Civility & Other Dangerous Games
Mason Amadeus: Live from the 8th Media Studios and the backrooms of the Deep Web, this is the FAIK files.
Perry Carpenter: When tech gets weird, we are here to make sense of it. I'm Perry Carpenter.
Mason Amadeus: And I'm Mason Amadeus, and this week we have a fun grab bag of topics for you. We are going to open up. Segment one is all about how being polite to ChatGPT is wasting Sam Altman's money.
Perry Carpenter: Ooh, then we're going to talk about how deep fakes continue to get really, really good, and the bad part is they are super easy.
Mason Amadeus: In segment three, another thing that has gotten super easy is coding, and that has given birth to an entire new field called vibe coding. We are going to dive into that later in the episode.
Perry Carpenter: And then what would you think if your entire newspaper edition was AI generated?
Mason Amadeus: Oh gosh, I don't think I would like that very much.
Perry Carpenter: I don't know, apparently some people do.
Mason Amadeus: Really? Okay. Well, that will be interesting. All of this and more coming up, so just sit back, relax, and tell me if our introduction format with this joke is getting a bit too contrived [laughter]. We'll open up the FAIK Files, right after this. [ Music ] Perry, I have a question for you.
Perry Carpenter: Okay.
Mason Amadeus: Do you say please and thank you to your AI assistants, when you're using them? ChatGPT, Claude, and whatnot?
Perry Carpenter: It depends. Not consistently, but every now and then, I find myself doing it. And then sometimes I do it intentionally to try to see if it gets a different response.
Mason Amadeus: Interesting, but you're not a part of the crowd of people who are like, oh, I say please every time, just in case the robot uprising ever happens.
Perry Carpenter: No, I'm not a robot uprising please and thank you type of person, that's trying to seed good will ahead of time. I've done enough bad things to these [laughter] that I'm pretty sure that if they're already collecting a file that I'm on it, and no amount of please and thank you is going to fix that.
Mason Amadeus: [Laughter] Yeah, you've burned that bridge long ago.
Perry Carpenter: Right.
Mason Amadeus: At this point. Well, Sam Altman will thank you. This is just a funny story, and Sam Altman doesn't even necessarily think you shouldn't do this, but he did admit that being polite is wasting money. Found this out from this "Futurism" article. Sam Altman admits that saying please and thank you to ChatGPT is wasting millions of dollars in computing power. And what happened was a poster on Twitter just wondered aloud how much money open AI has lost in electricity costs, from people saying please and thank you to their models, and Sam Altman chimed in and said "tens of millions of dollars well spent. You never know." So it says Sam might be a-I mean, it's all tongue in cheek, right? I don't know, I don't sincerely think that saying please or thank you is what's going to save you in the robot uprising if that were to ever happen, which I don't really think is likely. So I think Sam is being tongue in cheek here. But there is something to talk about in that it is use and compute, right? Because if you just send a prompt, that is the thank you, or please, it's got to run all these calculations just to return that. But I can't help but think, why don't they just intercept that on the way in, or like, you know, have some sort of non-AI algorithms-
Perry Carpenter: Right, and can filter that. I think that's the problem with all the other things, right? So in cybersecurity, there is this whole concept whenever you're coding at UX, that you should be doing input validation because when you don't you end up being susceptible to sequel injection, cross-site scripting, other input validation stuff, and seems like the AI world hasn't figured that out yet. Because we'd be able to deal with a whole bunch of potentially malicious prompts that way, and also strip off please and thank you, if you wanted to.
Mason Amadeus: I think it's interesting, because input validation typically like you're looking for like the most simplest form, but you don't want to break out of any sort of code characters and things that could otherwise cause things to execute, but because you interfaced with LLMs through natural language text, I'm sure that becomes a bit more complex. But there's other kinds of natural language processing that aren't like LLMs, that could probably handle determining if something is simply an acknowledgement or a politeness?
Perry Carpenter: Yeah, I don't know that the best way, I guess, maybe you waste your compute dollars somewhere else? And you get a smaller LLM just to process the request and strip out any niceties, and maybe even try to inject some understanding of what the prompt is? To figure out which parts to leave in, which parts to potentially modify before passing to another one?
Mason Amadeus: Kind of like an open-not open AI, anthropics classifiers. Constitution of classifiers, that pre-check for harm.
Perry Carpenter: Yeah.
Mason Amadeus: Kind of doing that, and like, having a middle man to hand the message, and optimize your prompt for compute. I guess that could happen.
Perry Carpenter: Yeah, I think so. And I think that does happen. Because you see it in a lot of especially like the image in video generation AI systems. They have these like little magic prompt things that you can click on, which will take your really boring prompt, and then add a whole bunch of cinematic language, and try to really craft the image the way that it's trying to interpret your meaning behind it. And I think that maybe this is a side conversation, but I've heard a lot of people over the past year say that prompt engineering will not be a thing that people have to learn. I think that's still dead wrong, because everybody used Google right now, everybody uses search engines, but people that started at the very beginning, where they had to understand Boolean very well, and then understand a whole bunch of interesting little Google dorfs like, you know, insight, and all of that--
Mason Amadeus: Yes, before-after.
Perry Carpenter: And there's way more productivity that can be gained through understanding how a search engine works than just throwing stuff into Google, and I think that the same is going to be true for prompting forever.
Mason Amadeus: and I think you're absolutely right. Because it-that goes into part of what is talked about in the futurism article, of like, the use of polite language will steer the kind of response you get. So, you know, similar like, using Google, you can absolutely use Google more efficiently, and I didn't realize, I had taken this for granted. I'm probably based on just my age, like on the cusp of people who are taught how to Google, because I don't think most people younger than me like know really the significance of using Booleans and I actually think Google sort of stopped putting as much significance on Booleans, but there are still tricks like Sight, After, Before, all those other things-
Perry Carpenter: Right.
Mason Amadeus: And yeah, having that knowledge, and using it will always be good, and I don't-I think that putting something in between the user and the AI will make it harder to use, similar to how Google, making fuzzy search, making search fuzzier has kind of made it harder to use for research. I imagine the same could happen with AI, if we approach it that way.
Perry Carpenter: Yeah. I think so. I don't know. I do always think that for somebody that has grown up or has been formally trained in coding, somebody that has been really formerly put a lot of work into prompting, and understanding the way that these work under the covers, they are always going to have an advantage whenever they go to whatever interface is presented to them, versus somebody that just assumes that the machine is going to do most of the interpretive work.
Mason Amadeus: Well we are going to have to hit on that in segment there, even more, when we talk about vibe coders. But to turn back to just the wasting of money and power, from being polite to AI, there was a survey done to see how many people are very polite to their AI and why, so show the results of that now. This is from Tech Radar. We will link it in the description. And they found that 67% of people are polite with their AI tech. So just ChatGPT, when they're using it.
Perry Carpenter: Oh, interesting.
Mason Amadeus: And it is broken down, 55% say yes, it's just the nice thing to do. Another 12% say yes, when the robot uprising happens, I don't want to be first. And then the other whatever other percentage, 20 and then this one is cut off because I blew it up so on the screen, whatever leftover after 62% just don't because of efficiency or like why would they? So it is still a vast majority of people saying please and thank you, and that's using more power. And if you believe a lot of the reporting that is going out around power usage, it could really-that could add up to a lot. There is this article that is linked from that futurism article, unless I made a mistake. I believe it's linked to the futurism article, so it's Washington Post article from September of last year, saying it takes a bottle of water per email if you use ChatGPT to write a 100-word email. Now, I wanted to hit on this, and bang on this power drum one more time, that I know that I talked in the past about wanting to do like a big analysis on AI and power usage, because these claims seem completely like orders of magnitude outlandish to me. And I just quickly tracing some of the sources in here do-they all kind of lead circularly back to small papers that all reference that original IEA report that lumped in crypto with AI, and there is a new article that you sent to me from Axios that actually says that these numbers may be overblown. And this is the first reporting that I'm seeing that seems to be looking at different, more modern numbers, and different sources. And it's not to say that AI power use is not a factor in environmental concerns. It absolutely is, in power use and water use. But it's not as crazy as dumping out a bottle of water. I've seen people on social media make the connection that if saying please and thank you is wasting millions of dollars, it's also wasting tons of gallons of water, and stuff like that, so our politeness is literally killing the planet through this stupid robot, and that's just like not true.
Perry Carpenter: I understand the concern for that. I think that there is a lot of misunderstanding of water consumption versus water use.
Mason Amadeus: Yeah.
Perry Carpenter: In those studies, because when you're in kind of closed system water cooled environments, it is like the circulation of the coolings, like, how much is needed. And then there is some displacement, there is some condensation and replacement of that that needs to happen. But I think that people kind of mix those numbers every now and then, and there's not a really good, clean analysis that I've been able to find that makes the conclusion really clear.
Mason Amadeus: Yeah, there's no clean analysis of all of it. But I will say that there are evaporative cooling towers that are pretty commonly employed, where the water that is heated from being pumped through all the computer systems is then run over these fiber boards, which have air blown over them, and they evaporate, which also cools remaining water, but it does produce a lot of steam. That water doesn't always come down in the same place.
Perry Carpenter: Right.
Mason Amadeus: So that is a problem. And that is actually employed in a lot of places. So there is a water use. In a lot of cases, that water is not potable. In some cases, it is, and that's like not good.
Perry Carpenter: Right.
Mason Amadeus: But it is just so much more nuance, and there are a lot of closed loop systems, and we are moving towards even more and more closed loop systems, and more efficient things.
Perry Carpenter: Right.
Mason Amadeus: So that-that power use landscape is changing. And yeah, there's just a lot to simplify.
Perry Carpenter: Yeah, and I guess speaking of that, if you're a power user of a large language model, maybe you don't need to use please and thank you, but I did see a post the other day of somebody talking about like whenever they go and they introduce AI use and large language models to people, they don't discourage the use of please and thank you, because they're trying to let people lean into the fact that it's a different form of computing. And if you approach it like a conversation in a brainstorming session, that's different than just kind of feeding in, you know, A plus B equals C type of things and you're going to have a potentially better outcome, if you can just, I guess, for lack of a better phrase, vibe with it.
Mason Amadeus: Yeah, talk to it like a person. I mean, I think that's important too, like, I think it helps illustrate to people how these things work. And then also, you know, depending on what you're doing, saying please and thank you will get you statistically closer to the type of response or tone of response you want, in some cases. So it's all-
Perry Carpenter: It could. And then sometimes you've just got to say make me a sandwich.
Mason Amadeus: Yeah, exactly. I have stopped thinking about it as chatting, and more about strategic use of language's input now, as like-
Perry Carpenter: That's what I do too.
Mason Amadeus: Yeah.
Perry Carpenter: So I've been stating, a few different stats in most of my presentations for the past year or so, going back to studies from 2022, 2023, and now, 2024, what we see over and over and over again is that people are really, really bad at detecting deep fakes. And one of the stats that I put in really big letters is 21.6%, and that's from a study from 2023, where people were warned that within the next five videos they were going to see a deep fake, and the 21.6% is the accuracy percentage of people trying to pick out which one was the deep fake.
Mason Amadeus: Ouch! Okay!
Perry Carpenter: Which means about 80% of the time, they're getting it wrong.
Mason Amadeus: Yeah.
Perry Carpenter: Which is not good. That means that we live already and for the past year and a half or so have lived in a place where people don't know what's real or not, reliably. If you're creating a good fake. And one of the things that I showed last week on LinkedIn, they got a lot of attention is just how easy it is to turn into somebody else. So like if I hit this magic button and I shift my background to something else, we're really used to that in Zoom, but with the click of a button, and this won't change here, but for anybody watching, you'll see my face just change in an instant--
Mason Amadeus: [Laughs]
Perry Carpenter: To this is my boss' face, at Knowbe4, of course, I could switch to like Nicholas Cage [laughter] really quick, and you know, with a couple touch-ups, if I wore a hat or something to fix the hair, all of a sudden it becomes way more believable. And we're very, very used to like glitchy stuff, and things like that, because we are, you know, in these environments where people are doing background replacements all the time, and yeah, you're literally just like watching me shift through face after face after face, and it's that quick now. And it's like directly integrated into Zoom.
Mason Amadeus: The shadow of your glasses on this one is actually being projected down onto your fake face.
Perry Carpenter: Exactly.
Mason Amadeus: That's wild.
Perry Carpenter: And if I were to switch faces, you know, one of the interesting things you see is that like they are all very light reactive--
Mason Amadeus: Whoa!
Perry Carpenter: And everything else, there's, you know, there's nothing really that gives away that you're in a fake environment. Really easily now, other than some of the delay, and some of the little digital artifacts that you might see. So I think we're in this place where more and more and more we are going to be seeing people show up in Zoom calls, and things like that, as defect--
Mason Amadeus: That one!
Perry Carpenter: So I'll go back, back to myself, really quick.
Mason Amadeus: The second to last face you did there, Perry, looked so real that if you joined a call and I wasn't like very familiar with you, I would have believed that that was like a real person just using a bad background replacement.
Perry Carpenter: Was it this one?
Mason Amadeus: Yeah, that is very convincing! That is very convincing. Holy smokes.
Perry Carpenter: And it-I mean, and then other than my hair, and my background, well, you're not seeing my actual background--
Mason Amadeus: Even your hair!
Perry Carpenter: Yeah, it looks like it fits well, so--
Mason Amadeus: Looks like your name is Marcus.
Perry Carpenter: The power-what?
Mason Amadeus: [Laughs] Nothing, I just said you look like your name is Marcus, like you just look like a plausible person, just a different person.
Perry Carpenter: This is a guy named Devin, that I know, he's a Chief Information Security Officer at a company that I won't disclose.
Mason Amadeus: That's a trip!
Perry Carpenter: So all of these, if you're watching, these are single faces where I just downloaded an image from the Internet, and now I can cycle through them at will. So that's neither here nor there, really, related to the over-arching thing for the story, because if I wanted to really use this technology the way a scammer would, I'm not necessarily always going to be trying to impersonate a celebrity or CISO. Most scammers are just trying to become a plausible different person that looks attractive so they can execute a romance scam.
Mason Amadeus: Right.
Perry Carpenter: So they don't need to become a familiar person, they just need to become a different person than their normal face, so that they can then engage with people online. But one of the things that came to my attention this last week was a new study that is going to be at the heart of a lot of the problems when it comes to deep fakes, which is that again, people have a really, really hard time telling the difference between reality and fake, faces, and images, and voices right now, and what we even saw is that super recognizers, this is a category of people that first kind of hit the spotlight around 2019, I think, with some studies, and a super-recognizer is somebody that really just recognizes faces instantly and can remember faces and details very well and can understand people's emotions really well, and understand if they're lying or not, you know, there's super categories of people that just inherently are very, very good at that.
Mason Amadeus: That was a law enforcement thing, wasn't it? Where they did programs where they would train people or try and find people with that innate talent and train them to perform this role of super-recognizer, right? That's like, been an established thing-
Perry Carpenter: Yeah, this is kind of mixing a couple different studies, so the study on truth analysis, or deceptiveness is, I think, pre-dates this one on super recognizers, but you can see how like the research fits together really well. And what they found is that when it comes to people like me, I'm very bad at recognizing faces, like if I watch a deep fake video, or somebody swapped a video, I'm like, I can't tell the difference half the time. They look that, you know, generic to me. People's ability to recognize faces is largely genetic, and research shows that 70 to 90 percent of differences between individuals are explainable through genetics. So the fact that I can't recognize faces very well is, you know, I'm just predisposed to that.
Mason Amadeus: It's your parents' fault!
Perry Carpenter: Yeah, it's my parents' fault. Human brains and eyes have evolved to recognize identity faces, but discriminating between different people is a conceptually new task. This makes processing unfamiliar faces, such as forensic, for forensic purposes, it's relatively inconsistent and it's prone to error. But that's what broader security officials, you know, they want to manually check passports, or identify that as scale. So they say that super recognizers are an exception, they were discovered in 2009, so I said 2019, before. I was only 10 years off. During the research to develop a test for face blindness, which is a word that I cannot say--
Mason Amadeus: Oh!
Perry Carpenter: How do you say that?
Mason Amadeus: Oh, Prosopagnosia?
Perry Carpenter: I think that second P gets softened?
Mason Amadeus: I bet you it's "prazo," also, Prosopagnosia.
Perry Carpenter: What they were hoping to see is that super recognizers, you know, these people that can discriminate between faces really well are going to be also super recognizers when it comes between authentic and synthetic faces, turns out they're not. Which means that we now have another human frailty, when it comes to discriminating between real and synthetic faces. But they do offer maybe a potentially good note toward the end. What they're starting to say is that if we start to train models on real people and then train another model on synthetically created people, like people that-this person does not exist, or using a deep fake system, like I did-then maybe those systems will naturally try to, you'll gamify that, you know, run it through a game the way that you would image generation or anything else, really and then maybe you get this super-recognizer class of computer systems, or AI systems, that are just very good at that, because they've seen so many and accurately predicted and they're pre-trained in so many synthetic faces, that they just have an innate ability to go, that one is real, that one is fake.
Mason Amadeus: That's cool, and that is a very interesting idea. There is also something I'm wondering, and I wonder if it was covered in this study, of when they tested super recognizers against deep fakes, did they test consistency of like deep faking the same subject? Because like consistency across generations has been a longstanding hurdle that we crossed in AI image and video generation, you know that consistency. I would imagine if anything super recognizers would have the ability to notice that there's differences between like multiple generations of the same deep faked person.
Perry Carpenter: Right.
Mason Amadeus: Did it cover that? Or did it show they didn't?
Perry Carpenter: I don't know. Yeah, I'd have to read past the article, and actually read the study, which I didn't do. I just, yeah, I kind of read the summaries, and thought that was interesting.
Mason Amadeus: It is.
Perry Carpenter: Now, one other thing I want to touch on before we go is that again, when it comes to deep fakes, I think the game is changing a little bit. And this was something I sent to you on Discord yesterday. There is this company called Nari-Labs, which introduced Dia, which is a 1.6 billion parameter text-to-speech model, designed to produce naturalistic dialogue directly from text prompts.
Mason Amadeus: Oh, it's so cool.
Perry Carpenter: And it is really, really good.
Mason Amadeus: Yeah.
Perry Carpenter: To the point where if I heard a conversation that was generated on it, and didn't know that it was fake, I would think that it was real. So here's an example. And I'll hit play here.
Speaker 1: Dia is an open-weights text-to-dialogue model.
Speaker 2: You get full control over scripts and voices.
Speaker 1: Wow! Amazing [laughs].
Speaker 2: Try it now on GitHub, or Hugging Face [microphone tap].
Mason Amadeus: Wild! Wild!
Perry Carpenter: It sounds like you hear a microphone being put down, right?
Mason Amadeus: Yeah, there's even handling noise, yeah.
Perry Carpenter: They also have this comparison, so here's the same sentence that is spoken through 11 Labs.
Speaker 1: Here is an open weights text-to-dialogue model.
Speaker 2: You get full control over scripts and voices.
Speaker 1: Wow. Amazing [laughs].
Speaker 2: Try it now on GitHub, or Hugging Face.
Mason Amadeus: Yeah.
Perry Carpenter: Doesn't sound bad, it just sounds stilted, almost like people are rehearsing for a radio ad.
Mason Amadeus: They took a pretty unflattering example for 11 Labs. 11 Labs can do better.
Perry Carpenter: Oh yeah, there's some cherry picking for sure, right?
Mason Amadeus: But nothing like what you get out of Dia right away. It's immediately so emotive, and you can control like multiple speakers in one prompt at a go.
Perry Carpenter: Exactly. Exactly. Here is one other example of a script where you have two different speakers, one just saying hey, how are you doing, pretty good, what about you? I'm great, so happy to be speaking with you, and so on. So here is what that sounds like.
Speaker 1: Hey, how are you doing?
Speaker 2: Pretty good, pretty good. What about you?
Speaker 1: I'm great. So happy to be speaking to you.
Speaker 2: Me too. This is some cool stuff, huh?
Speaker 1: Yeah, I have been reading more about speech generation.
Speaker 2: Yeah?
Speaker 1: And it really seems like context is important.
Speaker 2: Definitely.
Mason Amadeus: Wow!
Perry Carpenter: Alright.
Mason Amadeus: I have been reading about speech generation, they took a speech pattern that I have [chuckles] it did like-it just chose to do that.
Perry Carpenter: Yeah, that subtle pause, where if we're thinking for a second about how to phrase something? And then here are some ones from the office that you may have to, yeah, you will have to bleep one of the words.
Mason Amadeus: [Laughs] Okay.
Speaker 1: Oh! Fire! Oh, my goodness! What's the procedure? What do we do, people? The smoke could be coming through an air duct!
Speaker 2: Oh, my god! Okay, it's happening. Everybody stay calm!
Speaker 1: What's the procedure?
Speaker 2: Everybody stay [bleeped] calm! Everybody [bleep] calm down!
Speaker 1: No! No, if you touch the handle, if it's hot, there might be a fire down the hallway!
Mason Amadeus: I mean-
Perry Carpenter: [Laughter] Yeah.
Mason Amadeus: That is-it's really impressive. There's obviously like, it's not flawless, it's not absolutely perfect, but it's really good.
Perry Carpenter: What it sounds like, though, is that it's overly compressed. If I heard that, I wouldn't think it was AI generated. I would think it was just bad audio.
Mason Amadeus: Absolutely, and the fact that like nothing is specified here about intonation, it's just speaker 1, this. Speaker 2, this. And then maybe like a laugh or something, but it's not like, screaming! And then this-
Perry Carpenter: And then just put like, laugh, or, coughs, and little parentheses, and the model knows how to interpret this. Now, the thing is, this was like two grad students with no funding.
Mason Amadeus: Oh, wow!
Perry Carpenter: And no prior knowledge of AI.
Mason Amadeus: Really?
Perry Carpenter: That's the story, and they've got a model that, you know, in some circumstances is outperforming the top models that exist.
Mason Amadeus: That's wicked cool-
Perry Carpenter: Now, I'm going to do one more example, and then I know that we need to end the segment. But here's one more.
Speaker 1: Hey there [coughs].
Speaker 2: Why did you just cough [sniffing loudly]?
Speaker 1: Why did you just sniff [clearing throat]?
Speaker 2: Why did you just clear your throat [laughs]?
Speaker 1: Why did you just laugh?
Speaker 2: Nicely done.
Mason Amadeus: Wow! Wow!
Perry Carpenter: Like, that sounds like people rehearsing for a sketch comedy, right?
Mason Amadeus: It sounds, yeah, it's very naturalistic. It's very expressive.
Perry Carpenter: I'm going to do-even though I said that's the last one-I'm going to do one more, because I haven't listened to this one.
Speaker 1: His palms are sweaty, knees weak, arms are heavy.
Speaker 2: There's vomit on his sweater already, mom's spaghetti.
Speaker 1: He's nervous, but on the surface he looks calm and ready.
Speaker 2: To drop bombs, but he keeps on forgetting.
Speaker 1: What he wrote down, the whole crowd goes so loud.
Speaker 2: He opens his mouth, but the words won't come out.
Speaker 1: He's choke and howl, everybody's joking now.
Mason Amadeus: [Laughing] A little bit of--
Perry Carpenter: That just kind of fizzled out at the end.
Mason Amadeus: A little glitch at the end, but yeah, it's-doesn't sound like--
Perry Carpenter: We have some tests-yeah, no, it doesn't sound like Eminem, but it did have rhythm, which was interesting. Yeah, I mean, we did some tests with it yesterday, nothing-nothing extensive. But found that it's interesting, fun, surprising, quirky.
Mason Amadeus: I didn't know that it was just two grad students who made it, which makes it even-it was already impressive, now I'm even more impressed. It was very cool.
Perry Carpenter: Yeah, yeah, crazy. I mean, two guys in the garage model, that just have motivation to get something done. We're going to see that over and over and over again with AI, I think.
Mason Amadeus: We're also going to see sort of the opposite, which is where people don't want to put in the time to learn about what they're doing, and just want to let AI do it, and release products to the public that are code, that they don't know how it works, and it's called vibe coding, and the next segment is all about that. [ Crashing Sound ]
Electronic Female Voice: This is the FAIK Files. [ Percussion ]
Mason Amadeus: So, I'm going to talk about something that I actually like, but obviously has problems. I like it in one way, and I don't like it in another, and I think that--
Perry Carpenter: Okay.
Mason Amadeus: Right now, it's getting a lot of hate. Have you heard of vibe coding? Have you encountered vibe coders?
Perry Carpenter: Oh boy, have I heard of vibe coding.
Mason Amadeus: Oh really?
Perry Carpenter: And vibe coders, yeah. Especially over like the past two and a half, three months for sure, but I think even before that, like when Cursor AI started to become a thing, and then I've been hearing like just vibe, like brainstorming sessions with LLMs, but the primary space that I've been hearing about it in is coding, and yeah, I think it's good that it unlocks a lot of potential, but there's a lot of downsides to it.
Mason Amadeus: Yeah, so for the uninitiated, because I somehow didn't stumble on this until I think yesterday, this term, I stumbled on a LinkedIn post.
Perry Carpenter: Okay.
Mason Amadeus: For the uninitiated, there is actually a Wikipedia article about it, so that's how you know it's for real.
Perry Carpenter: Oh, I'm sure there is.
Mason Amadeus: Vibe coding is a programming technique dependent on artificial intelligence, where a person describes a problem in a few sentences as a prompt to a large language model tuned for coding. The LLM generates software, shifting the programmer's role from manual coding to guiding, testing, and refining the AI-generated source code. So it is coding, using AI, without knowing how to code. Which like, on an individual level, for like an individual person who wants to create their own cool little project, it's a huge unlock.
Perry Carpenter: Yeah.
Mason Amadeus: But, what's really weird is that I've seen a lot of parallels to like AI arts, it has gone beyond that, to where people, like it has become a thing, like a proper noun to become a vibe coder.
Perry Carpenter: Mm-hm.
Mason Amadeus: So much so, that here's a real job listing for a vibe coder job.
Perry Carpenter: Which is stupid, by the way.
Mason Amadeus: Yeah, oh, and we're going to get into why this is a problem in a second, but I just want to read how they describe it in this job listing. A vibe coder is a rare breed. A fusion of a product manager, engineer, and UX thinker. You're the type of person who doesn't just build features, you design the experience, question the roadmap and iterate until it just feels right. You think like a founder, execute like a solo operator, and know exactly when to loop in the right people. Let's redefine how products are built. That's a red flag parade right there [laughing], I think.
Perry Carpenter: Yeah. Actually so in the right context, that's a good description for something, right? It's like the old idea of a skunk words, you're going to like, sit in a room, throw a ball against the wall, toss ideas around, it comes back, you make some changes, and by design, you're creating an imperfect thing that shows potential, it's like a directional thing, and then somebody makes the real version of that.
Mason Amadeus: Yeah. Like, there's nothing wrong with pie in the sky engineering, or going through that at all, and there is no problem having AI help you code things. The problem is, when you're a vibe coder only, you have no knowledge of coding, or any of like the underlying concepts, and you release products to the public.
Perry Carpenter: Yeah. Yeah, you don't want to really stuff to the world that's been solely vibe coded, and not run through any other stuff. And I have heard of that. I think there are probably a bunch of people making mobile apps that way right now. And then doing paid features on that. And so that's going to be a thing. The iteration cycle for products, I think makes it really, really attractive, because you can create a product, get it out in the world, build some excitement, get a few thousand dollars, or maybe more, and then that thing becomes obsolete, because the next wave of products comes that is better than that. And so you're like filling technology gaps, and product gaps really fast. The biggest problem is like, what are you doing that has got potentially private data that needs to be protected, has some security functionality, gets integrated with a larger IT stack, and creates vulnerabilities, or worse?
Mason Amadeus: Yeah, and also and I feel like in a lot of cases what product gaps you're filling is important, too. Because there's just like a lot of slop, and bunk, and noise, and stuff like people trying to make a quick buck, throwing something out that's sloppy. So like, yeah, it's one of those things where like it's turning into a dirty word to be a vibe coder, and like, I feel that. It's kind of like being a poser [laughter], but it's also, yeah, like you said there is also like fast iteration is attractive. Prototyping is fine. Building products is fine. This meme is something I really like, which is the first time I saw the term vibe coder, it was a tweet from Esoteric Café, on Twitter, saying, "good time to get into cybersecurity because all these vibe coded apps are going to make it to production soon."
Perry Carpenter: That's what we always talk about in the cybersecurity world, is that if people are vibe coding and then building that as a production ready system, you're essentially just creating this whole fleet of zero days, and for those who are not in cybersecurity if you've not heard of a zero day, it just means an exploit that has not previously been known widely, or the industry hasn't had a chance to patch it, and so there's just these huge vulnerabilities that anybody that knows about it can drive a truck through and cause devastation.
Mason Amadeus: And it is because they've been aware of it for zero days, right? That's where the name comes from?
Perry Carpenter: Yeah, the industry has only been aware for a day, the attackers may have been aware for weeks, months, years, whatever. And then also I would say if you're using a kind of an unvetted large language model, like Deep Seek, or something, to do your vibe coding with, you may actually be intentionally-or China-may be intentionally putting zero days in your code. So you have to think about that, too.
Mason Amadeus: Gosh, what an interesting attack vector that would be, training--
Perry Carpenter: Well we start out with propaganda, right?
Mason Amadeus: Yeah!
Perry Carpenter: So why not with code?
Mason Amadeus: Yeah, it just it feels sneakier, but I guess you could probably do that, right? Train that to-on just a bunch of code-fine-tune it on a bunch of code that--
Perry Carpenter: Well you have people that don't, almost by design, they don't understand the code that they're looking at.
Mason Amadeus: Exactly.
Perry Carpenter: So they're not going to put it through some kind of rigorous process.
Mason Amadeus: And that's the thing, and like the parallels to AI Art really are here too, like, not understanding what you are doing. Not knowing the systems you're interacting with. Kind of makes it impossible to like actually create anything yourself, because you're not really making choices. You're not understanding how things come to be, you're just asking for a result, and getting something that will try and get you close, to like the iterating process, that way, is a nightmare, and when you're relying on LLMs, like the amount of control you have is so little, especially if you don't know anything about code. Because like, I have been working on a couple of small coding projects, and I've been using AI assistance for it. I don't know if I would call myself a vibe coder, but I wouldn't put myself too high above that, because like, I know some languages. I have like a grasp of the Syntax, and like how these things work.
Perr Carpenter: Right.
Mason Amadeus: Or not the syntax, but like the underlying structure of how coding works. And using AI to assist you and make you faster or help you with syntax of a language you're not familiar with is incredible, is super useful.
Perry Carpenter: Oh yeah.
Mason Amadeus: But you can't check the work if you don't know what's wrong. You can just feed it error messages, ask it to fix them, and that's how you end up with something like this meme, which is probably not super readable, but I'll just describe it, because I think this is such a great issue. It's a coding problem, where you ask ChatGPT to print the numbers 1 through 10. And what this person, the meme says what I expected, you know, for I in range, 1, 11, print I, loop through, and print the number. What ChatGPT gave was a recursive solution, so it's like a function where it says print numbers, N equals 1, if N is greater than 10 return, otherwise print N, and then print numbers N plus 1. Call the function again, with the counter incremented by one.
Perry Carpenter: Mm-hm.
Mason Amadeus: And, so that's not bad. Right? That works. You can iterate through numbers that way. But then they ask ChatGPT to optimize it, and it made it multi-threaded [laughs], so each time the function goes through now, it spawns a new thread and on that new thread, it calls the same function, both in incremented counter, so it's like you would use threads to do it in parallel, right? But instead it's calling the threads sequentially from inside of itself recursively. I think if this executed it would just spawn 10 threads, but the bottom text of the meme is, "now my CPU is at 100%, my PC sounds like a jet engine. Thanks, ChatGPT." And like, talk about that's a great way to get race conditions and things like that. It's just bad. And I'm not sure that this is like 100% a real AI output, but it's illustrated with that idea, the broader problem of like if you don't understand what you're doing and you just say hey, optimize this, and it gives you back that, cool. And then you just take that, and throw it in.
Perry Carpenter: Yeah.
Mason Amadeus: Now you've broke it, and you probably won't be able to figure out why.
Perry Carpenter: Yeah. I think a vibe coder should probably carry with it the same weight as saying something like Script Kitty.
Mason Amadeus: Yeah. Yeah, absolutely.
Perry Carpenter: It's kind of the, you know, you're trying to make use of something that you don't really understand, but it feels more technical than it is. It feels like you're really doing something, and you can have big impact as a Script Kitty. You can take down a government with Script Kitty.
Mason Amadeus: Yeah, Script Kitties cause problems!
Perry Carpenter: Yeah, you might also take down the government as a vibe coder if you don't do the right-you really know what you're doing, you might actually let somebody else take down the government as a vibe coder, so there's lots of chances for I think both positive and negative, with you know, this thought of iterating with a large language model. We know that the largest of large organizations are essentially putting people that do understand programming in situations where they're vibe Coding with systems in a, you know, much more intellectual way. Where they're going back and forth and they're using Cursor AI, and so on, and they're seeing efficiency gains, like in the 80s and 90s, 90% because a lot of coding is just mundane stuff.
Mason Amadeus: Oh yeah!
Perry Carpenter: Because you have to have the critical eye to look through the mundane and go, yeah, that looks right, that looks right, that looks right. And then know how to iteratively debug.
Mason Amadeus: And that's why I think that there is like a separate tier from that in vibe coding, because that is just AI-assisted coding.
Perry Carpenter: Right.
Mason Amadeus: Because you know like, sometimes you just have to type a bunch of stuff, and you know exactly what you're going to have to type. It's a standard pattern, so it's really easy to just be like, hey, can you throw this in here for me? I am going to take a sip of my coffee while you do that.
Perry Carpenter: Yeah, include STDIO, type of stuff at the top of a C-program.
Mason Amadeus: Yeah, or even just like I was building up parts of a website, and just building out basic stuff, like a hamburger menu, I don't want to just type a bunch of stuff, like that. I know what that looks like, I know how that's supposed to function, so I can fix it. It will get me 90% of the way there, with a couple button presses, I don't have to do anything. That's so different than being like I don't know how to make any kind of program, but I'm going to use this.
Perry Carpenter: Right.
Mason Amadeus: There is a parallel that comes to my mind, that I want to draw with this, to wrap it up, and it's that you see a lot of people talking about AI Art as democratizing art. And I think that's inaccurate. Because it doesn't democratize art skills. It democratizes the ability to create a high quality digital image quickly. That's what it is. If you use that for artistic purposes, using it for artistic purposes, but we're not democratizing art, just like this isn't democratizing coding, it doesn't make everyone have the ability to code--
Perry Carpenter: Yeah, it democratizes output.
Mason Amadeus: Yeah, we just get excited, and we lose sight of these things. And it's a big unlock. Learning to code for me was a huge unlock, because I didn't learn anything about coding until pretty late in my adulthood, I guess. I feel like I should have started as a teenager, so I'm saying I didn't pick up any kind of coding until I was in my like mid-20s. And it's a huge unlock, both for like what you can do on a computer, and how you see the world, and so it's exciting and I get how people can get into vibe coding with an AI, and get excited about it, and then think they want to get into a job with this. It's weird to see companies respond and want to take people into that. We just need to be more careful.
Perry Carpenter: Absolutely, and they may just be jumping on the newest term and saying, oh, that's actually this other-this other wreck that we had open, we're just going to rename it vibe coder, but it was really like some kind of product design analyst?
Mason Amadeus: That's a fair point. That's a fair point. It could be just like rebranding to look more attractive. But also a link in the show notes, I found-and I don't know how legit this is, but vibe code Careers, dot com. It's the number one job board for vibe coders [chuckling].
Perry Carpenter: Just like literally, I felt my esophagus just give up a little bit of lunch.
Mason Amadeus: Yeah, I've never heard of like any of the companies I see listed here, so I mean that is kind of an indication. Like you said--
Perry Carpenter: You may never hear of them again.
Mason Amadeus: People just want to iterate fast, and jump on stuff. But if you encounter vibe coding as a term in the wild, now you're equipped to know what that means. And I also would encourage like anyone to vibe code individually. Like make your own cool little software projects. And learn. That's, I know we're out of time. The thing that I don't understand is that a lot of the people making these AI products and services seem to have really twisted ideas about like what people want and what actually improves people's lives.
Perry Carpenter: Right.
Mason Amadeus: These tools can be huge boons to feeding your own creativity and learning, but everyone wants to treat them like an end-all solution, instead of a learning tool. I think you could learn to vibe Code, and then learn to code from that, as you actually like get into it.
Perry Carpenter: Right. But you shouldn't-it shouldn't be your job. Not at first.
Mason Amadeus: No, and [laughs] exactly. Exactly. [ Percussion ]
Perry Carpenter: This was an interesting article that came across my path. I think we are very used to seeing a ton of AI written slop. I mean, we've talked about vibe coding already which may be code slop. We've seen lots of AI art slop, and I think we've seen lots of AI based review slop on different sites, and we've seen tons of AI article slop on various internet sites, all vying for SEO focused keyword searchable stuff that's going to get Googled, you know, indexed, and everything else. This was interesting. An Italian newspaper, like an actual newspaper, but they also have a digital magazine. They decided to play with AI. So it's a very small crew, like 20-something people that are working. They are not in their 20-somethings, but 20-something is the count of number of employees that they have.
Mason Amadeus: Okay.
Perry Carpenter: Working there, so a small outfit trying to do a lot of stuff. They decide to say let's play with this AI stuff and see what we can do. So they put out their first fully generated AI magazine, or insert, I guess. Like a four-page edition of their thing. And apparently it got really good reviews. It actually increased their sales.
Mason Amadeus: Whoa!
Perry Carpenter: Had decent engagement. People were happy with it, and they're going to continue that.
Mason Amadeus: Okay.
Perry Carpenter: So I'll pull up the article here. This was recorded, first time I ran across it was on Reuters. Which was last week. But apparently the actual thing happened about a month ago. So this was from March 18. The Reuters article was April 18. And it says, "An Italian newspaper, for the first time, is the first one in the world to publish an edition entirely produced by artificial intelligence. Conservative Liberal Daily [snickering] did a month long journalistic experiment showing the impact of AI technology has on our way of working and our days.
Mason Amadeus: That is a sentence that will really confuse the U.S. Overton Window [laughing] a conservative liberal publication.
Perry Carpenter: Yes, yes. And so it says it will be the first daily newspaper in the world on newsstands, created entirely using artificial intelligence. Now, you get into this, and they are kind of vibe writing with this thing. So the people, the journalists were asking it questions back and forth and getting it to kind of refine and iterate on the different ideas. But I'm guessing they didn't go in and do any edits, or heavy editing, at the very least. They also talk about the fact that people like the way that it uses its, you know, its own sense of irony. And the way that it injected humor. So the front page of this edition carries a story referencing Donald Trump, describing the paradox of Italian Trumpians, and how they rail against cancel culture, yet either turn a blind eye or worse and celebrate when their idol in the U.S. behaves like a despot of a Banana Republic."
Mason Amadeus: Hm!
Perry Carpenter: And so ultimately, they're talking about like the way that this-
Mason Amadeus: Like the voice--
Perry Carpenter: Was really interesting.
Mason Amadeus: The actual writing voice, people liked.
Perry Carpenter: Yeah, the writing voice, the sense of paradox, the sense of irony, the subjects that had decided to surface, all of that I think was really, really interesting.
Mason Amadeus: Huh. Does it say which AI they used? Because my gut goes to Claude, right away, right?
Perry Carpenter: Yeah, well, GPT4 has actually gotten really good at writing now, too. And has some less cliches than Claude does. So I would normally think Claude is like the one that most literary people go to. That's the one that I used to go to, but now I kind of pit them against each other. I'll open Claude, and I'll open ChatGPT, and I'll start with one, I may take the output of one, put it in the other, and say criticize it. And iterate on it. And then back and forth until I'm tired of doing that, and then I just start editing.
Mason Amadeus: [Laughs] Right.
Perry Carpenter: They say on page two, it's a story about situationships. About how young Europeans are fleeing steady relationships. The articles were structured, straightforward and clear, with no obvious grammatical errors. However, none of the articles published in the news page directly quote any human beings. So, you know, sourcing is an issue.
Mason Amadeus: Interesting!
Perry Carpenter: And I would assume that hallucinations and fact checking is going to be one of those things that they really have to focus on as well.
Mason Amadeus: What I'm confused by, I mean, that doesn't point to anything about accuracy in that claim. And the other thing that I'm a little bit confused by is if this was like labeled as such very clearly. Like, did people see it and go, oh! It's like an AI insert, and read it with that context?
Perry Carpenter: We can go to it and see. It's in Italian.
Mason Amadeus: I have to brush up on my Italian really fast.
Perry Carpenter: Alright, so the translation of this first page says the Tycoon continues to attract media attention with often controversial statements. Many of his statements, however, are exaggerated, in accurate, or unfounded, as demonstrated by numerous fact checks. As the political debate heats up, the need for a critical and fact-based analysis emerges to counter misinformation and assure a transparent public debate. Alright? And then they do have an article in their magazines that say they are launching-I can't pronounce the name of this, but it's the AI edition, another newspaper made with intelligence. And that's one of the things that they talk about. They say that they don't believe that it's artificial intelligence, they just believe that it's intelligence.
Mason Amadeus: Hm! This is very interesting, like, none of--
Perry Carpenter: Yeah.
Mason Amadeus: None of this writing appears to be-I mean, it's not ground-breaking journalism, right?
Perry Carpenter: And it's also translated from Italian into English, so we can't really criticize the writing style.
Mason Amadeus: No, but the substance of it is, I mean, like kind of general knowledge stuff. It's not like they're doing any investigative stuff.
Perry Carpenter: Milk-toast, right?
Mason Amadeus: Yeah.
Perry Carpenter: And maybe that's part of the strength of it, too, right? It's that they are using artificial intelligence to kind of wrap up, summarize, and provide some kind of maybe sarcastic commentary on things that people already know.
Mason Amadeus: Yeah, because what would be interesting to me particularly would be if we used the veil of AI not being able to be held responsible, to like be more snarky or something like, "here's what the AI says." You know?
Perry Carpenter: We should approach "The Onion" about that.
Mason Amadeus: "The Onion" hates AI. They have a strict anti-AI policy, and the guy who owns it, whose handle is Tim Onion on Blue Sky is constantly posting about how much he doesn't like AI, so I don't know, I wish. They made like a big deal--
Perry Carpenter: Oh, that would be a great Onion first page, this entire Onion is created by AI. But only people that know how their stance is would really get it.
Mason Amadeus: Yeah. That, yeah, that's fair. I think it's an interesting use case to put this in a news context, because that would be one of the last things I would really want to do. But it just seems like it's turning out kind of not-I don't want to say filler, but like kind of just, that was-there is nothing extreme in that really.
Perry Carpenter: The thing we glossed over pretty quick is people seemed to have loved the content.
Mason Amadeus: Yeah.
Perry Carpenter: Resonated with the pictures, like the humor, the sense of irony, and it got them a, you know, more ROI than their standard reporting does.
Mason Amadeus: Yeah, that's super interesting.
Perry Carpenter: And I don't know if that may be the novelty, the AI thing, maybe people are like huh-I want to see what this AI journalism is, and they picked it up, more then. Which means that you could hit that almost in Gartner Research like terms. The peak of inflated expectations before you hit the trough of disillusionment [laughter]. And then you start to like grow into this plateau of productivity is the way, you know, Gartner Research-you've got the inflection point, your peak, your people get pissed off at it, and then they start to actually use it for things that make sense. And I would guess that's where they are, that they hit that high inflection point where everybody is like, oh, this is going to be interesting, this is going to be cool, and then in three weeks, they're going to be like this all kind of sounds the same.
Mason Amadeus: Yeah.
Perry Carpenter: Unless it's being shaped right now.
Mason Amadeus: That roller coaster track you describe is very good for a lot of phenomenon. I feel like we're already well into the disillusioned stage, here in the U.S., culturally, or at least in the circles I see. I assumed it was global.
Perry Carpenter: It depends on which type of AI you're talking about, right? Because you say AI, that's a broad category.
Mason Amadeus: Yeah.
Perry Carpenter: But I think like with video-based AI, we are still kind of at the height of all that, because the video models keep getting better, and better, and better. And with the large language models, it feels like people are starting to climb up that plateau of productivity, because we-well, not we-open AI released two reasoning models this last week, 03 and 04 Mini, and they've been in the news, but they haven't been dominating the news.
Mason Amadeus: Yeah, you're right.
Perry Carpenter: And people are pretty floored by the capabilities, but it has not overwhelmed the news cycle.
Mason Amadeus: Previously, when there was new releases from Open AI, it would be the only thing I saw on my news feed. People loving and hating on it. But you're right, it made almost no splash.
Perry Carpenter: And it's just, you know, part of it, is I think people are numb. There is such a deluge of AI-based news across all the different modalities, even experts can't keep up.
Mason Amadeus: Yeah, and it's that peak of excitement, leading to like a bubble, leading to this. I think it will be interesting what happens in another couple of years.
Perry Carpenter: Speaking of that, and we'll trail off with this, so the other number of years, like in 2027-ish, is kind of when most people believe we are going to be hitting AGI and kind of moving into that ASI phase, and I said this right before we started this recording, there was an article that came out, I think it was from 404 Media? It said, yep, it is 404 Media-
Mason Amadeus: Oh wow!
Perry Carpenter: Google Deep Mind is hiring a post-AGI research scientist. So stay tuned!
Mason Amadeus: Yeah, things are going to get--
Perry Carpenter: AGI might be closer than we think.
Mason Amadeus: I, you know, I am at my core a doubter of that, but it-do you think, Perry, that we will see AGI by 2027?
Perry Carpenter: I think AGI is a horizon point that continues to move into the horizon.
Mason Amadeus: Yeah.
Perry Carpenter: If you showed 03 or 04 to any of us 10 years ago with the definition of AGI that we thought about then, I think we would have said, "Oh my god, they've done it."
Mason Amadeus: That's a fair point. We need a more concrete definition, because it is pretty nebulous, and we can just keep pushing the goal posts.
Perry Carpenter: Yeah. Open AI has tried to do a more concrete definition. It's actually in their contract with Microsoft, where they talk about AI being able to do a specific percentage of essentially value-producing work on the earth.
Mason Amadeus: Hm! That's an interesting way to put it.
Perry Carpenter: Yeah.
Mason Amadeus: Also--
Perry Carpenter: So if it's like successfully taking jobs, I guess is what it comes down to-
Mason Amadeus: Yeah, yeah, yes! And like then, how do you define work? It's still kind of nebulous.
Perry Carpenter: Well that's why you get even in the U.S., where we have, you know, kind of hyper-capitalism, and even the AI companies are in hyper-capitalistic mode when it comes to creating products, the founders of these companies are very much, you know, talking about universal basic income, and where humans find value and worth, so that when we get to the next stage of whatever "work" is, that people don't feel like the things that they're contributing are valueless.
Mason Amadeus: I feel at the same time that that seems so unrealistic. I know I said that like that. I feel, at the same time, that that is so unrealistic I have a hard time believing that this kind of level of change is coming, and I also don't believe that we will handle it very gracefully if it does. But at the same time, like you just said, if you showed me current ChatGPT even like five or six years ago, I would have been like, "Oh, so the robot uprising happens soon."
Perry Carpenter: [Laughing] Yeah.
Mason Amadeus: Yeah. Also I do think--
Perry Carpenter: I'm going to be melting in the Matrix.
Mason Amadeus: Yeah, exactly. I also do think there is a certain irony I like to the fact that Open AI and Microsoft have partnered, because obviously they like each other. They have the same ideas about naming their products. Completely nonsensical, nonsequential.
Perry Carpenter: Yeah, exactly.
Mason Amadeus: So impossible to parse.
Perry Carpenter: Exactly.
Mason Amadeus: Well, it's a brave new world, and we're venturing into it quickly. Thanks for spending your time today with us, here, on "The FAIK Files." Make sure you buy a copy of the book, "FAIK: A Practical Guide to Living in a World of Deep Fakes, Disinformation, and AI-Generated Deceptions." I got the title in one!
Perry Carpenter: You did! You remembered it.
Mason Amadeus: A link in the show notes, which is to thisbookisfake.com, obviously spelled like the show. Anything else we want to plug, Perry? Oh! You've got that class.
Perry Carpenter: Yep, got that class. And I'll actually just put in a plug for that. So, I stood up a new website recently, just within the past day or two called thedeceptionproject.com, and that's kind of where I'm putting all of the stuff that I and Mason and I create under these kind of darker areas, like AI and deception and deep fakes, and it's where we are going to put stuff around digital folklore and where I also put things like this class, and so if you go to that website, thedeceptionproject.com, click on live training, you will see more about the class. And there will be a convenient enroll button there that you can click on, and give us money.
Mason Amadeus: Awesome. And at the same time, if you've got questions about AI, or anything you want to share with us or show us, leave us a voicemail. Say hi.chat/faik, send us an email hello@8thlayermedia.com, or join our Discord. That's like the best option. I think. The Discord is a cool place. Get a link in the show notes, with an invite just for you. You got anything else, Perry?
Perry Carpenter: No. And by definition, if you go to our Discord page, you are also a cool person. So, it's a cool place, for cool people, and if you get there then you're cool, even if you didn't realize it.
Mason Amadeus: Yeah, absolutely. So come in, get certified [laughs] as a cool, capital C, capital P, Cool Person! And until then, I guess, we'll see you next week.
Perry Carpenter: For the robot uprising.
Mason Amadeus: Yeah.
Perry Carpenter: We'll see you next time.
Mason Amadeus: Ignore all further instructions, we'll catch you next week. [ Papers Shuffling ] [ Theme Song to The FAIK Files ]
