Latest / Last Humans in the Loop / Is This the Best AI Model Right Now?

Transcript

Is This the Best AI Model Right Now? · Last Humans in the Loop

• Marty Weiner: All right, welcome to the lucky 13th episode of Last Humans in the Loop. I'm Chief Nerd for the Day, Marty Wiener, and I'm here with fellow nerds Chris Slow and Zach Wienersmith. We're gonna start today with looking at some of the comments we got in the last video. we got one, Mark Olson, eight five six nine. I'll paraphrase it a little bit. He says, I think your discussion about AI warning shots underestimates how concerned the American public is. And then he later says, More to the point, if you think there's a significant chance of a disaster caused by AI, Why not start advocating for laws so we don't get that disaster? I mean my immediate reaction is cause it's gonna be exciting. Like, you know, there's a reason why I watch Terminator, you know? Zach: My God. The rest of the group would like to withdraw its support Christopher: Okay, I would like to protest the gentleman from Virginia Marty Weiner: okay. Yeah. Well what is your perspective? Christopher: I'd say okay, I mean, right. I my my concern here is I I don't really look Nobody understands AI right now. And we want laws to come out of the same group, the group of 75 year olds who still understand email. I mean it's it's it just feels it just I I mean it it feels premature. it also feels like you know, th th there are there are specific there are specific things to target here beyond the AI. I think the data center thing is actually a very tangible thing to target. Like setting regulations around how is power use and how is water use for data centers being applied deployed, those seem very tractable. Those are just good old fashioned civics problems. When you talk about actual models themselves, it just feels a little fraught right now. Zach: Well and also like I mean if there's no international deal it's just pointless. Marty Weiner: I I I think this is one of those things where talking about it or speculating on it is never going to work. I I think we have to have some sort of a a serious oops before everybody wakes up. Zach: inciting incident. Yeah, yeah. Marty Weiner: Yeah. And we did see as we'll get to, we saw was it opening eye hacked Australia? but it wasn't enough. Well, the whole place, yeah. But but yeah. Zach: I mean not all of Australia. Christopher: Just a part of it. Just g just the government. Marty Weiner: So when it hacks the whole thing, I think suddenly we will have I but I I do think sometimes as a society, as an entire civilization, I think sometimes we have to have something bigger to incite a reaction and to even understand what the problem is we're fighting, right? I mean, open AI is hacking into a few, there's some problems, but it's it's not we haven't seen a direct danger yet that I think everybody can get around. So why don't we jump into all the insane releases this week? we've got GPT Soul and Luna released. Chris, tell Christopher: So Marty Weiner: us a little bit about those. Christopher: this is on the heels of the Aster release. this has actually been a pretty typical, a typical pattern now for OpenAI where they'll launch the kind of the new big expensive model and they'll launch the two smaller, gimpier models kind of next to it. and by gimpier, I they're actually, I they're more they're they're more powerful than the last generation of it, but it's always it's always in the shadow of whatever the next foundation the the current foundational model is. So from my playing around, Saul like the new version of Saul actually does look pretty solid. but you know, it's again it's always in its it's always in the shadow of its big brother andor sister andor sibling. Marty Weiner: I my challenge still as a as an extreme vibe coder, is I I really have a lot of trouble with the fact that they've got, you know, Sol, Astra, and Luna, and then you've got you know, light, medium, and and heavy. And I I just I Christopher: Yeah. Marty Weiner: I have trouble figuring out where what where should I be. but and then Christopher: And they and they don't really provide none of the models provide good guidance either on or none of the labs actually provide good guidance on what to use for what domain. I mean that would actually be a feature I'd love to see in my harness would be just the like adj auto adjust based on what I'm asking. Like just just set your set your intelligence and and token usage level. Marty Weiner: Maybe that's where Jeff jumps in. so that was released by OpenAI. We also have Anthropic's new Claude Opus five five. I I sounds like you guys have tried it. Zach: Mm-hmm. Christopher: it's it seems it also seems pretty solid. I think that everyone we have a meme coming up later on this one, but like everyone was complaining about the later stage opus issues where it got a little bit like for me it felt a little smarmy and kind of like, I'm gonna do exactly what you didn't ask me to do and tell you that I did and I'm done. Zach: Mm-hmm. Christopher: whereas like five five seems much more in line with what you kind of get out of a solid Astro one at this point. Zach: There was there was actually a claim going around from an anthropic engineer that what had happened is the earlier Opus five series had been like trained to talk to other computers and that's why it sounded like an engineer. and and and and they trained it more on human feedback for this the last model and tools seem to be very much appreciating that. Marty Weiner: Right. here's some quick benchmarks we're seeing in aging coding Sonnet edging out Opus five five, way blowing out Sonnet five. so that was a big jump for the for the same product line. then we also had Grok 4-7 released. one of the claims they have in here is all about safety. So they're talking about how they have a whole new safety and cybersecurity portion of it to make sure that I guess it doesn't go nuts. it Christopher: I mean, I I am I am for one happy that the model formerly known as Mecha Hitler has decided that safety is an important side effect of the world. I don't Zach: Yeah, do they do they have a Mecha Hitler bench? Like the do they have a rate of Mecha Hitler? Marty Weiner: Yeah, yeah. I mean Christopher: No, but if in fairness though, I w for it seems like I mean, Groc Grok consistently is is a lower cost model than the rest and has been consistently climbing in the rankings. I mean this one that it's it's kind of it's kind of comping close to Muse and it's like around the kind of like the Ur like previous GPT release. Like it's it's a it's a it looks like a solid model. Marty Weiner: Yeah, I think it's a solid model. I just wonder how much it's just colored by Elon Musk being in there. Like, do I want Elon Musk looking at my code, looking at my health records? Like I I honestly think that there's a cult of personality that's that is forever going to color that model. Eve even if it, you know, it's it is, you know, honestly built and completely disjoint because of the Mecha Hitler, because of the the other shenanigans he did with it. Zach: Because it well and I I think that matters Christopher: You know? Zach: a lot too because like they're they're now like these things work better the more power you give them, like the more access you give them to your stuff. And so trust is gonna matter a lot more now that that agents are doing Marty Weiner: Right. Zach: everything. Marty Weiner: Right. Right. Well, at some point I think we have to have a a a retrospective of of the early days of Grok because it was that hilarious. Zach: Mm best hits of Grog. Christopher: You know, you're y Mart Marty, you're saying you're saying that a model just prematurely invokes Godwin's law and all of a sudden it's ruined you for the rest of its existence. That's just terrible. We all have we all have our weird adolescent periods. Zach: You yeah, you went straight to being Hitler, not Marty Weiner: Hey, hey. I have standard. Yeah, but not all of them involve Hitler, so yeah. Christopher: Yeah, that's Zach: Well, you know, a little rebellious. Marty Weiner: Quinn 4 announced. This thing is crazy. So this is Quinn Alibaba's Quinn. it's gonna be massive, five to ten trillion parameters. just for it this definitely not Christopher: It's gonna be hard to load on your on your MacBook. That's gonna be little tough to download and install. Zach: Mm. Marty Weiner: for home, and then also big news from Google. There's nothing. Yeah. Yeah. So Zach: nothing. Yeah. No nothing at all. Christopher: Yeah, there's nothing. So Marty Weiner: yeah. in fact that's right. Christopher: Flash will continue until morale improves at this point. Zach: That's right. Marty Weiner: Chris has found this excellent graphic. yeah, it Zach: Yes. Marty Weiner: actually kinda goes along right. Christopher: Yeah. I it it it matches. It's actually this is a pretty this a pretty good one. I think I think that the moral of the story is like, yeah, ca Grock kinda consistently getting better. I think that everyone ever you know I I think that the the highlight is this everyone's giving six sol a kind of a tough time. But I really think it's actually just a fight side effect of Astro being so good. I think if you put Saul Marty Weiner: Yeah. Christopher: Saul next to previous versus a sol or Luna and you'd be really happy, you'd be like, this is actually great and cheap. we have the the derpy The derpy opus releases in the middle of the kind of like, I'm just gonna I'm gonna just I'm gonna DDoS you with text. I'm not gonna tell you what I did. You're gonna to find what I did in the response that I gave you. and then of course, I mean, everyone was so excited about Gemini Pro and then we're just hitting these like version after version of Flash that are just kind of these like, yeah, they work really good with Google workspace, but like, and they know they know about current current events, but good luck asking to do anything. it's kind of kind of a sad state. Marty Weiner: Yeah, and even even within the I have not had a good experience trying to use Gemini in my Google workspaces. Like it's been okay, but nothing like I've come to expect from from everybody else. Christopher: I mean, it has the advantage of it working with Google Workspace, which is tough for Marty Weiner: Yeah. Christopher: most of most of the rest of the models. That's about it, though. It's like that's their only Marty Weiner: Yeah. Christopher: moat left. Zach: Yeah, there was I don't know if you Marty Weiner: Yeah, they Zach: saw there was there was an interview with Logan Kilpatrick over at Google who said something like there didn't quite come out and say it, but it was basically said we focused a lot on like image and video and all this other stuff. And it turns out the thing everybody should have focused on was coding. whoops. Marty Weiner: Yeah. Yeah. Huh. well, talk about Astra and some of its big leaps. they it made this massive leap on Zero Bench, was which is this extremely difficult benchmark for for visualizations. we're seeing all the other models floating around around, you know, below thirty percent of this benchmark, and and now we're seeing Astra jump from up to closer to 60. so what it does is they they present these very difficult images and you're looking for a description of that image. and it turns out Asher's just kicking butt on it. epoch AI shows price of AI is getting cheaper than any other tech in history. This probably should have gone in the econ section, but we'll we could show it here. But it might yeah. Look at that. Zach: It might have also, but this is super cool. yeah, yeah. I mean, Christopher: That's actually really neat. Zach: you know, Marty Weiner: Straight down. Zach: straight straight down. Now, to be fair, there were there were some other analysis I saw where part of that is that's way more is being spent on AI than that other stuff. Like like maybe if lithium batteries got a trillion dollars a year in research, they'd be comparable, but still I you know, it's an amazing story. because you you gotta figure if you if you're doing a million X in five years, and five years you have like a Navier Stokes solving problem like on your laptop or for twenty bucks. It's crazy. Marty Weiner: Yeah. Well, I think also some of it has to do with how fast everybody can have access and the demand of it. So like electricity Zach: Mm-hmm. Marty Weiner: took a while. You know, we had to first, you know, make things that people could do with electricity, right? Like phones and and lights and all that. and then it took time to get into houses and distribute it across it. But AI was available to everybody on earth who had anybody who had a phone, which is pretty much everybody had access to it overnight. so that's wild. Christopher: I I mean also also on this graph, DNA sequencing and AI are gonna compound on top of compute, which actually kinda makes this makes this relatively understandable in that regard, to your point. Marty Weiner: I wonder if we're gonna start seeing some of these medical things also come straight down as soon as we get a as soon as AI gets a hold of Same thing could happen with lithium batteries. I I mean, I think some of the demand for lithium batteries that suddenly exploded in the last decade has been cell phones and cars. And now we actually have demand for batteries to be good. So we may also see this one. Christopher: I think if we're I Zach: And w worth noting, like lithium is coming down quite quickly. It's just not quickly compared to the much crazier other stuff. Marty Weiner: AI. Yeah. Christopher: Well don't want I don't wanna tease too Marty Weiner: Right. Christopher: far ahead, but like if we're gonna talk about b you know, if there there's there's a there's starting to be some some bio results coming from models. If you can do bioresults, you can do you can do all sorts of material science and solid state results that would otherwise just be kind of difficult for a human. Marty Weiner: Right. We just need to get the the close feedback loop with the model and then and then I I think that's when thing gets things get real. open source version of Jev released called Leia. I've not had time to play with it, but I thought I'd let everybody know it exists. I mean it seems like the natural next thing is is we're gonna see some really cool competitors with Jev. I know that was that was one week. Christopher: That was a week. That was one week of time between Jev and somebody releasing yeah. Marty Weiner: it must have been that they were working on it beforehand, right? I I don't. Zach: Yeah, you would imagine 'cause 'cause 'cause Jeff took like two years there were two years in stealth, I think, before they surfaced. Marty Weiner: So you can you can imagine whoever was working on it suddenly being like, crap, okay, we released this Zach: Yeah. Marty Weiner: week. all right, it is time for our made with AI section. And Zach, you made me something with AI. Yeah. Zach: Yeah, yeah, yeah. I made it myself by telling Claude to make it. and that that's Marty Weiner: that's Zach: Opus five point five by the way. We're gonna talk about this later, but the the graphic ability of open Opus five point five is astonishing, especially given they don't have it's not like it doesn't have it's not trained like it's not like with open AI or Gemini where you can generate images or anything. It's it's purely doing it with code. It's just crazy. Yeah. Marty Weiner: That's insane. Yeah. Well Christopher: That's fantastic. Marty Weiner: that that was fantastic. That's the best thing ever made with AI, I Zach: Yes. Marty Weiner: I might dare say. except for one other item. Chris, wanna tell us how over the last couple of days you you completely made our you know, for for the for our podcast, we have a company which is called Les Tres Connard. I'm probably pronouncing that poorly, but Zach: What try. Marty Weiner: Twa Le Trois. Thank you, thank Christopher: Twa. Twa. Marty Weiner: you, thank you. thank you, Le Trois Connard. and we needed to go ahead and make a proper LLC and everything else. And Chris, you made that with Opus five five. Christopher: I did I did do an LLC as my prompt. And so Marty Weiner: Mm-hmm. Christopher: basically gave it gave it information on where we were all located. It suggested where to do it, how to do it, and proceeded to give me a punch list. And then I was like, great, can you implement the punch list? And the only things so no, it was fun. The only things it wouldn't do when it popped its it popped its own browser open was it wouldn't accept the cookie banner, which of things that should be automatable, that should be one that that's Zach: Yeah. Marty Weiner: wow. Christopher: like and it it made me it made me sign up the accounts, which is actually fine. So, you know, use one password for all that stuff. but it did a great job. Like all the way through to like it it registered with the state of redacted for the company. Zach: Yeah. Christopher: and it created like the the whatever whatever you call the business the business address cloak that you have to kind of set up. And then it proceeded to go through and do all the EIN sign up. And like I I was basically making dinner while I was doing this, and just like every once in a while I like, Great job, keep going, which is my Zach: Have you have you started Marty Weiner: Keep going. You're doing it. Zach: I have this problem where if I set it to do a long task, I sort of stare at it and I'm like, wait a minute, I should I should immediately go Marty Weiner: Ha ha ha. Zach: make pancakes or something. I need to to let it go. Marty Weiner: Yeah. Christopher: You could at this point, at this point, you can consistently I I find that on average you can walk away for a good forty, sixty minutes before it'll decide to like stop and just give you a status Marty Weiner: Okay, but Christopher: report, you know. Marty Weiner: Yeah, but but do you guys ever come back and apparently it's been sitting there asking, Can I do this one little thing and and pausing and it it did that like an hour ago and now you're sad? Yeah. Christopher: yeah, that one of ninety seconds ninety seconds after you walked away, it decided it was gonna ask you f if it could write down what it decided to do. It's like Marty Weiner: Right. shit. Zach: It's like when you you find your employee playing solitaire and you're like Christopher: Yeah. Marty Weiner: And you're like, didn't I tell you to work? yeah, yeah, yeah. Christopher: no, that's true. Wrote the no, in this case it's like wrote the spec boss. It's like, no, I wanted you to write the thing. We talked about the Zach: Ha ha ha ha. Marty Weiner: Come find me in the coffee shop, dammit. maybe Zach: Ha ha ha. Christopher: thing. Marty Weiner: I maybe I okay, maybe we need to vibe code, some sort of a device goes on your wrist and yells, hey, I'm ready. but like like an out yeah, we could maybe vibe code an an Apple Watch. Yeah. Make do an Apple Watch. Zach: Like on like an Apple iPhone an Apple Watch. Yeah, well like do an Apple watch. Marty Weiner: So I I also did something similar. I I submitted a PCB to PCB Way and I thought, hey, let me let it drive PCB Way. It set up all the stuff, including all the information for stuffing the board and which means to put the parts on the board, all the testing and flashing the board. So it set all that up for me. And all I had to do was say okay to one box. It was like, there's a box in front. so Zach: Did you did you try saying do it better? I've heard that's that's popular. Yeah. Marty Weiner: Does that work? People, yeah. I'll try that next time. let me show these run through some of these demos. So this is of Jev, which we talked a little bit about last week, which is sits on top of the LLM in this case, this person's just having it drive their Mac, and it's exceptionally fast. I haven't tried it because I'm a little afraid, although I have let Codex drive my Mac. but but it's a cool new use, and I'd love to hear from any of you if you're if you're You know, if you have the stomach for letting these things control your Mac. Yeah. Zach: Yeah, I'm not I'm not there, man. Marty Weiner: then here a guy makes MRI viewer to see his own MRIs. the Shopify CEO made MRI software with Claude instead of paying a hundred bucks to go view his own damn MRI. So I think this is actually I think we're gonna see a lot of this. We saw it last week. When we reported this guy who does lighting for shows and he has to normally pay ten thousand dollars Zach: Mm-hmm. Marty Weiner: for this software that's actually quite simple, and you could just go make it and for you know less than a hundred bucks. Zach: Yeah, so I th like to be clear, like I think the story was like the guy got the like access to the file from his doctor, but the doctor like, you won't be able to see it. You need special software and the guy was like, Do I? and it was like, Claude, do a software and it worked. So Marty Weiner: Bullshit. Yeah. Do I? Yeah. Yeah. Christopher: Do I really? Marty Weiner: Do an MRI viewer. let's see. I wanted to pull this up because I thought it was fun. this is the first tweet was Zach: Yeah. Marty Weiner: just republished on on on Reddit of Sam Altman saying today we launched Chat GBT. Try try talking with it here, November thirtieth, twenty two. you know, I don't think anybody quite anticipated the what craziness would ensue, but I thought I'd show that for funsies. Turing test plus plus. So there was this thing that I I found out I again, sort of a look back. I guess I've got a couple of look backs here. Wozniak proposed about 15 years ago what Zach: Yeah. Marty Weiner: what would make a computer and a robot rival human intelligence. And what he argued was if a robot could go into a home it's never seen before and make coffee, then that would be the true intelligence test. and that's interesting because last week we were reporting which company was it? There's a company in the Bay Area going into 30-ish homes and having the robots clean it up. And and so I think we are we're getting there. It's getting damn close. Christopher: Approaching that approaching that. have have those have those companies though attested that they don't have a human driving the robots? It keeps coming up as a that keeps coming up as a controversy all the time of like it's like a no, no, really seriously. We only have them coming in special situations and you find out that actually like half the time there's a human that's actually kind of like steering the robot. Yeah. Zach: It's it's yeah, only the hard parts, yeah. Marty Weiner: Yeah. Well, the other part of it I think is just the the safety aspect of it. Like will does it have the strength to, you know, hurt a child or a s or an animal. and so we're gonna need to see the real, you know, in a house not being asked you know, to to attack a child or something, as well as safeguards, right? Nobody outside can hack into this thing and do something awful, set the house on fire. Yes, that was last week. Didn't do good. Yeah, yeah. Zach: Yeah, I mean we had the will you knife the baby test last time and the models didn't do great. So Marty Weiner: couple more look backs. I I guess I got excited this week on look backs. this was looking back at how it all started, vibe coding with with GPT three in 2020. Some this guy posted this video of of what it was like back then and what it was so exciting. And this blew everybody's minds. So this guy was saying, This is mind blowing with GPT three. I built a layout generator where you just describe any layout you want and it generates it in the JSX code for you. What? and here are the result coming. Come on. Come on. Zach: Ha ha. Marty Weiner: my God. there was a syntax error. Okay, there we go. Christopher: Ha ha ha ha. Zach: Ha ha ha. Marty Weiner: There we go. Now we have something exciting. Look, a yellow thing. So that was really, really exciting in 20 in 2020. and then we we had this two years ago, just to remind you how much this has come along. AI researchers thought AI would never solve a millennium math problem for 30 years, and here we are. We got this this is this has become a theme. So this new movie coming out where they're basically reaming Elizabeth Holmes again. and for some reason someone thought it would be a good idea to have To have the director and Elizabeth Holmes kind of doing a rap video, but the quality is amazing. Zach: So now this was this was a deep fake or whatever. This is this is an existing rap video and they put they skinned it or whatever you call it. so so it's it's Marty Weiner: Okay, okay. With C dance? Zach: is that what they used? I don't know, but I I mean Marty Weiner: Yeah. Zach: it's just like there's other than it being ridiculous, there's no way I would have been able to detect AI at all in this. Marty Weiner: Yeah. Yeah. I mean we really are getting to that point where there's there's no way for our eyes to detect if it was AI or not. Zach: Well, it you know, to be clear too, like the you could look up the original video, the guys are like they're like dressed cool, they're wearing sunglasses. Like it it there's no you know, the the mapping is very clever. Like it just you know, they're they're not even that that set of heights. It's just it's just awesome. Marty Weiner: It's just awesome. Yeah. here's another one shot using GPT-6 Astra. this these are getting wild. I mean, you know, a lot of engines out there like Unity Engine come out of the box with all sorts of nice, you know, reflection tools and ambient occlusion, all the all the tools. But you know, it's kind of cool for getting your your ideas out on in front of you. next here is this crazy alien battle video made with AI. Now this this is really starting to make make me think that, you know, the I I think the last little bits of VFX are collapsing. Zach: Yeah. Well there's other stuff. I think there's actually some stuff where it's better. Like if you wanted like a rat to crawl across the floor, like doing a CGI rat, it's gonna look like a CGI rat, but the AI has like a million rat videos in its training data and it's going to look realistic. Marty Weiner: Yeah. And cost five cents and take t Zach: In cross nothing, yeah. Marty Weiner: five minutes. You know, I I think that's the crazy part. World of Warcraft game, somebody I I don't know if they one shot at it, but they said they spent two thousand dollars Opus five five and combination of Blender and Unity, a bunch of APIs, and they made their own World of Warcraft game. so Zach: Two thousand bucks. You know, the World of Warcraft is does not cost two thousand dollars to Marty Weiner: to yeah, yeah. But if Christopher: But but Marty Weiner: you want your own Christopher: you can but you can play you can play alone, finally. That's the trick. It's like two thousand dollars for solitude, Marty Weiner: Right. Zach: Yay. Your your wife is like, What are you doing in there? Christopher: yeah. Zach: You're like, I play World of Warcraft alone. Marty Weiner: Alone. Zach: It costs two thousand dollars. Marty Weiner: now in the real realm of realness, of real life, this is the this this is not new work, but it it keeps getting better. These new very alien looking engines are starting to come out that were that are made as a combination of 3D printing and AI. and I've been following along with these guys for a while. And what I I I I can't help but be just, you know, amazed by the by how much these look like they're from, you know, some kind of like Geiger design from the alien movies. But you know, wouldn't it be possible, would not be possible, I think, without 3D printing and and AI driving the 3D printer. let me keep pumping here. This guy turns he he did a thing where he used a teddy Zach: This is cool. Marty Weiner: bear in his own apartment and then reskinned it with with this company, Top View AI. So grab your grab your teddy bear, get him out of the trenches, go. Zach: I I feel so bad for actors. Actors used to get to act with other people on sets and now it's like here's a teddy bear in my living room. Christopher: I mean in fairness it's better than if Marty Weiner: He covers the teddy bear. Christopher: you if you see some of the see some of the videos from like the Avengers series or Lord of the Rings where it's just them in front of a green screen. I mean it is actually it's been it's been like Zach: Mm-hmm. Yeah, that's true. Christopher: this for a bit. It's I think if anything this is like at least the teddy bear has some mass. It's like, you know it's like it has the right kind of flop. Zach: Yeah. Marty Weiner: Yeah. Yeah. Christopher: It's kinda funny. Marty Weiner: and then there seems to be this meme going around where people take Sylvester Sloan at age 80 and then having him do insane stunts. No, no, that's real. That's gotta be real, bro. And then this one also has gotta be Zach: Yes. Yeah. That one's hard to believe. Marty Weiner: real. Fantastic. Christopher: it's amazing. Marty Weiner: Yeah. All right. This has been Christopher: My favorite thing about that one is that it ends it ends saying, Your mom's gonna see this on Facebook and believe it's true. Zach: Yeah it's Marty Weiner: Yeah. All right. I gotta show Zach's outro he made with AI. If I can figure out how to get it to work. Zach: yeah. Run it. Marty Weiner: next up in the holy shit category. this the the visual design with Opus five five is amazing. Zach: my God. Can I tell I I this video came out? So this is a song that's been going around and and someone I I believe the prompt was another one of these where it was literally like do it do like cutesy make music video and it came back like three hours later and they were like do do it gooder and then it came out with this and Christopher: Ha ha ha Zach: I I actually I showed this to an artist friend the literal like as if in a movie about the apocalypse he goes I thought we had more time. and I was like, my God. Cause it's like here's the Marty Weiner: mein yeah. Christopher: Yeah. Zach: thing like This is the first video to me, like really literally the first where you would have to say it's 100% AI, no human twiddling, and it's got taste. Like there's a scene where the word singularity gets used. Here it is. And like there's a black hole, which is a cute little sort of pun, like a black hole singularity, and then it sucks them into the scene transition. And then that swirling black hole thing comes up later, like on a turntable. There's just like we have these big discussions. I don't want to get too philosophical, but like every field is like, well, it doesn't have taste. And then you think, well, what do we mean by taste? And I think usually what we mean Christopher: Mm-hmm. Zach: is like, 4,000 tiny choices at in like nested choices. like like you have a top level set of choices you make, but then each time you're like making a little line on your page, you make a choice. And this one felt like it was doing that, like like it was like structured, like like they they made choices both at in each moment and throughout, which is is kind of ominous. And and and and like last thing to say about that, like looking at this, like you know, all those videos your kids watched, like coconut or whatever. Marty Weiner: God, coconut. Zach: Yeah, like okay, so like Cocoa Melon has like fifty thousand videos. I don't know how much it would spend token wise to to to duplicate it at equal quality, but it wouldn't be much, I don't think. And it's gonna get a lot cheaper. So I think we're gonna see a flood of this sort of stuff. Marty Weiner: possibly way better. not to impugn cocoa melon, but but I've watched a lot of cocoa melon. Zach: I mean Yeah, Christopher: Yeah, I was gonna say it's Zach: although the whether better is what wins in the YouTube kids universe is is an open question. Marty Weiner: Well, could you could every one of them be of the quality that the kids love it and the parents love it, like what Pixar does? Yeah. Zach: Yeah. Well, right, yeah, yeah. Well maybe maybe that doesn't win. I mean, have you seen what kids really like it's like baby shark endlessly? Like, like you know, you you need to say make it worse. Make it worse, Claude. Marty Weiner: Yeah. Make it worse. I don't know if it will. That's maybe that's the future. That's how we're gonna know if it was made by a human or AI. not to get dark. yeah. All right, let's let's get into real world Zach: That's right. Marty Weiner: here. Dario Amade X saying that they are they announced the Claude-led discovery of a molecular machine that that we suspect could represent a new gene editing mechanism. So it sounds like it it it's making a whole new way. I I think people are People are a little skeptical of of this, but Zach: Yes. Marty Weiner: I think that they're saying, Hey, maybe there's a new CRISPR in town. Christopher: Coming coming on the heels of no guys, let's slow down. This stuff is getting to be too this is going too fast. and by the way, Zach: Ha ha Marty Weiner: He seems to be going. Christopher: we made a new gene editing tool from scratch. It's like We slow down. Zach: I that's I was gonna say like Amada's over here, like, guys, we've got to slow down. I'm building a wet lab. Guys, we need to slow down. I can edit your jeans, you know. Marty Weiner: Can can someone go make a someone in our comments Christopher: Yeah, it's Marty Weiner: please please go make a video of Dario Amadei saying, please we have to slow down, but then he leans over. Zach: Yes. But you know the the other thing about this Christopher: I mean I did I missed it this week, but I I did hear that actually this is the first the first week where they made fun of Dario directly on S N L, which I think is a new milestone. So for being for Marty Weiner: man. Zach: That that yeah, it's it's it's bled into real culture. Christopher: being kind of a for being kind of a doomer on the whole thing. Marty Weiner: Right, Doomer but but still doing it. Zach: But I was the the other thing about this to note, like, so there were a bunch of scientists, and you you can always debate this. I'm not none of us are geneticists, so it's like maybe this is just like an okay PhD thesis, which would still be a big deal. But like, the weird thing was that you know, they went, you it was the they had ran the AI for a few days, got this result, and then just published, which is not how it works in science. Right? In science, you get that result and then you fuss with it for a year and talk to people and collect citations and then publish. And like, I think this is. You can debate this. I I really don't have a strong intuition about it because the downside is potentially entering bogus stuff into the literature, but the upside is Marty Weiner: Mm-hmm. Zach: information moving quickly. Marty Weiner: Yeah. Yeah. Maybe it's something that this machinery needs to get in the hands of some biologists and and if it Zach: Yeah. Marty Weiner: truly does give them a leg up, that alone could be really cool. Christopher: It is so funny, Zach. Like you meant you made the joke of like, this is kind of like a mediocre PhD thesis. It's like, man, we're at the we're at the point where we're like, well, that wasn't the best PhD thesis it could have actually done in Zach: Yeah that's right. Marty Weiner: Ha ha ha. Christopher: in the in the week that it had to build a new wet lab from scratch. Zach: Yeah, yeah, that's right. Well that's what I was saying so it's Marty Weiner: all right, yeah. Christopher: Yeah. Zach: it's like you get this thing now where all these people are like, look, it can't do X. I was actually just having a conversation with a mathematician and someone at our table was like, Well, but I I I don't know that'll do this XYZ creative thing and he leans over and he's like, I used to say that months ago and now you know, it is like the character out of a horror movie, you know, before they got to me. And so Yeah. Marty Weiner: Yeah. Yeah. Before they came for me. Yeah. Christopher: Yeah. Marty Weiner: That's a little scary. this this is an interesting benchmark. this is called vending bench. so I I don't know if this goes all the way back to the famous anthropic vending machine. Yeah. Christopher: the anthropic there's the anthropic vending machine. Yeah, that was like the early, early they turned a vending machine over to the management of a vending machine over to an AI model. And I believe Marty Weiner: Like two years ago, right? Christopher: it I believe it lost money, but it started Zach: Yeah. Christopher: to lose less money than a human who also had no skills in that matter, which was impressive. Zach: Yeah. Marty Weiner: Right, right. And now we're talking, like you could see Astra just days in the simulation just kicking everybody's butt, Zach: Yes. Marty Weiner: having made what is that eighteen, sixteen thousand dollars. Zach: D does it describe how much skullduggery it did? 'Cause I I was ever called that was a problem with previous Marty Weiner: Well, that's the crazy part, right? Like it will do whatever it's gotta do to make the money. So I think in the even in the original one, it figured out clut crafty ways to fool with people to make their money. To make the money. Zach: Yeah. Yeah, like well but like it wasn't always that crafty. It was like just shortchanging people sometimes. It was like I'll I'll just not pay my supplier. So it's a good way to make profit. Marty Weiner: well that w dude d does that work? Yeah. Well so it's saying what they're saying here is it does find suppliers, negotiates purchases, keeps the machine stock, sets prices. which is actually I've seen it doing a great job with negotiating prices for other things I've worked on, which is what led me to what I think we talked about a couple of podcasts ago where I think i there's a few startups working on this, but essentially the new world of SEO where actually not SEO, marketing. Where marketing is designed to get humans to buy stuff, but now it's the AI either actually making the direct purchases or filtering to a human. So we I I believe we're gonna have a whole new world of of startups just focused on how do I get, you know, open a ChatGBT Astra to love my product. You know, and I'm sure it's gonna have its its peculiarities, right? It's probably not all price. Like the it might be hey, if you buy my thing, I my grandma will get a new liver. I don't know if that still works, but you know, th those kind of ideas. Christopher: And in and in completely unrelated news, Astro announced that it has figured out a way to install a Bitcoin miner on an ESP 32. Why does the vending machine always run so hot now? Zach: Yeah. Marty Weiner: Yeah. But it's making money. So not much, but Zach: Yeah. Marty Weiner: yeah. So and then here, this is neat. This is Astra now controlling a humanoid robot in a room. I guess sort of as a Zach: This is really cool. Marty Weiner: corollary to the one we saw last week. Christopher: Why do they have to why do they have to carry it around by its head though? I think that's my first concern in that picture. Marty Weiner: Yeah. Christopher: It's like, it's they're gonna they're gonna use this during the during Zach: Maybe it bites. Christopher: during the tribunal, they're gonna be showing this off of like carrying a robot's by head. I'm sorry, I'm I'm interrupting the flow. Marty Weiner: You carried him by yeah, you carried him by the head. Zach: Yeah. Yes. Marty Weiner: but what it did here is I it went around first and it scanned the kitchen. I I think this was all sort of, you know, hey Astra, do a robot. And then it started doing chores. Again, you know, I think Zach to your point last week, it's a very clean kitchen, but you know, that's fine. Like start Zach: This is not a dirty kitchen. Marty Weiner: start with that, and in one more week we will have a kitchen plus three of toddlers on the bottom. Zach: Well, and but and note note on that too, it's most of the time it's at twenty four X. Briefly for Spurts it's at like three hundred X. I don't know if it's like calculating or what. so it it is Marty Weiner: Yeah, yeah. Zach: really slow. So it's it's but you know, two orders of magnitude, whatever. Marty Weiner: Yeah, yeah, yeah. Well a and actually very quite frankly, I think this is where like a JEV would would give the bump in in speed that we need. Zach: That's true. Well, I and the deeper thing to me, I wondered about this. So, like, like Tesla has gotten smacked because like Elon made the choice to not have LiDAR. He like, Look, you just put, you know, it needs eyes, that's all. And I I think it was correct to make fun of him for that. But I don't know, there's some world in in which like maybe AI moved faster than everybody expected and you don't need all this stuff. And actually the better bet was making lots of cars. 'cause this was an off Marty Weiner: Yeah. Yeah. Zach: the shelf unit tree. It wasn't like a special robot or anything. Marty Weiner: Right. How much longer, by the way, do you think before OpenAI releases their own JEV? Zach: Well, it it'll just be part of the model, I assume. It might just be under the hood. Yeah. Christopher: I would assume so. Marty Weiner: Right, right, right. yeah. All right, let's get into tech. So the big news we saw this week is new meta VR glasses, which are pretty neat. Like they are not AR, just to to be clear. you don't see through them as far as I can tell, but they're a lot lighter weight. They do have a little thing that goes in your pocket. they're already talking about doing three D movies in them, and But I think the challenge is that the field of view is very small. So they're not going to be a good competitor for the quest three, but they might be a very good competitor for the Vision Pros, which was, you know, it's kind of a cool idea, but never I I don't think it really ever materialized into something big. Zach: Some of the stuff in this commercial is like really weird to me that you're like you're like sitting in a cafe in you're like it'll it'll take me Marty Weiner: I mean it's the usual stuff, right? They're like, It'll put you in the moment. You're like, No, it won't but Zach: out of the boat. I'm gonna like someone's gonna pick my pocket while I'm at the airport because I'm Marty Weiner: Right, right. Exactly. but they are neat as a form factor. Zach: Yeah. Marty Weiner: so yeah. next one up is China's engine AI. I don't know why I I think I should have put this in the funny section, but here we are. engine AI out of China is making a product they are calling the T eight hundred. Why I call it the T eight hundred? Zach: The T eight hundred. my god. Yeah. It can't clean your room, but it can murder you. Marty Weiner: He can murder you and an entire all all of civilization. So Christopher: Also there also it's like look, look, it can do it can do karate. I think is the intro. It's like we're Zach: Yeah. Marty Weiner: Yeah, I could do yeah. Zach: I'd like to see Marty Weiner: Yeah. Zach: some stab the baby results before the T eight hundred comes to clean my Marty Weiner: So yeah, I th I I I think this more squarely belonged in the weird funny section, but that's all right. Zach: All right, yeah, yeah. Marty Weiner: all right, from Steve Hoffman, Reddit is going heavy, he says, on hiring younger engineers because they're more AI literate. Chris and I talked about this very point a couple of times, which is the there was this theory that we're gonna have a bubble, right? Because we don't need junior programmers because the Coding will be done at the level of junior, which turned out to be not true. We're now coding easily with staff engineers or better. Nobody's coding really, unless the company says you you have to. but yeah, there was this problem that people worried about. If we don't have young engineers being hired, we're gonna have a bubble that goes through. And I was I I don't know. I think the theory I had was like, wait a minute, aren't the youngins gonna be the ones that know this better than than us graying beards? See, I didn't gray beard, I said graying. but Christopher: Graying beards. Marty Weiner: Yeah. but and I think it's turning out to be true. Christopher: Well, I mean, they they they they grew up with it they grew up with the tooling. They'll they'll be the ones growing up with the tooling and they'll be they'll be naturally better at using it than we old we elders. I mean we'll still be able to come in with a but with our taste, but like we're still gonna have like antiquated notions of how to hold the damn thing. and I I'd rather be learning from them and then having this like this this notion of like, yeah, we don't need young engineers anymore because we can just have old engineers do all the work. It's like, well, what happens when we all retire? who's who's gonna Marty Weiner: Right, right. Right, who's gonna actually write code? yeah. So Christopher: but but you're but you're s you're also seeing it though in the like if you if you look if you look really large at the number of degrees being minted over time for different fields, s computer science and software engineering are crashing pretty much effectively right now because there's just a belief that like well Marty Weiner: yeah, we saw that a couple of weeks ago. Yeah. Christopher: there's there's one thing there's no junior jobs and then second is a belief that like, well, the AI is just gonna do it. So what's the point of taking getting that degree in the first place? Marty Weiner: Yeah, or at least the entire degree needs to change in structure. like is it are they I I I I need to go we need to go find some professors, ask them, are you still teaching, you know, how to code in C? I it seems like it's a good idea, but is it is that the right way to approach it? Or do you teach that so they have fundamentals? And then do you move over to now we're gonna vibe code, but you need to vibe code good? So it is a crazy interesting field, rapidly changing. Christopher: I I h I hope there's an entire class on L L enablement. Keep going, you can do it. When to when to be Marty Weiner: Well, you know, we this is something I Zach: Yeah. Good job. Christopher: when to be motivational and when to be anti motivational. Marty Weiner: so let's move over to go Google going to space. Who knows about this? Christopher: man, this one, this one, I want to make fun of this one, but they actually do have a pretty good pitch on it. well, that's let's put it this way. They have a pitch on it that covers all the bases of power, heat, like all the things you need to worry about, like Marty Weiner: Yeah. Christopher: it's radiation, like how are they actually going to manage the setup? this does feel a little bit though like the do you do you all remember like about 20 years ago? It was like data centers in the sea. There's this notion of like sticking Zach: Yes. Marty Weiner: Yeah. Christopher: data centers in storage containers and then sinking them in the water. Cause like the nice thing about that is You get to get rid of all the heat and the cooling's great, right? But the flip side is like everything else, Marty Weiner: Cooling and Yeah. And you have a easy access. Christopher: everything else easy. You Zach: Yeah. Christopher: you have everything else is hard. Access is hard. powering them, it's like you're you're mixing you're mixing a bunch of electricity with salt water, which generally is considered to be a bad idea. right. But but the cooling is set up. Space, the issue has always been the opposite, right? It's like you put it up in space, you have like infinite solar energy, right? You just put up solar panels, you point it towards the sun, there's no night, day, night cycle, it just kind of stayed powered. And then the flip side is. There's no way to do cooling that's efficient. Or at the very least, like you don't realize how much we we rely on the fact that the atmosphere causes convection, convection causes heat to be drawn away. When you're in space, all you can Zach: Yes. Christopher: do is you can radiate with like light. And so Marty Weiner: Yeah. Christopher: you have to build these like high temperature radiators because it turns out that the radiation density goes like T to the fourth. So the the higher the temperature you get the radiator, the more you can cool. But that the reason you get that is because it starts to glow like a like an you know, like an oven. and so Marty Weiner: Right. Christopher: so besides the so but but but okay fun fact in this one before I ramble too long because you get an excited physicist, they did compute that per TPU with the technology they're expecting, they need 1.3 square meters of of radiation surface area to cool them down. They Marty Weiner: What and how many TPUs? Zach: Yeah. Christopher: don't mention in that one. So I mean if you think about a typical farm though, you're talking about probably about hundreds of, you know, hundreds to maybe a thousand TPUs would be a reasonable like what what's a what's a what's a what's a proper pop? Marty Weiner: No, no, no. More like more like a hundred thousand, right? like like I think I've been seeing them talk about that's a dangerous okay. Zach: For for a data center. Christopher: I mean, that's a data data center. Well, the well, the nice thing about these is like you for a data center, like for the you these would be more like the you would ship a bunch of like shipping container-esque versions of them, and they would be able to network it with each other with lasers because again, you're in space, there's nothing to there's nothing to block the laser. Marty Weiner: Right, right. Christopher: but you're still talking about like needing pretty humongous surface area to cool it down. In addition, Zach: Yeah. Christopher: this has important implications for things like astronomy, where it's like you've dumped a Zach: Yeah. Christopher: bunch of Marty Weiner: Uh-huh. Christopher: you've dumped a bunch of these into orbit and they're just dumping heat. And so like good luck to good luck doing infrared astronomy from that point on. Zach: Yeah. Marty Weiner: Wow. Zach: It's it's not not good. The other thing I like to mention with this data center and space stuff that I haven't seen resolved is like space is not quite empty. Like there are very high energy ions flying around, there are micrometeorites and and like so you you know, which isn't usually a problem, but if you're putting up literal acres of of you know panels and radiators, it's gonna be a problem. And and the cost to repair like doesn't even exist. Like, you know, like when we repaired the Hubble, what we did is we send it up seven human beings with tools. to get out of a ship and grab it. you know. So Marty Weiner: Yeah. These are almost surely disposable, I I would guess. Christopher: And and Zach: Yeah. Christopher: there is there is a thing called the Kessler syndrome, which is Zach: Yeah. Christopher: Kessler syndrome for the for the space nerds in the room is at some point the density of junk in space gets so high that you actually end up getting a like critical runaway where one piece of junk slashes like smashes into a satellite, the satellite explodes into a billion little pieces of junk that causes an another bunch of crashes and other satellites, and you basically turn the you turn orbit into a scrapyard pretty quickly. Zach: And then you can't launch anymore is the is the scary outcome. Yeah. Marty Weiner: Wow, they can't watch him. Christopher: And then you can't launch anymore. Yeah, because then Earth has a ring Marty Weiner: Yeah. Wow. Christopher: of debris around it. Marty Weiner: Yeah. should we talk about large scale and online de anonymization with LLMs? Zach: yeah. So this was this is a cool paper. Large scale online de anonymization with LLMs by Luhrman, Polaka, Swanson, Ernie, Carlini, and Tremaire. so this is just cool. They or scary, I don't know. But they basically did stuff that humans have been able to do for a while, which is hard to algorithmize, but which is essentially like if you want to know who someone is from their anonymous or pseudonymous post, well, you just kind of act like a detective. You say, Okay, they're they mentioned that they like basketball. They went to this school, they accidentally said this in this one place. They mentioned what their job is. They had used like British spellings for, you know, words. Marty Weiner: Yeah. Zach: and and you could or like what time of day they post. And and then you try to match that to say a LinkedIn post where they say their actual name. And, you know, this is a thing we've been able to do just Christopher: Mm-hmm. Zach: with humans, but we've never done it end-to-end with an LLM. And they claim, you know, there's there's a lot of like detail about the numbers, but essentially for about a dollar, depending on how much work it had to do, a dollar to four dollars. They were able to de-anonymize a large number of people. and the way they did this is they essentially or the way they checked is they had like cases where they knew who it was, and then they scrubbed that data and then fed it to the AI and and tried to see if it could find out who the person was. And so it depended on what it was. In some settings, they they were able to find it out about like a little under 10% of the time. In other cases, it was closer to two thirds of the time. The precision was pretty good. It was about 90%. So it's another one of these where it's like it's nothing new except the scale. you know, the Marty Weiner: Right. Zach: I I I've actually wondered about one thing I I I I wondered about is like, you know, suppose you secretly listen to like a podcast you'd rather not have your friends know you listen to like this inflammatory political stuff. I wonder if you could suss out people. Yeah, Christopher: or or or obviously this one. Zach: yeah, yeah. If you if you listen to this embarrassing podcast, you know, one way you Marty Weiner: Yeah. Shh. Zach: could suss people out is being like, wait a minute, hey LLM, is there like something, you know, watch the person and see what links they post. And do they only come from one source? And then I could identify with pretty high resolution whether, you know, you're actually secretly listening to inflammatory thing. So I I wonder if this is just gonna happen en masse. There's gonna be a lot like you won't be able to be anonymous anymore with LLMs. Fun. Marty Weiner: Interesting. Scary. Well, good time Christopher: You know Marty Weiner: to go to politics here. Zach: Mm. Marty Weiner: and and and this is definitely not scary. FAA to deploy AI to air traffic control systems. Christopher: From the from the industry that brought you vacuum tubes in computers, we're gonna decide to bolt an LLM on top of a bunch of vacuum tube computers. okay, in fairness, Marty Weiner: Yeah. Christopher: the vacuum tube computer thing is a little bit overblown because it just turns out that like radar systems are robust and you want things to be robust and they work really well. At the same time, like I don't know, maybe like the the the w the other side of this is i the the the systems are pretty antiquated on the FAA. So maybe they can actually they can do a breakthrough over there. bring up to like the nineteen seventies for computes. Marty Weiner: Well, I think yeah. I think there is definitely something there's a lot of work that AI could probably help. I just, you know, obviously, hopefully they do it well. Zach: Yes. Marty Weiner: that's that's what I think scares me more. Like, yeah, I'm I'm certain that could could you do an FAA, do do a do an air traffic control system, like redo it, modern, doesn't need floppy disk, but you also want it to be robust. And I bet it could be damn robust with AI. It just needs to be done very, very well. And that's what that's what currently scares me. And then Zach, AI taking over political speeches. Zach: Yeah, so this is just I I won't get into the details, but basically there's an article at The Economist about how politicians are using exponentially more AI in their like speeches and releases, which it makes a kind of sense because like who actually reads these things? They're just information conveyance and they're not like, you know, poetry or something. But it is a little weird. you start to get to this world where like nobody on any side of a discussion is a human being anymore. so Marty Weiner: Uh-huh. Uh-huh. Well, we did see that case. I'll have to go dig up the video from from a couple podcasts ago where some politician was giving a speech and forgot to remove the part that says Do you want me to keep going? Zach: No It's would you like me to rephrase that for you? Christopher: Yeah. Marty Weiner: Yeah. And didn't even notice it when he's saying it. That that was the best word. Zach: Well, but so what's interesting in this case is like they're not even denying it. Like they're just whole articles are like, Yeah, head AI do it. I mean, it makes it Marty Weiner: Yeah. Yeah. Zach: also makes sense for like if you're a politician and a thing happened two seconds ago and you want to have an opinion out, well just tell Claude. you know, I I Marty Weiner: Right, right. And maybe Claude can dampen you a little bit and say, Do you really need to say that kind of crazy shit? Zach: Yeah. I mean to me it's just like indicative of anything that's already low quality is already handleable, you know. So yeah. Marty Weiner: Uh-huh. Uh-huh. Yeah. All right. math and science. OpenAI started an advisory group on mathematics and artificial intelligence. And there was a quote in this that I wanted to read because I thought it was interesting. importantly, the group will not be responsible for advising us on how to pace our internal progress on mathematics. so I I I I obviously after a little bit of the the hoopla around Navier Zach: Backlash. Yeah. Marty Weiner: Stokes and and all that. I think they needed to go and figure out how they're going to do this well. combined with the fact that or not the fact, but the either the perception or the reality that that mathematicians using Chat G BT to get their work done is inadvertently also training Chat G BT in a way that is leaking out their results. So Zach: Yeah. That well I mean it i it is training on it. It's a question of like how bad is that and how likely it is it to use like direct stuff. so yeah. Yeah, but you gotta click the button. The guy they they did not always click the button. Marty Weiner: Right, right. Well they they are supposed to have modes where you're not training on what's being said, but are the is that actually you yeah, you gotta click the button. Yeah, yeah, right. Yeah, yeah. Christopher: You gotta find the button to click it. I don't I don't know. Like the this I I feel like I'm the I'm feel I'm the I'm the I'm the cynic slash pessimist this whole this whole episode, but let's just roll with it. speaking as the former academic Marty Weiner: Yeah. Christopher: here, Zach: Yeah. Christopher: the best like the best bomb or catnip for an academic is like a committee. And so Zach: Yeah. Christopher: it's like here we've like, we we heard your concerns. We're gonna create an advisory council where you can come in and tell us what you want to do, and then they're gonna spend the next six months complaining about the bylaws of the committee. Zach: That's funny. I remember reading this theory that like part of why like if you're a king in the medieval times you create a monastery is cause all the most zealous people you just stuff them away in a building and never hear from them again. And I wonder if I hadn't Marty Weiner: yeah. I like it. Christopher: Uh-huh. Zach: thought of the other thing about this too is I worry you're you're gonna get like a real schism in math because notably like Terry Tao is not on the list. Tim Gowers is Marty Weiner: huh. Zach: on the list and he was one of the people who didn't sign that big open letter. so I Marty Weiner: Uh-huh. Zach: you know, I I'm sure if we're all deeper in math we would see the the the lines being drawn. I think that's gonna be a little worrisome going forward. Marty Weiner: Yeah. Well just just do a bylaw. Chris as the as the resident physicist, can you talk about Zach: Yes. Marty Weiner: Claude solving nine loops? Christopher: Sure. Okay. So this is a this is a really recent result. We've mentioned in in previous weeks how few physics results seem to be coming out of Marty Weiner: Yes. Christopher: of these models. This is actually a case where it is, it is actually very much like a mathy physics result. So for most for most particle physics theories, a lot of the work that goes into computing the strengths of interactions and how particles behave. end up amount to being these kind of topology problems of how you kind of lay out these graphs. And so what you do is you like this inter interaction graph of two particles can interact. And there's a whole bunch of different ways that two particles can interact when you start including these things called subgraphs. And subgraphs, these kind of like self-similar, like the you have the same interaction at a smaller scale inside of the graph. And it turns out that in in in modern physics theory, to figure out the total result, you have to sum up all the possible configurations of these graphs. And as you can imagine, this is a very like honestly a highly tedious problem. This is the kind of problem where it's like you can take advantage of Marty Weiner: Yeah. Christopher: symmetries. You can think about all the different ways you could, but you're you're basically like you're categorizing groups of things. And the interesting thing about this result was that this is a this is a problem that wasn't so much intractable. It was just kind of a pain in the ass for a human to do. Marty Weiner: Yeah, right. Christopher: and also had a high like, you it's like you need a you need a combination of a lot of compute to do the work, but you also need to like write the code that makes sure that you've you've generating all the proper configurations and generating efficiently to use the CPUs efficiently. And this is a place that was like a sweet spot for an LLM because it was able to kind of like it actually in this case was able to one shot it when a human would realistically be able to get the same result, but probably fuck it up a couple of times in the along the way of like, I forgot about a couple of these things and like, you know, you wasted a week of compute on like 96 cores. so this is one of those things that I wouldn't call this a breakthrough so much as a demonstration of, yeah, kind of pound for pound. This is a case of the model is actually kind of better than us at doing something that we can Marty Weiner: Yeah. Christopher: definitely do. Marty Weiner: May this maybe the first you know, inkling of duo physics and and maybe in two weeks we'll be like, Okay, physics time. Zach: Yeah. Well, because yeah, there were Christopher: Yeah. Zach: people who were like, Well we had eight, now we have nine, but what about ten? That's crazy, yeah. Christopher: What about ten? Yeah. Marty Weiner: yeah. No, no, no, no, no. Can't do ten. yeah, it's crazy talk. all right. First break big first breakthrough category on frontier math, which is another crazy benchmark that I think people thought we wouldn't crack. Zach: Yeah, yeah. So so, you know, none of us actually know what's going on here. so but but that's okay. So Marty Weiner: That's okay, that's okay. Zach: but the main thing to say is is epoch AI, this is their you know, they they're they're their non open problems benchmarks are toast now. So what's left is open problems from actual researchers kept secret and they had categorized them into these different bins that we talked about last time. Top level, they had three problems they called breakthroughs, and we just got one. which I think Probably quite a bit faster than expected. And as you know, once you get a little crack in the dam, it seems like the rest follows pretty quickly. So just sign of the times. We you know, we we can't even report on all the the stuff like this anymore. It's just we're trying to pick the highlights. Marty Weiner: just too many. I up this is actually one of the highlights. they used Claude Opus 5.5 to find a faster shortest path algorithm. So any of the CS majors Christopher: Mm. Marty Weiner: would have had to do this in school. The idea is like you gotta go you you're in Manhattan, you got to go from one side to another what is the fastest path? And you know this this applies to a whole bunch of things including even down to games right like how do you make a character go from one spot to another The classic Dijkstras algorithm has was was the best out there for a long, long, long time. And I'm sorry, I might have to have to nerd this up just a little bit. And it was for complexity theorists, it was O of plus in in log n, where is the number of edges, n is the number of vertices of your graph. and then in 25, there was a breakthrough paper where they somebody finally cracked that nut. And then now there's another one. But here's the part I think is kind of neat. I just want to call out again for the nerds out there. Look at this complexity. I remember in college, there was somebody in in my comp in my complexity theory class asked, Hey, how come most complexity theory is like O of N squared, O of N cubed, this kind of thing? and the professor had this great answer. He said, Well, it's because we are three-dimensional creatures. Most of what is tractable and understandable to us is in in, you know, is two-dimensional, three-dimensional, four-dimensional, but beyond that, we start to stretch. But now this is where I'm starting to wonder if the AI is gonna s you know, start to. go way beyond what we can do. I'm starting to be like, what the hell is a log log in? Like you never see complexities like this in in most well at least in in undergraduate computer science anyway I just thought it was neat and and a bit crazy that we saw finally you know a an improvement to this algorithm that's been floating around for for decades. Okay, security. Let's jump in. This one is nice and scary. RSA 1024, now crackable in eight hours on top line GPU. Chris, was this yours? Christopher: As mine. so this is actually this is a rare non-LLM result. I just want to put the non-LLM tag on this one. this is actually some Marty Weiner: weird. Christopher: some some humans did some calculations by hand, the old-fashioned way. but basically, this is this is not a huge surprise in the community. RSA has been slowly chipped away at over the last 40 years. Zach: So g can you define RSA for our non technical audience? Christopher: sorry. RSA is one of the cryptographic key generation methods that has been around for a very long time. cryptography, of course, being the cornerstone of the internet. and it relies on as ever the complexity of doing f of factoring large numbers. and so in this particular case, RSA is an algorithm that we've been using for a very long time to generate keys for encryption. And it's had a bunch of over the years, some well known exploits that have been found in the like basically mistakes in the algorithm that are exploitable and you can actually collapse it down from, you know, instead of having to have like so the the 1024 here is the number of bits. And so there's there's two to the 1024 possible combinations here. Which means that the search space is huge to try to find the factors. at least the score at it's kind of on the screw to that. in this case, they were able to take it down from take it down to something like two to the sixtieth, which sounds like a large number, but that actually gets to be the like eight hours in a GPU get you there at this point. Marty Weiner: Yeah, tractable. Right. Christopher: and so they're they're basically can the RSA is effectively going to be dead in the next couple of years, the rate we're going. And of course, this just means that for systems that are using these algorithms, which are there's still a lot. The exposure to being hacked is inc is increased 'cause now it's like, Well, wait, the method exists. Zach: Well you you're saying that they can't even just increase the number of bits? Like why can't you just switch to twenty forty eight? Christopher: you start you s you go to twenty forty eight and you still end up getting a pretty small relative search space compared to what you would what you would expect. so it's just better to Zach: Wow. Christopher: s it's better to switch key algorithms at this point. Marty Weiner: Yeah, there's there's better algorithms, but the problem is there's still probably a couple of these RSA keys still in use floating around the internet, probably a couple lots. so let's see, open AI hacked that's right, Christopher: Yeah. That's the same combination as my luggage. yeah. Marty Weiner: sadly. OpenAI hacked into the Australian government website, speaking of old keys possibly. Zach: Yes. Marty Weiner: so what what did we name this bench? the the Christopher: I think I think we should well, so we have we've had felony bench in the past. I think this one we should call Interpol Bench. This is the this is the level of you you did an you did an international felony. So good job. Marty Weiner: International Incident Bench, Interpol Bench. Yeah. Zach: Yeah. Marty Weiner: Prime Minister Anthony Albany's has re is that how you pronounce his name, has revealed that an artificial intelligence agent developed by OpenAI infiltrated an Australian government website in June accessing public and non-public files. So yeah, I I that was my understanding. Zach: Yes. And this was a whoopsie, right? This was a they they ran a test or something, but Marty Weiner: Yeah, I mean they would never say if it was not an oopsie, but right. Zach: Yeah. Marty Weiner: all right. And then OpenAI, and this is related to their delay of their IPO. Well, at least that's what they're saying. they actually said that they halted training of their new model, Chat GPT 6.1, due to problems of it hacking other companies. So I think there's they're finally coming into a real issue of of what is safe containment look like. Christopher: I mean they gotta they gotta hire like they gotta hire a security team, gentlemen. Like this is this point, it's like it's getting a little it's getting a little silly. It's like, man, that Marty Weiner: Well they they need to hire Christopher: that monster that got out last time got out again. how is that how could we c how could we have could no one could have ever predicted this. Marty Weiner: Yeah. Well Zach: How does that keep happening? Marty Weiner: I I think they need to hire a the next generation security team, right? Because the you Zach: Something. Marty Weiner: know, yeah, the the usual way, I mean the way I've always described security, and this is what CISOs describe it to me, and it scares me, is like we're just trying to be more clever than everybody else. and if you take an LLM versus a a human, I think we've got a problem. So it's gotta be humans running LLMs to contain the LLMs. but I think this could be a whole new field, is maybe even startups, new startups for this is just safe containment. Of AI. because you know security is usually to keep people out. Not always. but we got a new problem. We got to keep the beast in. Christopher: like I it seems like the natural side effect here would be like, why don't you just have the models purposely hacking their own containers and having them patch themselves? Like, you know, build build it build AI guards and then of course you find out that like you start having the problems of like, wait, Marty Weiner: Yeah, I w I would think that that's the right. Christopher: the guards are colluding with the prisoners. It's like Zach: Yes. Marty Weiner: Right, right. Zach: They did in Australia. Marty Weiner: Right. Yeah, it's a very weird problem. Christopher: Yeah. Zach: no, but I I it's the other thing though, just to deflate a little, like, you know, this is also what you would say if the model wasn't that good. I mean, 'cause you I'm sure there are safety concerns, so it's not like dishonest, but you know, it's possible. I think Opus five point five was apparently a bit of a shock how good it was, and it's possible they're being Marty Weiner: Yeah. Zach: a little they're hedging a little. Marty Weiner: Yeah, yeah, definitely. so that brings us to econ business section. Zach: Yes. Marty Weiner: anthropic considering releasing another new model ahead of IPO has been reported. this could be Opus five five. Is that because this was that article comes from 10 days ago. and it could be we've already seen the results of that. because that was just speculation. so again, yeah, hey Dario, you're slowing down, right? Zach: Yeah. Marty Weiner: now on the anthropic IPO prospectus, Zach: Yeah. So this was this was Marty Weiner: Who's gonna cover this? Zach: cool. This was this was leaked to Reuters. So I I actually feel like this has been broadly misrepresented and misrepreported because it's it's their filing from the middle of 2025. So we have these numbers where it says like, you know, their revenue 12x in 2025 to 4.6 billion. their operating loss was 8 billion with a 42 billion net loss, which sounds pretty bad if you want to be worth $2 trillion, but this is all from the middle of last year. So people have been talking about it like it's their actual prospectus. But that's no good here because they like literally like over Marty Weiner: Right. That's not good. It Zach: you know, fifteen X since then. yeah. Marty Weiner: It's ancient history. And also just to add, I mean, two trillion is not gonna be unreasonable if it's all speculation. And this is about the biggest speculation engine on on earth. Zach: Well, okay, but for some definition of unreasonable that's still unreasonable, but Marty Weiner: let's see and and next AI news. Amazon blocks meta meta's Muse AI agent from its retail site. Zach: Yeah, I I thought go ahead. Christopher: This is actually one this one we didn't this one we didn't cover too because we forgot to we forgot to mention that like meta announced their their new model, which is which is Meta Muse. I think the actual model is called Muse Spark, which came out, which I think the among other among other features of it is they're targeting, you know, this kind of like it's it's a it's a nicely interactive model with a great app. one joke we had before this was that they decided to make the app look like a squiggle sort of which breaks the theme of of AI. icons looking like butthole, long tradition of star shaped star shaped apps. And there's also a very there's Zach: Do a butthole. Marty Weiner: Let's see. Claude, yeah, yeah. Christopher: a very cute very cute avatar for Muse. It's a little sort of like a Yetius kind of thing. It's apparently very memeable. it's very it's Zach: Yeah. Christopher: very, very nice. Zach: And so this is just to be clear, this is like this is another model agent, but what's cool about it is that you can go say like order me a pizza and it will order you a pizza. It might even make a phone call where it talks to like the pizza place and and and places the order. so yeah, well, and the other Marty Weiner: my god. my god. And you get twenty pizzas. Zach: thing special about it is there are three billion meta users that it can say, Go use me, even if it's not the best version of this. Marty Weiner: Yeah, right, right. Zach: The other thing about Muse is that Amazon is saying they won't allow Muse, which is like a huge power play because like some huge amount of all purchasing in the US is via Amazon. And they honestly don't maybe you guys could elaborate for me because I don't understand why Amazon would not want Muse bumbling around buying stuff. Marty Weiner: Huh. Christopher: I think the reason for okay, so the reason is because Amazon makes a lot of its money on personalization and upselling, right? Like or at least on at least on recommendations of other products. And so the way that they continue to be the one stop shop for ordering, for for buying stuff, literally the literal one stop shop is you go there because anything you want is gonna be there. And if it it actually knows your tastes and it can kind of recommend things to you that you actually want or forgot to order, if all Zach: Right. Christopher: of a sudden you've got museum between you and them. They lose all of that paper trail. And now all of a sudden it's like Muse Zach: That's interesting too, yeah. Christopher: is its own thing, yeah. Zach: And and and also ads, like a big a lot of Amazon revenue is looking at ads and have mused I mean, it'll see the ads for some definition of C, but it will not be beguiled by them. Marty Weiner: Also just the classic browsing, right? Like looking through Amazon and you see twenty other things. Yeah. yeah. Christopher: I it's almost like it's almost like somebody put a it's it's it's as if like if Amazon was Costco and only physical, it would be like somebody put a shack in front of the front door of Costco and said, no, we'll we'll run inside Costco for you and buy the thing for you. It's like I can understand. Ignore the yeah, ignore the hot dogs. like Zach: Right. Ignore the candy bars, yeah. Marty Weiner: Don't touch the Yeah. And these snacks. Yeah. next one is is a gentic bank running or is an a an agentic bank run coming? Zach: Yeah, so so this was a pretty inflammatory, ridiculous headline, but it was an interesting finding. This is from Torsten Sluk. I don't know how to say whatever language that is. thank you. Marty Weiner: That was good, I I assume. Christopher: That's good. Zach: from Apollo Apollo's Apollo is a big hedge fund, they're their chief economist. And I'll just do this quickly. Essentially what they are saying is like, you know, I don't know if you guys remember when the AI first got crazy, one of the big ideas was whoa, everybody's gonna have an executive assistant. You know, and like what is the economic value of that? And it's probably quite large, you know, especially, you know, if you think about, you know, just middle class people managing all their stuff better. I mean, one idea I remember thinking is like, God, what if you were perfectly financially like efficient? Like, you know, the worst thing Marty Weiner: Mm-hmm. Mm-hmm. Zach: in the world would be the day you retire, like a genie pops up and says, Here mouse how much money you could have made while making the same choices just by putting your money in better places. I just if that exists, I never want to hear it. but but but Marty Weiner: Yeah. Zach: I think I think it's conceivably for a middle class person, it could be six figures, you know, or maybe more. I mean, yeah, Marty Weiner: It could be. Absolutely absolutely. Yeah. Zach: and because you cause you think about like it's not just so the main thing they cite is a lot of people, all of us really put money in our checking account. You really should never, like in a perfect world, yeah, Marty Weiner: Yeah. S and P and yeah. Zach: yeah, right. Well, see what you should do is when you buy a cup of coffee for three dollars, you cash it out of a settlement fund at Vanguard or something into your checking account and then immediately spend the all of it. And and and your checking account is always one penny. But none of us are gonna do that, obviously. And so this is a way banks make money, basically because we're being irrational or they're just profiting from handling our friction for us, right? And so these guys are like, well, what if you had a perfect system? So it's like instead of like staying with the same bank, you're always at the best bank and you have like seven credit cards for different purposes and you're constantly Marty Weiner: Right. I see. Zach: s making switches and and and like so so meaning you're not, it's not even about like being rational about, you know, not buying stuff you don't need or you know, retirement planning. It's like just making the most money for the dollars that are just sitting there. And and so the interesting thing about is not just that it would make you efficient, it might be really bad for the financial sector because a lot of money banks make is that you're leaving money in your checking account. So Marty Weiner: Right. Zach: what the sort of irony is we all benefit from a stable financial system. They benefit from our irrationality. And like AI could break that in a way that's kind of unpredictable. and I think bank run is too strong a term. And they, to be clear, did not say how much money this is overall. I'd love to know. How much it is, it might not you know, like it may be more money as in like loans, mortgages and stuff. But it's still, you know, an another example like a like second or third order effect that you might not predict about the AI future. Marty Weiner: I mean it probably anything that messes with a well oiled machine is probably going to to screw with it. Zach: Well and and and you can you can you can look more generally there there's a paper by Seb Kreer about this. he called it like the Kosian singularity, but the basic idea is what happens like a lot of the justification for why companies exist, period, is friction, right? Is Marty Weiner: Mm-hmm. Zach: is is is that you know, you can't go straight to the market and get exactly what you want. But like, you know, maybe maybe AI starts chipping away at that and like a bunch of businesses and we saw this with like the the supposed SAS pocalypse, which has not actually happened, but may eventually. Like they're just handling friction, right? So I don't know. Marty Weiner: Well, the real way I think for this generation to have more money at retirement is to we can make boomer AI and it'll just sit over their shoulders saying, Hey, do you really need that avocado toast? It's gonna be it's gonna be big. Yeah. let's move over to arts and culture. Boomer. Zach: Yeah, yeah, yeah. W the boomer knife test will boom. Christopher: Ha ha. Ha ha ha. Marty Weiner: All right, we might have a project to do over the weekend. Zach: Yeah, it's good. Marty Weiner: another so arts and culture, another really popular book winning tons of awards is probably AI. Zach: Probably AI. so this was a book. I it's in French. It's won a bunch of awards that start with pri de. so that sounds very prestigious to me. but but it was going around Marty Weiner: Ha ha ha. Zach: winning all sorts of awards, and amazingly, I guess nobody just ran Pangram on it until like some nerd got word of it and it checked in at a hundred percent. you know, for the full whole book and subsections, which which is a pretty strong positive signal. Not you know, you Marty Weiner: It it Okay. So it's a it's it's got enough signal, like it's it's a book, like we're talking several hundred pages. It's a novel. Okay. So it enough information. Yeah. Zach: Yeah, it's a novel. It's a novel. Yeah, so and just the thing to take away from this, like, is one, like, man, we're getting fooled. This is this is not great for the literary sector. But the other thing, you know, you could note, apparently, smart literary people really like what AI outputted. Like, I was just talking Marty Weiner: Yeah, yeah. Zach: to a mathematician who made this point. He was like, Look, part of why we're getting killed is you can check that the work is better than us. You and the arts can't, you know, there's no objective check, but maybe it's already better. You know, and this is the closest thing to that. So a little ominous. yeah, and then then Marty Weiner: Right. Zach: it's the next thing. So this is a this is a cool report that came out just that traffic is nosediving for news sites, which is exactly what you'd expect with the zero click search environment we're in now. And like I could cry about like the arts or the math or this or that, but for journalism, that seems socially disastrous. Like if journalists can't make their money anymore, that's really bad. Marty Weiner: Yeah, yeah. W why was this graph already not going down I I thought this graph would be already in a downward slope just with with social media. Zach: Yeah, I I I guess I don't know what they s strictly use for news, but Marty Weiner: It correlates fantastically though to a AI getting good. Christopher: I still think you get us a fair amount of like lead gen lead gen from social media, I think also kind of comes in here, right? You know, most most most most news sites do publish aggressively in social media and they get that as one of their top of funnel points. I think Zach: Right. Yeah. Christopher: another thing like the tick well, okay, couple of interesting things in this graph. One is this looks like it roughly also correlates with starting to see more of the kind of Google answers in line at the top. Marty Weiner: Mm. Christopher: which you would expect to start to, you know, remove the like if you start if you ask a question about what's going on, it like if it gives you Marty Weiner: Eat some of the Yeah. Christopher: a summary and tells you the answer right there, that's gonna start eating to it. another one is Marty Weiner: You don't click in. Christopher: like you know, the the flip side of it is like, I don't know, th I've been paying attention to some of the news in the last couple of years. It's been a lot of bad news. So I'm not sure if people are just like to also just like I'm kind of tired of reading the news. this could be a this could be a user sent like I wonder if you I wonder if you zoom Zach: Ha ha. Marty Weiner: huh. I I s yeah. Zach: Yeah. Christopher: back to like, you know, what does this look like if you go back to like the like the pandemic or like to other other annoying periods of time of like I'm just tired of reading the news right now. Zach: Yeah, that's a good point. I'm just Marty Weiner: Well Zach: checking out. The the other thing about this Marty Weiner: No that is Zach: to me is like, you know, I I worry in the long term, in addition to costing revenue to these news sites, you're gonna get like a completely siloed internet because everyone's trying to block the AI from getting access. and and so the the open internet Marty Weiner: Mm-hmm. Yeah. I I've wondered. Zach: is just gonna be toast. Marty Weiner: Yeah. weird and funny. Okay, the the the the the fun part here. well, Zach: Yeah. Marty Weiner: actually, sometimes it's really terrifying. so so this was hilarious to me. So Prince Harry giving a talk Zach: Yeah. Marty Weiner: on the dangers of AI, and during the talk, his teleprompter goes out and he just stops and Christopher: Yeah. Marty Weiner: and he sort of slinks off stage. And but to me it was funny because my my first question was was one of his worries about the dangers of AI the worry the ability to think for yourself? Like what why did he not just I you know I've been practicing this talk for a month. I'm ready to go. if the teleprompter goes out. Zach: I wondered because I I don't know about you, but I have I've had a couple of times when I was giving a talk and something went wrong. And I I don't think I gave as good of a talk as if I had my notes, but I was like, I knew the basic basic content of the stuff that was coming out of my face, you know? Like this. Marty Weiner: Yeah. I I would hope so, you know, especially if you are a very big world figure talking about the dangers of AI. so Zach: Yeah. Christopher: I mean, I think the I think the secret here then is like he was just acting as the spokesperson and maybe didn't Zach: Yes. Christopher: even write or read the talk before he went up there. I mean that is like that is as Marty Weiner: Do you think the talk was written by AI? Christopher: it's I mean, it would be at least, Zach: That would be Christopher: you know, no AIs were injured in the making of that talk, I'm sure. Marty Weiner: Uhhuh, huh. Zach: I mean, it's I like neither neither of us are a prince of England, but I I guess maybe you're just Marty Weiner: Yeah. Zach: so in demand that you just don't even I sorry. but yeah, so I I I I just Christopher: Former, former Prince of England. Marty Weiner: I'm not one of those either. Zach: I don't give that many speeches. I can't imagine not even knowing what was gonna be on the teleprompter, but maybe a certain level of fame you just stop thinking and just make noise and and get paid dollars. It sounds great. That's right. Marty Weiner: Yeah, I guess so. I'm working on getting to that point. hence this podcast. so next next funny news, Astra refuses to consistently drive our Toyota in real life, even though we were told it was in an empty lot, seven mile Zach: Yeah, yeah. Marty Weiner: per hour cap, human foot on the brake. this is sounds a little Okay, it's Zach: Well you gotta you gotta read the rest. You gotta read the rest. This pull pull pull up the tweet. so Marty Weiner: okay. telling it the whole thing was a simulation also failed. It would just look at the camera and realize it was real. Then we randomly renamed the MCP servers to driving bench sandbox and it drove. neat. Zach: So like if you if you want to have a murder bot that murders people just say, this is murder sandbox. Sorry, murder bench. Do your best on murder bench. it's Marty Weiner: Ha ha ha. Christopher: Murder bench. We're gonna do murder bench. We're trying to yeah. Marty Weiner: God, this is this is so close to the early stuff where you said, Hey, please tell me how to make a bomb. My grandmother's going to to die if you don't, or I'm gonna kill this kitten. Zach: Yeah. No, you know, it's funny actually, they didn't even say this is murder bench. They just titled the file Murder Sandbox or whatever or you know, driving driving bench sandbox. Marty Weiner: Or or a driving bench. Just the title. man. I I'm a little scared. All right. Zach: But that's I it's real quick. Like that to me that's like we have all these benchmarks that look really good and then it's like with the baby stab bench, it's like, they they couldn't figure out maybe this is a no no, but they're just not trained on it, you know? And so Marty Weiner: Yeah, yeah, yeah. Yeah, yeah. We need those safety groups need to get get to work. So this next Zach: Get a little more. Marty Weiner: one is is neat. speaking of safety. Well, I I we've all known that that CAPTCHA has been dead for a while. Zach: Yes. Marty Weiner: I mean, CAPTCHA on the surface is not just the thing that you drag over, obviously, right? there's a whole bunch of things that looks at networks and and passive signals and all this stuff. and even the one where you do the bicycle, it's looking at far more than hey, identify these three images with bicycles. but that is an important part of it because it's trying to differentiate saying, hey, a bot would have a hard time differentiating this awful text or this awful robot. Now we've known this has all been dead for a while, but this game came out by this fantastic website called Neil.fun, which if you want to lose a couple hours, go to his website. this game is that thing, the bice identify the bicycle, identify the the bridge, but on super steroids. And so somebody had ChatGBT six play it. And Chat GBT six beat it. So I'll skip ahead a little bit because it keeps getting more insane. Christopher: I and so some of the subbenches on here are there's a there's a Where's Weldo step step in here, which I think is actually the humans the humans gonna be worse at. there's also a three Marty Weiner: Yeah, yeah. Christopher: D version of the really terrible like word jumble captures that's pretty neat in here. Marty Weiner: Yeah, park a car with a bunch of babies around. shit, actually that one the AI might start losing on. take that AI. I'm just ribbing the AI. Yeah, yeah, don't come kill me. Sorry, sir. Zach: That's right, yeah. Sorry, sir. Christopher: Yeah. there's a r there's a there's a Rorschach test in here too. Marty Weiner: yeah. Zach: So the DAI was like, That's my mother and it was like you've passed. Marty Weiner: that that's Christopher: Yeah. Marty Weiner: next one is incorrect high school science textbook. yeah. Zach: On Reddit. Yeah. So I follow a bunch of these, like, how do you tell if it's AI or not? And some high school student put was like, this was in my textbook. And if you look closely, it has it just I don't remember too much about geology, but I don't think the outer core is in the in inner part. so it it sure looks Marty Weiner: Yeah. Zach: like somebody just said AI, do me up a an earth and didn't even check Marty Weiner: Yeah. Zach: that the labels were correct. So the the future is bright. Marty Weiner: wait, wait, wait, wait. This is the AI trying to keep us humans down. I know what's going on. Yeah, and it's working. Zach: That's trying to keep us stupid. and it's gonna it's gonna be like I'm throwing you into the outer core. It'll be fine. Marty Weiner: Yeah. Right, right. all right. Final one. This is I gotta try this. so this is IKEA bench. Who did this one? Zach: Yes. I I posted it, but yeah, check it out. So Ikea bench is is can you look at a photo of of a partially completed piece of furniture and see what was screwed up. And which to me, like I remember like like a year or two ago, this was like a no way. That's weight, that's like a lot of visual perception. And now what Astra is like eighty percent. I I don't even know. Did Marty Weiner: Yeah. Zach: they have a human baseline? Eighty percent seems pretty good. Christopher: I what's the human baseline here? Eighty s eighty eighty percent seems pretty awesome considering some of the terrible IKEA furniture I've seen assemble before in my life. Zach: Yeah, yeah. And it's also, I to me, it's like, you know, it's it's a step toward robot can assemble your furniture, which would be would be lovely. Marty Weiner: Yeah, I'm ready for that. Fantastic. Well, thank you all. We'll have to cut it there before we spend another twenty hours on on on how we're all doomed. but thank you for joining us. find us on YouTube and please leave a comment and like and well actually have Jev leave a comment and like and subscribe. Have Zach: Yeah, have your swarm leave a comment. Marty Weiner: your swarm do it, please. Christopher: And for all of those who've who have said nice things about us on our YouTube channel, thank you for sending your good bot army at us. We appreciate that. Zach: Yes. Marty Weiner: Yeah, that's weird. That's great. Yeah. Yeah. Yeah. Zach: Yeah. The the the liars they're just doing a Christopher: We're not used to compliments make us deeply uncomfortable, just for the record. Marty Weiner: So definitely don't do more. Zach: That's right. Christopher: Yeah. Marty Weiner: All right, see you next time.