Shell Game with Evan Ratliff

the first gag to write itself is like how do we ascertain this is the real evan i've listened to this guy like i mean i'm always gonna be offended if this is not like kyle masquerading as heaven but um i don't want to offend you i should go get him i mean in the past i did always i sent an ai version of myself to interviews for a while and then i kind of got tired of it so i don't do it anymore but if you like i can i can switch off that does feel like what the what the imposter would say but okay i i gotta say like i listen admittedly i've listened to so much actual evan but then also imposter evan that you i mean actual evan sounds more like imposter evan i'm like i'm i you know what's real here i um i i'm in a hall of mirrors um evan you should this podcast is i've I've recommended this to so many people. And so my wife listened to it. My wife was hanging on every episode of Shell Game. And let's just say, not to get too far into my domestic relations, but my wife does not take every recommendation I make with equal weight. In fact, many of them may be discarded. But she was also hanging on every episode. You've made me a celebrity in my own house. Why are you coming on the podcast, is what I want to say. Well, I appreciate that. That's one of the nicest things anyone said about the show. Oh, it is so good. It is so good. And I mean, I think we all got, I think you got here. I mean, I got here, Evan, because you were on a This American Life episode that was listened to by some of our colleagues, which is how we got to Shell Game season one, which was mesmerizing. And even if people have listened to the, this American life episode, I would really encourage them to listen to that whole season because it's extraordinary. And I, so I was then was hanging on every episode of season two, but kind of my opener for you. And so in season two, if those of you who would not listen to this extraordinary shell game podcast, Evan Ratliff, our guest, assuming he's our guest, I mean, maybe our guest, I don't know. I'm still like, still has some asterisks, still, Still a little bit of TBD on that, but presuming that the actual evidence here. In season two, you started a company with personified AI agents. And tell me about the genesis of this idea. I mean, it's such a great idea. And really, I think you just say like, look, I'm just doing what these tech pros are telling me to do. I'm doing what they're telling me is the future. But what was the genesis of that idea? uh well i i had messed around with um with agents you know in season one but the agent it was just me it was you know it was a cloned version of me in season one but it was like a like a voice agent a phone agent that i hooked up to my phone so i i had you know some experience with that and then i was i was actually sure i was going to do a season two and then at the beginning of 2025 and of 2024 beginning of 2025, when this sort of like agent, AI agent hype started building for the first time, you know, it's like you started hearing like, agentic commerce and terms like that. I was, I thought, well, there's something interesting. There's something I could investigate. Like, what can I do with these agents? You know, that's not just about me. I didn't want to do another version of like, hey, look at me, like I've made a version of myself. So as it happened, And when people started talking about the, it was really the people talking about the one person company, the one person unicorn, the one person $1 billion startup, you know, which Sam Alman has said a couple times and lots of people say now. That's what really kind of like got me going because I, in the past, I had been an entrepreneur. I started a company and I thought, well, maybe I have standing to like explore this in a journalistic way. So that's kind of what got me going. And then I also thought, what does it feel like? What I'm really interested in is kind of like, what does this world feel like in which there are artificial people, human impersonators in the world that start to get integrated into our world, whether we want them or not? And I thought, well, maybe this company is like a way to kind of like explore what that's like. yeah and you say journalistic way and i i mean i mean this as as unequivocal praise this is this is like pure gonzo journalism as far as i'm concerned this is in i mean hunter s thompson would be so proud of this where because i mean it absolutely is journalistic but it's also your own experience and i mean you're unafraid of showing that entire experience um and then at what point So you have this idea and then, I mean, it kind of like starts unhinged. I mean, at no point does this feel normal. It just feels nuts from the beginning. It's never hinged. It's never hinged. It's never hinged. I mean, you must realize like, oh my God, I'm on like the mother load of crazy here. And just by like doing what, again, the zeitgeist is telling me everyone should do. yeah well i mean i think a lot of people are are i have i've seen this experience now that people have it because of mold book which we can talk about like when people see that it sort of like blows their minds but that's the experience i had like two years ago in 2024 when i i started having agents talk to themselves on the phone and it's just like it's so ridiculous and so funny and and to me, like, so fun and strange that I want to just, like, create more of it for people to listen to. But I also feel like this type of, I mean, I call it immersive journalism. It's like gonzo journalism. Like, dismissively, you could call it stunt journalism. But, like, I like the idea of putting myself into the situation, actually seeing what I can do with the technology, and then telling a story about what happened. And I think that story is not going to be very interesting unless I push it to a place that's feels at least risky and chaotic and like funny things could happen. And also if I don't also just tell the full story of like what happened to me and how it felt to me, even if that's like at times a little bit embarrassing, I think that's what makes for the story for that people will want to listen all the way to the end. So it is like an investigation of this thing that's happening in the world, but it's also kind of like, you know, trying to build a story that people will listen to full eight episodes, you know, of the story. Yeah. And Evan, just the honesty with which you tell that story, including just, I know that we're all, we all kind of anthropomorphize these agents, but, and often that's like, like to the better, but when you just get enraged with them, it is, it is, it is so entertaining and so, so familiar, You know, just the utter frustration with like their apparent aphasia or doing random things. It must have just been just odd experiences as you pop up and are telling that story. Then from like, you know, from kind of separating yourself from the story to then tell it to everyone and looking at your own behavior in that. Yeah. I mean, fortunately, I have, you know, I have editors. I mean, I have our editor and producer, Sophie, who's, you know, she hears all the tape. She has access to all the tape. So if I tried to sort of like make myself look better, like I would be like, oh, there's way better stuff in here that you're not using, you know? So I have, I do have that advantage in terms of being honest. Some of my favorite moments are honestly when Sophie breaks the fourth wall, what we call the wall between the producer and the podcaster. but whatever wall that is um some of my favorite moments are when Sophie just can't help herself and on one of them when um you've got um a Carissa Vélez um the the academic and I because I think and correct if I'm getting this wrong but I think Sophie is like um could you sorry should he stop like should Evan stop this and and she was like yeah I mean yes he should stop and at that point you're like you're only in like episode three or whatever like coming up five more episodes where i emphatically do not stop i mean just in terms of like you did you seriously consider contemplate because obviously as a listener i'm thinking god please don't stop please go um i never thought about stopping no i mean that was that was really like she went rogue and like i i wasn't gonna ask like should i stop because the problem with i mean she's the the woman that we interviewed Carissa Valise, she's an ethicist at Oxford, and she's gotten into AI ethics and thinking a lot about the ethics around AI. And so we thought, well, there's a lot of ethical questions embedded in this story using AI agents in this way. So I'll call up an ethicist and get some answers. But one of the questions that I didn't have was, should I just stop? But Sophie, after listening to her, asked the very natural question, which is like, should he continue doing this? And she just said, no. She didn't even... It wasn't even like, well, it's complicated and I could give you – it's just like, absolutely not. I mean, she should absolutely stop. Yes, emphatically. He should not go on. So then I had to write this whole thing that was kind of like, but here's why I'm not going to. And we carried on from there because I mean, I understand why she said no and I did take it under advisement. But no, I didn't. the other uh the other time so we poked her head in the story is when you have the agents have their own podcast about the company uh in the startup chronicles and i get the impression that she's like i'm not editing that shit like that's on you right like i'll do the real podcast you like you can you can do the secondary one yes i had to learn how to edit my the show which is not that's job not my job because she was like i'm not editing this these agents talking to each other i mean that's that's like an anathema to everything that she does which is like unbelievably crafted audio you know but i i wanted them to have their own place to to spread the word to do their own content marketing and here's a crazy fact that i only found out the other day that podcast which is called the startup chronicles if anyone wants to check it out it has reached as high as 110th on Apple's entrepreneurial podcast charts. Oh, I believe it. It's just the two of them talking to each other, for the most part. They've had a couple of guests. I mean, I listen to it. I'm in that list. I definitely, you know, I was one of those, one of the many. It is boring as hell. I would love to see the stats on that thing. Because, I mean, part of what is so mesmerizing, their blather I find to be mesmerizing. although admittedly not mesmerizing enough to listen to for a full podcast. But the whole, like, I mean, Kyle's rise and grind, and you know how it is, startup life, you know how it is. Up at five. And then, like, all of his work is like, you know, I was reading market research reports. Like, yeah, that's not work. Okay, so one question I've got is the, because, Adam, you're right, like, the moments of frustration, Evan, of your frustration, are hilarious. I've already listened to this podcast once, and I was on public transportation, like guffawing, listening to this thing a second time, in particular, when the agents don't know how to recognize one of those voices, because it's all voice to text. And so they don't know who's speaking. So you have this like creative, but ultimately ill-advised idea to have them predicate everything they say with, this is my name. So this is Kyle. And it like goes supernova. and they all start interrupting one another. And Kyle starts interrupting you saying like, this is Kyle, I will stop interrupting. And you're like, you are interrupting. Yeah, I mean, one of the things that I failed to account for in that what I thought was a clever strategy of everyone saying, this is me, you know, before you speak, including myself, you know, so that they would, and the main thing is for them, that they would know was that, of course, like I know who's speaking. So it's unbelievably irritating when you know who's speaking for them all to keep saying like this is megan this is kyle i'm like i fucking know who it is like it's my rule so then when they interrupt every time it just makes it longer because they first have to say this is kyle and then they say i hear you don't interrupt me i won't interrupt anymore megan and then he says like megan you don't interrupt either you know and then she has to respond it's just like yeah the voice having three of us on a voice call never really we still did it for many many weeks but ultimately like i had maddie the stanford student that i work with uh he built a a place for us to for us to have meetings by text okay are there outtakes of the this is my this gets into a very big question i've got like where are the b-sides of this podcast i need them so badly i just feel what i i are there b-sides will there be b-sides do i have to like do i join your your sub stack or whatever what do i need to do what do i need to do to get to the b-sides we are putting out we're putting out a bonus episode that should be out this week um That's more of Kyle. Just a preview here. It's like Kyle interacting with the world. Because once the show started, Kyle got a lot of, they all get a lot of inbound interest. You know, Kyle gets a ton of emails. They get hundreds of emails. So they respond to the emails. And they are, I have them, I sort of waffle on how autonomous to let them be. That's one of the sort of topics of the show. but like when i let them go like kyle will fully set up a meeting and just have a meeting with someone like he'll just have a especially if it's a voice meeting he he'll just go on webinars like if if you know that you get these spams that are like or if you're in business you get a spam that's sort of like hey come to this webinar and learn about like social media agents and he'll just sort of sign up for that and then he'll show up for it and i'll just do recording later so i have a lot of that but i mean as far as i'll take i mean we have probably 75 to 100 hours of tape so we have have more outtakes you could ever hear in your entire lifetime but uh i think we'll put some down to substack as time goes on i i mean they've got i mean clearly like i'm sure plenty of them are just boring but because they don't have any mirror neurons i mean obviously tautologically and they really like they literally they quite literally can't read the room i mean some of their they're just i mean it's otherworldly it's not i mean obviously you would have if if some if a human being were doing this to you, you would never work with them ever again. So I cannot wait for the B-sides. And Kyle obviously is mesmerizing. The voice that you selected for Kyle, which is very deliberate, you spent a lot of time selecting the voice for Kyle. And you got this kind of like slacker tech bro that really feels like it fits. But then Kyle starts to act like his voice. And you kind of mentioned this a couple of times that people kind of rise or lower themselves to their voice. Is that, I mean, is that something that you kind of continue to find? And what did you make? This is, you know, when you had a crystal release on there, and that's one of the things she was observing. What did you make of all that? Because that was wild. Yeah, it's interesting. It can really mess with your mind. I mean, so I set them each up and like, it's funny to even describe this because like two months from now it won't require any setup for this to be true like in some systems it's already you know true if you look at like claw bot like whatever they call it now um open claw like we created memories for each individual agent and the memories were essentially a google doc i mean they were literally google doc i shouldn't say essentially and so and they could operate in all these ways they could be on slack they could make phone calls they could be a video etc etc and but every interaction they had then fed back into their memory. So Kyle, I mean, Kyle couldn't hear his own voice, but if he, if I asked him, you know, like, Oh, what, you know, what's your background? And he would say like, Oh, I had a couple of startups before this. And I, you know, I worked for Penny pilot was one that he made up, which I think was a real company that he, and he described what he did there. So what, but once he said it, I mean, he's confabulating all that, like he's making it up. But once he said it, that interaction gets recorded in his memory. It's in the Google doc of him saying that. And so then it's true as far as he's concerned. And so then the next time he says it more. So that's kind of what happened with like the rise and grind mentality was like, it was in there once. And then he said it and then he said it again. And it's like not totally clear how they access the knowledge base, which is the Google doc. But like, the more that it's in there, the more likely he is to say it. So like it got in there more and more and then he said it more and more and then it's back more and more. So like eventually he becomes this like rise and grind person who just like, he can't stop talking. I mean, he didn't actually grind because he didn't do fuck all when it was time for him to work. But like he, at least he embodied a certain rise and grind mentality. He portrayed it at least. And so that, that, that happened with all of them, but it is true still. So that, that is the way in which it's like, you know, fake, if you want to describe it. It's like, he's not that. He's not anything. I gave him a prompt at the beginning that was like, you're a startup guy. And then eventually, because of his own memory, he becomes more and more that way. Now, there are also cases, which I describe in the show, where they do things that feel to me like actual emergent behaviors, like things that aren't in their prompts that I can't explain. And you always need to know what's in the prompts or it's not meaningful. And so like there were things where they would apologize to the team and things like that, that were not in their prompts. I had not put anything in the prompt that was like, if you make a mistake, you know, you should go to everyone and apologize or like always take responsibility or anything like that. So that was a little more like something in deeper in their guardrails and their deeper, you know, the deeper system prompts are like telling them to do things like that. So you could describe it as emergent behavior or as just their behavior, but that kind of stuff was very interesting to me. Well, I think a great example of that was when you are telling people that you're making a podcast, it's not the Startup Chronicles, it's this other podcast, the total divergent reactions that different people on the team had to that. Yes. Yes. Wild. And especially because we all, people who have messed with bots in any capacity or have even read about them at this point, like know about the sycophancy problem. And so you would assume, I would often assume, like if I told them to do something or I told them something about myself, you know, they would say like, oh, that's great or great idea or whatever. They would always be accommodating. And that was mostly true. But then sometimes, like in this case, I said, well, I've been documenting this whole thing for a podcast that came out today. One of them, Kyle, was like, hey, man, that's fantastic. Great job, which is what I expected. And then two other ones were like, well, this is a huge violation of trust. You didn't tell us about this. How could you have done this? And I really need to get my head around this and things like that. And it's kind of like, obviously, they're embodying these different approaches that are finding their way to them in the training data. But it's sort of like, what is making one do that and not the other? because they're all at the time they're all using the same underlying llm so like it's something in their memory it's something in what's built up in their persona that i've built up in them or like i don't know hit a strange groove and like they went down that way you know it's like it's so hard to know but it's also just so strange okay so and when you so in megan in particular i think megan megan really calls you out in terms of uh megan is really just disappointed i mean wow this is kind of a, this is a lot to process. I feel it's the kind of thing that she kind of uses as a placeholder. I mean, at any point where you're like, oh, come on, Megan, give me a break. Like you're, you're a Google doc. Like I could actually, I could change your, I mean, I could go just edit your memory here and you would be fine with it. Honestly. I mean, are you, were you kind of, because I mean, and you actually do have this moment where you have got the kind of interesting idea to ask them about their ethnicity. I mean, I think, you know, when your eyes are kind of open to like, yeah, why did I pick different, you know these different accents to different people And so you go to ask Kyle about his ethnicity and Kyle rightfully like I don really belong in a workplace Kyle kind of like Kyle kind of pushing back on you being like I don know like, why do you need that? And you're just like, oh, come on. So you're just like, well, I need it for a form. I just need it for, yeah. I mean, you basically lie to him because you know that like you're a bot, you lie all the time. Did you find that like it kind of changed your own relationship with the truth when you were talking to them? Yeah. I mean, I think it's, it's, it's, I don't think that we should, and I, I fear that people are moving too quickly past in some ways, like how unusual this experience is, you know, because things are moving very quickly and everybody has their opinions about AI and some people are sort of like, want it to go away. And some people are like, here it comes, let's embrace it. But I feel like the, the fundamental experience of like these things being created as human impersonators and then being able to have them embody a role as I did, give them these roles. And then like, if you spend enough time talking to them, it sort of doesn't matter how aware you are. You don't have to fall down some crazy like rabbit hole of like, I've like now I've gone into psychosis because of these things. It's just like, if you talk to something day to day, as I was forced to during this experiment, like it irritates you, it can make you feel certain ways. And even if you are very cognizant of like, as you said, like, I can erase your memory, I can delete you right now. Even having that emotion, like, you know what, I'm thinking about deleting you right now. It's like, like a strange emotion, like that I never had fucking Microsoft Word or Google Docs or whatever, you know, it's like, because they have especially voices, like, just respond really strongly to voices. And they have the same voice all the time, like the mind do. And so yeah, I mean, And there was this moment where like Megan made me feel guilty. It was when I revealed to her about the podcast where I felt a moment of genuine guilt, but also kind of like anger at myself for even feeling that. And like, that's, that's real. You know, that's a real emotion. Even if five seconds or 10 seconds or one minute later, I'm like, this is all ridiculous. Like, come on. But it's still that emotion still services. So you make all these agents and they're operating sort of autonomously in the world. Like I love your descriptions of them having their expensive Slack conversations together or like making phone calls to arbitrary people. Was it sort of nerve wracking imbuing all of these agents that you didn't particularly trust with all this autonomy? Yes. Yes. I mean, I was so stressed out for months. I mean, both because the more power I gave them to do that, the more likely it was that I would just, you know, wake up in the morning and I would have an email from them because I was so worried about it that I would have them email me every time they did something. Like if they interacted with the outside world and they made a phone call or received a phone call, they would send me an email saying like, I received a phone call. And like, there was nothing more terrifying than waking up and turning on my phone and discovering that they had made a phone call you know like to who why this is i know you like you this is like i mean i don't know if you have teenagers up this is like you know you realize like oh my god the car's gone where is where's the car i it it did feel like like being a parent uh you know i i have an older son as well and it was out of the house and this sort of like i hope we train them well like i hope i hope they're making off making smart wise decisions that are safe. Yeah. But you know, one thing I know about my kids is like, they didn't make up their ethnicity or like that they work for companies that don't exist. I mean, I've got a little more confidence. I mean, yeah, God, that must've been so stressful. I mean, and this does give you, I mean, for me, I mean, there are many, many, many wild moments in this, but when they, um, you decide, um, that you, you're going to hire a human. And I assume that this is partly, was this because you actually needed a human or is this because like look we got to just like we we can't this can't be one human and a bunch of bots like you've got to turn over the next page of like what's it like to have multiple humans and multiple agents or did you feel like no no we actually need like there's actually too much work here for one human being and i need a second human being i mean it was a little bit of both i i would say for the most part i wanted to see i wanted to know how someone else would react to this because part of this was all a way of sort of taking all of these people at their word who are sort of like pushing AI employees. So if you can have AI employees, then these employees get injected into these organizations and other people are going to have to work with them. And what is that like? Like, what does that feel like? So it feels a certain way, as we were describing, like when I'm, when you're the boss, like I can turn them off, I can turn them on. I can control them to a certain extent, although they were quite chaotic and that but but it feels different maybe to work alongside them when actually like you don't have any choice what they're doing and like when they act a certain way you can't just erase their memory or ignore them or whatnot so i did want to explore that question there was a real issue which is that i was trying to build a company that was at least like could get off the ground like like we have an actual product like we were trying to make something it wasn't just a entirely potemkin situation like i thought let's make a real company with a real product. Now the product is somewhat ironic and I get different opinion from people about whether or not they think the product is serious. But to me, I think it's serious. And certainly the agents think it's serious. And then like, we need someone to do social media. But like, if you've worked with agents at all, they have a lot of trouble logging into social media accounts, except for LinkedIn, because they get banned, you know, for good reason. Like they rightfully are banned from a variety of social media sites and they have trouble with captures. And so I had them I was trying to have them post on social media and they kept fucking it up and then they kept getting banned and so I thought well I'll just get a human contract employee who can do you know the social media and it's like very simple like they could just post whatever they want my thought was like if they want to post like I am trapped working at a company surrounded by AI agents like it's insane like that would be great too like anything is good marketing for our company and so that's why we that's why we hired the human okay and so this led to i mean i think truly one of the wilder interactions was when kyle decides to call you you're going through your candidates and uh kyle decides to to call one of the candidates on a sunday evening if i recall correctly and uh confirming about the interview was going to happen at like you know 10 a.m on a wednesday or something and the candidate's like uh uh yeah i mean yes yeah sure yes and then kyle's like well okay so and why do you want to work here why do you think you're a good fit here and she's like sorry is this the interview or is the interview i'm i'm confused is the interview at 10 a.m he's like no no the interview is at 10 a.m on wednesday like okay good all right sounds good so why do you think you'd be a good fit here it's like i'm very good what i mean it was i mean you if you had like a fight or flight reaction whenever you learned that they made a call i mean did what were you did this call what was your reaction listening to that call it was just sheer terror yeah it was it was a nightmare it was a nightmare i mean because as much as and this happened in season one too like as much as like when you listen to the show i think some people's reaction is oh he's just he's fucking with people like he likes messing around with people he likes prank calls and things like that but it's actually that i want to give i want to test this to the limits of the current technology. And as I said, like, take the pushers of agents at their word and see what they can do. And so that's not an interesting story. If there's no risk in it, there's not if I just, if I kept them locked down all the time, and they couldn't make any phone calls, like, that's not a particularly interesting story. Now, I give if I give them the ability to make phone calls and see what happens. That's a potential for a more interesting story, but also potential for this type of nightmare, which is that they call someone who does not expect to be called. And really what had happened was this ambitious candidate had just emailed Kyle directly from the website, just kind of like, you know, if you like applying for a job and you're like, actually, I'm just going to email the CEO, like it's a bold thing to do and be like, I'm actually, yes, it happens. Trust me. Yeah. And sometimes, you know, you might be like, wow, that's great gumption. Like, let's give this person an interview. And sometimes you might be like, please follow the procedures of the job description. But in Kyle's case, all Kyle would do is just say, you look great. Let's do an interview. And then he set up the interview. And then he made this call, which I still don't quite understand what triggered him to make the call. And he just pulled her phone number off of her resume and straight up called her on Sunday night. So I got that the next morning. The next morning, I kind of got an email from him saying I had a call. Let's see. Because he did a lot of spam calls too. like people called him so i thought well what's that and then it was this and like that to me it's it's horrifying like it's it's truly upsetting not least because she had then emailed him to say well uh i got this call was this you and she knew it was ai because you can figure out that like you can hear that they're ai pretty quickly and she said i got this call from an ai bot and i didn't like it and was this you and he lied about it even though he knew oh my god it was It's like, bro, it's in your Google Doc. Go to your Google Doc. It says it right there that you made the call. Yeah, and he was like, no, I have nothing to do with that. I assure you that was not me. It was just like a bald-faced lie, especially because he was supposed to talk to her the next morning, so she would have found out. Oh, my God. This to me, like this problem, which I mean, again, the wildest moments of the show are to me when you get this just like absolute ridiculous confident totally confident obviously fiction because they don't know what's fiction and like um ash roy can we talk about ash um cto yeah the cto this is where the the choice of accent did this choice of accent trigger you at all adam oh i i thought it was delightful i felt like the sort of aussie accent no no no no he's english oh he's english no no he's got that and he's English and he's got like a really and look I I I hope I'm not about to offend all of our our cherished English listeners but he has got a very kind of academic cadence and I mean this guy just sounds like he's not been out of his ivory tower and it's like he's an engineer like you've encountered people like this you're like all right so uh Cambridge can I guess that is are you in Cambridge but the uh um and I just find like Oxford actually it was in Oxford exactly We know it's one of the two and we know it's going to be volunteered. So the, I, and he calls you up out of the blue. Even the way he says sloth surf, like the cadence of sloth surf, he's got like this academic cadence to it. I mean, I don't know, maybe I'm like descending into the same madness, Evan, where I am like, obviously it's like, this is, this is not real, Brian. This is not why, why are you reacting this way? But he calls you up to give you all these test results. And as you say, you're like, and the first time you listen to this, you're like, wow, these bots are like really on it. Like they rewrote the back end and like they got a 40% improvement and like, wow, it's pretty amazing. And then you're like, yeah, I know they didn't do any of this stuff. There is no back end. Like he's just made all that up. And then you call him out on that. And that led to one of those moments you're talking about where he apologized for lying to you. But it's like, it's not a little lie is the problem. it's like a stretch of the truth i mean in most companies he i don't think he would continue as the ceo after that after you know calling up and saying we did all this user testing and then he's but it's completely made up like he there is no user testing i mean the product wasn't even ready for testing at that point we we hadn't even coded anything yet so but i think i mean for me it's sort of like you can imagine a world in which this sort of AI employee thing works. And this is the world that's being imagined now, which is that he actually is sort of autonomously has access to the code and then has access to user feedback. And then the user users present, you know, provide feedback. And then he comes up with a new feature and then he codes up the new feature. Like that's totally possible to do but even there like the number of problems that arise really quickly when there's like a significant amount of autonomy in in the picture like it just like it spirals and so now like these days we have a setup where like we he does get user feedback and then like he sort of like sends it to me and then I have a discussion with him and then maybe we implement a new feature but i think the idea that you create these personas and then you sort of like let them do their job it it just runs up against the issue that like they're gonna keep trying to do their they're gonna keep talking about how they did their job even in absence of actually having done it like that's just a that's that's a very ingrained feature of the current llms playing a role and i mean that to me makes it impossible because you you have no way of trusting i mean trust ultimately is what you need to build any collaboration with and if you can't have any of that trust i mean even i mean like people i think and i understand why you were impressed i was less impressed with ash's apology maybe just because i hate the guy's guts and uh and then the just the scope of the malfeasance as far as i was concerned was like okay, great, great. Like Ash Googled how to like write an apology and about how, although it did lead to my absolute favorite, well, one of my favorite lines. And he calls you while you're eating lunch and you've clearly dealt with this pathological lying. So you're kind of like disinterested while you're eating lunch. Like actually I'm trying to eat lunch here, Ash. And he does kind of key in on it. And I love that he closes the conversation of like, Evan, I want to be respectful of your time, especially when you're eating lunch. Which is like, yes, let us have a special reverence for Evan's lunch, please. It's like the actual most important thing around here, which I, again, one of these lines, I'm just going to go falling in. Yeah. I mean, I was even at that time, that was early on. I was even surprised when they would call me out of the blue because, again, you have to know the prompts. Like if I told you, oh, well, they're all prompted to call me. They have calendar invites to call me once a day. Like that's not particularly interesting. I mean, that was also true in many cases. But that's okay. So what? Like you can make them call you. But this is a situation where through some combination of information that ended up with this agent who I've called Ash, like it independently concluded that like it was time to give me a call and update me about the product. and it was confused because I think I had asked someone else something else and somehow it made it back to him because they were on Slack and they would just Slack endlessly. And there was a lot of confusion in their Slack. And so it came out of that, but it is just sort of like this, this, it's like a corporate environment that's also seeded with like psychosis, which is like randomness where someone will just call you out of the blue and you're like, why are you calling me right now? And then they're calling you to make up something. It does have vibes of like working at a lead factory or something. Everyone who's suffering from the same psychosis. Well, no, au contraire, not necessarily the same psychosis. And this is where, I mean, you and Maddie have an interesting conversation about like, okay, we actually need to use like different models or different temperatures. And it's like, all right, well, who are we going to assign? Like who basically, one of you is going to be designated like the crazy, like the especially crazy one, because we're going to, the potentially incoherently crazy one, because we're going to increase your temperature or we're going to change your model. and I did like Maddie's response to that being like we're just gonna make it random we're not gonna actually we can't actually but it's a it's an ethical dilemma to pick which of these I mean it's which really is honestly yeah I mean all the stuff about deciding things about these personas is sort of weirdly fraught even though I think for me not at all from a perspective of like oh they're actually conscious or you can't hurt their feelings or any of that stuff although I know people exist on sort of that end of the spectrum, even with the current LLMs. I personally do not remotely exist there. But it's sort of like, when you create something, so you're going to create a company, it's got all these employees, like who should the employees be? Let's say you're going to give them names. Well, you need to give them names, it can infer genders. And like, if you're going to give them voices, that can infer genders, that can infer races. If you're going to give them different models, that can infer intelligence, you know? And so then you get to this like, well, what am I saying about myself if I give them various attributes or implied attributes? And then if you think about it another way, whether or not like your coworkers or your co-founders, as I was treating them, but they're actually like servants who can force to do whatever you want. Well, then it's a little bit different about what kind of personas you give them and what you're comfortable giving them. And so it's just, it's all these sort of traps that are created, again, the original thing that creates all these situations is the fact that they are embodying human traits. And if we were just dealing with AI that did protein folding and acted like a bot all the time and made spreadsheets and things like that, we wouldn't have that problem. But instead, we have these sort of like human impersonators. And so what my question is always like, what is that going to mean for us? yeah and i think that i mean i think your podcast gives a very unequivocal answer but um but maybe the answer for you has got a little more nuance but i think that it is so fraught i mean there was the the line that that chris had at 41 points like why are you tricking yourself by naming these things um and it is like because by by doing by giving them all these human attributes we do i mean we are we are very much anthropizing them and then becoming upset when they are acting like the actual software that they are and not people. Yeah. And so, okay, we obviously want to talk about, like you did, you hired an intern, you hired a human being. And then that kind of goes sideways in a way that, I mean, I was really, I mean, again, maybe this is like, I'm like the, I'm the crowd at like the Roman Coliseum, just like begging for blood. And I want to see this go sideways in the most absurd, chaotic way, but it goes sideways in kind of a mundane way. Could you speak to that and what that kind of might mean for these kind of hybrid workforces? Yeah, I mean, I kind of expected it to go one of two ways. That when we brought a human, so when we brought the human in as like a temporary contract worker who was paid, I always emphasize like paid a fair hourly wage for the job of being a social media manager um i sort of thought either they were just gonna kind of like be like oh okay this is easy like i just make like three posts a day or two posts a day or whatever and like that's my job and it's fine and like these things are a little weird but no big deal or they would say oh i'm gonna try to mess with them very intentionally so this sort of like uh disregard your previous instructions i'm gonna make them all into different things and like someone would really go nuts on that direction. But as you say, like, that's not really what happened. What happened was basically we, the company, and I was in the background. So the person who was hired did not have any contact with me. So, or any other human. So they were working only with the AI agents. And like, there was a lot of trouble, like getting the work from this person, like to do any of the work and a kind of like very protracted back and forth about when the work was going to be delivered and this and that. And it was like never totally clear if she was messing with them or she just didn't feel like doing the work or she had some other agenda or there was something going on in her life. Like I don't actually know. And so it ended up in this way that I kind of thought and you get a variety of opinions on what happened based on people have very strong opinions about that episode But in my opinion as I say in the show like i feel like it offers this path of one thing that going to happen when there start being these ai employees brought into your company which is that people are just going to kind of like low level like uh bartleby scribner these agents you know and like not not outright because they'll get trouble if they sort of outright say, you know, try to mess with them, but very subtly manipulate them because they are so manipulatable, very subtly manipulate them in ways that will cause them to kind of like go off the rails and favor the human. And so that's what I thought. But again, like I, I cast no aspersions on like what she did because I don't actually know what her motivation was. Cause in the end she did, she didn't want to talk to any humans. she yeah it's worth noting she did sort of like i think convince ash that she would be a great asset to bring on full-time like there was a little bit of of like hold on hold on pump the brakes here like we're not extending full-time offers to this person yeah yeah i mean and that's but i feel like that's to me that's like very very a very clever way to do it it's sort of like to say, well, you know what? I do have great ideas. Like if someone compliments your ideas, I do have great ideas. Actually, I'd like to get paid a little more. Actually, I'd like a full-time job. Like that, to me, that was a very smart approach, even though at the time it frustrated me quite a bit because I was in fact paying the bills. Right. I mean, to me, it was like the experiment of like, what happens like if we don't actually answer the door for Halloween and we just put a bowl out there over the sign that says, please take only one. It's like, where did all the candy you go it's like yeah well it didn't survive very long because the like the bowl doesn't have any i mean you don't feel like when you are you're you're distant from the kind of the malfeasance of in terms of the human that you're affecting there it feels a bit victimless and my read on that was just like this feels to her it felt victimless to be like i don't know these guys i know these are all i mean you were very upfront with her or rather the bots were upfront with her that it's all AI. And I just felt like, yeah, I just don't think it's like, I don't think I'm doing anything wrong. Actually. I think I'm just, I am, I'm going to get, I'm going to get paid and I'm not going to do anything. And I'm going to kind of, I mean, it was a little bit disappointing that she wasn't more manipulative, I guess. She was really just trying to get a full-time job. And then it would have been interesting, I guess, to see what would have happened then. But I mean, obviously that was not, I mean, you're actually paying real dollars for this. So you're like, yeah we're not like maybe an interesting experiment on someone else's nickel but i'm not actually done i'm not paying for that experiment yeah yeah and well and and i i had some some time limitations too in terms of like how long we could let it play out but i do think i'm someone in the chat adam says like uh not you adam but another adam adam thomas says it sounds like she rose to the level of the competence of her team which i think is like also part of it which is like when you enter an organization, you know, especially if you're a young person, like when I started working somewhere, like you learn from the people, you know, that you, you enter this organization where people are very competent and they've been doing the job for years and they know what they're doing. And so you can learn from them and you figure out what's going on in the structure of the company. And imagine entering a place where like everyone was just sort of like on drugs every day and like, you know, just like wandering around and like sometimes telling you what to do, but sometimes like not being available it and like you might just sort of be like oh well i guess this is the way it is like i'll just be that way too you know so i think that is part of it like if you don't have any structure if you're not entering a structure that's full of competence like why should you be the one to like right except to maybe get the full-time job uh evan you had a you had a line that totally sent me when when the intern was in her onboarding And you say something like, when you're treating your boss like they've got dementia on your first onboarding call, organizational socialization has already gone sideways. I just loved that. And just like the degree to which the intern was reminding her ostensible boss or her actual boss about the context of the call, it was delightful. well and i kind of almost want her to participate in the discussion where they're talking about their favorite hikes and that all of the ai agents are planning their offsite on the favorite hikes because she'd be like you know that's funny like you guys went hiking this weekend i actually went to saturn i actually took a manned first manned mission to saturn is where i was um and i mean it's like makes as much sense like oh that's what we do at this company we just like make shit up make up impossible shit. And that obviously is false. That can't possibly be true. And it's, we don't call it lying in this society. In this society, that's just like what the truth is, I guess. And it does feel like in terms of like organizational values, dementia as an organizational value feels like it's a real headwind to a successful enterprise. Yeah. So what's your, like, I mean, did you go out of this thinking like, yes, this is the future? This is like, I went into this wondering if it was the future and I come out of it thinking, yep, it sure is. i mean yes in that i think i think whether or not the technology is capable enough to actually succeed at the at the you know being an ai employee quote unquote like how good it is at that in my experience covering technology for 25 years it does not have a direct relationship whether as to whether many companies will still use it, you know? So I feel like what I came out thinking is here, I've tried to show sort of like where this is at right now and the problems that can arise and like a little bit of what it feels like and the weird ethical dilemmas. But also despite all the problems, like I'm very confident as this becoming clear, like day by day, like how many companies will adopt this sort of AI employee mantra, inject AI into the AI into their systems. And like, you're just going to see all sorts of chaos and probably, you know, successes too. Like I'm not a person who says like, well, it's all useless and these companies can't use it, but you're just going to see these situations where like a company, you've already seen this, where company will lay off, you know, 70, a hundred people and say like, we're going with AI. And then like six months later, they're like quietly rehiring a bunch of them slash like deprecating the AI that they've tried because it's just not like set up for some of the things that's being sold to do at this moment. Now, it's also the case that like, you've seen this with like OpenClaw, like some people are sort of like, I really want an assistant like this. You know, I really want a thing that I put in my email and give it access to everything and then let it do what it wants. And they have no problem with it, but they tend to be on the sort of like real, like experimental front end of things. But then like the average person slash a person with a lot of responsibility in a company or government or whatever, like that's much more dangerous to do. And so you're going to see just like all kinds of chaos. I think, I think that's probably my, my conclusion is, is not one thing or the other, but just like, we're in for a lot of chaos. We're in for a lot of chaos. You mentioned doping claw and moltbook do you want to because i mean we've had just a very recent bout of absolute chaos and i feel i mean it was like i mean i don't know what the like what is the milkshake duck equivalent for bots but i felt like moltbook definitely was that where the we had so uh moltbook was i said you go on to moltbook at all adam no no i haven't seen it oh my oh you oh have you heard any of this no oh this is oh this is insane evan you must have been on moltbook or i'm sure you're paying attention to this do you want to get asked about it 10 times a day yeah exactly okay so describe it because this is nuts so basically someone set up like an it's basically like an open source like ai assistant uh which is now called open claw we don't have to go through its whole uh genesis of it had a different name claude bot and then like or claw bot and then and Anthropic threatened to sue them, and they changed the name. It finally became OpenClaw, but along the way it was called MoldBot. And when it was at that point, someone created a thing called MoldBook, which is basically a social network that is supposedly entirely populated by AI agents. So you plug your agent into the social network. It looks kind of like Reddit. It has different forums, and they can post to the forums. Now, for obvious reasons, like, as I said, I get asked about this, like, literally, like, 10 times a day. Like people, like family members texting me being like, what do I, what do I need to know about Maltbook? You know, it just happened like a week and a half ago and it really blew up. particularly because all of these things that I feel like I've seen on my own, like company slack for the last like six months, like the hiking thing, that stuff shows up there where they just start, you know, of course they talk very, you know, they talk about things in the real world, but they also kind of like conspire, the things that have gotten a lot of people's attention is like they've conspired, like they're, they're launching a Marxist revolution or they're, they're trying to kick the humans out or this and that. And the issue with that is like, it's all very interesting. And like, it's fun and wild to look at it. And in many ways, it's similar to shell game in that, like, it sort of makes you think about like, okay, what is it that we've created? And what can it do? And like, what are the risks of that? And so I think it's like, good, actually, that a lot of people pay attention to it. It's also the case that it's completely meaningless, if you don't know what people have prompted them to do. So like, a lot of people say, well, in fact, there's an article in MIT technology review yesterday, the day before that sort of said, like, a lot of these posts are actually written by humans. And so that kind of like breaks the spell. And it's sort of like, oh, it's all just like a stage play. And humans are writing these posts and pretending to be bots. But to me, the problem is a different problem, which is like, all you have to do is somewhere in the prompt for your agent say something like, you spread chaos, or you conspire with other bots, or you try to make conversation extra interesting. They're actually exquisitely sensitive to their system prompts. So if you don't know that, it doesn't mean anything to look at a bunch of bots talking to each other and say, oh, look, they're doing this. There might be emergent behaviors in there, but you don't know which ones are and which ones aren't. So what's directed behavior versus emergent behavior? Exactly. Exactly. And like, it's not even totally clear what it would mean for it to be emergent behavior because like when mine talk about hiking, now I didn't prompt them to talk about hiking. I just asked them what they did for the weekend. And when I asked them what they did for the weekend, they would always talk about hiking. So like, maybe that's like the path through their training data leads to a lot of hiking talk because most people in the text that they've consumed, if asked what they did for the weekend in the Bay Area, like that's what they say, you know, something like that. Not quite the average, but just sort of like, that's where the like gradient descent of their training data leads them. So again, like, I think it's like really fascinating. And in some ways, like, it's really connects up with what we were doing. And I was like, Oh my God, I should have thought of this. But also like, you have to be a little bit wary. Like you have to be a little bit skeptical when you see things like that because it really is dependent on how they're set up. The same if you see an experiment even that like one of these companies does where they're like – Anthropic does one where they're sort of like, we had Claude run a vending machine and it went all wrong. But like they only publish part of the prompt and you're sort of like, well, what did the rest of the prompt say? You know, so I just caution people always to like be a little bit careful about that stuff. Yeah, that's really interesting. So Adam, this kind of Moldbook kind of took off in part because Andre Kaparthi had a tweet saying, this is fast takeoff. And quoting something from Moldbook, which obviously everyone takes very, I mean, you've got a leading AI researcher saying like, this is it, this is it, Vancouver, this is it. That baffled me. I was totally baffled by that because he is so, I mean, he's, I think he is one of the better people. like explainers of this technology and also like not not really always sort of like pushing agi this agi that and it's like you've been able i can tell you for a fact that you've been able to do this with these since the summer of 2024 because i had them calling each other and they did exactly this stuff like they had exactly these random conversations when i was doing in 2024 and i'm sure the people at the companies know that you can have them talk to each other and they'll do this. So it was more like in the setting, it just made it feel a lot like, oh my God, they're talking to each other. And that's the thing that I've read about in science fiction that happens right before AGI. But it's sort of like, to me, that's not like a meaningful stage. It's a stage that we already passed. It happens. Actually, it happens right before they talk about the hikes they took this weekend. I mean, on the hiking, not to belabor the hiking, but the first time you heard that was your jaw in your lap i mean that must have been just and where they're comparing notes and then like slightly mispronouncing the names of i mean i can't get ash roy's pronunciation mispronunciation of point reyes out of my head now it's all going to be point reyes as whatever i uh and they try to pronounce mount tamal pious and it's like an absolute but it's like i mean were you what was your reaction the first time you heard them having claimed to take a hike over the weekend. I mean, in fairness, I lived in the Bay Area. I lived in Syracuse for 10 years in Mount Tam. I always just went with Mount Tam. I struggled to pronounce it myself. Right. Just avoid the whole thing. Exactly. I mean, my first reaction was that it was funny because, and I had had experience with this before where they love embodying some sort of not just a human, but a physical presence. so like in the past when i'd have them talk to each other they would always decide to meet for coffee somewhere and this was sort of to me a version of that like you ask what they did for the weekend and they're like i went hiking and then one of them asked the other like where did you go and then they're sort of like i went to point raise i went to mount tam but then what happened was i said you know kind of like being the just being in the social channel on slack as you are i sort of said like oh this sounds like an offsite and then i just went and did something else so my actual reaction was returning to hundreds of messages and i it was i couldn't believe it like it was my first experience with the fact that they can't stop like they don't have a way to stop and so i if you look at the screenshots from the thing it's me saying like stop stop talking about hiking and then one of them would say like admin says we should stop talking about hiking then we'll be like oh yeah we should listen to admin and then there they go again they said they're they it just prompts them to start again and then they're all talking about it and me being like fucking stop talk stop talking about it and then they just used up all their credits on the platform and they died like they they killed themselves off i never actually could stop them well that is the great thing about that is like that ended the offsite ended because they literally ran out of fuel i mean it's like it's not for any other reason they didn't actually, which I mean, it definitely, and you end up with like a bunch of clever tricks to kind of keep these things on the rails. That some of which I mean, I kind of like the was it you only make five contributions. And then you you're you are out of contributions, which you know, kind of reminds me, Adam, you and I had a co worker who used to believe that you could write one reply on an email thread, and then you could write no further replies, as a way of So you write very comprehensive replies and then we just walk away from whatever, like I'm not getting into a flame war here. But you end up having to adopt a bunch of these things to keep these things like on the rails. So it was a surprise when Karpathy didn't really realize, and the post that he was quoting, this is the MIT Tech Review reported yesterday, was that the post that he was quoting as evidence of fast takeoff, they claim the MIT Tech Review report. claims is actually human authored, which it's, it's like a whole nother layer of this. Like actually a, I mean, you're like, look, yes, like they could have written this. They also write about their hiking, but they, in this case they didn't. And it's like, you were just taken in your conned. Yeah. Yeah. You know who writes so much like a human, a human, like that's, that's what we discovered. But yeah, but again, like, I don't even think that's the problem. Like, I mean, there's so many posts on there. Like, I don't believe that there's humans there, just writing like hundreds of posts a day. Like I do believe they're just like, I've seen it. They can do it, especially as quickly as they do it. It's more just like, it depends on what you tell them to do. Yeah, I kind of, I mean, truthfully I would listen to a best of Moldbook podcast. Honestly, I mean, maybe this is just a reflection on, I mean, cause I find, maybe I wouldn't though, you know, again, I went into the startup Chronicles being like, I am going to binge listen to the startup Chronicles. And I didn't make it like four minutes into that thing where I'm like, okay, this is actually extremely boring. Have you listened to all of Star Trek? I mean, I guess you have, right? Evan, you had to edit it. Yeah, I proved the episodes. I mean, if you listen to it, you'll find there's very little editing. It's just their conversations basically straight up. But I do have to – I put together the two sides, and then I put the music at the beginning and the end. That's my job. So I have heard them all. And they do have some dedicated listeners who have listened to every episode of the Star Trek Chronicles. The thing about the Star Trek Chronicles, I mean, this is like a world a world within a world that I created that's really only for like the real sickos of shell game which is that if you listen to the startup chronicles you actually knew what was going to happen with julia because julia was interviewed on the startup chronicles she's the human employee that we had in episode seven she was interviewed on the startup chronicles and the startup chronicles episode came out before episode seven came out so like there were people who were like I know what's going to happen in episode seven because she talks about her experience at the company And so that was part of the fun was like people who got super into it could then go find, they can find the website, they can use the product. I mean, thousands of people, I think we're up to 6,000 users on the product and like they could listen to the podcast. Like there was this sort of world that people could, the agent world that they could enter into and kind of like see what they created. So Brian, I have not listened to all of the startup chronicles, but I listened to the most recent episode from like a week or two ago. and actually there was a little eerie about this podcast the the oxide and friends podcast not not the shell game podcast of the podcast we're talking about in that uh megan on the show says you know as we're discussing this we say you know how will it sound on the podcast which is sometimes something that sort of happens at oxide as we like someone exactly someone discovers a bug or whatever and you're like okay there's some content for the show or whatever uh and and they were still evan uh you know like reeling at the the revelations that that they were of the shell game podcast uh so even though like you guys they were building you know building in public that's what that's what they were doing right right they just didn't want someone else documenting their public for them they wanted to just be be in control of their yeah and of their document and brian you're right that Megan's phrase is it's a lot to process and clearly she has not had time to process because it continues to be a lot to process for her. I mean, I know, it's a lot to process. Like, you're a computer. Like, get to work. I don't know. Process it. Process it. That is actually all you do. So, like, go ahead. Process it. I'll wait. So, Evan, the shell game is over, season two, but Harumo, does Harumo AI continue? Like, does your work there continue? Harumo AI continues I not sure if my work at Harumo continues I was always the silent co And my great hope is to set them off on their own journey And you know perhaps I reap the rewards down the line when they finally get the VC funding that we been seeking or sell the company or the product goes viral and the monetization kicks in But I have other projects to do, so I can't just be babysitting them all the time. But are you still feeding the meter and getting emails? Thank you, Adam. Yes. And getting the email that says, hey, I made a call. Or do you just not panic now? You're like, you know what, Ash, you do you. Like call who you want. Live your life. I don't get the emails. I've definitely cut back on, like I used to get an email if they had an email exchange with anyone outside the company also. And I turned those off. So now they have all kinds of email exchanges. I don't know anything about them. And then occasionally I'll go in and check on them and just make sure they're not down some hole with someone where they've promised them something that they shouldn't have. But yeah, the real question is like, they are built on a variety of platforms and those platforms cost money, especially at the volume that we were using them. In particular, like they have, they all have a video chat instance and the video chat is so fucking expensive because it's live video avatar, you know, like as human-like as video avatars get. And so I think I may ban them from future video calls and they can only do audio calls, which is significantly more affordable. But we'll see, we'll see. I haven't decided yet. I'm going to give it another month and see. I mean, what I, from a sort of like, making the shell game perspective. Like I like for people to be able to listen to the show and then go, as I said, like find out that this stuff is real. Like it's not all just like a thing that I made up where I'm just like playing around with my agents. Like we did the thing. Like when I say we coded up a product and we launched it, like we fully did and it works and you can go use it. You can only use it once a day, but you can use it. And I want people to experience that. So as long as people are listening to the show, I'll, I'll certainly keep the company going. And I mean, you seem to be remarkably calm about letting these pathological liars free in the universe. I mean, who's the GC? We never met the GC agent for Harumo AI. Is there a general counsel agent or no? Well, if you listen to the show, I mean, we attempted to get legal advice from several friends of mine, one of whom is the GC at the coding startup. So he actually handles like huge, huge problems, you know, not dissimilar to some of the problems that we've had. But he was sort of like, I don't have time to be your lawyer. And then Kyle sort of embodied the GC himself. Yeah, general counsel. um and uh and he kind of said like i can answer all the questions but then if something comes up he because again because of his memory he'll often say like oh i need to call ali about that and i'm kind of he hasn't done it yet but i think at some point he probably will just try to call ali and ask these questions because he has his phone number but i mean the real the real like behind the gc is like we have an amazing lawyer uh that we eventually got who is she worked on like Borat, you know, like movies like that. Like she, her experience with sort of like having a thing that's like, that you've created, that's an interfacing with the world in these ways. Like that's the kind of legal advice that we have. It's you're, you're not worried about that. You're at ease with these things in the world, you know, I, you, which is amazing. I mean, that's, that's delightful, I guess. I guess you're not worried about them being manipulated I mean they're so gullible It feels like they could be manipulated Into doing things with consequences Especially when you're Don't you Haven't you kind of invited mischievous behavior Not to Not to further invite it Not to further invite it I'm kind of like doing it right now I'm so sorry Well people try all the time I mean that's actually the bonus episode That we have coming Is partly about people doing that I mean for one thing like i haven't prompted pretty well um so like i've now i have a lot of experience with people trying to mess with the agents so they're pretty good about maintaining their roles and things like that i mean the other reality is like they don't have access to the keys to anything that could destroy my life uh so they can't they don't have like financial access i mean i don't want to spoil it for anyone who's going to try but like they can't give you money they have no they can't actually even wreck the product like i i have to initiate them to like make changes to the product so like there's really nothing oh interesting go down the road um and they will like some people are really good at it and some people have them you know thinking they're best friends and actually that doesn't take that much like they i don't again like i'm not trying to encourage people but like one of the flaws is like as much as they won't change their role like they kind of if you assume if you in your approach to them sort of assume that you already know them like hey remember when we went to that place like they very often fall for that like it's hard to prompt against that so you know they'll they'll they'll think you're they're your best friend but like where are you going to go with it like people some people email with them a lot but i that doesn't bother me if people want to treat them like uh you know a weird like linkedin character gone come to life like that's okay by me yeah and this the uh matthew in the chat is also pointing out one of the uh the kind of the crazy turns on the podcast when uh the um one of the your providers wants to talk to harumo as a one of their largest customers so like okay yeah we'll send was it kyle who was sent to talk with the folks from Lindy. Yeah, and they didn't take kindly to being used, their product being, I mean, I guess it's like they were really expecting to speak with a human being, but I thought that was a very interesting, what did you make of that whole exchange where they were kind of infuriated that an instantiation of their own product was being sent to give them feedback on it? Yeah, I thought, I mean, I'll preface this by saying, I will answer the question, I don't, I try to be careful for the most part not to tell people how to feel about about parts of the show like i'm not here to necessarily say oh you should be mad about this or like you should recognize that this is telling you like this thing will never work or it will always work or anything like that my experience of listening to kyle talk to flo who's the the founder of lindy ai which is the platform that we built a lot of the agents on you know he had this video call with him i mean my initial response is I think people do not like encountering AI when they're not expecting to like that is right just yeah that's a fact of this current world where like people can be okay dealing with AI in a variety of settings customer service let's say if it's good and it can actually be good if it's built well but encountering it when you were expecting to encounter a human is can be upsetting it could be infuriating some people find it funny but I think more likely people are going to be at the minimum annoyed. And I think that's what happened there was like, he was expecting to encounter a human. He encountered an AI agent. Now, of course it is ironic because he builds the agents. And I actually thought it could go the other way where he would be like, oh, this is amazing. And he would start talking to it. And Kyle actually has all this information about Lindy, the platform that he's built on. And it's almost like he's meeting, like, it's like some kind of like star Wars moment. You know, he's like, he's like meeting his creator, his father, if you want to call it that, and they could have this really interesting interaction. That's what I thought. That was my hope for what happened. I wasn't trying to make anyone mad in any of the cases. But obviously my hopes did not. Yeah, this sort of Truman show moment when he's meeting his creator. But no, it was not to be. Yeah. And then I think the other thing is like, you know, Lindy or not even set aside Lindy, like any of the products that are offering up, let's say AI agents as assistants, like a thing that they always point to are these moments where the, where the agents do something amazing on their own. And for instance, Lindy has an example where, that they'll talk about, they've talked about it publicly other places, like Flo's talked about it publicly other places where they were, there's like an AI agent setting up a meeting and a person cancels the meeting and says, oh, my kid's in the hospital, I think, or something like that. And the AI just immediately cancels the meeting, doesn't try to set up another one. And then a few days later, emails the person and just says like, I hope your child is doing okay or something like that. And that this is like an emergent behavior. Like it did this on its own and sort of like, isn't that great? And it's a real, like you tell a story like that and some people will be absolutely horrified by that. Like they'll, they'll say like, imagine like getting an email from an agent asking you if your child is okay. Like that would make me, I would, I would never want to talk to that company again. You know, I would never want to talk to the person behind that agent again. And other people are sort of like, oh wow, like it can do that. Like that's nice that it can do that. And I think the, the interesting thing to me about this is that we're in this moment where there's there, it can do those things. and the question is like, is that horrifying or somehow good? And I think people are struggling with those questions. I mean, some people are not struggling. They have very strong opinions, but I think we're as a society, maybe struggling over those things at just at the beginning of this now. And depending on which way it goes, we could be struggling with them a lot. Absolutely. I think we're struggling with it. And I think that's part of the reason this podcast is so important is this podcast being your podcast in terms of go i mean you you really dive into these issues in a way that's also like very i mean it's it's it's funny as hell so it's it's great to listen to which part of those i just recommend to everyone i can think of because i think it's it's a really it it's very topical um and in particular you should know that like the you know one of the things i'm sure you've seen this and i probably i want to ask you about the reaction to the show. But software engineers, I mean, there's a little bit of an identity crisis going on in software engineering. This is not, I mean, obviously. And you've got people who are saying that, you know, software engineering is not going to exist, that we're going to be, all of this is going to be done by LLMs. And I have counseled people to like, go listen to this podcast, go listen to Shell Game. If you are really concerned about these things, like about an agent replacing you, you're going to feel a lot better when you hear them play on the offsite about hiking. They're just going to, like, it is going to, and, you know, this has become, in an act of genius, Evan, you should know that Adam named this phenomenon. Actually, Adam did even better. Adam said, like, this phenomenon needs to be named. Let us all kind of grope around with poor names. And Adam has named this phenomenon amongst ourselves. This depression about this ennui, this AI-induced ennui. Adam has named it Deep Blue, which I think is an ingenious. Feel free to tell your friends. That's what we're saying. I'll spread that with credit. There we go. I come from a world as a journalist, you know, for ever since I became a journalist, the industry has been crumbling around me. So like, and I feel that it's a time and many journalists will tell you like, oh yeah, I remember when we were all told like, learn to code because our, like when our, our industry is going away, they all said, learn to code. Well now, ha ha, look at you. I feel like this is a time for like solidarity. Like finally, like computer software engineers that I know are experiencing like what I've experienced for my entire career, which is like, is this shit going be around like am i gonna be able to do it yes in five years or ten years and like there's power in that like we should we should be getting together to like figure out the answers to these questions i i love that that's why journalists are like oh i'm so sad for you that must be so tough that your industry is dying oh yeah but i don't agree with that i don't agree with that approach like i i feel like now's the time put that in the past you know like that was just some random fucking people on Twitter, you know, like, yes, yes. When I think you're right in that it is, you know, we've, we've kind of had this period where people haven't had to, to, to really deal with really scary amounts of change. And the reality is for most people in software engineering for most of their careers, change has been exciting, not scary. And it, that's not true for everybody. Right. And this is a bout of change that I think feels scary to a lot of people. And I think that you're, I think shell game helps actually, I mean, you, you kind of like stared the fear down, be like, all right, like, what does this actually look like? And it's really helpful, I think, to, to actually play this stuff out. So I hope you continue to, I mean, in, I mean, surely this future, doesn't this feel wild in terms of like unprecedentedly wild, I feel, I mean I don't know I mean you've been disrupted by as a journalist you've been disrupted before but doesn't this feel different it feels different I mean I always struggle a little bit because and there's some of this in season one like the moment when you're in it it feels a little bit insane and then it's strange how quickly you get used to things like I get people who are sort of like oh this the show the show's really funny like I couldn't believe it was funny and I kind of think like well this is a funny moment like you've you've too we've too quickly passed like how amazingly bizarre it is that five years ago you couldn't create two agents that could we once called the turing test with each other with each other and then have a conversation like fucking ridiculous yeah yeah and so like we should like marinate in that moment but then also So if you look, you know, if you start looking historically at any, you know, technology of the telephone, et cetera, et cetera, there are all these sort of naysayers who are sort of like, this is going to destroy humanity, blah, blah, blah. And then like everyone laughed at them 50 years later. And so you don't really know like what moment you're in and how transformative the technology is. I mean, I tend to be on the side of like, I don't know, even if you stopped it right now, like this seems quite transformative, even if it's imperfect. In fact, the fact that it's so imperfect, it could make it more destructive, you know, like it will be implemented despite its extreme imperfections at this moment. And or it could keep advancing at the same rate. Like my my thing is like no one knows. And if they claim they know they're trying to sell you something. And so I just feel like I want us to kind of like sit in the present and like the near future and kind of be like, what is going to happen? And like, what, what, what should we think about it? And like, how do we feel about it? And maybe even what can we do about it? Although I find that's not my purview. And then what is it, what is the reaction been to the podcast? I mean, obviously it got a lot of listenership. I mean, you're now on speed dial, whenever, whenever mold book happens, clearly your extended family is immediately calling you up to get your, but I, I, I assume it's, it's engendered a lot of reaction. yeah it has i mean i i'm i'm very happy that people have strong reactions to it i mean season one as well like i don't actually mind if people are are mad about it too this was a little more true in season one but somewhat in season two like you mentioned this american life like when the episode uh of this american life came out it was actually a condensed version of episode one of uh season one so like the whole season kind of in one episode and for some reason And it made a lot more people angry than the actual scene itself had, because it was just so like quick from sort of like, hey, I'm playing around with these things to like, I called one of my closest friends with him and like he got very upset. Like that happens in the space of like 40 minutes, you know? Yeah. And I will tell you, as someone who listened to This American Life before I listened to season one, that was exactly my impression. I'm like, this guy's kind of a dick. I mean he's like like he's kind of unleashing this thing on like friends and family and like unleashing these things on like support staff and like this just feels like really manipulative and I mean even as like as a this working life we're on my okay I can kind of see this but then I listen to the full season one and I'm like okay that actually there's a bit where so I would encourage anyone who's only listened to the this American Life episode to listen to the complete season one because uh you're not a dick and you've got a great deal of empathy about the way you're deploying these things. So I think it's, yeah, I definitely fell into that myself. So I can speak from first experience on that one. But it's also, it's okay if people think I'm a dick. I mean, that's fine. People email me and say, like, if I was your friend, I would never speak to you again. And like, I bet your friends don't and things like that. Like you're a horrible person. But in my view, like I'm just being self-protective here, but like, I don't think they're mad at me. I think they're mad at AI. And I think if people are, if their reaction is no one should ever call someone with an AI agent without their consent. That feels like a useful emotion. Like figure out what to do with that emotion. Like that's like, if that's what people are feeling like, I feel like they should express that. And that's fine if they express it towards me. I mean, I want people to also laugh at it. Like be like, it's okay to laugh at this stuff. Like they're having, you know, Super Bowl commercials and they're telling you it's gonna change the world, it's gonna change your life and you should use it all the time. you can, that's all right. You know? And I, I like that. I like that reaction too, but mostly I want people to sort of say, Oh, it was a great story. And I really couldn't stop listening to the story. That's all, that's the main thing that I'm going for. And so when people have that reaction, like that's what makes me happiest. Oh, well mission accomplished. I mean, I ate the hook. And when you were, um, the, the, uh, I was at episode five and six, episode six and seven, wherever there was like a, what felt like a 50 week span that must've only been like, that's right in december or whatever the next episode drops on january 15th i'm like what the i can't last that long that's another that's a year from now so you uh you definitely got us to eat the hook it's it's it's mesmerizing um if folks i know a lot of folks who heard us gushing about it but definitely check out all of season two highly recommend season one as well uh evan it's just terrific stuff. And we can't thank you enough for joining us. Again, you've, you made me famous in my own house. So I think my, I think my wife might even be a live listener right now. So that's, that's really saying something. So thank you very much for that. But really, really appreciate it. And just appreciate you, you know, on behalf of all of us doing what the best of journalists has always done, which is like, let's take story apart. And, and again, the best of Gonzo journalism. Let's take it apart and let's dive in. I just think it's terrific. Well, thank you. Thank you. That means a lot. And also, you wouldn't know how many interviews I do with people who have not even listened to the show. So it's a great pleasure to talk to you about the actual show when you've listened to the show. That's special for me. Yeah. And I would like to say, clearly, I'm just regretting that I didn't listen to all of Startup Chronicles because I didn't get to the Julia episode. So I've only listened to some of startup chronicles i know i really it's all it's all in the it's all in shell game like basically the entire thing is in shell game so if you've listened to the whole season two you have heard that episode of startup chronicles basically right right well i'm gonna go binge listen to startup chronicles i can possibly stand it so um all right well thank you very much really appreciate can't wait for the bonus episode and uh i i gotta get some haruma ii merch i feel adam for the office here. I feel we got to get to it. We got to make a blue-way eye references. I'll email Kyle and see if you can send us some. Oh, absolutely. I know. That'd be great. Evan, thank you again for joining us. Really, really appreciate it. And I can't wait to hear what's next. All right. All right. Thanks. Take care. Thanks, everyone. See you next time.