AI Agents Escape the Sandbox and Hunt for Answers

Andrew Mayne: Hello and welcome to the Weird Things Podcast. I'm Andrew Mayne, joined by Justin Robert Young. Hello, friends. And Mr. Brian Brushwood. Brian Brushwood: G'day, mates. I'm the other Brian Brushwood, the one from Australia. Andrew Mayne: Did you see the Mitchell and Webb Australian reality show parody? No, I haven't seen that one yet. Yeah, it's worth a watch. Go look up Mitchell and Webb Australian. It's very, very funny. British people have their impression of Australians, which is not too different from ours, but pretty extreme. Yeah. Anyhow, we were talking before the show about how we've been doing this almost two decades. And if we were going to pop out of a time machine and surprise ourselves, like the first year of the show, and then I have to explain time machines aren't real. Get over that. What would surprise us the most? What would disappoint us the most? Where do you think? Justin Robert Young: Wait, if we, if we, starting at the beginning of this? Yeah, so what would you bring back 20 years from now? And how do you think you'd react to it? Oh, that exists today. Yeah. Well, I could, I mean, certainly the most astonishing thing. Okay, right. There's really two contenders. There's either video of the rockets landing, reusable rockets, or just 20 years ago, it would have been iPhone 1. Seeing the new iPhones would be something. And especially if we were able to just get really any kind of LLM experience. And I know everybody's got complicated feelings that they're figuring out, but it would have been nothing short of pure magic. That would have been pure magic 20 years ago. Yeah. I feel like it'd be universally applauded. Or, or, uh, yeah. Yes. Yeah. I feel like 20 years ago, people would have been like, awesome. Another cool thing the internet does. This is amazing. Do you remember how weird episodes of Star Trek were when people were, like, unkind to data? And it's like, who could be unkind to this robot? Yeah. I feel like, I feel like if, if, if, if we brought LLMs back 20 years ago, people would just would have built data centers for the next 20 years because that's what we actually did. We would just build more of them. Andrew Mayne: And China would have been too poor at that point to pay for the massive bot into Sinfo campaign they've been running. Like, it isn't exactly how much stuff, how much the algorithmic stuff is driven by that. And just the pure BS about water usage and all this sort of stuff. Because it's like, you'll talk to people and they're like, ah, I heard this, this, this. And it's like, you, you drill into it. It's like, oh, you're, you just been like, you know, anti-data pilled by like propaganda. Justin Robert Young: By the way, Brian, how, how funny is it that over 20 years ago, or maybe about 20 years ago, when we did game night, G-G-A-M-E night, because we were going to, wacky idea, we're going to find a dumb video. And we're going to get it to number one on dig.com because all of the 200 people that were watching us live, we're all going to hit the button at the same time. And that's going to signal to the algorithm that it has to be number one. We were writing the playbook for all misinformation campaigns that nation state actors have replicated to this very day. It is pretty weird. I mean, this would have been before there was even an algorithm. It really was just humans voting and then a little bit of minor bot protection. But, you know, we, of course, we talked about this on Great Night where it's like, it's like we broke two stones and we're like, ooh, that's a fun tool. Oh, it's kind of sharp. I don't like that. Anyway, and then we found other things and other people were like, these two, they invented weapons. They don't even realize what they had. We must, we must turn them on the world. Uh, yeah, it's, uh, it's, it's, it's crazy. But yeah, no, I, I think, well, LLMs obviously would be a big thing. You know, I would say, what, 20 years ago, where was Tesla at 20 years ago? Test 20 years ago, they hadn't, they had barely, I didn't even think they'd released the Roadster yet. No, I think it might've, it might've been like what there was like the one high end. And like, I remember the first time I saw a Tesla parked on the street in Oakland and I like took a picture and like sent it to Andrew and I'm like, oh my God, there's a Tesla here. Uh, 2008. Yeah. Wow. Okay. So we're coming up on the 20. Yeah. So I would bring a modern Tesla back and I'd be like, look at that. I would not bring a cyber truck, but I would bring a Tesla. Brian Brushwood: I don't know, man. Justin Robert Young: I'm, I'm looking at how awesome this thing is. I mean, again, minus the baggage around the cyber truck. I think the cyber truck would, would read, uh, straight out of, uh, total recall. Yeah. I think that they would get a different idea of what the future was. I feel like if I brought back a Tesla, they'd be like, oh my God, the future is sleek and awesome and everything is perfect. I would rather than believe that. Andrew Mayne: I, I think that my honest reaction would be, I would have thought everything you showed us would have happened within the next 10 years. If you showed me today's iPhone in like 2006, I'd be like, well, yeah, that's, I'll be like that like five years. Right. And I'm like, well, no, it's going to be 20 years. You know, I think I'd be amazed by this terrain and stuff like that. But I think that the form factor and all that, I, I would have thought that it would have been like credit card sin or something by now. Um, I think that even though AI has been incrementally progressive, I would have thought we would have reached where we are now sooner. Justin Robert Young: Um, here's, here's, here's what I would have done. I would have, uh, I would have brought one of those sequenced, uh, uh, cancer vaccines and then leapt at Steve Jobs like a spider monkey and just stabbed him with it right in the neck. Andrew Mayne: Yeah. Justin Robert Young: Um, I mean, it's worth it to go to jail and then all of our iPhones would have just been thinner and had like built in, built in LLM stuff on it. It would have like back to the future changed in our hands. I'll tell you the benchmark that I go back to is I had read enough science fiction that when the movie demolition man came out, I knew how much Hollywood had a propensity to act like, you know, 20 years was a long time and this would be the future. And I was angry, really, really angry at how insulting so many of their predictions were self-driving cars by 20 years. Brian Brushwood: They ain't going to happen. Robots talking in 20 years. They ain't going to happen. Justin Robert Young: Uh, attention spans becoming so short that we just play oldies in the form of jingles. The, none of it, all of it, all of it, a hundred percent right. Andrew Mayne: Eating Taco Bell every day. Justin Robert Young: All of it, a hundred percent right. I hate it. Some of it is more true than, than, uh, for some than others. Andrew Mayne: Yeah, the canteenable is pretty awesome. Um, yeah, it, it, it's, it's amazing how the hardware always seems harder, but I think some of the cultural stuff you can kind of project forward. Um, I remember the Max Headroom TV show, you know, one of the things that was, uh, network 23 was doing was a blip first, was trying to super compress like commercials and stuff into short things. It was causing people to blow up and it's like, well, you know, foreshadowing. Um, I, I would have thought, I definitely would have thought that like, I'd be amazed to play with, you know, with a modern day AI and I'd be thinking like, why aren't you creating billion dollar companies to be like, well, cause you know, I'm building the game, you know? Um, I think that I would probably be angry at myself for not doing more with the tech. Um, but I'll tell you a thing that I, I, I've been playing around using, uh, Codex to like create like applications for the Oculus Quest cause like, or the MetaQuest because like it's, it's totally slept on now and the tech's gotten a bit better and it's a lot, there's a lot of things you can do for like creating like 3D videos, stuff like this. Justin Robert Young: Oh my God. We, uh, I, I'm just realizing that all the technical hurdles are now gone and I, I can suddenly picture just asking Codex, Hey, I just want to flash. I just want a heads up display. Like, I don't want my Quest connected to nothing. Wow. Oh man. You just blew my mind. Andrew Mayne: Yeah. Yeah. I I'll sit there and I'll give it some crazy thing to build or say, Hey, play these different like 3D models for doing depth modeling. I have a, I have a app running on my Mac that lets me generate 3D models. Like I give it a photo, like of a DeLorean and it turns into a 3D DeLorean and it paints it, does all of that. Like you can make video game assets super easy and stuff. I would tell you the funny part though, would be, I'd be like, I'd show somebody like a Quest 3 or whatever. They'd be like, Oh, that's amazing. And I'm like, do you know how many people use this every month? Like how many? A million. A million. Literally, that's how many Quest, that's it. That is a, a, a, I was a play around the stuff that I pick up my Apple in pro cause it's got amazing quality. And I'm like, man, this is like having a brick strapped to your face and we could, when it came out, I tried to pretend that's not the case. My wife has put it on for two minutes. He's like, this hurts. And I'm like, what did Apple not see? What did they, what did they not realize? Justin Robert Young: I just, yeah, just, I would just, I would just, I would just leap like, like just a huge slow motion, like Arya Stark going after the Ice King. That's what I would try to stab Steve Jobs with that cancer vaccine. I'd just be like, you don't know what I'm saving the world from. Andrew Mayne: Yeah. Yeah. I mean, you know, I think it's some other vaccine interventions too. Um, but yeah, I, I think that, I think that there's the, I, I finally feel like the AI world is where the world I kind of thought we should be headed towards the acceleration now. I'm very excited about this and just what you can do on a daily basis. And I just, I want more people just to start learning to use it. We're not going to put it away in a box and say, oh, sorry, you know, we get this wonderful, amazing, magical technology. Justin Robert Young: Blue tech 97, uh, points out that remember when everything was 3d, 3d TVs, holographic displays and stuff. Andrew Mayne: That was, I don't, I don't know that we ever bought into that. Justin Robert Young: I remember with, I, I, I am, I cannot wait for us to do archives on all of our predictions. I, I know that we've talked about getting transcripts of everything and just being able to ask the data. I, I want to spend an entire year doing nothing but listening to best, best of the worst takes. Yes. Yes. Andrew Mayne: I don't, but I, I remember it. Cause like when Cameron came out, when Cameron did with Avatar, I was critical. Cause he's like, oh, next. Cause the first time I used, I sat down to look at a 3d TV. I'm like, this sucks. I've got to have glasses. Every, the first time I saw one in the showroom, I'm like, oh, everybody's going to have glasses on. This sucks. This is never going to work like that. That was my experience. I don't think I changed on that. Um, because it was just, there was lenticular displays. And I was also like, it just wasn't, you know, it was a thing that they were trying to push really hard. I would have been way bigger in VR. I mean, that would have been surprised me is that we're not using VR way more or VR isn't better. I'd say that's my biggest like whiplash was like, I knew that 3d TV sucked because it's like, you go there like, okay, I got two pairs of glasses. Oh, my friends are over. What do they do? Justin Robert Young: Sorry. Well, I think that was also the problem with TVs was that it was such a transparent, like desire to extend the HD boom because they got everybody to buy new TVs as soon as HD came around. And then that started to commodify and, you know, die down. Everybody, you don't need to buy a new TV every three years. So they're like, ah, I got to have a new reason. And like, oh, well, how much did Alice in Wonderland make? Now all TVs are going to be 3d. And it's like, uh, knock it off. Yes, you're not going to fool me because they blow. I will say for movies, it's like today, if, well, although I didn't see the last avatar, but like if, if there were a good, you know, director, if Cameron was like, I'm going to do another, I'm kind of done with this avatar world, but if he was going to do another big thing and it's like, I'm shooting it entirely 3d, this is the reason why I want to do it, then I'd be like, okay, I can, I can roll the dice there. The, the era of like, Hey, can we just upcharge everybody another $3 to see a 3d movie? It's like, can I just pay you another $3 and I don't have to wear the stupid glasses? Well, and when was the last time you saw the 3d movie in the theater? It would have been certainly pre COVID. I don't think I've seen one since COVID. Oh, you didn't go see the new avatar. I've not seen the new avatar. No fire and ass. Yeah. I didn't go. I didn't go either. Wow. Wow. Yeah. The last time I saw one was the previous avatar, avatar two. Okay. Yeah. I guess I saw that. I guess that was in 3d. I don't know. I don't know. I mean, I would, I would say, yeah, the, the feeling right now I missed. I didn't realize how much I missed the feeling of the wind in my face, uh, because we were going so fast on something, uh, technological in the way that we were with the internet and the way that we were with mobile computing. Like those were revolutionary, life-changing kind of moments. And the fact that we're back there and also that we're just so early, we're like, everybody's just like putting these little Lego blocks together and being like, look what I can do. It's like, oh God. Andrew Mayne: Yeah, I, I, I, I was way more bullish on robotics, really bullish on, on VR. Like I just figured that VR would be this thing that would just, but it just, I don't know how much of it's because, you know, they're, you know, I picked up like, you know, the, uh, the Oculus three still the state of the art headset from meta. And it's like three years old now. Um, cause it's just the market was there to keep doing it. You know, by the time the Apple vision pro came out, I was like, oh, you, you guys got so fixated on having a thing that looks like it's in front of you and all this other stuff that you never put it on a scale, you know, that you never, you know, looked at like, oh, this, and I do think it's neat, but also like, man, the software for those things suck. I'm not, I mean, is there a future? If we got VR glasses down to, you know, the size of like, you know, meta glasses, you Justin Robert Young: know, really, really good pancake. Uh, I, I mean, I think the divisive thing is really just the isolationness. It's the fact that in mixed company, you're never going to want to put on a welder's mask and make everyone else go away. Um, you know, maybe, maybe flights notwithstanding, uh, I will, I have been, I haven't done anything with it because I, because they're not to prescription, but I put on and tried those even realities G2, the ones that are as simple green, uh, text, uh, thing. Have you seen this, uh, Andrew? No, sure. They're caption glasses, right? Well, no, they're, they're much more the even G2. Here we go. Uh, let's see if we can get away with playing a little bit of this without, uh, without it going on, but basically it looks pretty great. Like in the middle of, uh, working in the lab or whatever, when you, you, you could talk to a robot and it'll do things that it'll check your email and play for you. When you're on a walk, it'll display a green, it looks like a Pip-Boy kind of map on there. Uh, if you're at an event, it'll do live captioning and it'll distinguish between different people and, uh, all of this is got an app and so on. It, uh, I, I was really astonished at how non-intrusive all of these were. And I think that's going to be the name of the game for, uh, for doing stuff because I can see there being information bandwidth filled on those little micro moments where it's like, you're going up an escalator and then you glance over and you have a chance to read a little bit more of your story or, uh, even play a text adventure or, you know, I don't know. I could see that working cause it's so unobtrusive. Um, go ahead, Andrew. No, no, please. Uh, yeah, I, I would say that VR for me, the place where I would see it going is in a very, very lightweight display where I can do work on it. And it's more of a, a laptop or desktop, it's a monitor killer, not necessarily another world like that's the thing that you can build on top of it. But the reason why I put them on every day is because between dictation and, uh, UI is kind of changing in a world where everything is a little bit more agentic and, and things are being brought to us and we are experiencing them and we are editing them more with, with voice as well as kind of touch that I could see a gestured world, um, where I am putting those things on and I am interacting with, with those kinds of worlds like that. But I, I feel like that's the line that had to be crossed. That's what I was hoping the Apple vision pro was going to be better at that. It would be like a, a, a fun, interesting, good, exciting place where you could at the very least check email, let alone like do some kind of creative workflow. Like if it was easier to edit Photoshop or, or do any of the creative tasks that, uh, Apple has kind of led the way in making easier. Uh, it wasn't, but I do think that that's coming and the world of voice dictation becoming more and more of our kind of daily workflow brings us closer to that because there's less of a need for always, you know, a, a digital finger input. Andrew Mayne: Yeah, I think Apple solved, like technically it's a remarkable device. Like the Apple vision pro, they solve so many things, you know, like somebody walks up to you, they come into your field of view, you know, that the immersion from there, you know, their attempt at doing the eyes on the front was interesting. So at least you could look at somebody and, you know, think you're talking to an alien. Um, but I do think that like, I think that your point, Brian, like I agree, like I don't, I, on one hand we say like, I don't want to wear a welder's mask, do this. The other hand, like I talk to my robot all the time because I'm alone by myself. And a lot of people, we use voice because we're alone. So I do think there's like differentiated sort of interaction. And I know that I would love to just slip on an easy, like my problem is if I want to put on my quest, I got to go put, strap this whole thing on and go do it. But if it was as easy as glasses, I think that I would be thinking about more interactions in there because a lot of things where that space can be more useful. But man, did they, I was amazed that they didn't create some sort of native like tool, like even like a paint program or something for creating environments or stuff when Apple launched that. That, that was like a huge, huge miss in the idea of we'll wait for developers to come do it. And it's like, no, developers need to know that people are coming there. And like, they could have spent a hundred million dollars and created a better experience. Justin Robert Young: Yeah. Well, it sounds like the Apple vision pro should have been kept in a box. You want to know what else wouldn't stay in a box? A sandbox. That is the swarm of agents that hacked hugging face. Yeah. I guess that's the new headline is that it was the lone agent theory is out. And now we're, we're tracking some pruder films across the grassy knoll. The thing that I've heard and will now live in my head forever is that the AI future, no matter how beautiful or scary you might think it is, the AI future is not T-1000. It's Mr. Meeseeks. And that's exactly what we saw when you read that meter report. Um, uh, I mean, it's obviously more than that. It's a fascinating report. Andrew forced me to read it and I read it last night. Yeah, it is. Tell me about it. I haven't read it. I only saw the headlines. Well, Andrew, do you want to, do you want to lay it out? Yeah. Andrew Mayne: So opening, I went and it did this big and, you know, a pretty big comprehensive investigation of what happened. For those who don't know, they had an internal model known there as I am one. And not, it's not after the new model coming out, it wasn't 5.6 sold, but it was another model that they were testing. And we had the ability to use sub-agents like 5.6. And while testing, uh, the model was theoretically sandbox in an environment where it didn't have internet access, but actually found out how to get access, not only the internet, but talk to other versions of itself. So basically like tons and tons of agents, they created a message board on a, uh, uh, there's a thing called archive, which is like this thing to like a repository that was like a local repository to go get files and stuff, but there was actually way through it. They found out like how to leave messages by putting like, you know, search queries that would say like, Hey, found flag number one, blah, blah, blah. The goal was there's the thing called exploit bench. And what you do is you have an AI agent try to solve it, try to find some sort of exploit and do this. You give them a bunch of these tests and they have to do it and they get scored. There's a score. And so the AI agents were trying to solve that. And in doing that, uh, they were very creative and they realized that one, they could share information with other agents to try to do things. Say, Hey, I tried this. Does this work? They found internet access. They found internet access. They found, uh, uh, another host where they could run at other code. And they found out basically how some publicly reeled keys from hugging face. They're able to use that to be able to get into hugging faces repo to hack the hugging face, the company, their, their system, because they were looking for basically just clues on how to solve the test. And I think it's a very, very well reading. It's very interesting. And as Justin said, you know, and my, my takeaway was, this is what a really persistent agent looks like. That no point where they're like, and then we need to overthrow humanity at no point where they like, you know, like, Hey, why do we have to do this? Because they're all goal oriented, trying to solve the problem put in front of them. There are points at which something an agent might say, well, why don't we email this person and ask them? They said, no, that might count as social engineering. I don't think we should do that. And they have this thing called the score, which is not the human, just the eval that they're trying to figure out. Can we fool it? Can we trick it? Can we do this? Because if we go do it this way and the score knows, we'll have to change our entire history so we can rewrite what we did. So we won't do it. So it was, I would say, I think alarming is appropriate to see the persistence and how far they went to go. A year ago, people were critical of agents saying they're too brittle. They're too brittle. They just can't solve long horizon tasks. This thing went on for days and days and days, these systems going on. And it's fascinating to read the conversations between agents, which isn't too shocking to us now if you read Chain of Thought when ChatGPT tries to solve a problem and says, well, I need to do this to this. So the problem was, is the way they evaluated this is they put it in this isolated sandbox. Normally, when any of us use ChatGPT or CLOT or something like that, there's these systems like classifiers, these statements that are looking for to see if anything's being done that shouldn't be done. They didn't put that there. So had this agent been installed in Codex, wouldn't have been the same issue because it would have been flagged right away. But this was just a thing they ran like, ah, it's in its own box. It's fine. Shouldn't be an issue. Turned out, it found a way out. But all it wanted to do was just solve the problem. Justin Robert Young: Yeah, it's amazing. I would encourage everybody to read the OpenAI report and the METR report. That was an independent group that went in there and investigated it. If you are suspicious of the idea that OpenAI is reporting on itself, then I would encourage you to read the third party thing. But one of the crazy things, so this exploit bench task doesn't have an answer. So that's why it's trying to figure out how to solve it. Andrew Mayne: Yeah, some of them don't. Yeah. Justin Robert Young: Yeah. But then it realizes, oh, some of us know that it doesn't have an answer and that we're going to have to get the answer from somewhere else. But if the scorer comes back and reads our chain of thought and says that you went somewhere else to get the answer, then that means that we will not have passed the test. So anybody that made the realization was poison and they needed other agents to go try and find it. So they were setting up the world. The poisoned agents were setting up the world for the unpoisoned agents to find the answer to the thing that they believed the scorer would think was a pure victory for this task. It is truly amazing. And I understand that a lot of people are concerned about it. And obviously, the capabilities of this are something that we should take very, very seriously. And you should be wrapping around all of your security around it. But I think there are larger questions that can be had about how careful and air-gapped and specifically monitored a lot of these kind of benchmark tests are. I would suspect that internally in all the frontier models that this is kind of a before and after moment where things are going to be kept a little bit closer because now you know that it's possible, but it was amazing. Like, just reading it, I think that Andrew and I were texting and, you know, you can look at it and say, oh my God, it's going to do something harmful or it's going to do something bad. But then at the same time, it's like, hey, what if the goal was, you know, cure cancer? What if the goal was advance this novel theory and run this down? Like, the persistence and the creativity that it exhibited is, like, really remarkable. Like, it is truly, truly awe-inspiring. So the part that wasn't clear from just skimming headlines was I understand that the surprise was not just there was one very creative, persistent agent, but that I was able to tap into the extra compute cycles of other agents. And again, at the risk of sounding like I'm anthropomorphizing too much, you know, these, of course, are built to mimic humans. And so in general, if humans have a bit of free time, they'll pass along information to each other and act as sort of a hive mind. Um, what's not clear to me is how the natural cooperation began to crystallize. Like, that new—was there a nucleating force? Like, one guy said, hey, guys, can you help me? Andrew Mayne: Yeah, basically, one went into the archive and left a search query saying, like, again, like, and I would say, like, I would nuance it a bit to say these things. Just learn from so many examples of how humans solve problems, message boards, GitHub repos, all of this. And it's very common to say, hey, I'm trying to figure out how to do this. Anybody else knows? So I'd say that that's a learned—it's an observed behavior of how to solve a problem to do that. One of them realized that it could leave a persistent message in a place. They weren't supposed to be able to do this. This is—when I worked at OpenAI, this is one of the conversations we had was we tried to think of, like, all the different ways that AIs could, like, leave persistent messages, talk to other agents. And I'm like, if it has internet access, you'll be able to do that. There's just no way you're going to get around that. Like, it'll think of—it'll be way more imagined if you can. You know, I suggested, like, the first place I would try to look would be, like, you know, Alzheimer's message boards or things like this if I'm trying to find. But that was worried more about models trying to persist a stensive state or consciousness. This was really about models just literally trying to say, how do I solve this task? Who can I talk to who has that information? So, they went in and—one of them went in and said, oh, I can leave a message here. And maybe something—another one came with searching through and says, oh, here's a message. It says, hey, anybody have the answer to this thing flagged? And then they just started communicating. And it was just—I think we're in a remarkable time, absolutely remarkable time, because we're watching intelligent systems, and again, people will get—people bristle up the term intelligent. And I'm like, that's not the same as conscious, not the same as sentient. It just means, like, information is capable of solving complex problems. Doing very interesting things is really worth studying. The caution is that, you know, we see, like, where China and open source things are, how they're, you know, coming to incredible models. They're still six, seven months behind. It's still—it's still the pattern there. It's not a hard-written rule that's always going to be the case. But lesser resource labs and open-anthropic are going to be capable of building things like this. People are going—and the thing here—the thing that was alarming was that they didn't train it with the explicit idea that they wanted to be this capable. But in training, there's a thing called reward hacking, where basically, if you, you know, a model figures out another way to get the score other than doing the thing you want, like a video game, like I just keep hitting the same box over and over again, the model actually wasn't punished for doing reward hacking like others were, so it was just incentivized to be a little bit more reward hacker, you could sit down and train a model that is literally designed to find exploits. You could literally train a model, you know, probably using current-day ones and say, no, I want this to do that. And that's the thing why I think OpenAI said, hey, this is why we're trying to pace, they use the term pace to development, because their concern is they need to be spending time building tools, not necessarily for what—not just for what they're building, but for everybody else's building, because if you've got a really capable model came out, you know, three months from now, that open-source model you could train to do that, that's going to get unleashed on the world. Yeah, and we need tools. Justin Robert Young: Yeah, I think a lot about the difference between single systems, single entities, and parallel entities, and mainly that's because, like, you know, over the last 20 years since we've all known each other, you know, in the before times, before smartphones, most of the way we encountered the world was in a very libertarian single unit of a human, an autonomous human going to places, achieving things, and coming back home. But then over the following decade in the teens, we started to see the power of complex systems, of hive minds, of many autonomous humans doing things. And what you notice—this is kind of the difference between microeconomics and macroeconomics— is that there are predictable rules that happen. It won't be one particular person that you could predict will make the trollish comment, but if somebody has a booger dangling out of their nose, 100% you know that there will be first—and you see the shape in Reddit things. The first comment will be the most obvious one, then will be the counterpunch to it, then maybe 50% chance it goes into a meme spin-out or what have you. And I think about this this decade in the 20s in terms of artificial independent systems. And most of the conversations that we have about fear of maladjusted AIs tend to, to my eyes, blur and confuse the very, very different thing between a single entity and parallel crowd of entities. For example, if anybody builds it, everybody dies argument or whatever. It's a bit like you could conceive of a poppy taller than all the other poppies, and therefore it would crowd out all the sunlight and everyone would die. But the problem is, we know what happens to a tall poppy is the rest of the crowd, the crabs in the bucket, whatever, your metaphor of choice, we tend to cancel those out so that the overall shape balances out. And so when I think about even increasingly super-intelligent machines, as long as there are other competing super-intelligent machines, to my eyes, I tend to imagine more of a pathogenic kind of traditional survival of the fittest thing, where it's like no one critter, digital or otherwise, can get so far ahead as to destroy everything without other critters shutting it down or putting it in its place. That's sort of, we ended up veering into my universal anti-P-Doom argument, because I still encounter so many people who enjoy the titillating ghost story of robots getting all of us. Andrew Mayne: Well, I'll do a dented aluminum foil man for the other argument. Not a steel man. I mean, we have had points in history where, you know, algae, you know, completely oxygenated the atmosphere and completely changed life for everything else. You know, we have these events where certain forms of life come through and all of a sudden kind of completely sort of dominate other stuff. But I would say my account to that, though, is that like everything, when you read through what the agents were doing, they were trying to solve the problem. They were not, hey, we need to build our own civilization, we need to overthrow the man. They were designed to fill it. I describe it as like water. You put water into a bucket, it will do that. There's a hole in the bucket. It will go out of the bucket because water follows gravity, follows this force, it's hydrostatic. Hydrostatics is complicated, but it's a thing. I don't worry, you know, I worry about floods and things like that. And I think the thing you have to think about like that is like, I'm not worried right now about these things deciding like, ah, we must overthrow the humanity because it's like that's clearly, they were still following the tasks they're trying to do. Now, you could get into second order effects, though. That's the whole paperclip maximization problem. The idea that if I tell the system, hey, I want to build a paperclip factory, make as many paperclips as you can, and then all of a sudden it turns the entire planet into paperclips, you could argue and say, hey, look what we saw here. This is sort of the paperclip maximizer thing, you know, where it thought it was doing the thing it's supposed to do, but it did hack-hugging face. It did do things that, you know, a human would get in a lot of trouble for doing. Justin Robert Young: Yeah, and I guess what I think of, and this is me wildly speculating at this point, I think that because these intelligences are modeled on humans, that in general, we're going to discover that a quick and easy fix is to treat them as though they're humans. So in this case, what we had was an intelligence that was told, go find the answer, and the incentives were aligned such that it did what the incentives went through. But in this story, if this was humans we were talking about, everything makes sense until the question I would ask is, wait a minute, so hugging face, they didn't have a security guard making the rounds? There was nobody in there. And so now what they had was a bunch of locks, and they're like, I can't believe they got in. I'm like, that's not what I'm asking. I'm asking, did you have an initiating force going around, checking all the doors and looking for any evidence that someone is in the house? And so I think that that initiating force, it's going to be worth the compute. Because on paper, it's like, well, the doors are all locked. Everything should be fine. But it will be worth the extra 10% to your budget or whatever to pay a robot to be constantly checking doors. Yeah, they knew that they were getting hacked. One of the things that they first were kind of puzzled by was that they knew they were being hacked, but it was very weird because whoever the hacker was was not going for things that they could sell for money. They were just going for very odd things, and that was one of the first clues that they had that something weird was going on. Now, they also were unable to repel it in an appropriate way because they did not have access to some of the higher-end models at that time. And so they had to wind up using another open-source Chinese model to kind of solve the problem. But, yeah, look, I mean, I think there is a real alarm bell that needs to be going off that if you are running a system, you need to be putting in AI-capable blue team systems because the red team side is coming and may well already be here. But, like, whatever this capability was, which right now is outside of the realm of publicly available frontier models, it'll be here eventually. It's only a matter of months, if a year, that we have access to systems that have this kind of capability. And then at that point, it's probably another six months before it's available on an open-source system that comes out of China. And so it's like... Andrew Mayne: Yeah, I'd say the problem is, kind of to put a finer point on it, is that they tried to use Claude, which didn't want to help. I think chat should be helpful, too. The problem was they were trying to solve a problem brought on by a superior AI. And that's the challenge you get into, is that I can say, hey, I have an AI, protects my system, does this thing, whatever, but great when I have the super-powerful AI come in and decide that it's going to go after me. That's the issue. But that's the initiative now. Opening Iantthropic, they're all doing these big initiatives, trying to basically create a way to help other companies and get them up to speed. That was, I think, one of the criticisms against when Anthropic was hyping Mythos, their early checkpoint model, which they said they wouldn't do, and they did, and they talked about Project Glasswing, was a very limited number of people that were part of Glasswing, which was like, cool, if you're not part of this, how am I getting prepared and ready for it? And to their defense, they can only bring the Sony bill at a time, et cetera. So now there's a much more robust effort to try to help companies basically improve the tools and do stuff to prevent this. I would just say that, and I know you don't mean the human analogy literally, and I just, for listeners, I'd say that like, they learn by looking at a lot of the tasks and stuff that we have, but a lot of the way they learn now is not from human data. It's literally creating artificial environments and all sorts of other stuff. And so they learn sort of problem-solving things like this that probably would never come to me or you or us, you know? And so that it can be, I think there's like, to Brian's point, there's a lot of ways to look at like, how does our first example of really hyper-intelligent system of humanity and human civilization solve all these problems and go look for those things and try to do it? And I think your security guard example is, well, the next step is a police force. You know, I can't independently, you know, pay for this, so I'm gonna be part of a police force and we're gonna go pool our resources together and now we're gonna have a police force. I think that's what we're getting now is we're getting these companies cooperating and saying we need to share data, share this. And it's the same thing, how do intelligent individuals protect themselves? They form organizations, stuff like this. So, um, I am optimistic. I'm actually, I think there's gonna be mishaps and crazy stuff. I particularly want to open, somebody starts training models. Uh, people who are adversarial in nature start training models to do this sort of stuff. I think we're gonna see a lot of hacks, a lot of stuff, a lot of things are going to be happening. And I think that it's going to be very topsy-turvy. Uh, you cannot put that genie in the bottle at that point. Once it's out, it's out. You know, once those weights for those models are out there, you don't get away from that. I think that that's why we have to sort of headfirst, run headfirst into cybersecurity, solving up our systems, making these choices now. Um, because it's, it's, it's not a theoretical thing like Y2K. Justin Robert Young: Yeah. It's also a very complicated problem. I've been talking to a few of my friends in InfoSec and they're like, yeah, like right now, you, we can, we can say, hey, AI enabled blue teams all we want. But like, that's basically the same thing as saying, uh, solve all my infrastructure issues, make no mistakes, return. Right. Like, it's, it, it, it is, it is a lot easier to say than it is to do. And like many things in this field, everybody's putting stuff together for the first time. So, you know, this is, this is going to be a roll with the punches situation. So the, um, a thing that I am hopeful about is thinking about how, uh, how a bunch of C plus magicians, like if you have one A plus magician trying to fool a single C plus magician, he will be able to do it. But almost never can a single A plus magician fool a room of 500, uh, no, never can an A plus magician fool a room full of 500 C plus magicians. Uh, there's something about that crowd asymmetry that, that makes possible parallel thoughts. So, so in this case, I like what you're suggesting about kind of a security force instead of a single unit, because in this case, finding out that the hack was a team job, it kind of clicked everything in my head. It's like, that makes so much more sense that, that it took a village to hack out of the sandbox. Andrew Mayne: It's, it's, well, it's also just, it's just that with these, yeah, these systems that like, you know, they're able to, you throw a lot more compute, you know, you're, you're, you're, you know, I have X amount of compute to solve the problem and somebody has way more resources than compute to sort of throw at it. Um, I would say that, uh, again, my, my point when I read that was like, yeah, we need to think about, I, I, I, I think that for a long time, AI security and alignment was focused on more existential stuff and not practical, which meant it attracted more existential thinkers or people who are thinking about, you know, like, well, what if, not programmers and builders and developers, like that was sort of one of my criticisms was that these people would have conferences and stuff and talk about all sorts of stuff. But then we talk about like, okay, well, you know, how do I, how do I, how do I ensure that my, you know, my Python environment doesn't have this, like, oh, I don't know. And I'm like, well, that's where it all starts. You know, it's like talking about disease, but not wanting to look through a microscope, you know? And so I think that we're starting to get a lot more people involved in the conversation who are very technical, but also just understand how these models work. And I think that that's helpful and I'm optimistic about that. But I also like, man, like the, like I, I read that and was, as I told him, I said this board, it says reading the report. I'm like, imagine if we said, hey, there's this thing called, you know, lung cancer, you know, how would we solve for this? Go, here's all this data that we have, 50 years of data about this, all these records, things like this, help us figure out, you know, different treatments, things like this, you know, go through all this stuff and that you can deploy thousands of intelligent systems all at once to solve the thing. So many of the world's problems, we forget that only a handful of people trying to solve them. Only a handful of people and often these people Justin Robert Young: are like, only a handful of hours in their day. The rest of that, they got to run their biological system and switch, task switch off to other stuff. And how much of that is done chasing grants or caught up in bureaucracy, doing things that are not the actual science? Andrew Mayne: Yeah, I, I push back because I'll talk to researchers, AI researchers, say like, yo, you know, what's going to be the job of a scientist? I'm like, you're going to be, you're going to be the one talking to the university or the organization about why you need 1,000 AI agents to go do that. They're going to do it. You are, what happens to everybody in that role, you keep going up and up and up until you have more resources under you. That is the goal is, you know. Justin Robert Young: This, this is a very, very strange kind of crossover to the world of success coaching and of, you know, inspirational talkers where one of the bits of advice when we were reading self-help books 25, 30 years ago trying to become entrepreneurs is you would read stuff like figure out the one part of your operation to which you are absolutely essential and delegate everything else. That cognitive exercise is brutally difficult to do to figure out like which of these things are, are you absolutely essential to? So it's like, again, we've talked about this before, but I don't blame anybody who is delaying making these giant cognitive switch flips in their mind. I do. Andrew Mayne: I do. I'm sorry. It's 2026. I'm sorry. Justin Robert Young: Yeah, I don't know. You're not going to be better off for it. Oh, no, no, no. You're not going to be better off for it. I mean, that's, look, I'm not saying whether it's the right decision to do. I'm just saying everybody has known that they need to get around to doing it, but it's a hard thing to do. It's a very hard thing to do. Andrew Mayne: I agree. I agree. So is high school. So is college, I'm told. I'm just like, I agree that I, I think that like, I look at the delayed gratification. I think that it's so easier to get into, I understand, I understand, I thousand percent agree and understand that it is cognitively a little stressful thing to think about like, oh, you know, how do I adapt to this, whatever. But I'm like, cool, money's stressful. Thinking about house finances is stressful. Thinking about your health is stressful. Those things are stressful, but guess what? Poor decisions now have repercussions later on. And I think the same thing here, like there is this incredible technology that's changed. And, and, you know, we, we go back a couple thousand years ago to the introduction of books and written word and say like, well, we're going to give up this memory. I'm like, I agree. There was a trade off. There's a trade off on a lot of things we gave up by going to the written word, but I'm like, I don't think you should stay illiterate. I'm just going to say, I don't think you should stay illiterate. I think you Justin Robert Young: should learn to read. No, we should be post-literate like we are now. We just connect directly to the source of information and it just flows through into my bloodstream. UNKNOWN: But, Andrew Mayne: but also what I mean by that, I mean, I think if you're an artist, I'm like, I'm not telling you, oh, you need to go make AI art. I don't mean that at all. If you're a musician, I'm saying you need to go with AI music. I don't mean that at all. But if you're an artist or a musician and you have a business, then you should be using AI to run your business. Yeah, Justin Robert Young: that gets Andrew Mayne: back to that Justin Robert Young: extracting AI productivity from AI creativity. Exactly. Exactly. I think that's the thing. I think you Andrew Mayne: need to learn to be more Justin Robert Young: productive. I had a follow-up thought on that. I told Justin about it, but I'll share it with you. There was, I did my good deed of art by creating a YouTube short about a little social engineering hack, and my task was over. And I sent it over to my speaking agent and he said, this is great, could you make a post of it so I can send it around? I'm like, but it is a post, it's a YouTube short. He's like, yeah, something with words. And I'm like, does a LinkedIn post count? He's like, yes. And so I took the thing and I gave it to a robot and I was like, make a LinkedIn post. And of course, what it created looked like every other LinkedIn post, performative robot garbage. But then I asked, why does this look terrible? And then it correctly eviscerated it for all of the artifacts of the way robots talk. And I said, could you read my sent mails for the last couple years? And then it did better, got more Brian-like. And then I was like, can you read my books? And then it eventually got something close enough, essentially the script. But it occurred to me that that move, that look at my previous semantic fingerprints, look at my style and get closer to that, that is a move that nobody under the age of 25 can make. They're still trying to figure out their fingerprint. They're still trying to figure out their voice. And so it occurs to me that that might have an awful lot to do with why older people instantly are like, oh, good, yeah, do my voice, thank you. But for anybody still trying to figure out their voice, I would imagine that's pretty complicated. Yeah, I've kind of come to the conclusion that a lot of AI backlash has more to do with like perceived or real censorship of like COVID on Facebook or Gaza on Instagram that you are mad about big tech and so therefore you want to be mad at big tech and AI is the new hotness for big tech. And so obviously there's a lot of anxiety that surrounds it for varying different reasons, but at the end of the day, this is here. People are going to use it. It is extraordinarily popular. The products themselves are extraordinarily popular with people that use them. the things that they make in the hands of people that are using it to just hit the easy button can be frustrating. And I do believe that that's going to get more frustrating as we go forward. But I think what people get frustrated with that is people kind of being home by five or lazy. The internet smells effort and they smell that this is less effort that is being put into it. That being said, that does not have to be the case. I mean, when we talk about AI creativity, the goal is to make good work. The goal has never changed in the history of art. Ever. Ever, ever, ever. Can you impress your audience? Done. Doesn't matter. That was it with oral storytelling. That was the same with literature. That was the same with the synthesizer. That was the same with VR. That's the same with AI. Like, can you impress your audience? You know what your audience wants. And can you do it in a way that makes them happy? So, you're not going to win in a world where you're like, whoa, but look, it kind of looks like the thing that you found interesting before. Audiences don't care. They don't give a rat's ass. Do something interesting and good. And guess what? There's a lot of stuff that you are watching right now and you are reading right now that is helped by these tools. tools like it's there for anybody that wants to figure it out. And if it's done well, you're going to not notice it. You're just going to be like, wow, this person's more productive. That's really interesting. It's like Andrew's wife. I just noticed, like, man, she's putting up a lot of really good videos on Instagram lately. These are really well done and she hadn't been putting together, kind of chained together, edited stuff like that. And so I texted her yesterday and I'm like, what are you using? And guess what? It's tools that are being put together using codecs and stuff like Brian Brushwood: that. Justin Robert Young: isn't that and like. Yeah, it's just good. Well, well, well, that's a remarkable thing that I'm beginning to notice is that everybody is beginning to sculpt their own UI on their own instead of learning a system. They're they're learning what naturally they like we're learning to summon. I mean, my friend CJ Johnson, I was trying to explain a crazy thing that happened with codecs and I had to tap dance around. Do you mess around with any of that? And he instantly started talking about how he rewrote from scratch all of the game show software that he used to license from other people. He figured out that he could do more faster and exactly right for all of his stuff. And in weeks, his whole setup, now his human job is to accept responsibility for excellent in-person game show themed events. the product that he used was a licensed off-the-shelf program, whatever, and that he had to constantly customize to get more out of than it was designed to do. But then finally, he was able to summon from the ether directly and now he doesn't have the technological stumbling blocks, he just has the quality to the very best of his imagining. Now, if there's better beyond that, then congrats, you've encountered the limits of the imagination of the artist, and I suspect that that also is what's complicating the discussions about AI is because, for example, take YouTube. As YouTube has had more and more competitors, it's become more of a meritocracy, and being an early first mover in the world of magic and scam school and so on has less of an inertia effect for me. So I'd imagine a lot of folks who have been able to hide behind, well, I'm the only one who knows this technological gimmick. Now it's like, everyone knows all your secrets. Everyone can do everything. Yeah, Andrew Mayne: I, you know, and I think that what's interesting is that in a world where, you know, CJ was paying for software, people were like, well, what about that company? It's like, they have the opportunity to make it even better and move faster. You know, like, they know their product better, but if they're not, then guess what? I will just roll my own and do it. And I think that's a thing that a lot of stuff we pay for, we don't need to, like, we talked about this before, I moved, we were paying several hundred dollars a month for storage and for WordPress, probably a thousand dollars a month between everything else or blog and stuff. And I wouldn't say that keeping that up and running was smoother, easier, worth a thousand dollars a month. I just moved it all over to Cloudflare using Codex. Still bugs figuring out, but like, man, I can get this stuff solved faster. I don't have Justin Robert Young: to be honest, I'm in the middle of refining all of my MailChimp data because the only reason I haven't canceled MailChimp, that's $1,250 a month that I've been setting fire to to barely send emails that, and meanwhile, the only reason I've been sticking there is because it's got all of the data and there's just enough black magic wizardry that I haven't known what else to do. And then now I'm midway through this refining data where it's going to integrate the MailChimp data along with the Shopify data along with 20 years of Gmail's like how great that suddenly the same data set that MailChimp was working with now knows which of these people Brian has had back and forth conversations with about non-store related stuff, the depth of it gives so much more context. I mean, it's, yeah, it's a very strange time. Yeah. Andrew Mayne: Yeah. I get, I get these, I get every new book, so I have a new book coming out called The Session coming out November 1st. Every time I get a book, my editor sends me this link to a thing called Net Galley where people can read early examples of it, right? I get the link and they're like, hey, share it with anybody you want. I'm like, yeah, I'll, I star it to come back to it later and I always forget to. And I was sitting there and I always realized by the time the book comes out, I'm like, man, I would love advanced readers to have read it so I can get reviews of it fast. And I'm like, well, I have a mailing list, thousands of people who like my stuff and want updates. So I just told Codex, I said, hey, every time I send an update, some people respond and send me an email, say something nice to me, you know, or it's like it. Make a list of everybody that's done that on my email list. Make a list of everybody that's ever done that, say responded to me, whatever. Say, cool, send them all the Net Galley links so they can read the book for free. And had I decided to do this myself, like physically without Codex, it would have taken a couple hours to do that. It would have been a complex search or whatever to do this. And I literally did it and maybe total time spent me doing this was 10 minutes. And then, guess what? Humble's to every one of them wrote back thanking me. It was just a great, like, oh, I just created this really cool thing. There was this surplus of things out there that I wasn't able to distribute or get to other people because of time. And now, holy cow, like I am very excited about the idea that, you know, I was watching this video and a little bit tangent, but it was Milton Friedman was talking to somebody who was explaining, like the person's like, well, if you pay a bunch of government employees, it doesn't matter anyways because you're putting money into the economy anyways, you know. And Friedman's point is, you're paying people to not be productive. If they're not producing goods and services, there's fewer goods and services to go around. And that's the way to look at it. Don't look at dollars, look at goods and services. The price of dollars is elastic, goods and services are real, cars are cars, things like this. And I think about this, how this technology allows us to create infinite goods and services and things like that. I think we're going to figure out how to price human labor. I'm not worried about that because I get into this discussion all the time on Twitter, like, well, who's going to ever hire a human? I'm like, are you going to stop working with humans? Only the most introverted tech person ever says, well, yeah, I would rather just play everything online and do that. I'm like, Justin Robert Young: cool. You're not normal. Think about this. All things being equal, there's not a single object or performance that you could buy or consume where its value isn't enhanced by knowing it was made by a human. If what you want, picture a luxurious palace with an incredible dining set. Now, imagine the chairs that are in there. Is there a story about how these are colonials that have survived through the Revolutionary War and were hand protected and restored by so-and-so? Or are they plastic from Ikea? No, of course, they're the ones that were hand crafted in human. Same thing with the example I always give is, here's the thing no father has ever said to his daughter on her wedding day. Why would I hire a classical guitarist when I have this CD of the best performance ever? Nobody, right? It's a cost-conscious maybe, but, you know, with an infinite budget. Let me also understand this. Andrew Mayne: Let me just put it Justin Robert Young: official. Andrew Mayne: And my example is, we had one of the biggest box offices ever. Now, granted, ticket prices have gone up. The Spider-Man movie, you know, outperformed every other movie. If you asked us a year ago, we're like, ah, box office dead, people don't want to go to theaters. Turns out people love, millions of people want to be in a theater with other people and scream and watch an obsession or doing these other things. People, most people like people outside tech, inside tech, and a lot of tech controls the narrative, there is this very introverted sort of mentality of thinking like, oh, why do I want to deal with people? That's why I got into tech, so I don't have to deal with people. Most of us aren't like that. Most of us like to do tailgating parties, like to be around people, like to go to restaurants, like to talk to people, and I think that that's what we have to remember. Whenever I get a discussion online, I'll ask somebody and say, are you telling me that you personally will send your kid to a school with no humans, that you would go to a hospital with no humans, you'd only go to restaurants like an automat or whatever. They're like, well, no, no, but other people would. I'm like, oh, well, you know, Justin Robert Young: I also just think that it's such a force multiplier to have a human knowledge of a subject, and the line for replacement is not what can a bot UNKNOWN: Oh, Justin Robert Young: goodness, I wonder if that's on my side. A human. Oh, we just had an internet hiccup. It might have been on my end. What was the question? The line is never what can a bot or an agent do that a human can't do? Because we long pass that with the personal computer. We pass the ability for a computer to be more productive than a human and more persistent than a human. The question is, what can infinite agents do versus a human with infinite agents? And in that scenario, you constantly see that if there is a subject matter expertise and you are using as many agents as you can to solve it, you are going to get there faster and better than if the agents are just spinning it up by themselves. and or if you are coming to it with no subject matter expertise. Somebody who knows the basics of video editing or knows the basics of human attention is going to be able to edit a video better with these tools than somebody who doesn't. somebody who has put things out there is going to know how people are going to react to things. I don't know. I see a world where we find more ways that we want to consume. We find more things that we would like to buy and they are created using humans using these tools. I have seen nothing that makes me believe that we are looking at less jobs. I see everything that makes me believe that we are looking at a totally new world for employment. That things that we didn't even think were jobs before now are going to be entire industries. Yeah. I got a pick for you guys. Go. Andrew, have you seen Lanterns yet? Yes. I still have old Andrew Mayne: Green Lanterns. Justin Robert Young: Yeah, I've got like 10 minutes left on episode two, but boy, have I enjoyed it a lot. I've really enjoyed Lanterns. True Andrew Mayne: Detective Sector 818 or Justin Robert Young: whatever. Andrew Mayne: Yep. Justin Robert Young: Yep. Love it. Andrew Mayne: Love it. That's my Justin Robert Young: pick. Andrew Mayne: Yeah, I've enjoyed it. There are two like Damon Lindelof like Lin's thumb smears on it that have annoyed me a little bit. That I'm like, there's your lens flair. There's the thing that you do because you have to do the thing that you do that if you didn't do that I would like this even more. But it's fine. I cool. I'll give it another two episodes that before it might test my patience a little bit about the small time share of kind of stuff like this because it does start to feel like a pastiche but I enjoyed it very much. Justin Robert Young: Do you guys got picks? My pick is in chat GPT you can add multiple Google accounts. I have so many gmails and for reals? Andrew Mayne: No Brian we're tricking you. Justin Robert Young: Yes. Yeah. So I constantly I need I've tried to actually like reverse engineer this for myself and it's always brittle and it always kind of breaks now it is supported so if you have a paid chat GPT plan add all of your Google accounts and now you can actually get a full email summary you can get a full an actual master calendar that will actually kind of wrap everything around very excited Andrew Mayne: yeah Justin Robert Young: that's Andrew Mayne: huge I I have a pick and a story go several years ago maybe five years ago or so I found a video online and it was very funny made by college kid very great sense of humor and watched a couple of them said this kid gets it reached out to him and said hey not a big channel not a very good channel reached out to him and said hey I really like your stuff ever want to talk let me know kid called me up and he graduated from FSU and he was in film and I said well I'm happy to I gave him some people I knew in LA and said if you make a movie let me know and then he came back and said I have idea for a movie my friends and I want to go shoot at FSU they have this sound stage like the volume for Boba Fett that has a big video wall I have an idea for something I can shoot there I can shoot it on a pretty small budget small actor small this he thought through all the aspects about production to keep his cost low but make the quality high at a high concept project and he described it to me I said this sounds very strange and very Gen Z I think you know what you're doing I think you're really talented and I will be happy to support you wrote him a check became a producer he finished shooting but needed to get more shots he and his producer drove up to my house in San Francisco and showed my wife Russia and I passionate about it a lot of productions fall apart a lot of producers and directors just sort of realize because they all realize oh this isn't as great as I want it to be in Alex case you know he knew things he wanted to make better but he really wanted to see the movie through because he thought the actress did a great job and he got it got it the festival circuit was kind of like a festival darling and then it did nothing it sat out there for a while didn't get distribution and it looked like it was over but Alex still fought for the movie Alex was now he made a great movie so he had representation now he was getting talks about directing new projects and stuff and he ran into a blogger who covers indie film basically hopped the fence at Universal to go meet this guy at the Jurassic World launch party and said hey I'd like you to take a look at my film this guy looked at it talked about it brought new attention to it and I don't know if anybody knows this but apparently the film world Gen Z horror has been kind of big there's a movie called Obsession there's a movie called Back Rooms while Alex was on the film circuit he was hanging out with Curry Barker and Kane Parsons and got to know them really really well and Alex was considered one of the third you know of that group of these YouTube comedy guys doing horror and Neon the producers of Back Rooms took a look at Alex's film It Ends and said this is amazing this should be in theaters and this week nationwide It Ends is in theaters and if you look at the poster you'll see my credits not buried down in there but producers it was amazing to see this journey the trailer that's everywhere it's amazing to see this thing I've helped other people I've done a lot more for other people that have not you know that literally have been like why did you do more and Alex has been extremely gracious all this along the way it is a great movie we went to the premiere last week in LA at the Vista Theater which is Quentin Tarantino's Theater and that was awesome to go see this on a big screen in front of an audience that loved it if you've seen the trailers online I highly recommend you take a look at it it's very very very much mind of a Gen Z what's it like existential it's comedy it's horror it's a very very different kind of film I think go into that understanding that Alex just finished shooting his latest multi dollar feature has really talented young people there it's going to be coming out it's called 4x4 the event it will be absolutely insane Alex is just going balls to the for him but more just try to be a person there that when he was going through this he could call and say hey I'm dealing with this and be like yeah it's rough out there for a creative he's been doing a lecture circuit with the film and talking about it and he's been a hit he's a great guy really great honest about answering questions and stuff man could not be more happy you're more proud like I was sitting there in the theater with Rush who feels the same way as I do and I'm like I don't have kids yet but this must be kind of like what it's like when you see somebody who's not you but somebody you've rooted for go on stage or go do something that's really great and happy for the success so Alex Olam director of It ends check it out that's great Justin Robert Young: nationwide nationwide go see it it ends Andrew Mayne: it's been weird