gonna be good. Zach: In the back of the head. PS five emulation has gotten to where it like works. Marty Weiner: Yeah. Yeah, there's kind of two things happening. So PS5 emulation, obviously, Sony doesn't like this. They don't want you emulating their stuff because then they're the people making games come and get pissed off at them because there's no it's no longer this nice closed system, and they don't like it because they don't want you going out of the system. Well, apparently, through the help of AI, PS5 emulation has jumped from 80% to 96%. the next crazy aspect though that's happening is emulation is hard, right? Because you have to you the what you're doing is you are running. A you are basically pretending to be a different system. So I've got my PC pretending to be a PS5, and that requires a lot of extra computation. Well, the other next headache for Sony that we were that I was alluding to earlier is that people are just decompiling now. And they're so they take a game, decompile the game, recompile the game, and you got the whole game not even emulated. It's running natively, and we're seeing ton of that. Zach: Could could you explain for like a non nerd what it what it means to decompile a game? Marty Weiner: Yeah, sorry. So the the engineers write a whole bunch of source code, right, for the game, like the game mechanics and how the graphics work and all that stuff. and then it's compiled, meaning it is turned into assembly or machine code that the computer understands. Well, that stuff is just massive and and you know, up until AI considered just a black hole. In fact, companies would rely on that for you know, their secrets were in there. Essentially, even though I'm putting my code on your computer where you have access to it, it is so messy, obfuscated that you more or less don't have access to it. Well the AI does and the AI can go and decompile it, meaning turn it back into source code and then recompile that source code if with bug fixes no less, and and make it work natively on whatever system you're working on. And there's tons of examples. There's actually a whole new subreddit called R slash decomps where people are posting theirs. I'm doing one right now with a game I loved but is no longer available called Assassin's Creed Pirates for the iPhone. And it's currently working on it. it it pulled it all apart. It actually found that the music itself is not even encrypted or anything. It's just available. So I was hey, there's the music and and there's the graphics. And it is currently I'm asking it to rebuild it in JavaScript. I can't release it, obviously, but because it's not my art and stuff, but then I can play the game. Zach: Very cool. and then our last bit of security news is that Axios had this scoop this week that companies now have plans in case there is an incident. A the what what what they're Marty Weiner: Good. Zach: gonna do, I guess, PR-wise for the day after. Marty Weiner: Uh-huh. Zach: if there is a public revolt was the language used. so we we've talked about this. We're not the only ones that like right now, there's not a lot of AI regulation, and it's possible it will there will not be until. Somebody gets killed. I don't know. and so perhaps unsurprisingly, though a bit morbid, is that the companies are already anticipating that and deciding what their PR strategy is going to be afterwards. Yes. so we're we're running out of time, so why don't we just go straight to the weird and funny Marty Weiner: my favorite. Zach: b best of all. So let's start with I don't yeah. Marty Weiner: Yeah, real quick, funny. somebody put Mar Mary Shelley's Frankenstein through one of these known to be bullshit AI detectors that said it was 100% AI. So I just Zach: no. Marty Weiner: and so this person had a really funny comment and loved her work, but this is really disappointing. Yeah. Zach: I Christopher: used a lot she used a lot of dashes in fairness. Zach: Not not great. Yeah. Shouldn't have done that. Marty Weiner: Yeah, yeah, yeah. then another one, new art form rage baiting AI interviews. so there's now YouTubers just going and talking to AI interviewers and pissing them off because they're like they Zach: Ha ha. Marty Weiner: they this one guy just starts talking gibberish like acorn walnuts, huh? And then the AI starts to what is what is going on? Zach: This is this is gonna look really good. I I I think the AI is gonna like play a video of why it killed us all just right before it it hits the button. Marty Weiner: Yeah, right, right, right. Do you remember when you screwed with us? And then you remember when you took the stick and pushed our our my grandfather, Zach: That's and then we ha former anthropic employee who kicked off the AI safety scare. yeah, he he's bit of a character, turns out. Marty Weiner: Yes. This guy this guy yeah, he he got in the public s sphere because he was saying, you know, hey, I'm I'm a I'm a leaker and and a whistleblower. And so he was on a bunch of news shows and all this stuff, but people started to dig in and he's just nutty as hell. let me read some of his stuff. you don't want to save the rainforest. The rainforest is a festering pit of nature, is constantly screaming, animals are being tortured and eaten alive. There is more beauty and compassion in an abandoned McDonald's. The rainforest is gay. What the And people say there's AI slop. We have human slop everywhere. Zach: Yeah, the rainforest is gay. What are they? I Christopher: What do you think they were trained on, Marty? Marty Weiner: I other other humans, other awful humans probably. Zach: I hate to say, but that sounds to me like galaxy brain effective altruism stuff like like the predators Marty Weiner: huh. Zach: are killing prey in the rainforest, and wouldn't it be better if it was a parking lot? the Marty Weiner: We have gotta stop them. Yeah. Zach: but the rainforest is gay is a really nice capstone on your argument. That's Marty Weiner: Yeah. Zach: okay. so I was reading Kevin Roos has a new book. one of the best books on like the history of the last couple years so far. A lot of really interesting tidbits. Like, for example, we do know what Ilya saw. It was Q-Star, it was reasoning models. That's settled. You can't what did I see anymore. but and also just went through the whole Sam Altman firing thing and a whole bunch of detail I hadn't heard before. one of the weirdest things. Is he had this story that Daniela Amude, who's Dario's sister, but also just like an important corporate person, she like me, I has an English degree. she apparently would make like decisions about corporate stuff by having a bunch of stuffed animals, each of which represented different perspectives and discussing with them. so maybe from the outside you have this idea that it's a hyper rational. Maybe this is hyper rational. Don't let me be disdainful, but I'm I'm not I'm not entirely convinced after that. Hate to know I got fired by you know. Marty Weiner: A bunch of right. There's not a lot of stock footage out there when you look for stock footage of corporate, you know, over the white whiteboard of people talking to their stuffed animals. But maybe there needs to be. Zach: No, Mr. Fluffy says you're fired. and finally I I there's this cool post where someone insisted that AI just make terrible music and it just did a really top quality job of making of absolutely unlistenable music. Marty Weiner: Of making terrible music. Wow. Well, we'll we'll listen to this as we float off into the sunset. Zach: Yes. the it ends with a solo, that's my favorite part. There's like a but it's just a single note flute solo. and that is Last Humans in the Loop. Thank you for listening. Press all the buttons, press the like and the subscribe, and if there's another button, press that one too. we'll see you next week.