Latest / Last Humans in the Loop / Laptop AI = 6 Month Old Frontier Models, AI Cancer Vaccines, Test Scores Down, Bomber Drones + AI
Transcript
- Zach: Hello everyone, welcome back to Last Humans in the Loop, the the podcast that is less popular than our one clip of Kevin Sorbo that we put on YouTube reels or whatever. Marty Weiner: Boom. Zach: I'm worldwide celebrity Zach Wienersmith. I'm here with two former CTOs of Reddit, Marty Wiener and Chris Slow, fifty percent of which is my brother. we are I think we we're leading off with some some nerd almanac entries from Marty for for the day. Marty Weiner: Got some birthdays and all anniversaries here. Tim Cook took over Apple fifteen years ago. he's stepping down on September one, so that's finally happening. Happy ninetieth birthday to Margaret Hamilton, this is a fantastic picture of her next to her stack of software that she wrote as a software engineer at MIT instrumentation laboratory, where her team wrote and tested all the onboard flight software for the Apollo spaceships that went to the moon and the lunar module and later for Skylab. And then thirty fifth anniversary of Linux, as of August twenty fifth, That's right, and I do. Zach: For those who celebrate. Chris Slowe: So so do you guys think do you think that Tim Cook's gonna join Jeff Dean Jeff Dean's startup or is he gonna start his own? 'Cause it seems like that's Zach: Yeah. Chris Slowe: in the water right now of I can't wait to hear Tim Tim Cook's new AI startup. Zach: Ha ha. Marty Weiner: I'm excited too. I'm ready to invest. should we jump into models here? okay, so Zach: Yeah, let's jump in. Yeah. New models this week. Marty Weiner: I wanted to read this from this Twitter. his name is Udi Wertheimer. I don't know who needs to hear this, but Quinn 3.827B, so this is Quinn 3.8's model that can run locally, is ranked above GPT 5.3, so that's what, a year and a little bit ago of ChatGPT, Gemini 3.1 Chris Slowe: Yeah, it's about that. Marty Weiner: Pro, but everybody's better than Gemini. And Opus four point six, all of which were state of the art six months ago and it runs on a laptop. I thought that was just a really cool summary of of what's coming. So that's crazy the thing that running on a laptop is starting to be better than state of the art six months ago. Chris Slowe: I mean I've been I've been running I've been running a version of Quin sitting on an old GPU that I have sitting in garage and it's pretty good. It's honestly pretty good. It's I mean it's small model, but it it seems like the it seems like the shift now has been to kind of take like ex don't put as not like don't cram as much knowledge into the model, but kinda keep it keep it as kind of intelligent and smart, 'cause at this point, what do you need to have it be a walking encyclopedia when you have access to a walking encyclopedia on the rest of the web? You can actually look it via MCP. it's it's it's r it's amazing what these models are doing with the reasoning side. Marty Weiner: Right, right. they have this new template that improves coding and efficiency. They're claiming they're getting better than Opus V on coding tasks. So I wonder if it's the same kind of thing where if you reduce what it has to worry about, it can focus on certain tasks. I do worry a little bit I you know, the one thing I think that makes these models really excellent is that you can have a coder who understands everything about relativity, everything about Keynesian economics and all this stuff while they're coding and that is phenomenal, but sometimes you just need to write a little thing that runs a motor and you don't necessarily need to know who won the World Series in in the early nineteen hundreds. So anyway. Zach: It does seem like the future Chris Slowe: Hey hey. Zach: is like like an intelligent swarm where you start with a very powerful model and it dictates to a bunch of smaller models, like it picks exactly the right model for each task. Marty Weiner: Well and that's what I'm seeing a lot of harnesses do, and that's what I've been working on on on the math side, is is sometimes you wanna encourage a big, strong model to come up with the architecture of a big, strong, hard problem and then spun it out to cheaper, faster, more efficient models. Zach: Yeah. and then this this is big news this week. it according to Stripe, we're in the Singularity now. which I've been waiting on Stripe. I've been waiting to hear what what Stripe thinks. Marty Weiner: th they Once finally Stripe Chris Slowe: Mm-hmm. Finally, it's official, everybody. January first january first, apparently. Marty Weiner: had a word. Stripe by the way Stripe by the way has been you know, their IPO is on indefinite hold, so I think they needed some news, but Zach: Yeah. Yeah, but it's nice to be like beyond the horizon of the singularity and now it feels good. I'm thank you stri I'm waiting for Starbucks to announce the singularity, then I'll really believe it. that that's about Marty Weiner: I know, finally Yeah, finally finally thank you, Stripe. but which by the way, I I I'll just clarify what cr how crazy it is. It they didn't just make a comment on a blog, they told their investors this on January Zach: Yeah. Marty Weiner: first. so by the way, and that was already Zach: Yeah. Marty Weiner: nine months ago. So Zach: Hey. okay, moving on. NVIDIA's coding agent scores a hundred percent on ARC AGI three. So so maybe for context, ARC AGI is this test that a few years ago was considered very hard for AI, and then AI defeated it starting with the like post-chain of thought models. and now there is this supposedly very difficult to break ARC AGI three. I don't know if you've played it, but it's like kind of like little video games. Marty Weiner: Yeah, I I think the key idea is, you know those games we play as kids where you know, here is a circle and here's a bigger circle and here's an even bigger circle, what's the next picture? And you have to infer what is the idea we're going for and Zach: Yeah. Well so Ar RK GI three is a little more like those like move the box to the X and turn it pink sort of games, but yeah. Marty Weiner: Yeah. Right, right. Some kind of a game where you have to a l a reasoning game, a logic game. but yeah, I thought it was interesting that obviously we've been chatting a lot about benchmaxing the hell out of these benchmarks. And I think it's actually not even intentional. I wonder how hard it is to not benchmax. I was talking to John Wise, my math buddy, about this problem we're running into about how do you how do you come up with new math problems to train models with or to test out your your harness because you it's so hard to know with any other math problem if it's already in the brains of the model. Zach: Yep. Marty Weiner: So we're running into that Zach: Yeah. Marty Weiner: very serious problem. I think probably we're gonna run into that with Arc AGI three as well, but still it's a really interesting result. I mean back to your point that a while ago it was at zero percent and now we're at Zach: Yeah. Marty Weiner: yeah, okay, solve problem, done, go to the next thing. Zach: Done dun done. Chris Slowe: I think only criticism I've seen of this one was that it wasn't actually them it wasn't overall model, it was actually a model with a harness that NVIDIA also hasn't released or they're planning on releasing. Marty Weiner: Mm-hmm. Chris Slowe: But that to me feels like I mean that's we're we're getting really to kinda some pretty fine hairs if we're worried about the harness of things living inside of. I it's like Zach: Yeah. Marty Weiner: Right. Right. Chris Slowe: if the harness if the harness plus the model's intelligent, well that's pretty cool. That's like that's like what are we arguing about now? We did a we did an AGI. Marty Weiner: That's pretty cool. Yeah. Zach: Yeah, you're gonna start the singularity with a harness. and then lastly Linus Torvald is not only saying AI is okay, he's now using it. Marty Weiner: Yeah, he solved a what he described was solving a gnarly Intel GPU driver bug and he described AI as a tireless helper. So I think that's Zach: Yeah, so it sounded like like he still had to do a lot of work, but it was like a lot of the drudgery was handled for him. Marty Weiner: Well, yes, a and I think the interesting part is, you know, a lot of us, including myself, use AI to do greenfield work, like write you know, yet another website, right? And it's really good at that. But this is the internals of in of GPU drivers, and that code is really awful stuff and and very hard to get right. And so the fact that the AI was able to handle that and and get a good fix and and and actually fix the bug with Linus was interesting. I also think it's interesting as him as a figurehead for tech, basically saying, Yeah, AI is o okay. Chris Slowe: Also, L Linus is has a has a notoriously high bar for what he considers to be appropriate to ship into Linux. And Marty Weiner: Yeah. Chris Slowe: so if he's finding he's getting quality out of models that's up to his bar, that's a pretty Marty Weiner: Yeah. Chris Slowe: good that's a pretty good indicator that it's working really well. Marty Weiner: Right. Zach: Moving on to the tech and robotics stuff. So there was a Robot Olympics in China we've all been watching on Reddit. Marty Weiner: Yeah, I'll I'll post some videos 'cause they're quite incredible, Unitree's biped robot surpassed Usain Bolt's running record. Zach: Mm-hmm. Marty Weiner: and then it also they had another robot that jumps higher than any human at two meters. Record human record is one point eight meters. Now these are bipeds, but I just I kept thinking like, gee, I think we've had robots beating humans at speed and everything else since like nineteen forty-five or earlier. Zach: If i i if you count horses, we c we can push it back farther. Marty Weiner: Yeah, yeah. so Zach: you know what I I was saying, like so someone needs to have I would watch this old day if there was an Olympics for cleaning the kitchen. Like it could take like a week. I would watch every moment. Marty Weiner: yeah. Well, I mean, no joke, like, you know, folding laundry has been a problem Berkeley has been working on for I don't know, for for decades. Yeah, Zach: Twenty years. Yeah. Marty Weiner: and it's been continually getting better. But yes, an actual that would be more interesting. You know, having robots Zach: my god. Marty Weiner: run is interesting, but if I had five robots next to each other folding laundry and we got to the problem where that's b boring, you know, 'cause they're all so fast, we've got a robot in every household. You know, folding your laundry and genuinely making a difference. I know that sounds kind of cartoony, but you know, it's like going from washing by hand to having a w a washer and dryer. Chris Slowe: W What do you think the what do you think the switchover point is there? Because I think also like if you if you if you had a robot in your house and it was ten times slower than you at the task, Zach: Yes. Marty Weiner: It's okay, right? Chris Slowe: but it doesn't but it doesn't need to sleep and it's not like interrupting you doing something while it's like i if it took like an hour to clean up your kitchen after a meal as opposed to doing it in like ten minutes. Marty Weiner: Right. Was Zach: yeah. Marty Weiner: like the i robot or the I'm sorry, what what is the the little Zach: Rumba? Marty Weiner: Roomba? Chris Slowe: Rumba? Marty Weiner: And and that even in the beginning, before it had Zach: Yeah. Marty Weiner: any intelligence, it would just do loops, like little circles. And I think that the idea was, do you care? You're gonna be asleep. It could go for four hours, even though Zach: Yeah. Marty Weiner: it would take a human ten minutes. So I think you're right, like the the crossover point could be really slow. It just has to be safe, you know, in in the household and Zach: It it well and there's also like Marty Weiner: work. Zach: you can imagine like an initial use for like people with ten houses who are super rich who want to not have to pay humans. Although, you know, the cost there is gonna be you know that part part of why why I'm a little skeptical of near terms house butler robots or whatever, which by the way a butler's a chief of staff, not a cleaner. I've been complaining about this for a while. Yeah, it's is this is important. no, but but like, you Marty Weiner: you've been informed of that? Okay. Okay. Zach: know, as an economist will point out, like it i it's really cool, but like maids are cheap. you so you have to not Marty Weiner: Yeah. Zach: only, you know, you have it be good, you have to beat a human and this is a you know, a a considered low skill test that's not very well paid. So it it might be a while before it's actually useful. Marty Weiner: Right. The one of the things I've been thinking about, you know, robotics has been around for a long time and it has always been a slow, slow grind. But I think Zach: Yes. Marty Weiner: now with AI, that's what's been needed to to push it over the edge. And so I'm I'm thinking, you know, looking at these this robot Olympics is i i I I think it's a bit of show, but it is also interesting. and they are making improvements. but I where I really want to see robotics make improvements is in something that can help daily life. Chris Slowe: I I think the folding though is a good example though, because honestly, if you look at if you look at the things the robots are now really good at, they're they're aiming for and by the way, it's like this feels like complaining about the wrong thing, but I'm gonna do it anyway. It's like, okay, they can jump really high and they can run really fast, which of course when Skynet comes is gonna be problematic for all of us, but we'll kinda put that to the side for now. Th they there Zach: It's gonna great, yeah. Marty Weiner: Yeah. Chris Slowe: still aren't that many robots that can actually like deal a set of playing cards, right? Because it ends up being that that hands are really hard to make. Right? And so like, Marty Weiner: Yeah, yeah, yeah. Chris Slowe: you know, leg like legs Kind of the done thing. Now that's it, don't get me wrong. going back, go back 20 years in the Blue Poo reel from Boston Automations, where they were doing their Marty Weiner: yeah. Chris Slowe: like or Boston Robotics, where they were doing their like, you know, the Boston Dynamics, that's it. Zach: Dynamics. Chris Slowe: where they were doing like, you know, what turned into big dog, and it's like this kind of like originally it's kind of like plotting hokey robot, and now it's like it's doing parkour like it legitimately looks like Terminator style, Zach: Yeah, yeah. Marty Weiner: Yeah. Dancing. Chris Slowe: but but it's a Terminator style. Where its hands are still like knobs. Like they don't actually have hands on the things because that's the hard that's the hard robotics problem now, is getting all Marty Weiner: Right. Chris Slowe: that all that sensor density into one place. Marty Weiner: Yeah, I and Zach: Yeah. Marty Weiner: the example I always give about that that I find fascinating is that if you took your finger and your finger was the size of the earth, your finger can detect the difference between a house and a car. We're we're that good. And I don't think that there's any censor tech that's even close to that. So anyway, we we have dancing and parkour solved, so that's at least good. now on the flip side of it all, Zach: Yes. And Marty Weiner: we have this article from the New York Times. This is the on the absolute terrifying side. Russian AI guided drone killed Ukrainian university student and two others using AI to help it navigate. 'Cause one of the things that's been happening in the Ukraine-Russian war is you know, first they had drones, the U Ukrainians started using drones, so the Russians added jamming. So then the Ukrainians started adding in these w spools of fiber optics. So as the drone flies, it just unwinds the spool. But now if you add on that last mile, you know, you get the drone within range, you tell it what you want to hit, and it hits it with only using guided AI. There no jamming doesn't matter at that point 'cause you don't need a radio sign you don't need any communication. So that's quite terrifying. Zach: Quite terrifying. you what we were talking a little earlier about how this we don't know yet how this is gonna interface with the law of war. So there's this weird international norm we're Marty Weiner: Mm-hmm. Zach: all living in where like Russia and Ukraine have a war and everybody else is not in the war, but we are all of it. We're all giving weapons, money, all this stuff is happening. But if like a single US soldier acting as a US soldier were in the war, it would it would cause all sorts of havoc. But what's weird is you you might get to a world where the US could give like a literal army of robot soldiers that even look like I don't know if you would actually want them shaped like humans, but just just for poignancy, you can imagine an army of like human looking robots that behave like American soldiers, but they're technically not Americans. like This is gonna really push right up to the edge of what international law allows, and I don't think we know what the consequence is yet. so there's there's Marty Weiner: Yeah. International Chris Slowe: Well there was a there was a there was a parade, I think it was actually yesterday in Ukraine where they had an entire parade Zach: I saw that. Chris Slowe: of just of just the drones. and Zach: Yeah. Chris Slowe: it was it was a little I mean it was it was impressive what they have managed to do over the last couple of years. It Marty Weiner: Mm-hmm. Chris Slowe: was also a little creepy to see an entire parade of bots effectively. you know, be begun the clone wars have. Zach: well it ha and th th there's been talk now for a Marty Weiner: Yeah. Zach: few years about like just saying no autonomous killer robots and that seems to have just whooshed by. yeah. Marty Weiner: Wait, wait, wa but Right, right. Chris Slowe: yeah, it's it's a in the bets in the rear view mirror by a couple of years at this point. It's like Marty Weiner: Who's who said that? and then Russia's like, really? None? Zach: Yeah, it's Chris Slowe: Yeah. Zach: it's not good. let's see. Speaking of which, hopefully not the same. god. god. let's move on. so Chris Slowe: Send No target target target fell fell out of window onto drone. it's like sorry, too soon. Marty Weiner: Yeah. Zach: Tesla, Robo Taxis in Austin, may now be unsupervised? Marty Weiner: Yeah, they're starting to run them unsupervised or at least that's the claim. so Zach: Yeah. Marty Weiner: I when I see it then I'll believe it 'cause you Zach: Yeah. Marty Weiner: I believe they said that it was gonna be running a year and or two years ago. Chris Slowe: Did they get did they get permission on this one or did they are they just doing it? That's always the funny thing in these situations. Zach: It's Marty Weiner: I don't know. I do they know? Chris Slowe: I don't know. It's a Zach: Do they know my god. Marty Weiner: I would assume they have permission, but you never know with this stuff. But no, it's interesting. I I mean I've been a little bit of a naysayer on on Tesla's rollout just because they they refuse to do things like lidar and I feel like that's just kneecapping themselves for no good reason. but if they start to get things rolling and actually working in Austin, then that'd be great to have a competitor against with Waymo. Zach: Yeah. so let's talk about NVIDIA's Grok chip, which is not to be confused with Grok GROK. The mecha Hitler. th this Marty Weiner: The bad the bad mouth yeah. Yeah, Mecca Hitler. Chris Slowe: Yeah, this is the Mecca Hitler. Zach: is GROQ which has not claimed to be Hitler, it's a chip, not a model. Marty Weiner: Ha ha. Zach: as far as I know, it hasn't. I don't wanna assume. Chris Slowe: yeah, it it th the the chip actually the company actually predates Grok the the LLM by a couple years. I think it was founded in sixteen. so got got scooped up by NVIDIA a bit of time ago. their main pitch from from the Grok chip has been that they have they're building these LPUs, which they're calling language processing units. or at least the branding has been that. with the idea being that instead of being a if you if you look at the at at traditional GPU use for language models They're useful because, of course, GPUs are giant, massively parallel, high memory devices. And they end up being a really, really great as a side effect for for things like LLMs, which are giant parallel operations with huge memory requirements. The problem is that they're not designed for that. So there's still a lot of work that has to be done into making sure that the that the work ch the workflow is optimized for the for the hardware. This is the first time we're actually seeing chips that are optimized for doing LLM inference. And the main difference is that, well, first off, they've got just a ton of memory on chip sitting next to the next to the com to the the processor. So there's not really a notion of cache on these. It's literally just like the RAM is on the chip. So also somewhat expensive to build aside effect. That ends up being the most expensive RAM you can possibly buy. and the other thing that they do though is they're able to basically hand over all the work directly to the compiler. So there's no like there's no branching, there's no expectations, there's no assumptions built into the the way it operates. It's baked into the code, which means this is actually extremely efficient. so the real question's gonna be like do these end up being the being the the the the way to move from super expensive GPUs to probably also very expensive but probably a heck of a lot faster LPUs for inf inference, which ends up being most of the problem. Marty Weiner: are these also designed to work at the edge or are they more for the the big bad boys? Chris Slowe: These are designed to work at the edge. So the intent here is that, you know, if if you think about the the underlying problem with LLMs, inference ends up being where most of the compute time goes these days. baking the model, of course, is very expensive, but it's a one-shot kind of kind of long-term operation. Here we're talking about kind of ongoing. You need to have like your local context window loaded into RAM, and you need to have your compute nearby. And the intent here is to actually have one chip that can manage a whole bunch of parallel conversations. At really breakneck speed. So some of the demos from a couple years ago, I haven't seen I haven't seen recent ones, but when they first were announcing them before they got bought by NVIDIA, they were able to do like 100 tokens per second at a time when they were still trickling out at one per second typically. you Marty Weiner: Mm-hmm. Chris Slowe: know, the models have moved, the the technology's moved, but they're still incredibly fast compared to what you see, which means they're actually great for putting them at the edge, which also means you can start thinking of these as almost like a layered model approach, right? You have like the first quick model on the end. A lot smaller, but can like start doing things like, you know, routing traffic and figuring out like do I need to go to the you know, the full blown heavy model or can actually answer this here. Zach: seems like there's been a trend of trying to match the chips to the type of like model architecture that we do now. I guess there's some risk that tomorrow someone does a breakthrough and and and the whole system changes and these chips aren't as useful. Chris Slowe: Well, mean the hard the hard part here has been that that that's definitely that's definitely one big thing is that we don't know what the what models gonna look like in two years. The other hard part is it's actually really slow to build new chips. I mean we're seeing this right now, we talked about it last week with the with the cost of RAM and the cost of of GPUs. Ultimately that boils down to the fact that like bringing up supply and bringing up capacity to build chips is slow. in this particular case, like bring building entirely new architectures is just as slow because it's like you're not you're not just iterating on last generation GPU, you're building a t a new die from scratch. Which means you have to deal with now all of the debugging that goes into is this gonna work? Where are the quirks? And of course if you're shipping hardware, it's you can't exactly ship like a point release and say like, yeah, we'll patch that in the next version. Marty Weiner: This would be where it'd be great to get somebody who from NVIDIA on the show. 'Cause one thing I'm curious about here is I I bet these chips come together relatively fast. there's a couple of things I to back that up. One is there's ridiculous amounts of money now. Like I remember hearing that OpenAI was going to build their own chip and I've worked at companies building chips and I was thinking, okay, that's gonna be a hundred million to five hundred million dollars. That's a lot of money. Is that worth the investment? But they're also raising eight hundred billion. Like the n the scale of numbers has gotten to where building your own chip Zach: Yeah. Marty Weiner: is not that interesting anymore. The other side of it too is these chips, you know, at the end of the day, models are a big giant matrix multiplication beast with a lot of RAM. and We're not talking about really I don't know about this chip. I I this is where I want somebody from NVIDIA, but I don't think we're talking about substantially new architectures, new ways of thinking. I think we're talking about choosing your troops just right, right? I want a lot more RAM at this spot, I want a lot more compute in this area. And so I'm thinking that that probably leads to a much faster chip development, more like an ASIC model where you you you choose your pieces and and assemble them and and hit go. but I I'd love to learn more about that. I also wanna go Zach: Yeah, it does seem like Marty Weiner: I I wanna go check Terminator two and zoom in real tight and see if the word NVIDIA is secretly written on those chips. yeah. Zach: secretly on the the little chip. Yeah, I I was gonna it see it also seems like economically we're in a world now where at least for a while every chip is gonna be used no matter what it is. So like that that might have de-risked some of these moves people are making. but let's move on to we want to talk about math and science. So there's a really cool paper out by Terry Tao based on a talk he gave recently called Mathematics in the age of AI. anybody can read this paper. It's not a mathy paper. It is it is oddly to me anyway written like a math paper, like he's like conjecture four point three about what we should do says this. But like he's basically saying like, you know, what what is the the the purpose of math? And and he does say, you know, one one of the purposes we we do talk about is making theorems and figuring out what's true. But then there's all this other stuff we do. Like we train the next generation and we're like a reservoir of knowledge to transmit truth about the universe to other people, and this sort of stuff. So he he's trying to I think of Terry Tao is kind of like a one of those guys with just infinite cred. You know, like nobody would would think he's acting like to make money or to please a corporation or be political. He's just saying what he thinks is true. And so for him, he he he very gently says, let us assume that the models are going to keep getting better and better. What should we do? And he he lays out what what maybe for people interested in this, something people might find valuable is this what he calls a rule of thumb. And I'm just gonna read it because it's the main thing to take away, which is if the authors that is of a theorem, if the authors cannot convincingly demonstrate that they are able to give a clear expert level talk on their results, one that is correct and properly attributed, then the result should not be published. A proof that no human can properly explain should be viewed as incomplete, even it has even if it has been formally verified. Marty Weiner: Hm. Zach: so I don't know what do you guys think of that? I mean or maybe we get to a world the stripes as we're in the singularity. Marty Weiner: Well, I'll tell you a funny thing with the coding world is we've all stepped into this world. It started with, no, you still need to understand the code and we would, you know, do one line at a time. Vib coding starts, you know, I remember Shopify had some article that their senior engineers were no longer coding around February of this year. And then now we're in a world where you know, many of us don't even look at the code whatsoever. I I still look at the shape of the thing. But I don't look at the code and I still I I just look at the outside of it. I I mean we've all kind of stepped into more of a manager role. So would it be I wonder if Terry Tao would agree if we came up with some if there was some proof that was phenomenal that had a, you know, say P equals NP or something, and it was extremely difficult to understand, but the right shape of it was there, and other LLMs said, Yeah, that looks good, is that Is that valid? Zach: I mean that that's an interesting one to me. I mean, you guys are the C S people. My understanding is to say if you proved P equals N P then you could like use an application to be sure, right? Yeah. We we w Marty Weiner: Yeah. Yeah. I guess in that case the proof would be in the pudding because you know the one one of the things with P equals MP is it basically says if P equals NP, there's a lot there's a whole class of hard problems that are immediately Zach: Right. Marty Weiner: super easy to to solve. but yeah, I I don't know, I'm some big Zach: Yeah, so so so it's Marty Weiner: result. Zach: but well so yeah, but I think that's important 'cause it's like if the big result is P equals NP, then you you literally you don't have to know how it works. You just have to like be like, give me something I can run on the computer that I couldn't be possible unless P equals MP. But if if you're saying something like, I discovered some cool new shape in thirteen dimensional spheres or something, you know, that there there's no application for that. All we care about is the truth. And in so in that case it does matter. Marty Weiner: That's yeah. I I think to some degree, but a lot of math, even though that most that very esoteric one is a building block for other math. So even if you don't understand the result underneath you, you can still build on top of it. But y you're sort of you're you're taking it as truth, right? And so if it Zach: Yeah. Marty Weiner: turns out you're wrong, then some of your research may collapse. But there Zach: Yeah. Marty Weiner: is also a phenomenon with a lot of math where you keep going forward and you're No, things are working. So my assumptions must be pretty decent. You know, maybe there's a slight problem in it, but Things are working where I am, and so I I I don't know, I I I feel a little wanting to contest this result a little bit by saying, Yeah, at the moment that's probably a good idea, but but do we is there a point where the humans are just not gonna be able to keep up but the results are equally good? Zach: Yeah. Chris Slowe: I I guess like what I I I I treat this request or that's not request, I treat this guideline as almost like we want math to continue to be human discernible. The intent of math is to generate insights that humans can understand, not just like because if if the humans don't understand it, what's the point compared to the abstract reality of the math that exists? Right? Like it's almost like they're they're requiring that a brain be able to compute it, not just an LLM. Which I think is a is An interesting line in the sand 'cause I 'cause I I do worry that we've already crossed that line in a couple of places where we have these like these encyclopedic proofs that are just like effectively enumerate every possible every possible way Marty Weiner: Mm-hmm. Chris Slowe: this could be true or false and therefore prove that it's true or prove that it's false when you have like a hundred thousand cases. So it doesn't mean that gives any insight derived from that, but it's still considered to be perfectly fine math. Interesting times we live in. Zach: Interesting times, yeah. Marty Weiner: I yeah. I I guess it comes down to yeah, what it w what your opinion of the purpose of math is and that it'll be interesting to see that evolve. Zach: I I think that's it. I I think, you know, maybe maybe a way to close on that point is is I I had this experience early in this where I was talking to a friend who's a developmental economist that is like trying to make poor economies work, which is a very hard problem. And we were just kind of talking about AI and I was like a speaking basically as an artist, I was like, How would you feel if like your profession was just solved and like there was nothing left for you to do? And he was like, that'd be great. So many poor people would be better off and I like, This is a to th Marty Weiner: Well. Zach: I I I like it was a stupid question to ask a developmental ec economist, 'cause to him the result is the entire point. It's not it's not about the journey or truth or anything. It's it's it's it's fixing lives. so I I th I think how you feel about this stuff And what the point is depends on, you know, why you're here. Marty Weiner: I I gotta say the way you paused was interesting. You said, that'd be great. So many poor people. I I didn't know where you were going with that. Zach: so many more poor. all right, let's move on before I embarrass myself. So my god. it's a it's a beautiful future. okay. Chris Slowe: They can drive they can drive the robots that have to fold my laundry for me. Marty Weiner: It it's funny right now, yeah. Zach: Yeah. Moderna has a there was a big announcement this week that they have a cancer vaccine. So just to clarify quick because I didn't understand this, usually when we think about a vaccine, we think about a preventive. This is more like an injection that teaches your immune system, like a vaccine should, teaches your immune system to react properly to cancer. Marty Weiner: Yes. So this was really interesting stuff. So I think what they made, this is Moderna and Merck, made an mRNA vaccine that was succeeding in late stage melanoma trial. And I if there's a biologist listening, I apologize because I'm probably gonna screw this up. But my understanding of it is a lot of medicine is here is one thing that works generally on the population. It's you know can think of it as static. But what I think they're starting to push into is these pipelines, the almost like programming pipelines where you take out say a little bit of the cancer, what is that called? Biopsy. See, not biologist, not doctor. Anyway, I think what you do is you take that, you then s make a custom made vaccine for that individual that would target that exact cancer. So we're moving from static drugs to more like individually targeted drugs, which I think would is easily the future of of medicine, at least for the the time being, that AI would allow. You know, we we see this all over the place, right? AI allows you to have a very custom experience learning math, things like that. It meets you where you are. So if AI can leverage this up to the point where we can make a drug specifically for an individual that targets exactly what they need, that that maybe hits exactly the markers they need to to stop their exact cancer, this could be a huge path forward, especially just even building up the entire flow for how you get this going. Zach: Yeah. it's it's funny you mentioned AI for for math tutoring because the next thing we have on the list is this funny result from students in China, is that right? Yeah. Marty Weiner: yeah. so the idea was they did this study on learning penalties when kids use AI. They did they took a whole bunch of students in China and they looked at their homework scores, they looked at completion time of the homework, and they looked at exam scores. and I think you probably know where this is going, but homework scores have shot up. They're much better with AI, which makes sense, I think, in any world. So I've been using AI to l to relearn things like electro electromagnetic theory, but I am learning it for the sake of learning it. My incentive is is aligned with learning it. But a lot of school does not align Zach: Yes. Marty Weiner: with that incentive, right? especially I think in China, where they push the students exceptionally hard to to get good grades and you know to some extent true in the US. but most students, including myself in high school, are motivated for Zach: Yeah. Marty Weiner: a good grade, not to necessarily learn the thing. And they're also motivated to go play, you know, video games or or watch YouTube. Zach: Yes. Marty Weiner: before AI, it would take say, I don't know, average of sixty-five-ish minutes to complete their homework with AI, it kind of s the graph smushes over a little bit. It's more like forty-five minutes. Well, if I want to get back to playing video games, yeah, I'm gonna use AI. but on the other side of it, if I'm stuck on the homework and I have to go read the book, in theory I can actually use AI to help my help me learn it. so it's a little hard to tell. Now the proof is then in the pudding with the exam scores. The graph kind of smushes and you probably are gonna have some students who care and they're gonna still be at the top. there's a healthy number of people who have gone off the end. A very large number of people are now going into that exam and they're failing it or they're they're much Zach: Yeah. Marty Weiner: lower. And what that means to me is that we have a lot of students who are gonna fall off the edge simply because they're just using AI to get past you know, get through it all, right? so I think teachers are gonna have to come up with new mechanisms. this is nothing new, they're gonna have to come up with new mechanisms to make sure that students are learning this content and it may have to be some in house work or maybe more regular, you know, in house exams, things where you don't have the AI. Zach: Yeah, so we're we're already seeing the return of blue books in a lot of context. one thing that's interesting to me Marty Weiner: Mm-hmm. Zach: with this, so like you the bottom line story is everybody's doing better at homework and worse in exams, at least in this study. So Chris Slowe: Mm-hmm. Zach: one interpretation of that is a bunch of people are just cheating. but another interpretation would be that the quality of your learning is lower if you're using AI. So Chris Slowe: Mm mm mm. Zach: like like, you I I think we're still getting data on this, but but I've had this maybe you've had this experience. You're trying to learn something. You know the whole time AI can just give you a succinct answer. And there's a claim that, you know, it's possible that in some sense moves you along faster, but unless you really have a good grip on how your brain works and how to make sure stuff sticks in it, y Marty Weiner: Mm. Zach: you might be fooling yourself a little because like in in other words, like, you know, like you I'm sure you have a memory of being like in college and there was some impossible physics problem and you just ground on it. And that's not to say all that time was necessarily productive, but there's probably, you know, your brain doesn't want to remember things. And so possibly the struggle itself was part of how you got good. And so to the extent you eliminate that, that might be part of the problem. Marty Weiner: I could see that. I I mean a at this point, like for me lear relearning electromagnetic theory, I I know a lot of what I don't know. and so I can I can I can ask just the right questions and I can ask questions that get right to the heart of it and and I learn it really fast. But I could imagine, you know, if I went into a field where it knows it exceptionally well and I'm just starting from scratch, it might be a it might take some time to I don't know, get Zach: Yeah. Well Marty Weiner: at the meat of of where my problem Chris Slowe: Well the the hard part Marty Weiner: is. Chris Slowe: the hard part here is that you have to deal with the paradox of knowledge, right? Which is that, you know, you can't you can't know if somebody knows more than you. And the side effect of that when you're talking about learning is if you have some if you have a if you have an LLM or you have any s any sort of teacher here, and they end up being extremely confident, but it's indiscernible whether or not they are correct or not, because you yourself don't have any basis on knowing or judging if they're correct or not, it's a good way to actually do the opposite of learning, which is to learn the wrong things or like learn false facts. Zach: Yeah. Chris Slowe: 'cause you know, I a so I think it I think it what what's interesting about the new technology is that it's definitely giving us a chance to reassess like what's the point of the education? Like why are we doing it in the first place? Cause like for a very long time, you go back and you go back deep in history and or like and by deep in history I mean probably our parents or grandparents. And it was Zach: Ha ha. Chris Slowe: like and it was like math. The only way to do math is you sit and you just grind through doing math. Right. And like to the point where they would just teach you algorithm after algorithm that you would need to use because there's no other option except to do things by hand. Right. Like there's legitimately a hand algorithm for doing square roots that they don't teach in schools anymore because of course calculator do it for you. And now it's like, well, no, actually the point of math is to understand numbers. And so you see kids and they're kind of like doing these more like conceptual, like play with numbers, like you know, more much more like broader versions of it. What's the version of that for writing now? Right? Because Zach: Uh-huh. Chris Slowe: you can ha you can all of a sudden now have the elements like do a do a five paragraph essay. And it'll do a pretty or do a five-paragraph essay that sounds like a sixth grader. It'll do a great job. Right? Zach: Yeah, yeah. Marty Weiner: Yeah. Chris Slowe: probably a couple more M-dashes than you'd expect, but otherwise it's you know pretty hard to read. Some Zach: Yeah. Chris Slowe: part of the education is just kinda like go through the grind, but we just don't know which part is important for that. Zach: Anthropic had a pretty cool announcement that they made some improvements in protein folding. Marty Weiner: Yeah, so Claude on their on Anthropic's blog, they use they said that they use Claude to accelerate protein design. and this reminds me a little of some of Google's work with protein folding. they said Claude designed protein binders against fifteen targets and succeeded against fourteen of them. Between twenty-two and thirty-five percent of its individual designs bound successfully depending on the setup compared to ten to fifteen percent in typical protein design campaigns. I assume typical protein design campaigns we're talking about, you know, humans plus s the the current state of the art and ML. So if that is the case, that is a sizable jump and improvement in being Zach: That's pretty cool. Marty Weiner: able to create new proteins. Chris Slowe: And the important thing is that they're able to make their logo out of proteins, which I think is the milestone that was here. Zach: That's right. Yeah, yeah. Four fourteen Marty Weiner: Yes. Yeah, Zach: billion dollars was spent at Marty Weiner: the finally, marketing and biology together. Zach: last. there was that complete neuron map of a house fly and and they got it to be a fly in Minecraft. Yeah. Marty Weiner: This is quite terrifying and interesting. I guess the idea was that they were able to go download the entire neuron map of a housefly, which is available. and this little housefly flew around and thought it was in a real world and and even, you know, accidentally drowned itself a few times as it was exploring. Yeah, it's definitely in the oddly terrifying category. Actually I think I discovered this Zach: my god. I often Marty Weiner: on r slash oddly terrifying. Zach: I play Minecraft and I think it would be great if the the flies I encounter could actually die. that that would that would really improve the experience. Yes. Marty Weiner: That's right. Yeah. That I d I don't want him I don't want him to be four lines of code. I want it to be an actual suffering housefly. Zach: my god. yeah, it's it's getting weird. What's not weird, another hundred year old math conjecture falls to a counter example. I feel like we're getting we're getting bored of this now. Yawn. the the cotheod conjecture. Marty Weiner: Boring Yeah, yeah. And can anybody pronounce this the Kara Theodore conjecture? It's a hundred year old conjecture just just fell. so cool, yeah. Another day, another conjecture falls. Zach: yeah, another another one of those. and our our our last tech thing, and I'll I'll throw this to Chris is the quantum fridges. Marty Weiner: Yeah, finally we can get our bananas really cold. Chris Slowe: Yeah, okay, so qu Zach: Yeah. Chris Slowe: so I I'm just gonna I'm just gonna rant for a minute because the the the actual headline is IBM's new quantum fridges are nearly two hundred times colder than deep space and could pave the way for fault tolerant quantum computing. Units I don't know why we don't like to use actual units for things. it's like these fridges Marty Weiner: Ha ha. Chris Slowe: also weigh the they're the same weight as three refr three washing machines. it's like Zach: Yeah. Chris Slowe: it's like what's the I Okay. So the the long and the short of it is Marty Weiner: Whoa, wait, wait, wait. Three washing machines? That's awesome. Chris Slowe: Three whole washing machines. Units, we understand. so I think what they're trying to say here is two okay, so so space is about two point seven degrees Kelvin, which is two point seven degrees above absolute zero. Two hundred times colder means yeah, they're ten they're ten millikelvin, which is which is impressive for a fridge. But actually, like it looking deeper in this press release, it looks like they actually just used off the shelf dilution fridges, which have been around for a good twenty, thirty years. which actually can get down to two millikelvin, that's That's in in in in Imperial units, that's that's a thousand times colder than than deep space. the the actual breakthrough here was one, they managed to point these dilution fridges at their marketing department to get a purse release out of it. And two, I Zach: Ha ha ha. Chris Slowe: believe the actual interesting thing is that they have some sort of modular technology where they can basically they're trying to network now quantum chips. And so Zach: Mm-hmm. Chris Slowe: the fridges, of course, for to do something like this are huge. And the real breakthrough is they're able to basically have some way to dock multiple fridges together with a port that they can actually connect multiple multiple quantum chips together. that seems to be the breakthrough here. The fridge thing is you can buy fridges like this. Zach: I th that's interesting. So so you know the my understanding, you know, to keep a get a qubit to behave, is is you have to have it be extremely cold and not interact with the outside environment. And and part of the problem is to do anything Chris Slowe: Exactly. Zach: interesting you have to have like I don't I d I haven't heard the latest ratio. I i i e exactly. Well then you have to have a lot. Chris Slowe: You have to interact with the outside environment. Yeah, like the the using the chip you have to read off the chip. Yeah. Zach: Right. And and and you have to have a lot of them and and that's really hard. My understanding, maybe you know this, is like, we talk about qubits like it's a thing, but actually because of a fault tolerance you need like a large number of qubits to have a single logical qubit like in an algorithm. Chris Slowe: Yeah, so so for for the most part, no one yet has th as long as qu as long as quantum computing has existed as a field, at least a theoretical field, there have been attempts to figure out like how can we actually ensure that qubits live long enough to be useful. 'cause if the qubit only lasts for less than one compute cycle, then you're not really getting much out of it. and so there's been these algorithms called quantum cor error correction algorithms. Generally the smallest you need least like seven qubits that are working together to get one logical qubit out of it. I don't believe yet anyone has succeeded in running a quantum error correction all the way through. if that happened, that would be a major breakthrough. Zach: Yeah. All right, let's move on to I want to get some of the weird econ stuff. So there was this big thing where Anthropic had a a profitable quarter, which is crazy because they're just burning down money like it's a party. and actually I I so I was listening to an economist who said you could argue this was this is bad news for anthropic because it means like they should have been spending a lot more money on getting more compute online or training more models. They should not be running a profit yet. but it is kind of a sign of the times. you know, because like one of these big things we talk about is are we in an AI bubble? And people have this knee-jerk. desire to compare it to dot com, those companies were not turning a profit. the fact that Anthropic is already making money like this is kind of wild. So let me get to so this was really cool, and it's gonna take a minute to explain, but I think this is worth understanding. Epoch AI, which is one of the really cool groups that spends a lot of money and time and expertise making benchmarks and doing analysis, they basically want to claim that we are missing a big chunk of our GDP because we're not counting NVIDIA properly. So I'm going to walk you through very quickly why we're not counting it properly. so the basic thing to know is I was trying to think of how to explain this, but if a like Pixie showed up and put $50,000 in your bank account That does not actually show up in GDP because GDP is manu measuring transactions Marty Weiner: Mm. Zach: in the economy, right? And so if you get magic money, it just it doesn't nothing shows up. If the ferry said, you know, punch yourself in the face and I'll give you fifty thousand dollars, then it can be measured because there was a service, Marty Weiner: Yeah. Zach: right? so GDP is a way we try to account for the economy but can only measure things that actually show up as as transactions in the US or at our borders. So NVIDIA does this thing though. So imagine you have a chip. NVIDIA is really smart. They design a new chip. They call up TSMC. They say, we'll give you $10,000 to make this chip exist. TSMC makes it exist. The chip is now in Taiwan. That chip gets sold to Marty Weiner: Hm. Zach: an assembler in Mexico for $40,000. At that point, NVIDIA, like the giant pile of gold Jensen Huang swims in every morning, gets an a net of 30K, right? that does not show up in GDP. That's just magic pixic money pixie Marty Weiner: Mm. Zach: money because the transaction happened outside the economy when the chip went from Taiwan to Mexico. Later, when the Mexican assembler is done, they sell it to Microsoft or whoever in the US. Microsoft pays fifty thousand dollars and imports something that's worth fifty thousand dollars, which nets out to zero in GDP. So the results of all that tortured pathway is that. the IRS knows that NVIDIA made all this money, and like you know, stockholders know, but GDP does not actually count it. And the reason this matters is that NVIDIA is so gigantic. If you don't count that, they claim you might be missing 10% of the growth in GDP per year. Like it it's I forget their number, but if we say the GDP growth was two percent, it might have actually been something like two point two percent, which is is really non-trivial, especially if if you count it over a number of years. And then the last thing to say on that is they have this graph which is a little goofy. But it's worth saying if Nvidia just stays on trend, within like a few years, we'll be missing like 2% of GDB growth per year, which for an average year is about all of it. so it's this huge accounting bug, which is kind of fascinating, and it sort of extends into this problem we've had since the IT revolution of like it being really hard to count all the ways in which we're maybe getting richer that don't show up as transactions. Marty Weiner: It's a general problem with GDP and it probably Zach: huh. Marty Weiner: I guess it's getting worse in a more globalized economy. Zach: yeah, yeah. I mean this this is this classic problem of like there's there are ways in which we are richer that are very hard to measure. you know, so it's it's like like I I can I can do an amount of research on Google Scholar that would have cost a million dollars forty years ago for free, but that doesn't mean I go to the grocery store and things are cheaper. so anyway, I thought that was that was really interesting, that we might be missing more and more of the economy as we get globalized and we deal with more intangibles. Chris Slowe: I I remember I remember historically one of the biggest criticisms of GDP is that since it's only a measurement of motion of money, that Zach: Yes. Chris Slowe: you can actually come up with like easily come with scenarios of of really terrible situations where actually like that you'd have like huge GDP, which would be like you know, a a professional who goes through a divorce and he's going through cancer treatment. is actually probably going to be higher as like a g like a GDP hero because they're paying lawyers, you're paying a bunch of medicine but medical bills. Zach: Big win for GDP. Yeah. Chris Slowe: And like you've you've now created like two separate economic units out of one originally it was a marriage. Zach: That's right. Chris Slowe: As opposed to and and this is like feels like the oppos exact opposite situation. It's like here's all this money that's actually moving around. But because of accounting, it's not counted towards GDP, even though it's arguably Zach: That's right. Chris Slowe: some of the most important money motion you could see. Zach: Yeah, so th there's this classic observation that feminist economists started making in the seventies, which is like, you know, so the you know, at the time you would have said housewife, right? So it's like there's some amount of labor done by housewives that is literally never counted in GDP because it's done quote unquote for free. And if you literally took that housewife and just had her do all the labor she's doing at home, it's worth maybe, you know, fifty or a hundred grand a year. It's like like a a a a a median level job. And so, you know, there are these estimates that are like, well, actually we're not counting two trillion dollars of labor. The only thing I would say is like You know, and then we should move on. But like economists talk about GDP. I think because every time someone tries to come up with a better measure, it basically tracks GDP 95%. Like we have to have some way to measure the economy. Like you can imagine a world where you could have a a a robot over everyone's shoulder that like counts how much stuff they've made, and then you would have a better number. But but ultimately we have to have some way to dip into the stats. and and this is maybe the least bad way. And it's also the one we're stuck with. Alright, so so there's a a Stanford study that found it looks like AI is hitting entry-level jobs the hardest, so You know, if you're an econ nerd, what you've been looking for since 2017 is for AI to show up in the top line macro stats, and it has not. It's not really obvious that it's there in productivity or GDP or whatever. but if you look at more fine-grained stuff, there are possibly some signals. So maybe we can throw up these graphs. They're a little hard to read, because they've they've got a whole bunch of different curves on them that are saying stuff. But the basic finding is the these this group bend different jobs into what you might call like more AI exposed. they do call it that. AI, high AI. AI exposure versus low AI exposure. When we look at the more exposed entry level jobs, they're not just down, they're like non-trivially down, like close to twenty percent down in terms of hiring. Which which I i i i is that does that comport roughly with what you guys are seeing out in the field there? Marty Weiner: I I've would say in my f area with my startup we have probably slowed hiring than what we would have normally done. it's definitely a different equation than I would have used ten years ago and it's definitely Chris Slowe: This is Marty Weiner: a much fewer heads Chris Slowe: This is this is also pretty endemic too of across across the industry, the n the number of junior engineers especially being hired is actually h at historic low. because a lot of those tasks that would have ended up being pretty easy to give to a junior engineer to do, which is kind of like grunt work, has been offloaded Zach: Yeah. Chris Slowe: onto LLMs. And so now you're in this you're in this kind of trap, which is which actually drives me a little nuts because like the hard part is like where are we getting our senior engineers in five years if we keep down this path? you need to you need some you need some time Marty Weiner: I'm also Chris Slowe: and seat to get some taste. Marty Weiner: Yeah. I I'm also a little unclear on if they have specified which fields we're talking about, because I would guess that plumbing Zach: Mm-hmm. Marty Weiner: has not been affected by that. Zach: Yeah, I I need to look through their documentation, but they they you know, they they talk about this idea of tacit knowledge. Basically if there are jobs where there's stuff like you're saying, like taste or just stuff that's like diffuse, then you're less exposed to AI. Whereas you have something where like the information can be put in a book, then you're in trouble. Marty Weiner: Yeah, they're they're talking here a little bit about jobs like accountants and auditors and receptionists Zach: Mm-hmm. Marty Weiner: and in information clerks were among those judged most susceptible to AI automation, while jobs like chief executive and registered nurse were among those using AI augmentation most often. Zach: Yeah. Chris Slowe: I also believe one of the major disrupted job families right now has been paralegals. 'cause if you think about what the job Marty Weiner: Yeah. Chris Slowe: a paralegal is, like ultimately they're about like kind of boilerplate and contract summaries and kind of like a lot of the grunt work for lawyers Marty Weiner: Do an N DA. Chris Slowe: do an NDA, which is you know, Zach: Yeah. Chris Slowe: it's it's it's it's a little hard to justify that if you have to if you can just ask an LM to do it. Assuming the LM is actually going to not hallucinate an entire court case. Zach: Yeah. Yeah, Marty Weiner: Sure. Zach: but but but the price is there. That's part of the problem. Like that paralegal better be pretty good if they're gonna beat an AI on price. speaking of which, by the way, so so Marty sent this that you can now good news, you can connect Chat GPT to your finances. so you you guys gonna do it? what was it? Do a rich. Do a rich, please. Marty Weiner: Yeah th that's gonna do it no. do Chris Slowe: Do a stock portfolio. Do a rich. Marty Weiner: a drain my bank account accidentally. Zach: So let me say I'm sort of I've sort of torn on stuff like So on the one hand, like I don't know if anyone's estimated, but if you can imagine literally everybody was perfectly economically optimal in terms of like Marty Weiner: Yeah. Zach: how they allocated their money, that must be worth trillions, right? And and not just trillions, but like Marty Weiner: I think so. Zach: life improvement. Because it's like if you're rich, you already have good allocation because you pay somebody to do it. But if you're like a regular person or even middle class, you know, I mean we can argue maybe there's no better allocation than just index funds or whatever. But but even then there's stuff like, you know, debt, how much debt do you want to have? Like like are you on Marty Weiner: Well well even e Zach: Yeah. Marty Weiner: even knowing all these little things that nobody we're not really taught in school, right? Like if you're rich, you deal with money, you work with money, you learn money. If you're middle class, you know, it's do you know how a mortgage works? Do you know that sto that the SP is probably just a fine investment for you and don't let one of those hedge funds come and, you know, mess with you? I I wonder if those would be extremely valuable just to the average person. Zach: Yeah. I I Marty Weiner: maybe maybe I wouldn't connect it to my finances, but but maybe maybe it's not a bad idea, especially if it came back Zach: Well I I Marty Weiner: and said, would you like me to invest your money? I don't know. Yeah, like yes, go and invest my money. I don't even know how that works. Chris Slowe: little it's probably a little hard for us to to talk about this too 'cause like I'm you all of us all of us are are at some are young professionals and you know, probably have like Zach: Yes, yeah, true. Chris Slowe: have have dealt with mutual funds and you know, in I in Zach's case he has that giant mattress full of dollar bills. I mean like but all of us have some sort of a solution for long term wealth. Zach: That's right. No, I have Ram. I just have Ram under the bed. That's Marty Weiner: Yeah. But b right. But but not everybody knows what brand of mattress to choose to stuff your money into. Zach: right. Chris Slowe: Yeah. Zach: No, but you know, I was gonna say the the other thing about this is that like where I actually wonder if it could be really useful, so you guys know like one of the pieces of advice you're given, even if you're just in index funds, is If you're not in like stable stuff like bonds or like literal gold or whatever, you need to you have nerves. Like if the market goes down, you don't sell. And and so Marty Weiner: Yeah, yeah. Zach: that that would be one area where, you know, offloading to a robot might not be the worst decision. Yeah. I don't know. Marty Weiner: might be stabilizing. Yeah, there were a lot of people during the two thousand eight when it dipped to about fifty percent. If you stepped off at that point, you would have been in financial ruins. and and I think having some of that stability to say, hey look, we we've seen worse. Don't panic. Yeah. Zach: J just have the robot say don't panic, chill out, this'll be Yeah. So Chris Slowe: It happens during the Or or the or the reverse of like you've you you're you you need to rebalance your portfolio is another one of those things, right? It's like these stocks have done really well, these haven't done as well. You sell the winners every once in a while and like recoup the coop the gains. Like the both both sides of it work. Like don't Marty Weiner: Right. And and take a tax incentive. Chris Slowe: don't think it's gonna go to the moon when, you know, it actually Zach: Yeah. Yeah. but let's let's keep rolling 'cause we're running out of time. I want to talk about prompt kitties. Marty Weiner: So, I this fantastic blog post by this guy saying that he had an Amazon tablet and Amazon kept shutting it off and he's a hacker. He goes and he's working on it. He just wants to use it as his own little screen. and he basically spent two hundred sixty six bucks on four AI models to go in and get it working, and and find an exploit that allowed him to unlock it. and he described himself as a as what I think is kind of a new and and interesting term. He called himself a a prompt kitty. He didn't really know what all he did there, but in the end it it worked and he was Zach: Yeah. Marty Weiner: able to get his tablet. Zach: Hey, it was only three hundred dollars. all right, let's Marty Weiner: Whereas a normal tablet costs a hundred bucks. But that's not the point. That's not the point. It's cool. Leave him alone. Chris Slowe: It would have been three hundred three hundred three hundred ten if we bought a new one. Zach: move on to arts and culture. So this was a really cool article out of Seesaw the group at MIT. I I I I read through this and I think I understand it. I'm trying to gonna try not to get into the details too hard, but essentially what they did. Is so there's this thing where suppose you run a model and you can have it make images, you don't exactly know how attribution works, right? You don't know like, did it totally steal from just this one artist or whatever. And so what they did is they trained a whole bunch of models on different chunks of a corpus. And the reason you do that is so later you can say, what happens if we remove this one artist? What happens if we remove this image set? Right? So they have this big set of like stuff from artists, but also just regular pictures and things. And and because they have all these modules, they can effectively ablate people out of the training data or or see what it look would look like if you did. And the result of that is essentially the bigger you make the model, the easier it is to remove any given. An artist with little to no effect on the output. So let me give an example, like a narrative example. Suppose I trained a bunch of these models on like a thousand manga artists, or like I said, I had a a a corpus of a thousand manga artists, and then I trained a thousand different models to to get different chunks of that corpus, so that later I could just switch off a single artist by by comparing two models. Right. So what I could do is say like like there's a great manga artist named Shiguru Mizuki, I'm sorry about my pronunciation. you could probably get stuff that looked like he had drawn it without having him. And more importantly, at the end of this, you know, once you've trained this, you could say manga model, draw me a doggy. and once you've done that, you could eliminate any single artist from the corpus and it would still draw you roughly the same stuff. And so the question is what why is that? And so the r authors of this were willing to s to say maybe it's creativity. there's a much more boring explanation which they also give, which is something like this that is when you're doing that training, The model extracts elements from many different artists, right? And and there's a within a large enough corpus, you have all the elements you would want to produce something. So I'm using this manga as a random example, but it's like if you look at it, it a single artist who's just part of a milieu, like was influenced and influenced people, whatever elements they have are going to show up in somebody else. Right. And so then the question is what this means. They had this term I found kind of funny, which was attribution decay, which means something like, How much we have to credit you, artists, keeps going down as we make the model bigger until we don't have to credit you at all philosophically, because we can remove you from the model without loss of quality. and so I thought that was kind of like a bit of an accounting trick. I think it is legally correct though, which maybe is all that matters from the perspective of like a CEO or something. I don't think this will answer any of the complaints from artists about how the training is done. I thought it was mathematically interesting. Marty Weiner: I wonder if this has been litigated. you know, there's a term c transformative for if something is differentiable from from another piece of work or has has diverged enough and I wonder if this is getting at kind of that kind of terminology. Like if I if I make something that clearly looks like Mickey Mouse, that's not Zach: Transformative, Marty Weiner: transformative enough, but if I, you know, add on several layers and eventually it doesn't really look like Mickey Mouse, it could be argued that I'm I'm sufficiently far away. And I Zach: Yeah. Marty Weiner: I don't know if that's been litigated, 'cause if we said, hey, yeah, you've got one one millionth of artist A, can artist A claim any anything from it? Zach: Well well and I so to be clear, th they want to claim you you can get lower than one one millionth. You make the model large enough, Marty Weiner: Right. Zach: literally anybody can be removed and the model is just as good. And so in a sort of philosophical sense, nobody gets attribution. yeah, yeah, yeah. Well Marty Weiner: Right. Or everybody has a little bit. except so so insufficient to not be worth to be attributed attributed. Zach: yeah, yeah, yeah. Well or you know you could say it's sort of like the model extracted all the elements. I need to duplicate your style so I don't literally have to use you. Therefore you do not deserve credit. which is like seems to me Yeah, yeah, Marty Weiner: That's a little different though. Zach: well, I I would say it's kind of like sort of mathematically interesting. I don't think it answers anybody's concern, but it was it was an interesting Marty Weiner: Right. Zach: paper that's been making the rounds. let's let's move on. Val Kilmer is coming back to movies. Marty Weiner: Yeah, and it looks real bad. but yeah, AI Valkilmer in a new movie. which is inter I'm sure it's gonna get better, but this one just looked bad. Zach: Yeah. Marty Weiner: The new the movie, by the way, is As Deep as the Grave. coming out in twenty twenty six. Yeah, I know that that's too soon. Too soon. Yeah. Zach: God. That's ooh. Gra the grave is not so deep, for for yeah I Val Kilmer. could you talk about Chris Slowe: god, that's on little on the little on the little on the nose, isn't it? That's Zach: so it's I g I guess I watched a little of it. My impression was essentially it was bad because it was kind of stiff, and just didn't quite feel human. Marty Weiner: It looks stiff. The the audio performance looked bad. It reminded me of what how bad was it Rogue One, whatever that Star Wars movie that came out with the Peter Zach: Yeah. Marty Weiner: Cushing resurrection. I like, Come on, AI, we we already have better than this. So I I think it was just poorly done. Zach: Yeah. Chris Slowe: Well that was also I mean that was also just traditional that was traditional CG when that one came out too and it was just kind of like the usu and Marty Weiner: Right, right. Chris Slowe: and I think they also had like a in that case they had a voice actor who sounded a bit like like that actor, but but yeah, you you would you would think at this point that Marty Weiner: Peter Cushing, yeah. Chris Slowe: you would be able to I mean they they already have issues with deep fakes for a long time. You think they would at be able to do as good as a deepfake? Zach: Well I see. Marty Weiner: Well, I I've seen better content come out of r you know, guys doing weekend projects on Reddit. Now I know it's different when you get up to a more scaled project, but I was surprised at how oddly bad it Zach: I mean Marty Weiner: was. I I Zach: Wha what I got to wondering about so there's this this analogy I always think of, which is like for like I I'm a vegetarian and when you make fake meat, you can do an okay job on something like ground meat. Basically, like the lower quality the initial product, the the better it is for faking. Marty Weiner: Yeah. Zach: But if you try to fake steak, you just completely fail. And I think part of the problem is like if you're doing like a song with someone's like moving a lot and there's a lot of noise over the voice and no one's like holding a note for a long time, you can get away with a whole lot with AI. This was like close-ups of emotional scenes, which to me are like the equivalent of Like that's that you know it's it's gonna be very hard to do that with AI without anybody noticing anything peeking around the corner. I wonder if that's part of the problem. well let's move on. so we only have a couple more things. there was AI concept art for Spider-Man brand new day, and I think the main thing Marty Weiner: Yeah. Zach: to observe about this is it got printed in a book and looks really quite bad. we we were joking, yeah. Marty Weiner: We were trying to figure out i i yeah, it looks like it looks like AI quality dolly of two thousand twenty three. And incidentally, it looks like maybe Spider Man Brand New Day was being r written and developed around twenty twenty three, so that might be might be you know, just Zach: Yeah. Marty Weiner: pointing to the time and space it was made. But it's it's a little more embarrassing that that they didn't you know, they didn't read the room yet. People are really upset about Zach: Yeah. Right. Marty Weiner: AI. Don't put it in your con in your your book about, you know, the the movie. Zach: Yeah, well Chris Slowe: I mean I was feeling pretty nostalgic though looking at those pictures. I mean it felt like it felt like the original like what like like getting Marty Weiner: I know. Chris Slowe: chance to see the original like Will Smith eats spaghetti meme you know, just in a book in like physical format. It really takes you back to twenty twenty three. Zach: Really takes you back, yeah. It wa and it was like Marty Weiner: Yeah, Zach: melty it was like melty cars in the background. It was just not even questionable. It was just very obviously AI stuff. Marty Weiner: Right. And the ladders are are just a mush of of escape route of w what do they call those? The the fire ladders. Yeah. Zach: so we we're a little past time and I want to get to the last stuff. So real quick, Dr Dr. Dre came out as pro AI. I don't know, you know, d I might I'm not Marty Weiner: Yeah. Zach: an expert on Dr. Dre. I I su surely in some communities he has more pool than George Lucas, but I I he has a reputation, I understand, for caring a lot about money in addition to the arts. So I don't know if it's Marty Weiner: Well, i a and and a sort of the voice of, you know, a huge portion of music, Zach: It's true, yeah. Marty Weiner: you know, w with people like Eminem and Snoop Dogg and others coming out of him he coming out of him. That's not the right way to say it. Zach: Coming out of him. w when he butted off Eminem back in the nineties, that was so my God, Chris Slowe: They no, it's it's it's like Zeus. They sprung fully formed like Athena from the head of Zeus. Marty Weiner: Right, right. Zach: this Marty Weiner: That's what I meant to say. Exactly that. Chris Slowe: Ha ha ha. Zach: is so c Marty Weiner: so but it but he but he's like Linus Torvald of of that community, right? I if I I'm just gonna keep talking. anyway, I'm trying to say Zach: that's Marty Weiner: that he has a massive cultural impact. So I Zach: Yes, yeah. So let me let me just I will close that by the the quote I wrote down. He said, I had a discussion with people a few days ago, they were against AI and I'm like, Okay, you sound like the person who would have been against the drum machine when it came out or synthesizers, right? It's a new tool for creativity. Marty Weiner: Yeah. Zach: Some people are afraid of learning new things. So Marty Weiner: It's I completely agree. We're we're just replaying the Chris Slowe: It's a very reasonable statement. Marty Weiner: exact same conversation that we we keep having over and over. We we're we're look yeah, and Linus Torvald. Yeah. Zach: You you and Dre are just yeah, it's it's okay. Yeah. Yeah. Chris Slowe: Marty and his good buddy Dre, who, you know Zach: okay, so now we're just gonna close out with the weird and funny stuff. So I guess this is Marty, you're just gonna play the video of of s Steve Jobs rising from the grave to interview or introduce Terminator Pro Max. Good, good. Chris Slowe: I have to say, having seen this video, it is a much better performance than the latest from Val Kilmer, so I'm really impressed with Steve Jobs, Marty Weiner: Ha ha. Zach: But Chris Slowe: in this particular one. Zach: I do want it reminds me of the nineties when it was like Marty Weiner: Yeah. Zach: you see this stuff and you're like, Why didn't you hire somebody's awkward like teenager 'cause they would have done a better job. but well Marty Weiner: Yeah. Yeah, 'cause this is probably made in a weekend, but on stable diffusion no less. Zach: speaking of great uses of AI, well Marty you wanna talk about the three lawsuit? Marty Weiner: I love these. These are so great. so apparently an expert witness used ChatGPT to to write a defense for a a deadly explosion lawsuit I guess against 3M and they use the phrase show how 3M is zero percent at fault. So wha what is the phrase we're using? Do a what would that be? Do a defense. D do it not our fault. Yeah. Zach: D do a not our fault. yeah. Chris Slowe: Do it do a defense. Do a defense. Marty Weiner: but I don't know how that came out. I don't know if you have to expose all your embarrassments in the court case when they they come and see why why your legal briefings are so bad. Zach: I mean i in in my experience when this stuff gets exposed it's always really obvious. probably there's a lot of other stuff just slipping through and we don't see it. It's like the toupee effect. You know, you only notice when it's bad, but yeah. and then Marty Weiner: Right. Right. it's just just embarrassing. Zach: just to to Chris Slowe: Well I think I think the the worst the worst part about this one though, honestly, is like I don't know why they would be keeping their receipts like this. It's like if you're gonna if you're gonna ask an LM if you're Marty Weiner: Yeah. Chris Slowe: gonna ask an LM to do your job for you, why don't you do it on your like personal account and then don't take Zach: Press the delete button. Chris Slowe: screenshots of it? Like just you know, just pretend I totally Marty Weiner: Ask Chris Slowe: I came up with this myself. It's like yeah. Marty Weiner: Ask AI to clean things up, you know? Zach: Yeah, yeah. And it comes back and it's like actually it is your fault and then you're like, No, try harder. You can do it. You can figure it out. Marty Weiner: Right. Well well this is all very childish. Do we have any more ad you know lofty news highbrow news Zach: Yeah, I wanna I wanna take this somewhere I wanna s Chris Slowe: Some highbrow endings. Zach: yeah, so there was some good news. so you know Marty Weiner: Ha ha. Zach: there's this big concern now that AI is gonna take all the aspirational work. and w we found out this week. I'm Marty Weiner: Uhhuh. Like plumbing and you know Zach: just gonna read the wired headline. These quote masturbation consultants were hired to pleasure themselves with AI. so and the best part of this, so this was a company that I don't I don't know exactly what they do. I I don't know anything about that. I don't know anything. they hired 10 gentlemen to do this study Marty Weiner: Gentlemen, yes. Zach: where they had to use the device once a day. four of them dropped out. There was there was a high level of attrition. I d I don't possibly my my theory is they formed a union. they wanted better better benefits. Chris Slowe: Yeah Marty Weiner: Well and they were being paid well, right? Zach: They're paying no, they were getting paid two thousand for for once a day. Possibly the problem was they the the need to cut back like that. I can sympathize. No, but okay, so this this was an actual thing that happened. made the rounds of around the internet for obvious reasons. I was I was I was Marty Weiner: Wha wha which AI company do we know? W was this for chat was this for open AI? Just sorry, I'm checking. Zach: Are are you asking for the website? Is that Chris Slowe: Asking for asking for a friend. Zach: Yes. Marty Weiner: Asking for I I Chris Slowe: What did you say the requirements were of that job, Zach? Can you go into some details? Maybe provide a URL? Marty Weiner: yeah. Are they still looking? Zach: n All I'm say I I just I can't believe there was a forty percent attrition rate on this study. I I guess there was some actually pretty crazy requirements for like filling out forms and stuff. So maybe it was like I not worth the effort. Chris Slowe: I mean, you know you've you know you've had a pr you a problem with your business model. When you've Marty Weiner: Well you say AI to fill off the form, Chris Slowe: when you've alienated your masturbation consultants, I'm not sure what your business business model is. Zach: Consultant. my god. All right. Well, having now spent we Chris Slowe: On that note Zach: spent more time on this than we did on protein folding, so maybe it's time to to to say thank you. thank you for listening. if you have any questions or comments, please post them on on YouTube or wherever else we post. press press the like button, tip your waitress. thank you. This has been Last Humans in the Loop, which will keep being made until we don't need to make it anymore.