Latest / Last Humans in the Loop / Robots in Your Home, AI in Your Job, and Inside AI's Mind

Transcript

Robots in Your Home, AI in Your Job, and Inside AI's Mind · Last Humans in the Loop

• Zach: if you had a laundry folding robot tomorrow, it would not be that useful. What it needs to do is put the laundry in the machine and then take it out. and then fold it and put away and now you're really saving time, you know. Marty Weiner: While while also not accidentally stepping on the toddler, who's coming and shaking it, or the dog. Zach: Well well, not smooshing the toddler. Yeah, yeah. I mean, you know, there are trade-offs. you know, I don't know. I mean, I'd I'd Christopher: Or the dog or the small dog. Zach: accept a lot to get my laundry folded, Marty Weiner: Welcome to the last humans in the loop, your weekly AI podcast. Today we're going to be covering topics like robotics. We'll look at how far we've come and what's coming next. We've got math occlys, physics occurly, job occalypse, and we're gonna look at is it real and what are we seeing? we're going to be looking at can we see inside the AI's mind now? And then finally, we will end it with our AI news lightning round, So let's get into it. I'm Marty. I'm here with Chris Slow, everyone's favorite former Reddit CTO, other than me. and Zach Wienersmith, mega super mega worldwide celebrity and fashion god. So Zach: That's right. Hmm. Marty Weiner: Let's jump into robotics. I sort of repaired a little robotics section here to kind of take you all on a fun little tour of how far it's come, even going back to the 30s, where we are as of what I'd call a few years ago, and how the LLMs have basically ripped it all apart and are destroying everything we've worked on in an amazing way. so what what I want to start with, guys, here is We do have robots that are doing things for us but mostly in what I call controlled spaces, right? You know, 3D printers, CNC mills and lathes. These are, you know, you have a mill going around carving out a piece controlled completely by a computer. We've had those since the sixties. we have pick in place where you have when you make circuit boards, you have a little circuit board and something goes and puts all the little pieces in and then it runs it through you know, things to heat it up and get all the parts to solder. We have car assembly lines. We've had that since the 50s but what's been really hard for forever and possibly up until today or you know the last couple of months is that when the space is not well controlled, like you know, a car assembly line, right? Like the car comes along, it is fixed in place, the robot comes along and it puts a screw in a known location. when we don't have that kind of controlled space, things get hard. really fast. So, you know, you imagine like in the sixties we had the Jetsons and when you have robots walking around. Well, that robot has to deal with gravity and changing situations, such as a child on the floor or the laundry on the floor, and so the problems go from zero to infinity exceptionally fast. And I think that's why we don't have robots really in our houses yet. We have a few. You know, we have Roomba. And that's kind of it. So let let me take you on a little fun tour because there there's some hilarity along the way. this is my favorite of of all of them. this robot is called Electro, the smoking robot. It was at the 1939 New York World's Fair. And it was pretty much all a farce. it was controlled from the back. Somebody they did some kind of recording, someone was in the back, and it even sounded like a robot. So I'll play that for you in post. and then here's the best part. Look at this. Okay, let's find it. look, look, he's gonna smoke. Zach: Man. Probably would have saved a lot of lives if we Marty Weiner: He's wait, where is he? yeah, yeah, there it is. Zach: offloaded smoking to the robots. Marty Weiner: That's true. but look, he put had to put the cigarette in its mouth. But let's see. Christopher: Yeah. Zach: Do we have smoking robots today? Can Unitry smoke? Christopher: pretty sure I'm pretty sure that there is a there's a s there's like tester robots for determining how much nicotine and tar you get. There's like but there's been inf there's been some infamous there's been some infamous like scandals about about this. I mean yeah, 'cause you know, he's being Marty Weiner: look. He's lighting the cigarette. it smoked. Christopher: he's being polite. Zach: my Christopher: He's being polite. He's helping his friend who is a robot. That's just what he's showing. Marty Weiner: That's right. and then we actually had useful robots for a long time, even all the way back to the 50s. So automated assembly lines, again, controlled environments. but outside the controlled, it breaks really fast. One of my favorite to look at is Boston Dynamic is making progress on this for since even the 80s. So here's you could see some of their very early 1983 robots, this sort of bouncing balancing thing. And I'll just speed along as they keep making progress. Here's one that runs and does flips connected to a little arm. Here's one without the arm now running. And this is this is already in the starting to bump into the 90s. and then they they keep making amazing progress, even having it start to run on a treadmill. This is 2011. It's crazy how far we've come. This is 2011, and it looks so old. This is the famous one of the guy who's gonna be the first to be killed by the the AI apocalypse Christopher: Yeah. Marty Weiner: because he keeps Zach: Yeah, yep. Marty Weiner: screwing with this poor robot. And Then they get it fully automated, running outside. And then there's doing gymnastics. and and even like these fun little dance shows. but I you know it's one of those things where this company's been working on it for 40 years and surely they know a lot. But I think what's gonna happen here is the LLM's gonna come along and be like, yeah, done. So we'll look at that in a moment. we're seeing robots entering the home, that that can actually walk around, use the door, use use an uncontrolled environment. Christopher: do we have any is there there's been some there's already been some amount of controversy around the home robots where it's unclear how much of the time they're being spent driven by some sort of a remote human as opposed to actually being driven by Zach: Mm-hmm. Christopher: a model. Because that's that's the thing about the like most Marty Weiner: Yeah. Christopher: of the traditional robots here, they're they're there's no intelligence behind them for the most part, right? It's mostly just like there's a servo and there's a sensor and it's like put the arm in this position, which means that if you're standing in between the robot and its position, you're probably in for a bad time. Whereas like Marty Weiner: I think Christopher: here for a home robot, you want something kind of soft and squishy, but therefore you want humans probably involved in loop. Marty Weiner: I'll have to go dig it up, but I believe that Tesla kind of got in trouble for that early on when they had some sort of one of their robots Christopher: That's our number two. Marty Weiner: and it turned up it was being driven by somebody behind the scenes. It's like, well, that's kind of cool. Yeah. Christopher: It was the Optimist the Optimist Launch Party, in fact. Yeah. Marty Weiner: Yeah. that's we've had that kind of thing for for years. I mean, I we they sort of had that Disney had things like that in the 60s. Now you didn't have all the articulation, so that was an important jump, but but what we're talking about here with robots, the important part is no, you can actually Drive the robot with AI. Berkeley had this great project in the 90s and 2000s and even further along just for folding laundry, which turns out to be an insanely difficult problem, which sounds surprising, but we humans are are really good at understanding fabric. And you can see this robot is 30x speed and it's just sitting there. trying to understand the shape of the damn thing. All right. It's almost cartoonish. I mean, it's a big advance. It's great. I don't want to diminish it, but but this is how far it is from being in the house. and then we jump to now and we're seeing they have this figure two robot is doing it at full time speed. So it's just folding laundry no problem. It's funny to me now they're showing it 0.4x speed. They're actually slowing down so you can see the detail Zach: huh. Marty Weiner: in it. give it another little, you know, a few more iterations and and we might start seeing this in the house. I think the obvious big problem is safety, having a big, chunky, meaty thing that has strength to it. could hurt somebody or break something. The robots that we do have in the home do include a few things. We've got the Roomba and we've got a thousand knockoffs now. and they're they started out actually early on as this robot that would just go around in circles trying to, you know, not even AI. It was just trying to figure out an algorithm that over time would work. And that was fine, right? Because you could leave it on all night and you don't care that it's gonna take four hours versus a human taking, you know, 15 minutes. But they got smarter but that's a that's about the extent that I could think of about robots in the home. I have 3D printers, but they've never really gotten to the point where they could really enter the casual home. we do have a bunch of toy robots, and everybody loves to give it to their nephew for Christmas. but what's funny to me about these is they I call them the Christmas tree toys. They never leave the Christmas tree because they're actually not that much fun, right? They go forward, Zach: Mm-hmm. Marty Weiner: they go backward, they maybe pick up a little tiny box and they put it on top of something else. And you're done. I actually got one for my kids and we built it. And I I think it's still in the very same box from last Christmas. I do think it's funny here. there's a bunch of cool robots, and then Google lists the United G one for $65,000 as if it's just a you know, click and delivery. Christopher: Ha ha. Zach: Yeah. Marty Weiner: But but even that is probably not that great yet without something great controlling it. And and I wanted to dig into why we are not there yet? Why has it been so hard? And there's a couple of reasons One is Anything involving a bunch of input. So you can imagine a robot's got audio input, it's got visuals, it's got all of the sensor inputs and the the the motor inputs. for a programmer, you know, remember when humans used to program? For for them to get it complete is is just a an intractable number of if statements and adjustments, and and it's even worse than that. even if you were to write out a whole bunch that worked. In one scenario, if you changed one thing, the dependencies between all those little tiny if statements makes makes it just a just totally attractable, right? You touch this if statement, now you have to change all these if statements. You change those, you have to change more. The only way to really approach it is with some sort of machine learning, and ultimately a a neural net is and transformers because of all of those little tiny if statements. You know, you could think of a transformer as trillions and trillions of little if statements that have been optimized. So I think that's part of the reason. It is just not possible to program them. the second one is the sensor capabilities of our hands is vastly beyond anything we've even seen today. there is nothing even coming close to it. If you look at the the ability for us to to detect touch, detect pain, detect friction, temperature. if your finger, the tip of your finger was the size of Earth. We can detect the difference between a house and a car. and we just have nothing even close to to that with with any of our our robots. so that makes certain interactions very difficult. One of the other examples is handing a cup from one hand to another is exceptionally difficult, and robots still, you know, are are not good at it because they partly because there's a lot of tight compute that we're actually doing in that interchange, including all the the sensor compute. and so until we we get to that point, which yeah, I don't know that AI is going to help us with that. That is Or I I'm sure it will help us with it, but it still may be years out to get to that kind of sensor capability. I I I always want to say like a decade out, but but but this stuff has been really compressing. but the other side of it, the programming robot to do something much more difficult than than we're able to do by hand, the AI has now tackled that. And so I want to show you a little bit about what's coming up next, because it's this is now getting insane. we now have anthropica, even as of July, is able to control a drone and have it fly around the space and understand what everything is. and so that changes drones from being human controlled, not knowing what it's seeing, not knowing what is a wall, what is the ground, to now being able to even navigate next to these people's heads with little spinning blades and it possibly be okay. Christopher: Or purposely flying Marty Weiner: we are also Christopher: into your eye, depending Zach: No. Christopher: on the mood of the L L at that moment. Marty Weiner: well, I mean, that's probably next, honestly, with the Ukraine war. Right now, the Ukraine drones have gone from being little, you know, grenade carrying Wi-Fi connected robots, but those started getting jammed. So then they went to they went to having these spools of fiber optics that would unwind as the drone flies. But we could easily get to the point if we're not already there where we have AI on board. I think the problem still is that the compute is still heavy and probably too much for the the the energy usage. but if you get to that point, you can say, hey, look for you know Russian tank and and blow it up. And then you there's nothing you can really jam there. but yes, i we could change it from avoiding humans to not avoid to specifically targeting humans. Let's see, who had this RuneScape thing? Was this you Chris? Christopher: this is this is me. So this is actually relevant to the conversation of like how how are LLMs going to drive robots? Because if you think about it, the problems that you encounter with driving a robot is a lot of visual information. And visual information is possible for LLMs to solve, especially multimodal models. They can go in and they can just they can they can describe what they can they can solve problems that that twenty five years ago were considered impossible, right? Like try describe to me what's in this picture. But doing that in real time, you know, the example of the drone is a good one, right? You don't have a couple seconds for the model to decide what it's seeing. It has to have other kinds of tools available to be able to determine what it's seeing more quickly and make reactions. And so I came across this video by Max Bitker, who was talking about a benchmark that he set up called RuneBench. And the idea is Runescate's this old RPG that exists from about 20 years ago. And he wanted to see if he could actually get LLMs to play it. And it turns out the LLMs have as much trouble playing with this game as they do. with real world situations, right? Like they click to move the character around, and then it's hard for them to figure out which character is theirs based on the other character moving on the screen, because at any given time, they're kind you know, they're a couple seconds out of sync. And so what he found was the the solution here is you don't what LMs are also very good are very good at is actually writing code. And so he turned the problem into one where instead of just having a look at the screen and click the mouse, he built a simple toolkit for. like an API for the LLMs to interact with the world, in addition to having access to the images. And that was enough for them to not only be able to basically they started building themselves little macros to kind of control their their their little humanoids in the space. And that actually made them able to respond in better than real time. and so I think the analogy applies here just as well for robots. You don't actually want to just have a vision toolkit and a bunch of motors, you need to have some sort of a mid-level abstraction almost. to be able to interopt with the robot and then all of a sudden you unlock a whole bunch of capabilities. Marty Weiner: Right. And actually I think we're starting to see that, Zach, you pulled up this article about the L LM's now Zach: Yeah. Marty Weiner: using off the shelf robots. Zach: Yeah, so this was a thing that was starting to happen this week, which is, you know, I'm well, I was I was telling you guys I'm gonna coin a term, which is the botter lesson, which is the bitter lesson for robots. Christopher: Yeah. Zach: so you know, we have this classic problem, which is people come up with clever, you know, fine-tuned models to do a cool thing. And then two seconds later, they get obliterated by a gigantic model trained on a lot of stuff with a huge amount of reinforcement learning. And I hadn't thought about till this week. that this might have applications for robots. But what we've started to see is you have people who, you know, want to work on robots and there are there are models built for robots, but instead they're just saying, well, what if I just drop Astra six in and say, Astra six, you're gonna steer this robot, figure out how to do that. And it turns out they do surprisingly well. there are a number of of benches where like Astra Six and Opus five and a half do pretty good. I mean, they're still in this world we're in now of kind of janky cleanup or like. Clean up in the sense of move a pristine piece of garbage to another container. but but but like the fact that you can hit state of the art by this thing that was not designed to do it, that is essentially doing it through like okay visuals and coding. it for me, like it makes me wonder if we're going to see faster advancement than we thought. Because you you wonder if you get to a world where like, okay, when we we talk about all this stuff where Well, the robot needs to speak the language of forces. It doesn't have our exquisite sensory apparatus. But what if you just get an L L that's so powerful it doesn't matter anymore? That it can just like, you know, build build a physical model of the entire kitchen and how it's supposed to be and how everything in it works and just use pure reason to to get it to a clean state. so I don't know, that might be coming. you know, this this is this is a new development this week. Marty Weiner: And there was another one. Christopher: do a it's do a breakthrough but with physical physical capabilities almost. Zach: Yeah. Marty Weiner: Do a robot Christopher: which is which is actually kind of amazing considering the the the Boston dynamics example. I mean Boston Dynamics has been legitimately working on this problem for 40 years. And they're Marty Weiner: yeah. Christopher: they're notorious for doing this in like I mean I wouldn't call it hand coded, but they're definitely of the approach of like they would prefer to do this without neural networks for most of their history. They were basically going straight to like we're gonna we're gonna figure out how to walk programmatically. And they actually did solve the problem but it took 30 years to crack that nut. to the point where they could actually have robots doing parkour. now you potentially have a like a drop drop an drop an astra on a robot and I'm not sure it'll be at parkour yet, but you can imagine like an Astra seven that can do that. Zach: Right and and and and good. Yeah. Marty Weiner: Well and I guess it depends on what you mean by solve the problem too, because Boston Dynamics keeps making amazing things. I've not seen one be useful. Zach: Useful. Yeah, yeah, yeah. Christopher: Well it's I mean they're they're good on Cany Valley examples though. They're usually there's a canonical examples of Uncanny Valley. They look like a the big dog looks like a yeah. Zach: I think they'd they'd be good at cage fighting. Marty Weiner: Right, but I don't have a big dog that's helping me. I mean, literally like in my, you know, where where is my my dog robot that helps me carry my stuff when I'm going, you know, camping Zach: Yeah. Marty Weiner: or or or building something? And I think that's part of the problem is is they have made steady progress, but I wonder if they were making progress along a path that would never get there. And then Zach: Yeah. I I think that the the the the the fundamental deal is like to t to cleaning a house is like four thousand tasks. It's extremely hard to clean a house, right? And it involves dealing with all sorts of set like if you get a plate that has like some gunk on it, you visually recognize what type of gunk it is, what kind of viscosity it's going to have, what you need to do, and then you you know, you move it in just a certain way to try to clean it. And that's only one tack. Like even if you had a laundry folding robot tomorrow, it would not be that useful. What it needs to do is put the laundry in the machine and then take it out. and then fold it and put away and now you're really saving time, you know. Marty Weiner: While while also not accidentally stepping on the toddler, who's coming and shaking it, or the dog. Zach: Well well, not smooshing the toddler. Yeah, yeah. I mean, you know, there are trade-offs. you know, I don't know. I mean, I'd I'd Christopher: Or the dog or the small dog. Zach: accept a lot to get my laundry folded, but like, but but so I mean may maybe Marty Weiner: Right. Well Christopher: Yeah. Zach: the point is the LLM is actually building a model of reality, you know, in its in Marty Weiner: Right. Right. Zach: its weights. And and and that's, you know, something you can't do programmatically. Marty Weiner: And it does understand, well, except for a few examples we saw, but it it does understand that that there is a toddler and it should be careful. Now, whether or not it can be successful on that yet is the stabbing the baby test from the last a couple of podcasts ago. Yeah, yeah, that was neat. So and Zach: Right. Yeah, w with the the stabbing the baby test from a couple weeks ago is a little not super encouraging. Marty Weiner: that's the security side of it, right? Like, okay, well, there's there okay, maybe the robot can be safe as long as you know somebody doesn't hack into it and say, Hey, go go stab the baby shaped thing. so th I I Christopher: I mean, it is it isn't true. I I just just for just for fun, I just a little quick search and see if if Boston Dynamics has appeared at all in the Ukraine war. And it sounds like they've been used in very limited situations around humanitarian aid as opposed to actually like they're they haven't been used as part of the fighting, which is kind of telling because honestly there's been a lot of like that war is turning into the future of warfare with drones to the point where they have, you know, automated, automated not quite tanks, but they're definitely kind of like the automated, you know, ground vehicles, automated drones. And they're definitely more at the stage of simple off the shelf components kind of figure it out as opposed to forty years of investment in a bleeding edge piece of robotics, which probably costs several million dollars. which which shouldn't be surprising at the end of the day. Marty Weiner: Right, right. And a few of the other crazy things we're seeing on the frontier. So I think we just need to start having, you know, we we we do frontier for all sorts of things on the show. The robotics is now starting to hit. So we're seeing robots assembling themselves, which is crazy. we're seeing now open AI, yeah. What bad Zach: Good. Marty Weiner: could happen? OpenAI is now gonna start having their own humanoid robot that was announced this week. they expect to have a cool demo in 2027. it says I don't think there will be robots walking the streets for a few years longer. That's what Altman said. and hey, you know, we'll probably have it next month is is what what's gonna happen. Because the the reality is there's a lot of competition for this too starting to pop up. there's the figure one now two robot. there's the company we reported on recently that was doing this in in homes. So I think once that as that competition really starts to heat up. I think things will start to move exceptionally fast because I think we finally have the technology to drive these, which we didn't have until a few years ago. a few more just for funsies. We now have robots fighting in the ring. Zach: Good. Marty Weiner: for our amusement, which is definitely gonna get us in trouble during the apocalypse, when these things figure out how to climb that fence. yeah, yeah, I'm sure that's fine. Yeah. Christopher: Let's teach them how to fight. That's a good idea. let's teach let's teach how to rough us up. Marty Weiner: but they are they are strong. Like you'll see that they're knocking off parts. This one lost its head. either strong or weak. I Zach: Yeah. Have you noticed that like before all this there was all this stuff like, well let's not teach the robots to fight and let's not let them have the internet and let's not and then like at some point it was like, But it'd be cool. It'd be really it'd be really awesome. Marty Weiner: It's so cool. yeah. Yeah. So I think we need to now start having robot frontier every single podcast, because I think we're finally getting there. And I I guess I won't be surprised if we start seeing robots in in the household and in maybe next year. Zach: Wow, next year? You think it'll be that soon? Marty Weiner: I'm gonna say next year. let me show you another cool one. Pollen robotics. Zach: Man. Marty Weiner: Chris, Chris, I think you have one of these. Christopher: Yeah, I have I have the I have the Ricci Mini. it's adorable and it's pretty simple to use. It's also running by the way, the hardware is it's a Raspberry Pi, honestly, inside. It's pretty neat. Marty Weiner: Right. So this this this may still be in the world of toy, but it might be finally toy that is actually kind of fun. This micro duck I'm showing on the screen here, is there one that's coming out in a few months? you could pre-order it now. and what's what's wild is even as of a few years ago, a little tiny robot walking on two feet, not super awkwardly, right? Like not big lumping feet that that obviously are cheating, was hard. Disney actually came up with one a few years ago that they had walking around at their parks. and now it's sort of like, yeah, just do a walk and and Christopher: Yeah, we just just do a walk. Marty Weiner: and I think the other fun part about what LLMs bring to the table is they can truly role play, like they can they can actually act as a character and actually be fun as opposed to a very simple program. I I remember there's one computer scientist told me once if you know if you're making a game and you want to have AI in it, which well or you want to have an NPC, you want to have a computer controlled character, you can get it pretty interesting off of about one page of code. But to go beyond that, it suddenly jumps into hundreds of pages of code or if not thousands to have it be an actually more interesting character. and now you can have the LLM just be a character, be a cute little duck, and and control this thing. and so I think that also changes the on the entertainment side. I think we could start to have more robots just for fun and entertainment. So anyway, the the robot uprising is coming and we will be either entertained or Or controlled. I think we should move to another fun topic. Zach: Yes. Marty Weiner: the apocalypse. So we have the apocalypse everything. Christopher: Speaking of apocalypse. Marty Weiner: Speaking of apocalypse, we we have a section here we want to cover all of the various apocalypses, the job apocalypse, the physics and math or I guess they're calling it the the math occlip, the physics occlys. we're starting to see well the the robot occlys is coming soon. First I thought we should just for funsies. guys, what does apocalypse even mean? Christopher: well, as the former recovering classics nerd, apocalypse comes from the Greek of apocalypsis, which actually means uncovering. apo means out of and kalyps is a verb that actually means to cover. and so what's funny about this is that up in you know, it if you if you look at the origins of the word, it's the Greek equivalent of the word revelation. Which is also why, for those who were also recovering Catholics like me, you'll find old Bibles where they actually refer to the last book as the apocalypse rather than as revelation, because it's actually Zach: Yeah. Christopher: the same word up until the twentieth century. which is also Marty Weiner: That's wild. Christopher: kind of funny because for all the modern usage, whenever you start using c like ellipse at the end of it, it actually just means cover. And so to some extent it's Zach: Covering it. Marty Weiner: Cover. Christopher: actually kind of on the nose because you know, with the math acalypse, it the LMs are now covering math in a way that's really kind of impressive. Zach: That's right. Marty Weiner: And can it can it also mean like to cover with a like a blanket or like like to shield? That is the actual well, 'cause that's what's scaring me because Zach: No th that is the actual sense. Yeah. It doesn't mean like reportage. Yeah. Yeah. Christopher: It's the actual sense of it, yeah. Marty Weiner: they're coming up with math we don't understand. so I I did look up what is the w Christopher: maybe the maybe the tortured analogy would be smothering the math community. and Marty Weiner: Yeah, well, Zach: There we go. Marty Weiner: yeah, right, right. And I I was curious, like, well, how did we get from that to what we how we use it today? And and according to the dictionary, it was that by the twentieth century the term had shifted just in common secular language, away from its meaning of revelation to violent world ending catastrophe or global disaster. So do a little more research how we got there. Christopher: Th there's the there's the overtone of the overtone of it because because the last book of the Bible has been referred to as Revelation or Apocalypse, people are more familiar with the Latin form, which feels much more like reveal. reveal, great. Apocalypse. That sounds like a dire word. It must be about the end of the world. which I think is the way it came about. Must be bad. Marty Weiner: It must be something bad. Yeah, wild. Well, why don't we start with the decline of entry level technical hiring at anthropic. Zach: Yeah, so this was cool. So Anthropic actually had a couple econ related papers or articles out this week. One was called Employment at Anthropic. And so the basic idea here is maybe we should look at Anthropic as a bellwether for what might be happening in the broader economy. And so what they found is that hires are down, headcount is down, and open roles are down. and and that is that is accelerated over the last year pretty drastically. Marty Weiner: Is that for juniors or or all across the board? Zach: So so so the basic finding is is perhaps not surprisingly that fewer junior employees are getting hired overall at anthropic. and and and then they have this cool graph about agents doing more of the work over time research organizations. They have nineteen times the human hours. and so they're they're trying to compare like how much work in human terms the agents are doing. And it's just going up and up and up and up. it looks like exponential growth. This is a a nice little logarithmic chart. and and then I pulled one more Marty Weiner: Mm-hmm. Zach: one more graph about this, which is is that they're showing how fast productivity is growing. And again, perhaps unsurprisingly for engineers, it's just gone wild. you know, as as you move down to more traditional roles, they're doing more, but but it's it's like a modest boost. but Marty Weiner: I see. So hires per recruiter has increased one point four X, but not like engineers, which is closer to five. Zach: Right. Yeah. Yeah. And you know, I mean it's like it would be it's hard to imagine a world where you like it's desirable to 10x the number of hires per recruiter, right? You know, so it's not exactly apples to apples. but it's still crazy, right? And so the the Marty Weiner: Right. And I do see c contracts per legal employee is on a nice little exponential there. Zach: Yeah, yeah, yeah. And and so, you know, what's interesting about this really so it's interesting that it's happening at Anthropic, but you would you would expect Anthropic to be the most AI forward company maybe on earth, right? Like they're they're probably pushing out to all employees. There's gonna be no stigma inside the company. There's probably the reverse stigma about using AI. Marty Weiner: Mm-hmm. Zach: And so, you know, the the question is, you know, are we going to see results like this across the economy eventually? So that's that's the first paper they had out about their internal experience. Then they did this really cool paper. Where they took a survey of just regular Americans, just like a big poll. And then they also took a survey of people who are internal to anthropic. And they just compared what people's intuitions are about what's going on now and what's going on or what's going to go on in the future, right? so let me just go some of these graphs with you. So what AI can do today. So, so you know, roughly sympatico, there is a bit of discrepancy. Generally speaking, anthropic. People think AI is a little better. Interestingly, with some categories like Nobel Discovery, regular people actually think it's doing better than anthropic people do. 12% of regular people think AI can already do Nobel quality work, whereas it's only like four percent of anthropic people. But generally speaking, for like quotidian tasks, anthropic employees think AI can do a lot more. 86% routine emails, which kind of surprised me. Like, do you guys use AI for email? I feel like that's extraordinarily rude, but don't let me Marty Weiner: I I do not, but I know a lot of people who Zach: Yeah. Marty Weiner: do, both to read it and then a little bit to respond to it. I I was thinking the other day won't it won't be surprising to me soon. You know how we used to say, bear with me, I'm writing on an iPhone at the bottom of our our email messages. I wonder if soon Zach: Yeah, yeah, yeah. Marty Weiner: we'll be saying, Bear with me, I'm not using an L L Zach: I got but maybe it's just the kind of email I do like, I get that for scheduling or something, but a lot of emails like I don't know, it just strikes me Marty Weiner: Yeah. Zach: as really weird to be like, you don't get to talk to me, you're talking to yeah. Marty Weiner: Kind of rude, yeah, yeah. Zach: But maybe that'll change, or or maybe again, like maybe anthropic's the bellwether, like this is just what the future's going to look like. But but so where it gets really different is is by 2027, most anthropic people think AI is doing everything, right? All routine emails, almost all bookkeeping, or or most of them think bookkeeping is taking care of tax returns contracts. 58% of anthropic employees think in 2027, that's next year that AI can do Nobel Prize level work compared to 22% of regular people. and by 2030, basically, like anthropic employees see the world as saturated on all these benchmarks, whereas regular people don't. And so it's, you know, because we're talking about percentages, this is a little weird, but like eyeballing it, one way you could say this is that regular people anticipate linear growth over time, that just AI will keep like improving Christopher: Hm. Zach: by steps, and anthropic people see. exponential growth. everything just going to a hundred by twenty Marty Weiner: Well Zach: thirty. Marty Weiner: I I wonder if another side of it is there's a certain kind of person that's gonna work in anthropic. You're talking about a non uniform Zach: Yes. Marty Weiner: set, right? And out in the real world, I'm sure when you're doing this, you're gonna run across a couple of people who are like, Zero, zero, zero, zero, it all sucks and and Zach: Yes. Marty Weiner: and AI will die because people will realize it's it's all garbage. So I wonder, I'd love to see the nuance and the data on the on the split there. Zach: Yeah, that that would be nice. And and and and yeah, I think probably the people s who who apply to anthropology are not in any way typical people. They're very unusual people. so so that's probably affecting their judgment. Yeah, Marty Weiner: Well, also they're they're they're cool with AI. Zach: well, and and actually one other thing I wondered about with this. So I was reading Kevin Roos's new book about the last few years of AI. And one thing I think I would have thought of as conspiracy theory stuff before it turned out it actually happened is that they actually like when they got clawed code, they sat on it for a while. Because it it was almost like solving an optimization problem. It was like, well, let's let's see if we if we keep it to ourselves, we get this much ahead. Whereas if we release it, we get this much money. And so there's some kind of, you know, sweet spot. But they actually kept it internal for a while on that basis. So like when you see anthropic people like like there was this famous video of Dario Amade, you know, saying, I think in a year, ninety percent of the code will be AI. What we didn't know is that they were already using it internally. So it wasn't that impressive Marty Weiner: Mm-hmm. Zach: of a prediction. So I wonder Marty Weiner: Right. Zach: with some of these numbers is part of the difference is they're actually seeing, you know, a model ahead from the rest of us. Yeah. Marty Weiner: The next thing. Right. Right. Right. Zach: and then just a couple more cool graphs to go through on this. So, like they they asked regular US adults what's gonna hold back AI at their job. And the top thing people mentioned was customers prefer human. and Marty Weiner: Yeah. Zach: a little down from that was privacy. and then and then you get into more regular stuff like is is it reliable, is it legal? just is just friction, workers refusing to use it, and and and legal issues and that sort of stuff. but but but did seem to the the overwhelming things people cite is like I wanna talk to humans, I like talking to humans and I'm scared to trust anthropic or whoever with my data. Christopher: I think there's a possible, there's a possible coup here though, where because the because the LLMs themselves are so I don't want to say necessarily obsequious, but generally kind of like on that, Zach: Mm-hmm. Christopher: on that side of the spectrum, I can almost imagine that the customer service could flip really easily. Like I mean, I think we've all had the experience where you call up and you get like a really sur like you get the person on cut on the far side who's Zach: Uh-huh. Christopher: either had like a really bad day or just like doesn't like hates their job for probably justifiable reasons. You'll never you'll never get that with an LLM. Now that that doesn't mean you'll get your problem solved. and Marty Weiner: huh. Christopher: in fact, you gotta worry about like alignment and safety being a big problem. You're not gonna end up with those in those moments where the humans is gonna give you like a solid. But then again, like I don't think the even the humans have a lot of real like range to give you a solid anymore. They're kind of constrained on what they can do. So to some extent, like Marty Weiner: Mm-hmm. Christopher: I can imagine that flipping pretty easily once once the like once the voice models get just like a little bit better, they're pretty close right now. Like there there are there are moments where I've definitely had the robocall that feels very human, but they'll self-identify Marty Weiner: Yeah. Christopher: up front and then I can actually kind of hear it. Marty Weiner: I had one where I was pretty confident it was an AI, but I w not enough that I could just feel comfortable hanging up on them. huh. how about this autonomous how yeah. How about this? Zach: Huh. Yeah. Christopher: See that's the difference between you and me, Marty. I have I actually feel better hanging up on the human. Zach: Don't want to hurt the AI's feelings. Marty Weiner: Zach, you've got how autonomous will AI be? Zach: yeah, basically the the the short version is for almost any task, anthropic people think within a few years it'll just be AI top to bottom. There's this cool graph. By 2030, anthropic employees think AI will do 75% of all their work. But actually, to me, like equally crazy. If you take regular US workers, this is a cross section of US workers. people think about a quarter of their work will be AI. And it differ differs by the trade, right? And it differs by like the experience of people, how much they use AI, but still at the lowest was a quarter, and that was manual labor production and transport, right? So like every regular people are predicting an insanely changed world, right? Where fully a quarter of all the work of manual work is done by AI. and so You know, so they have this cool bottom line stat that I can I get I'll get into briefly what economist friends were saying, but like I'm just gonna read this. The median US adults answers imply GDP growth of five point four percent in twenty thirty, according to the model. The median staff members, that's anthropic staff members answers imply GDP growth accelerating to seventeen point one Marty Weiner: Wow. Zach: percent by twenty. If you believe that, you should just stop doing anything. You should mortgage your house, whatever, just get cashola. Go live it up because the future is so wildly rich, there's no point in doing anything anymore. We're all gonna be making ten million. I mean, like a world with seventy percent GDP. Okay, so so if people don't know, a great or a good year for GDP growth in the US, a frontier economy is three percent. Two percent is like a meh year. and then then on either side of that is is better or worse. Seventy percent does not happen. It it just doesn't happen. Okay. Th Marty Weiner: Right. Zach: this is the whole economy doubling in like three th what like three and a half years? I mean math, but like Like it's insane. Like if you really believe that, you should just quit, right? Marty Weiner: Right. Unl unless Christopher: Well here's a here's a Marty Weiner: you believe that that's gonna be distributed nicely to all of us. Zach: Well, you right, sure. Well but no no Christopher: Well I what Zach: even then, even then, right? So okay, I mean unless you want to imagine like an evil feudal apocalypse world, like just put your money in index funds with GDP growing twenty percent a year. You'll you'll retire four seconds later. Marty Weiner: Mm. Mm. Yeah. Christopher: I mean the the the the you know the going back to where we started with robots, there is there is exactly one quadrillion dollar field, and that ends up being human labor. Like if you add up all of the work that humans Zach: All of it. Christopher: do, like like say like like just take that entire sector of the economy, that is like a quad like it's north of a quadrillion dollars, which was I I believe like was actually cited as one of the reasons why Tesla decided to go into robots in the first place, was like if you wanted to like Like, say what you will about Elon, but he knows he knows how to like go all in on disrupting a field. and his intent here is like, why don't we disrupt labor with robots? Which does make it a little bit difficult to figure out like what's what's left for the meat bags to do. Zach: Like, so so what what an economist would say? Like, so because a bunch of economists were like, this is stupid. Like, even the low estimate or the high, you the low estimate, a mere five percent of years is kind of stupid. And the basic claim would be like, you're right, like if if robots can do a bunch of stuff and software can do a bunch of stuff, we're all more productive. The question is like friction. And and a, you know, like like are people actually going to adopt this stuff? But then also, you know, there's this old problem that which is like, why is it that If you hire a guitarist for a bar mitzvah, they get paid like 10 times what the same guitarist at the same bar mitzvah got paid a hundred years ago. So why are they making more money? Because they're just playing the same stupid guitar, maybe literally. And the deal is you have to pay them not to be in some other more productive remunerative job, right? And and so, you know, you can't underestimate this effect, right? When there are jobs that you need a human to do, they start costing more money. And so there is some world where you think it's gonna be You you know, we're all going to be trillionaires, but actually the the jobs that humans can do become the overall economy and you don't get this crazy growth rate. It's still a wonderful, better world, but but these these zany numbers that get thrown out by tech sector people, I think are are are questionable. I I hope it's true. and by the way, just as a little econ tidbit, one of the predictions, if you really think this is happening, if the average investor thinks this is happening, you should see interest rates go crazy. Right? Because if you think right, if you think the future is going to be Marty Weiner: uhhuh. Right, you start Zach: wildly rich. Start borrowing, right? Either to like Marty Weiner: Yeah. Start borrowing, yeah. Zach: start borrowing to make your life better or to buy part of these data centers. And we're not seeing that either. Marty Weiner: Uh-huh. Yeah. Right. Well, let me let me speed run through some of these other apocalypses real quick. and actually Zach: Yes. Marty Weiner: I I think what's interesting is is the word is bothering me a little bit because I think what it it it sort of implies is the death of these fields. And I don't think that that's a good word for it. I think a better word would be something like intense, exciting transformation. Because we'll show this one real quick. a couple of weeks back, it was a couple of weeks, a couple months, there was the big open AI. paper where it had addressed the quasi Riemann hypothesis took 200 pages, was described as AI slop, but who was reporting this one? Zach, was this you showing that they had reduced it? Zach: Yeah. so so I just I follow a lot of math nerds and and and basically Christopher: This a co this is cool result. This is a cool result. Zach: it was neat. So this quasi-reman hypothesis comes up, people are like, what is this confusing alien slop? And about four seconds later, some nerds had gotten it down to six pages. And what's neat and kind of fun is then a day later it was five pages and four pages, and now it's down to three pages and it's like basic number theory. Not no, I'm saying this like I understand it, but it it's it's what's interesting is is is just like that it's it's clear the joy is still there, right. that that you can you can take this confusing result and then the fun is like you know, in the way a poet does, right? You you you you see a sunset or something and you're like, Okay, how do I make this good? How do I convey this, you know? I think it's cool. Marty Weiner: Right, right. And what I think is neat is I I don't I need I don't understand this stuff and probably never will, but two hundred pages of little tiny, you know, conditionals and all that is is not very in there's not a lot of interesting what's the word I'm looking for? intuitions that have been created there or insights. Insights is the word I want. but Down to three pages. I I would love to we need to have Jonathan back on and and ask him like, okay, there's some cool insights here that that have actually been derived. But also, you know, this came from humans, right? It was the first one was by from the AI, and then the humans reduced it to three pages. So I that's what's really cool to me is we were talking we were predicting before for a while, math will be this cyborg thing, this back and forth between the AI and the humans, I think. but it is also it's yeah. Christopher: Which is which is how software development has evolved at this point too, and which is what you're seeing in other fields in the same situation. It's right, you know, you don't like the legal the legal example is a good one, right? Okay, so paralegals as a field are probably disappearing, or at least they're becoming much more constrained because it it's very easy to have an LLM write a contract now or to read a contract and summarize it. That doesn't mean that the work is going away. It means that the that it means a lot of the the toil is going away. Now, I don't think mathematicians see themselves as toiling, but I do think there's like an there's an intrinsic sort of cynicism in all of this that like somehow all the like If you were to cast this back 40 years, back when like everyone still believed in progress, you'd have this belief that like, my God, in like five years there's gonna be no more jobs. It's gonna be great. And now we're kind of like, my God, in five years, there's gonna be gonna be no jobs. What are we gonna do? It's like, well, it's Marty Weiner: There's gonna be no more jobs. Yeah. Right. Zach: Yeah. Christopher: like most maybe most people don't like their jobs. Like maybe it's actually Zach: Yes. Christopher: like a good result. Marty Weiner: Right, right. Well and Christopher: Well the trick of is like how do we feed ourselves in the process of doing this? That's that's the social contract is also not there necessarily. Marty Weiner: Yeah. And it's also Zach: Be nice to eat. It'd be nice. Marty Weiner: it's also a little freaky along the way, right? Because you what you understand to be your field that you loved is is gone. We we had some interesting examples from the math community this last week. Terry Tow basically gave in. He said we're seeing the end of the romantic era of mathematics, the age of heroes. Cornell Yeah. huh. Zach: Yeah. That I can I that video is worth watching, 'cause 'cause it really does feel kind of like a kind of grief, like he's describing a thing that's passed now. that that math is now going to be a sort of like giant group endeavor and not no no more heroes. Marty Weiner: Right, right. But I'm really curious to see what the next thing looks like. I wonder if he'll find new excitement. like especially like with the qu quasi. Yeah. Zach: Yeah. yeah, I think so. Yeah. Marty Weiner: there's a Cornell Prof talking about how everything is beautiful in it in its time, It's definitely worth a watch as he's you know sort of coming to grips, I think, with the changing of this field that he loves. and then finally, field medalist Richard Borchards had a interview and and a couple of interesting quotes that came out of it. He says it's fairly obvious that if things go like this in a few years, humans doing math research will be redundant. So I don't I just don't see that happening. I I see it evolving. And then he also says, I think what people are getting upset about is that they're slowly realizing that they are dumb and somehow blaming Zach: Yeah. Marty Weiner: this on AI, even though it's not really AI's fault. So I I think it's almost like also a grief of I've I've felt this too, you know. We build sort of our identity, especially like in the computer science world and nerd world on being the smart one in the room. And I I do feel like that's slipping away, right? If the AI can Zach: Yeah. Marty Weiner: explain it just as well as I can. Nobody's gonna come to me to say, hey, can you explain you know electronics? Zach: Right. Marty Weiner: They could just go to the AI model. And so my utility in the Zach: No, I I I Marty Weiner: community is going. Zach: the what it reminded me of, do ever see that Futurama episode where like Fry, who is the stupid character, it's like flowers for Algernon, like he becomes smart and and they Marty Weiner: huh. Zach: have this thing where they talk about like there's this instrument that makes pictures and music and only a few people can play it and they all suck at it. and I feel Marty Weiner: Ha ha. Zach: like that's like that we're discovering that's humans at math. Like we had a few people who could do math, but we all actually just suck at it. There's just there's just degrees of suck. and now Marty Weiner: Right, right. Zach: now there's an entity that can actually do it. and that's that's kind of upsetting. Marty Weiner: So I suck more than many, but but we all suck. Zach: Yeah, that's yeah. Marty Weiner: and then the physics occurpse really began. I think last week we started seeing a lot Zach: Yes. Marty Weiner: more, but we're seeing Chat GPT Astra solve 49 out of 50 unpublished theoretical physics research problems, which is what we saw with math, you know, a couple of months ago. Christopher: And what's what's what's what's better about this, the best part about this too is it not only solved them, but it actually uncovered mistakes and initially got some of them marked wrong. And then when they re-reviewed Marty Weiner: Yes. Christopher: it and they found out that actually no Astro was Zach: Yeah. Christopher: right and that the original the original result was wrong or that some of the assumptions were Marty Weiner: Actually was right. Yeah. Uh-huh. Christopher: wrong. Zach: Hmm. Marty Weiner: Right. Another one is autonomous Claude was used to find a hundred and four new eclipsing binary stars and came up with a new method to find them. This is not peer reviewed though. it was just some some enthusiasts talking about it. But if that's true, you know, it's it's starting to crack into astronomy. then Google Gemini four, you know, kudos to to Google for showing up. cracked a problem Zach: Yeah. Marty Weiner: in plasma physics. which Chris, do you do you know what this means for us and maybe fusion? Christopher: So so fusion, fusion is the is the problem in nuclear physics that is always 25 years in the future. We're always 25 years away Marty Weiner: Uh-huh. Christopher: from nucle from solving fusion. I actually was gonna do a at one point a a plasma physics degree. And at that point, 25 years ago, it was only 25 years away. the interesting thing here Marty Weiner: hey, we're there. Christopher: though, I know it but it must be ready to go tabletop. the interesting thing here is that the the ultimately the problem is the math is very hard. And doing these simulations very difficult because you're talking about a very like plasmas are very strongly interacting fluids and they're kind of the worst of both worlds. They're very hard to deal with and they're very hard to simulate. And so it sounds like that there is now on the math side of this and the theoretical side, Gemini is making progress on simulating plasmas in a way that actually could have a breakthrough possible for fusion, which is very Marty Weiner: Wow. Christopher: interesting. Marty Weiner: I mean fusion alone would change it'd be another you know, AI revolution size revolution. and then what about this quantum gravity drive that Claude audited? Christopher: So here's a fun one. a this is I think this is a good example of how cyborg in in the sciences and in math could work really well. So here's an example of a researcher who is a quantum gravity researcher. Quantum gravity is at the kind of like the edge of physics where we actually don't know what the answer is. originally had spent a good 10 years working on a result by hand, because we're talking about like probably 10, 15 years ago, and then decided to basically employ Claude to do an audit. on on on the on their work. And in the process, Claude not only did it like they they did this all in a very short period of time compared to much less than 10 years. And also Claude found a bunch of simple mistakes throughout their work. It didn't invalidate the result, but it definitely kind of like dotted I's and crossed T's and made the work much more like durable and more important as a side effect. So it's it's a it's a very interesting result. I think it's actually a good demonstration of like with our powers combined, we are better. Marty Weiner: Yeah. Well Zach: For now. Marty Weiner: what's wild to me is I feel like physics is also Christopher: For now. Marty Weiner: these kinds of changes are closer to changing our daily lives. I I still don't know when I'm gonna get a flying car. I'm a little upset at Christopher: Well, maybe not the quantum gravity result, but the the fusion result could be like if you can start to do interesting plasma physics results and come up with novel experiments. I do think that the the difference between physics and math here is that with math, you're kind of like invalidating the way that the field runs where you do have these heroes kind of sitting in a corner thinking really hard. With physics, when you have a new theoretical result, it's like, cool. Claude, tell me how to do this in an expert, like how to build an experiment. And then you actually have the fun of building the experiment and actually like testing out whether the result is true or not. Marty Weiner: Right, right. So those experimentalists are gonna have some work to do ahead of them now. so a Christopher: could be. Could be. Marty Weiner: few other crazy th news in the world of apocalypses, character date.ai collapsed. we were talking last week about the AI bubble and whether or not it existed. And I think at least some of the predictions that came out of that is that the overall thing is healthy. well, healthy is not the right way. It's moving along in a way that does not to me look like a bubble. But there are a lot of little ancillary things that I think will start to collapse. We see all these companies that that are popping up to sort of you know get the last little bits that are you know not tackled or make it accessible to humans. if you go look up C Dance, if you want to go use C Dance for making your own AI video, there's probably hundreds of companies out there that are like, come use our company to to make AI video, but they're all at the part that they're the same and they're just using someone else's technology. So what they really are is, you know, Chat GPT six with a small wrapper. And I think we're starting to see some of those collapse. on the broader economy, yeah, go. Christopher: Well and also with with with with character AI though, like I think the specific thing here was like I don't think this isn't this is necessarily an LLM issue. I think this is an issue of their product was making some questionable like there's it's b character AI was basically a chat bot that became really hot among teenagers and among people to like kind of have a have a bot they could talk with with all these personas. And much has been covered in the LM press over the last several several months and years. It was coming back with some very questionable things. and like definitely. Definitely amplifying some more self-destructive behavior. And they actually ended up getting crushed by lawsuits because their model was not particularly well aligned for that situation. Marty Weiner: Hmm. That's wild. on the broader world, we're seeing Kroger had an article of Christopher: The broader occalypse? Marty Weiner: the broader ocalypse. Zach: Yeah. Marty Weiner: Kroger fired 3,000 corporate jobs. a little hard to tell sometimes with those what the actual reason was, but they did claim that AI was part of it. And you know, that that's that's easy to believe, right? I mean, you how many accountants do you need if if the AI can start to do more of the accounting job? and we're also seeing call center worker decline. global call center Zach: Yep. Marty Weiner: employment is contracting. more than after more than a decade of growth. and and and rapidly declining. like it looks like we're seeing a sharp exponential Zach: Yeah, I mean like usually usually you don't get a smoking gun like this. You know what I mean? Like usually that you have like squint a little. Like this is like consistent four percent growth almost every year for decades and then wham, twenty twenty four hits and it's the opposite direction. Marty Weiner: Right, right. And then on the macro scale, the labor force participation rate. Zach: Yeah, I just wanted to put this in. So, like, you know, we are gonna we're using the term job occlipse. probably the best term to look at or the best graph to look at is prime age labor force participation rate. That just means of the people who are 25 to 54, what percent as like just a big dumb number, right? Not people who are looking for work or not, just how many of those people as a percent are in the labor force. And if you look at this graph, you don't see AI. so we're we're worth noting at you know, at least we, you know, we we hear all these reports about sectors getting messed up. But like on the broader economy, it looks exactly like it looked in nineteen ninety eight. So, you know, maybe don't get too excited. Marty Weiner: Right. Huh. It also doesn't really show the dot com but well I guess the internet coming along or mobile. Zach: No, no. I mean, like, so we can go through this. Like, essentially what you see. So we have data going back to I think like the late 40s. essentially what you see is women come women coming into the workforce over time. And that that hits the the the current rate around 1990. And then, you know, it wobbles. part of that's like people stay in school longer now than they used to. and but but basically, you know, we've been in this band from like 80 to 85 since women entered the labor force. And it it, you know, wobbles, it changes with recessions. You can see the great. recession really messed things up for a while. but otherwise it just looks perfectly normal. Those people are getting jobs. and and and wages are pretty good. So so, you know, it it's I would at least say from this graph, it's way more the economist picture than the tech nerd picture. Marty Weiner: Yeah, it'll be interesting to see what happens in the next few years. let's Zach: Absolutely. Marty Weiner: see, all right, Chris, can we see inside an AI's mind now? Christopher: Okay, so this is a really cool result. this is a this is a very long write-up. so Nick Camerata on his on Goodfire has a write-up. He's actually he's been an AI researcher for a very long time. And one of the questions that has been we've been struggling with with LLMs and with with with any kind of large model is basically like what what does each one of the neurons in the model do? and so it there's this this kind of a struggle of, and I I often joke it feels like we're you know, we've we've gone from doing computer science to doing like LLMs or neural network psychology of like trying to figure out like how does this big network do what it's supposed to do? And for a very long time it was assumed that models for any if you look at any given neuron in it and you try to identify what causes this neuron to fire. And just to keep it just a just for for context. A neural network really is a bunch of these little modules that are all assembled in and they're not real modules, they're software modules, that have inputs and outputs. And the all they do is they light up or not given their given their inputs. and that passes on to the next layer. So you have these layers and layers of these little switches that turn on and off depending on what the stimulus is coming in. And the belief is that there should be some the like j in organic systems, there are generally specialty neurons that light up for certain reasons. So for example. Your visual cortex is very well organized. And there's an entire set of neurons that are determined that will determine whether or not they see lines that are vertical versus horizontal. And then all the angles in between. And in fact, you can actually determine how good you are at measuring angles by looking at the density of what kinds of angles you have for neurons, right? So we're really good at determining perfectly vertical, perfectly horizontal, and we're kind of crap at around 45 degrees because it actually maps to we have a lot of we have a lot of a lot of specificity on those vertical and horizontal. We don't have much on the edges. The question is in an artificial system like a neural network that we've trained, is there the same kind of specificity? And for a very long time, the assumption was no. It's just kind of a mess. And so they had this notion of well, they found examples of a neuron that would light up when it showed when you showed pictures of a cat or pictures of a car. And so it's like, well, how do you determine Zach: Mm. Christopher: what exactly this neuron does if it's going to do those things? And the basically the the the assumption was. well, you know, for for given different given inputs, you've basically overloaded this network to do multiple things. And overload in the computer science term sense of like it can do two different functions depending on the inputs, right? Well, it turns out that that might be wrong. so first off, it might be an artifact of older models which were a lot smaller and where space is much more of a premium. And so there's an there was potentially much more, you know, almost compression going on where there's no place Zach: Mm. Christopher: to cram this knowledge, therefore you kind of have to reuse the same neurons for it. The second thing is that it might just be a a question of figuring out like what are the actual features that tip off that neuron. So, for example, in the case of the cat, what what the researcher discovered was well, actually, it's neither looking for cats nor cars, it's actually tricking off triggering triggering off of a sort of a V shape that's common to both. And so visually, cats and cars don't look similar. But there are subsets of there's a there's a set of features of a cat in a car that are similar. And they actually demonstrate like, if you just show it a V, it actually lights up these neurons. So now it actually pushes the problem from and then they they started basically going, okay, well, here's an example where we didn't find any weird, they call it superposition when you have two different roles here. We didn't find any superposition here. So let's pick a different neuron and see if we can find an example an answer for that as well. And so now it's gone from writ large, what's the psychology of this large tangled model to actually for any given neuron, it's practical to start finding out what's the story behind what this neuron does. And this will probably be something that we can actually assist with LLMs to figure it out. But it means you get you get you you can actually make this problem reducible, which speaking as a former physicist is like the freaking holy grail to solving these kinds of problems. Because now you can Zach: Yeah. Christopher: get like you say like you can s you can assemble it from its parts. You can say like, look, here's the section of the neural network that's actually acting like a visual cortex. And here's the section that understands like standing up versus sitting down as a concept. And so it's almost like every single neuron in the network actually probably does have a discernible purpose, or at the very least, so far in this new line of research, it was demonstrated that they couldn't find a counter example. Zach: And is that like like Marty Weiner: Wow. Zach: so, you know, in in principle then if we keep scaling up the model and and you know, the world itself stays at the same level of complexity, at some point it gets the like the grains get even more tractable the larger the model gets? Christopher: I think that's one of the beliefs here too. And I think also the the idea is that as you get more as you get more neurons, you also potentially get more abstruse concepts that will that will be represented in the network as a neuron or as a set of neurons. so for so another example and given in given in this in this again, it's like it's a it's a long read, but it's a really well written read, and it actually is very approachable, even for somebody who doesn't do this as a as their career. they give the example of you know, w to some extent, I would say. We we're we tend to be very literal. And the assumption was, it's it's mapping to car or cat, whereas it's actually mapping to small feature. They also gave examples of it was actually mapping to something like pose, like not how not what not what the image was of, but how the image was structured. So they gave the example of they found this pose example where, like, you know, kind of like laying down as a notion would trip a neuron off. Marty Weiner: Uh-huh. Christopher: Which is like a which is a concept that we wouldn't necessarily attribute to something you would have a neuron for, but actually makes a ton of sense. Like you're you're not talking about just just like literally looking at pixels, you're looking at like composition, which does get back into the you know, when you start thinking about things like the artsy side of things, they have notions of composition probably built into the network too, which is an interesting side effect. Zach: Yeah, so so like a lot of the talk about this was is this gonna allow us to solve super alignment? And I I assumed the idea was essentially, well, in principle, it's really hard to stop some machine that's infinitely more intelligent than you. But if you can look inside its brain all the time, then you too have power. Christopher: Yeah. Zach: so could you is is that true or is that just hope or is that, you know Christopher: I think it helps with it does help with okay, so for alignment, what makes alignment difficult, especially now with the models that are sufficiently advanced, is it's very difficult to tell when the model is just lying for you, lying at you, lying to you, right? It's like the it's like the bench maxing problem, right? If you just are if you're just trying to push a test and the test becomes your metric, it no longer becomes a good test. And so it feels the same way with alignment, where if you're gonna say, we want to make sure this model is safe and that it's not gonna suggest when you say like, I have suicidal thoughts. It's kind of like, that sounds like a good idea. that's like the basis of safety alignment is like, don't encourage the human to to do destructive things. don't don't explain to them how to build a bomb, right? And the way that they so it's almost like historic, like historically, the last couple of years, all the way in the deep past of 2024. The solution Marty Weiner: my god. Christopher: here was to effectively try to identify these paths and to kind of fine-tune them out, which is the kind of model equivalent of like a lobotomy. You're kind of removing concepts willy-nilly. Trying to identify like what section of the neural network do you kind of burn out so that it doesn't do things that are going to be unpleasant or off base or inappropriate. Here, what we're saying is actually more like, no, no, no, we can actually start to identify the pathways and the actual reason behind it acting those ways and remove those specific pathways. So to some extent, like I the lobotomy analogy is probably pretty on the nose here, where it's like historically, you know, you look at lobotomies, they were just like, we don't know what this does. We're just gonna cut it out, and all of a sudden people become adostile. It's great. it's like here you're saying like we don't want DOSL, but we do want them to like not, you know, do a do an anthrax, for example. Zach: Right. Yeah. Yeah. I mean what was coolest to me about this, like obviously we didn't we're not there yet, but there's this old question of how can a a weaker intelligence control a stronger one and it's like I I hadn't ever thought about like, well, you could build this other asymmetry, which is I have access to your weights and you don't have access to mine. which is really intriguing. Christopher: Which is very intriguing. And it's also it's it also does have potential implications for what does this mean for neuroscience as well, right? So like if you can start to do this for 'cause you we we're we're still actually fairly under emphasis on how how the human brain works, there's an assumption that there's the same sort of like superposition of of function. For the biological system, it might be true. In fact, there's actually a lot of evidence to show that, given how much you can cram into very small organisms, like, you if you look at look at like nematodes who have like, you know, 30 neurons. who can exhibit some really complex behavior, it's probably because there's a whole bunch of layered systems that are all being, you know, reusing the same neurons for different purposes. The point is that might not be true for neural networks, which actually means that they're more scrutable than we are. Zach: man. And and and would also imply that like if our brains were twice as big, we'd be we'd be able to understand them better. Christopher: Potentially. I mean, it also has Zach: Yeah. Christopher: implications for, I mean, the other side of this that I think is probably undersold is hey, if you have a what what is what has at least irritated me for the last five years has been that, you know, we finally got AI and it's ununderstandable. There's always this belief that you'd have this maybe a breakthrough and it's like, we finally understand intelligence. We finally understand consciousness. And it's like, no, no, no, we just have a bunch of weights that are all acting intelligently together. It's like the worst example of the Chinese room paradox of just like they're just like they're off, they're off being smart. Now we might actually have a way to go through and figure out like a little bit more of what is intelligence. Like what is the what is the ability to think in this situation? It's not just a bunch of weights, or at least it's an intelligible problem. We might need LLMs to help us understand it, but it would be great to get to the situation of yet LMs help understand it. And then you finally write the three-page paper like we did with the with the with the re with the Remont hypothesis. Zach: Extremely cool. Marty Weiner: All right. Shall we do news s hyper lightning round? Super Zach: Yes. Christopher: Super fast. Let's play it at let's play it at chipmunk speed for the for the podcast. Marty Weiner: fast. All right. Okay. Done. so fable five point five rumored to be coming soon. in fact, if it follows the same path with past shows, it will come as I'm editing the damn show. So I'll have to put a little asterisk on there. Zach: Good, good. Marty Weiner: next one security. 406 scientists write letter demanding AI be cut from the internet. So it's you know, it's our weekly news of somebody asking to slow down, which won't happen. Zach: Yes. Marty Weiner: but thought that was kind of scary. god, it's really bad. then well, Christopher: It's gonna really wreck felony bench though, I would point out, if they do that. Zach: That's true. Well we won't saturate felonies. Marty Weiner: speaking of terrifying, anthropic ran a test with self-replicating small models that could be put on Android phones, meaning no centralized locations. So aka you know. the what used to be called a worm, you know, right, right, right, which is scary. For funsies. Christopher: It was a it was an LM worm. They built an LLM worm. for for fun I mean for funsies. This is like this is why they're being told to unplug the machine from the internet because they keep doing these things where it's Zach: my god. Marty Weiner: And you yeah. Christopher: like, we're really terrified of what this can do. Let's build a worm net out of LLMs. Zach: D do you think the world is gonna end? Marty Weiner: Stimul. Zach: The world is gonna end 'cause Anthropic is gonna be trying to demonstrate that the world could end. And Christopher: Yeah. If I were gonna try and destroy the world, how would I do it? Marty Weiner: See, we told you so. Yeah, yeah. also other news. So you we saw all that you Ukraine launched a strike on a data center in Russia, apparently for the company Yandex, and immediately Russian bot activity came down eighty-four percent. So Zach: Yeah. Marty Weiner: including a lot on the the right wing crazy sphere, the the the far stuff. Christopher: the great the great part of this too is there's a really great Cloudflow article on how not only can you see the if see them go down, but you can see them moving to other data centers in the p in the course of the next couple of days. It's fantastic. It's fantastic. Marty Weiner: yeah, yeah, yeah. Start to move. Fake fix my system. Zach: Yeah. Marty Weiner: My money's going away. then DHS requiring civilians, federal authorities to uninstall Deep Seek, which makes sense. like Deep Seek is is open weights Chinese model. I I'm a little surprised that they have to that I know that was what really got Zach: I'm surprised they had it at all. That's crazy. Marty Weiner: me. then in the world Christopher: But it is but the flip side of it is like if I can play doubles advocate it's like it is an open weights model. So it's not like it's actually phoning home. So I mean I guess like if you're maybe Zach: Yeah, true. Christopher: if maybe if you're using if you're using the office shelf version that is actually phoning home, it makes a lot of sense. Marty Weiner: Well, the part I was more worried about is well, what could be baked into it? Like could we have a little bit of screw with democracy baked into it the same way that the Russians are trying to s, you know just s yeah, yeah, because you don't know it's alignment, Zach: it's like slightly shifting your answers. Yeah, yeah, yeah. Marty Weiner: so you shift the alignment internally, right? Zach: Yeah. Marty Weiner: all right, made with AI. I wanted to show you this really cool thing. So so D comps have become the cool new thing. There's even R slash D comps. It's this take we talked about it last time a little bit, but I'll talk briefly about it. you take an old game that you loved but has Not work for a long time, or you have to emulate, or you want to actually add something to it, the AI then actually turns it back into source code, meaning like human readable code. Then you can the AI can then help you make modifications and then it recompiles it, meaning it turns it back into something that can work on the computer. And there's a lot of implications from this. One is this cool group made a harness called Co-op Anything, where you can take any old game you love and it helps you make a co-op. So they did this with Ocarina of Time. Zelda Ocarino Time. So they made it so you could play the original Zelda Ocarino Time with a buddy. and so I I thought that was really insane. Then one cool thing is you know for playing these old games usually you have to use what's called an emulator meaning you know if it ran on the old Nintendo, the NES, you know modern computers can't run that code because it's it's a different you know assembly, it's a different machine code. So, what you do is you run it on an emulator, which then converts it live to your modern computer. Well, we don't even need that anymore. You can instead decompile it, recompile it into your native computer. So now there's a whole group called Mac Native Games because a lot of games weren't made for Mac. So they're starting to get those to work. Someone got tier Zelda Tears of the Kingdom running on a Mac natively with really good frame rate because it could run natively, which is insane. and then old PS2 games are being decomped and rebuilt so now they can run natively on on modern computers like Simpsons Hit and Run, Grand Theft Auto Vice City. so that that that whole world has gotten really exciting and you should go dig in. other fun stuff. so in the art world, which we didn't get to really dive into the art apocalypse, but maybe next time, San Diego Comic-Con banned AI-generated art from its 2026 art show after a backlash. I don't know. I'm gonna have to predict that by next year this will have passed, but we'll have to have a whole discussion on that. And then and actually leading in leading into that discussion, one of my main points is one interesting area that AI art is starting to crack into is the ability to go and fix what corporate America has screwed up. And the the one example I love is there's this this one guy is making new Old Star Wars. So you got original Luke, original Leia, original Han having Christopher: Ha ha Marty Weiner: adventures. And what's crazy, this video is really decent. Actually, I mean it needs a little bit more work, but I mean, compared to it's cartoony to even say that because if you compare it to what we had four years ago, I mean it's insane, but a little bit more work, and I'm sure they'll get there. It just feels a little stiff, the acting and the audio, but it's getting there. But the comments are what gets me. people are saying this. Star Wars stuff that the people are putting out better than seven, Christopher: Here's the Marty Weiner: eight, and nine altogether, meaning the episode seven, eight, and nine. People love this guy's stuff that he's making. It someone also said, Now this is good use of AI, not slop. One person said at four minutes and 31 seconds, and that's the point where tears are starting to wash down my face and won't stop. Somebody else said, Anyone Christopher: Mm-hmm. Marty Weiner: else crying their eyes out? Because what this video was was sort of like Luke and Leia finally finding, I guess, about their mom. and so they somebody was saying this is all the same person said, This is all I ever wanted. So this is crazy. Zach: Right. Marty Weiner: So finally, you know, Disney took a hold of Star Wars and to put it bluntly, screwed it up. And now this is people saying, No, now we could take our hand, we could take it back to the community. I'm sure Disney lawyers are gonna be freaking out about this, but I'm not sure how much they can control it or how much they want to control it. It could be this becomes a whole new, you know, world for for Star Wars that the fans actually always wanted, but you know. could not be created by by corporate America. let me finish with funny and we yeah. Christopher: I mean, speaking of s speaking of somebody who did actually enjoy the sequels but didn't enjoy the prequels, why don't they just redo the prequels finally? Like I would just like to find like if you want to talk about stiff acting, Marty Weiner: Can we just fix them? Yeah. Christopher: let's just redo Zach: Okay. Christopher: let's just redo the first three. Marty Weiner: Yeah. I I would throw out yeah some good money for for that alone. And I bet it could you know maybe with a human writing this a better script and then and then using this to actually make it into a video. Anybody else? Yeah, yeah. Zach: With anyone else writing the script? Christopher: Hey hey Claude, do a screenplay. Marty Weiner: we might have that ne by next week. and then finally, in funny and weird, these are fantastic. This guy he made it pulled in two AI job interviews, which by the way, this is a thing now. You have an agent doing the job interview. He pulled in two of them into his one discussion, his one video conference and walked out the door. And then these two AI agents basically spend the other the rest of the time trying to figure out what is going on and they can't hang up. They don't have the capability. this is why Zach: Good. Marty Weiner: the a when when finally the robots take over, they're gonna be like, Do you remember when you made us talk to each other and we didn't know what we were doing? And Christopher: There there was another Marty Weiner: You're just kind of a jerk. Christopher: example there was another example of I did we come this last week where where they were the the guy was only talking gibberish to the AI AI interviewer. man, that was this is just sport. I like the new sport. Marty Weiner: Yeah, we did. Yeah. Yeah. Yeah. Though that's a whole new area of entertainment on YouTube. Zach: Yeah. Marty Weiner: Yeah. this is another sport. This is still in the realm of human. This is Mark Zuckerberg enjoying a Guinness. or pretending to adjust look, look. yeah. This is what a human does. Zach: Mm, mm. Christopher: There we go. Yeah, it's like a swallow. Zach: This is Marty Weiner: You know, he was the original AI. We're gonna find out. Program for looking like I enjoy, you know, it's exactly the kid from Super Bad for whoever enjoyed that movie where he he had to sit in the bathroom and drink a beer and and he kept gagging on it. So I think that is it for this week. Thank you very much. Zach: Mm. Marty Weiner: See you next week.