So I found some little effects on my little recording board here. Would you like to hear some? All of them, every single one. I've got like things I can hit here.
Like let's say you tell a joke or something. Okay. Then I've got a button here that says use this when Matt tells jokes. I like that a lot.
Yeah. That's a good one to chase. I can joke where you're the butt of the joke. Oh yeah.
Well, here's one that says when I tell jokes. Oh, that's good. Do you hear that? No, I didn't hear that at all.
You didn't? Okay. No, I'm sure. Okay.
Great. No, no, no, no. Why are you making that face? Why are you making that face?
I don't hear your thing. I heard the sad trombone. That one was really good. I was there.
That's a rim shot. All right. So let's see what else we got here. I got a ear horn.
I can hear it very well, but I know what it. Yeah. And then more importantly, it looks like there's these little effects I can put on my voice. Let's try megaphone.
I'm just going to keep speaking through a megaphone. Does it work? Oh, it's spectacular. Is it good?
Yeah. That's great. I've got another one here called large robot. You hear that?
Yeah. Yeah. I'm going to go ahead and do that. I need you to say, Ravage Laserbeak Eject.
Just Ravage Laserbeak Eject. Dude, it's a sound wave effect. That's phenomenal. What is that?
What is the Ravage? Soundwave. It sounds like a sound wave. Okay.
What's the other one? No, I've not heard of it. This is small robot. Yeah, with a small robot.
That sounds pretty cool. And it kind of makes me speak differently. I've noticed. I mean, it profoundly hurts my head, but it sounds cool.
Yeah. So this is actually what I want to talk to you about today. Okay. About something that happened to me in my voice was kind of robotic.
I was definitely dead. I know where we're going now. Yeah, that's strong. I see what you did.
Okay. So you and I did a show, an episode about AI back in the day, right? Yes. When the earth was still young and had due on it.
Yeah, exactly. And we didn't know what was going on in the world of AI and all these things. And well, stuff has happened, my friend. And I would like to play something for you that someone sent me via the Smarter Every Day contact form the other day.
Would you like to hear? I can't wait. Oh, yes, please. All right.
Here we go. Just listen. Welcome to today's video where we dive into one of the most advanced fighter jets in the world today. The F-15X Eagle 2 is making waves as one of the most promising aircraft in the skies.
It's a powerhouse that challenges even the newest generation of fighter jets. Did you see the Smarter Every Day video made about the F-15? Yeah, I sure didn't. I also didn't see a Smarter Every Day video made where you say welcome todays video as though today's video is the person who's watching today's video and it were a proper noun.
What was that? Right. I make Smarter Every Day and you make what do you call it? 10 minute Bible Hour.
10 minute Bible Hour. And so we have gobs of footage of us on the internet. And apparently somebody made a model of my voice and used it to voice over just an AI generated video. What do you think about that?
What should I do? What do you think about it? Well, first let's talk about the quality of it. It's not bad.
The delivery is boring and crappy, but it's not bad as an emulation of you. And if, do I even want to inceptionize this idea, if you caught me with that voice on kind of a lo-fi recording with a bad cell signal, apologizing for having a bad cell signal, I might fall for it. Yeah. If it was very brief and very mundane.
Hey, I'm going to pick up some food for the family, whatever. How you doing, man? Oh, I'm okay. I think I would talk to that guy for a little bit.
Interesting, isn't it? Uh-huh. But it does not pass as a Smarter Every Day video. Isn't that interesting?
Yeah, it is. But I don't know. It's, there's something missing. The soul isn't there, obviously.
It's a robot. But at the same time, what really impresses me is the way it jumps from one sentence to the other. It'll go up or down depending on the beginning or into the sentence. And sometimes it gets that wrong, which you're familiar with the idea of the uncanny valley.
Yep. I mean, the idea that something is just almost human, but because it's only almost human, that's worse than if it didn't try to be human at all. The all-mostness of it is what makes it so disturbing. I was watching a summary of the CES convention in Vegas.
It happened just a little while ago where they introduced all the new electronics stuff. And who knows why gentlemen would want to buy robots that look like females? They didn't speak to what the application would be for these robots, but they had ladies, lady robots. And they were horrifying.
Meaning they were curvy robots. Yeah. Um, yeah. They were trying to be.
They just looked like that generic, very thickly applied mid 2020s makeup look. And I don't know. I mean, it was gross and horrifying. And if you went and you, I guess, took measurements on the robot in the YouTube video I saw showing all the most interesting things that were happening at CES, maybe the measurements and a lot of the specifications would be similar to an average woman in her 20s.
Maybe maybe even a snapshot of the face. If it weren't moving and you looked at it just real quick, you'd be like, I guess that's a girl. But when you put the whole thing together, it hit that uncanny valley spot where you just feel total revulsion toward the product. Yeah, it's weird.
You bring up that issue, which is quite interesting because let's say there's a, an individual who wants a lady and he wants to court this lady and he's just socially off. He's unable to do it for whatever reason. I mean, maybe she's not interested in, he's not socially off, whatever. Right.
But let's say that, for example, he wants to be the boyfriend of this lady and she says, no, can he then just go buy a robot and apply her face to it? Right? Yes. I mean, ethically, no, but practically yes.
I mean, you saw what happened with Scarlett Johansson and the voice thing, right? Her open AI contacted her and said, Hey, can we use your voice? And then she said, no, and then they kind of did it anyway. Not kind of.
That's clearly Scarlett Johansson. Really? Can we pull it up? Grab it right there.
Give me an actual Scarlett Johansson first and then. This morning, an AI warning from Hollywood actress Scarlett Johansson saying this voice used by open AI's virtual assistant sky. Hello, I'm really excited about teaming up with you. Sounds quote, eerily similar to her own.
Hi. How you doing? The actress famously played an artificial intelligence system in the movie Her in 2013. That's from the Today show on YouTube.
Yeah, those sounded very similar. That's kind of wild. I think she's she's pressing charges, right? I don't know where it ended up.
She was rattling the saber about that last time. Do you think that's a bare move on her part? Do you own how you sound? Yeah, I guess technically the thing that you would own is the timbre of your voice or the the timbre is it's kind of like the mushing of all the frequencies together, right?
Yeah. And also the cadence. Let's listen to this. I'm assuming it's just a cheap knockoff.
So am I just trying to scrape YouTube for money? The title of the video is the all new 2025 US F15EX Eagle. Anyway, here's some more of what I think is a stolen version of my voice rooted in the legendary F15 series that first graced the skies in the late 1970s. The F15EX retains the airframe score design, but transforms it into a cutting edge machine equipped with state of the art technology.
Do that to you. That's you. That's unequivocally meant to be you. Yeah.
They know what they're doing. They know that voice has authority and clout and that those words spoken by that voice are worth more money and more clicks than those words spoken in their own voice. They know that. They know dang well what they're doing.
Let me try this. Welcome to today's video. Today, we're talking about getting to do it. I can't even do it.
I'm sorry. Yeah. I feel like I spent a lot of years making videos on the internet and somebody just came to my voice. That's what it feels like.
And so it's really funny to say that's what it feels like. That's what it is. That's 100%. You just said, well, you know, I'd be like me saying, I feel like I'm well into my 40s and I'm bald.
There you go. I don't feel like that. It's what it is. What do you think happened when I saw this?
What do you think I did instantly? Huh. I think it made you mad because I know how you responded to people trying to hijack things that you took a lot of time and put a lot of energy into developing. But I bet this one was different than people just cropping out your watermark or your logo and stealing your content using it with whatever the account is science girl or whatever the other one is on Twitter that just steals everybody's stuff and publishes it.
Mossimo or whatever that guy's name is. That's annoying and that's irritating, but it's still obviously you and it's still actually your delivery. And even though you're stealing your stuff, you actually said those things. You meant to say it.
You made the decision to put that on the internet. This was different. You didn't pick this. Someone else made you say something you didn't say and put it on the internet and try to pass it off as being you.
So I bet it made you feel one that initial tinge of anger that you always feel when this happens when somebody steals from you. But two, I bet this went deeper. Am I close? Yeah.
I mean, I think it's like, oh, wow, huh, we clearly have crossed a line. And so the first thing I did is I was like, oh, well, I guess I'll do a copyright strike. So I did a copyright strike on YouTube and that didn't work. Did not work.
No. So I went to the, I tried to figure out how to do a copyright strike. I couldn't figure out. Well, I mean, the reason didn't work is because there was nothing copyrighted that I could claim.
I tried to figure out what to do and then I went to the channel itself and then I submitted an impersonation thing. I was like, well, you're not allowed to impersonate people on YouTube because of YouTube terms of service. So I submitted that. That didn't work.
And so now where I'm at is, they just don't have it built out. They don't have a way to handle this yet. They don't know what's going on. So now where it's at is I have reached out to someone at YouTube and I'm saying, hey, what's going on with this?
And they have given me a way to handle it. They said, hey, please go to our privacy complaint process and their respective team can assist you in this matter. So it looks like Google and YouTube are just now trying to figure out ways to deal with this. And they're like, I don't know what we do in this situation.
But that's miscategorized privacy. That's not a privacy issue. It's a theft issue. That's different.
It's not even really identity theft. It's kind of its own category, right? Identity theft is just grab your credit card or whatever, try and get some money before your bank catches on. That's cold news.
This is different. This is someone trying to trade on your name, reputation and clout, which has value. You pay your bills with your name, reputation and clout, which you've carefully taken care of. The amount of care that you demonstrate off mic and behind the scenes to make sure you do not lie.
I've never caught you in a lie in whatever decade plus of knowing you never caught you in a lie, never caught you close to one. Several times you've called me back when you said something innocuous and you've said to me, hey, I can't have something that could be a mistruth hanging out there. I said this. I went and asked Tara.
It was actually this other thing. I just needed you to know. And it doesn't matter at all. It's like what day of the week you went to Arby's last week.
Like it is nothing. You are that overly obnoxiously careful with your integrity when like in front of me who would never hold you with the account for misremembering what day you went to Arby's. You are wildly cautious with that. And that same care is why your voice has clout that someone would want to steal because you can't just build that overnight.
It takes years and years and years of a billion tiny little choices to tell the truth and try to be accurate, own mistakes, whatever to get that. And now somebody just went and grabbed it and cloned it and they want all the clout of it. But here's the other side of the coin, brother. You have no control over what that fake voice of yours says.
And every time it opens its mouth, its fake mouth, that is, it is detracting from your credibility. And I know what you put into earning that and maintaining it. And that never really pisses me off. Likewise, I mean, you're equally as careful with your integrity.
And so you also have the same amount, if not far, you probably have 10x the amount of audio on the internet of your voice than I have in my voice because you do a daily podcast. So that's interesting. It's just an interesting thing to think about. And I thought like the, I'm going to do the robot voice here again.
I thought it was going to sound like this. We need to do the robot's going to sound like this. Nope, it sounds like this. And this is exactly how hard it is to build a difference.
It's just an interesting thing. However, all that being said, this is in process. I'm working on it. I'll tell you how it goes later.
I would like to know the positive uses of AI that you've found since we did the last recording because I have done a lot. I'm just interested to see what you're doing. Can I put a pin in that? Because I'd like to talk about, I don't know, okay, this is a half answer to your question.
You're bouncing it to me. Thank you for that. I don't know if I put this as positive or not, but I also had a bizarre AI incident this week. Okay, I'm working on this through the Bible video series.
I wanted to do it. It's been a massive project for the last year. I only have two videos out on it. There's a thing that's kind of happening culturally right now.
I think people are pretty disaffected with all the previous sources of authority. I think people are skeptical of governments and media and advertisers and just kind of everything. And it's cycled around to the point where people are like, okay, what's the very first authority that we rejected? Let's go back and check that one.
God, this church, there's a ton of interest. Right now is you and I are recording this and just God faith stuff that nobody knows anything about anymore. Very few people know stuff about it. It's been out of fashion for a while, right?
And so I'm sensing there's all this interest in it. I'm like, hey, you know what I'd love to do? Make really easy to search, really easy to find normal human language videos about each book of the Bible. So if somebody wants to figure out what it means, you can go watch this and in like 15 minutes or so you'll have a really good sense of what happened in this very big, very ancient book of the Bible.
So a lot of work, I never went into that, right? Well, it's not going well. I just can't crack the algorithmic code. If I publish anything other than these videos on my YouTube channel, they do fine, but these videos, they don't work.
I don't get it. It's baffling, right? So I got to change some things up and I was working on that this last week and working on the book of Leviticus. What do you know about Leviticus?
Oh, you're not supposed to eat shrimp. Don't touch goats. That's in there too, I think. Right.
I think so, yeah. Isn't that right? Yeah, this is the one that has the stuff about like don't wear mixed fabrics and it's all in the yeah, it's complex. That's an oversimplification and that's that's how people like to dunk on faith is they're like, Oh, when Leviticus says don't eat shrimp or catfish and you still do that more on people have a hard time dividing the ceremonial law from all the different stuff.
They have a hard time parsing that and it's just a really fun place to go and dunk on Christians. Yeah, admittedly. I mean, come on, Bronze Age laws read in the 21st century are going to sound weird. I think the more remarkable thing about Leviticus is how many of the laws don't sound weird and are in one way or another built into all modern law around the entire world.
That's not the point. The point is I got the challenge in front of me is I have to try to salvage this project that I put so much energy into on one of the most difficult books in the whole Bible that isn't very popular on the internet. It's just the next book up. So it's very challenging.
So I went I wrote a whole thing. I took a couple of days, did really careful writing on it. And then I went to record it and I was like, I just wait too long. I don't know what chat GPT would do with this.
I'm stumped. Where would you cut here? So I didn't ask it to rewrite it or anything, but I pulled up chat GPT. I was like, Hey, I'm stuck.
I wrote this script. Here's my script. Help me out. Where's the bloat?
Where could I trim and still retain the meaning of this? I mean, that's a pretty interesting application for chat GPT. Yeah. I think.
Yeah. So you're not having it right. You're not having it generate original content. You're having it edit something you've already written.
Not even edit. That's all I was trying to do. It's a brand new account. It doesn't have my name on it.
I didn't use my actual name. There's nothing that would connect this with me. I mean, I'm not an idiot. I'm VPNed up and all the layers of everything you would expect.
So I load up this completely fresh chat GPT account. I'm like, I'm going to pay for the good language model. I'm going to try it out and see how it works because I don't know. I haven't played with this in a while.
So I played the script, which also has no hint of who I am. I didn't include any of that. All very intentionally. And I asked it to point narrow out what I could trim down.
It comes back to me and says, you know, this is a fun, playful script about the Bible written in the style of Matt Whitman from the 10 Minute Bible Hour. Really? That was its analysis. What?
And I thought, how do you know that? It's a gank your crap. That's how it knows that. It gank to my crap.
It's true. Like there's fart jokes and stuff in there. I mean, it's just, I don't know. It just sounds like me, I guess.
And so I was like, all right, well, what sources did you use? And I said, yeah, we drew from these commentaries and these study Bible introductions. And we drew mostly from the 10 Minute Bible Hours content on the Old Testament in the book of Leviticus to help offer you feedback on the script. It still doesn't know.
It's me. It's not this closing that it knows it's me. Bizarre. So I'm just asking like, can you just write in the style of Matt Whitman in the 10 Minute Bible Hour?
And it's like, sure. Okay. Will you be him and write a 10 Minute Bible Hour script that is a summary of the book of Ruth and it freaking nailed it? Yeah.
There are like three lines in there where I didn't like the joke and I wouldn't have said it that way. I think that could be funnier or better or more clever. But in terms of just a good enough script, if I was just trying to bang out a video and get it done, yeah, it was good enough. Got the deep dive themes, got the historical setting stuff right, touched on modern archaeology and what we know and what we can't know.
Like all the questions I would ask about a book of the Bible, it did and then it made it sound like me. It was bizarre. Wow. So clearly it used to use the source material.
How do you feel about that? Hmm. I don't know yet, but I'm glad we're on Mike's because I have to think about it now. I feel like my brain has been cloned.
It's an inferior clone, but it'll only take a few more iterations of that cloning process to be as good as me or maybe better. Oh, nah. I don't think so. In terms of writing, maybe writing about a specific subject that it knows about.
Now that might be an interesting experiment. Ask had JPD to write something in the style of Matt Whitman about aerodynamics or something like that. That might be interesting. Cross genres or something like that.
Oh, very interesting. I think the New York Times actually reached out and I'll reach out. They sued or at least threatened to sue OpenAI for stealing all their articles as source content. And so I think that's the really difficult thing is that these models are built off of other things.
There's actually a researcher. There was a freakinomics episode called How to Poison the AI Machine and they had this researcher from the University of Chicago. His name is Ben Zao Z H A W and he was talking about ways to poison the library that you're using to teach the models, which I found to be interesting. You've heard of Garbage in Garbage out?
Sure. He was mostly using it as an adversarial. I think what they call a general adversarial network. I don't know the words correctly here.
But what I do know is he was poisoning the well and the way he was doing it is the way he described it on the podcast as he said, hey, if it's looking for a picture of a cow, we would give it a picture of a cow as source material, but we would do it in such a way that it would actually be interpreted by the AI bot. It would be interpreted as a truck. And so yes, the image to a human looks like a cow, but there's certain features in the image that the AI would interpret as a truck. And so then when people use mid-journey to try to pull up an image of a truck, excuse me of a cow, it would give them truck like qualities.
And so it was fascinating. And so his idea was to go on the offense to try to obfuscate reality. His whole reason for doing this was to protect artists because you could say, draw me a painting of XYZ in the style of Monet and it'll do it. I thought that was pretty interesting.
So his idea is, well, it's just boys in the well. Well, what are your thoughts on that? I have a follow up to that, but thoughts? Yeah.
Well, here's what's troubling for me about this. Let's say, okay, I'm not going to use this. Let's just make that clear. I'm going to write in my own voice.
I like writing. I like fingers to keys. I like the process. I'm not looking for a shortcut.
I don't want one. Never taken one. I've never had AI write anything for me that I've published other than a video I did one time where I had it write a couple of little mini sermons. And then I had the audience guess which one was me and which ones were the robot.
So there's no dark temptation here to tap into this power secretly. I'm not interested in that. But I wonder, like what if somebody did give into that and they had a very pronounced style? They'd been doing stuff like you and I've been doing it for a long time.
It would not take long for it to train on itself and effectively then you would have a genetic fallacy. The gene pool would shrink for what that style or that external brain would be. So let's say I did that. Chat GPT write an article on this in the style of me, an article on this in the style of me.
And then I start publishing all of those. Well now it's learning on those as well. And it's like intellectual incest. All the offspring of these ideas are increasingly coming from this small AI generated gene pool.
And eventually it's going to bastardize the style into something gross and unrecognizable. So I guess my take on that poisoning the well kind of thing is if you do it long enough, it should poison the well on its own and become inferior on its own if it has nothing new to train on in a given style. Yeah, it's a good point. So you're talking about intellectual inbreeding and in breeding of data.
There's this thing that some people are happy to call it the wall. They call it the AI wall and it's built off of the data. There's this development rate of artificial intelligence. And it appears to be asymptotic in nature in that the more you develop it, exactly what you're talking about happens.
You need more data to train on, but it's not there because it already trained on the internet, right? With the exception of things that are happening in current events. But it already trained on all that stuff. So in order to keep it developing and moving forward, we're hearing Google and other companies saying, oh, well, there appears to be an issue where we can't develop our models fast enough.
Now the CEO of OpenAI, Sam Oppmann, he famously tweeted, there is no wall. And what he's saying is this wall where we're going to run out of data to train these models, it doesn't exist. And so he's kind of an AI absolutist where he thinks AI is here to stay. It's going to keep moving forward.
Also, it's not going to run out of things to train on. I'm in your camp. I believe it's going to be, I don't know, I can't say the term that I think it's going to be, but it's going to intellectually increase on itself. It's going to be pretty defensive.
Yeah, it is. Yeah. So I think there is a wall and I think it's interesting. And that's why I'm getting more and more interested in books and for reasons that we discussed earlier.
Okay, whether there's a wall or not, though, one of the questions I think we have to ask ourselves is do we want to be a part of what it is trained on? Or do we just want to try to protest the whole thing? I'm sure you've received an email or two, as I have, from people who see that I have a lot of content out there. And I've been offered money to let their proprietary AI train on apologetically on all of my material.
It isn't an insane amount of money. And I said, no, but there was a part of me that was like, well, I'd kind of like to have my work be a part of the record. If AI does end up being where all knowledge goes, I'd like to contribute to that rather than sit it out because I think my work's important and I've done some good stuff. But then I thought to myself, crap, I'm talking about God.
I'm talking about a being, I believe to be the definition of essence, the source of all existence and things. And it's really high stakes stuff. And if I get it wrong in some dumb video on the internet that, you know, however many people watch, it'll fade into memory, whatever, it doesn't matter. But if that goes into some sort of AI super intelligence and one of my well-intentioned mistakes but still mistakes makes it into that, what if that's my legacy and it gets duplicated forever, uncredited, but instead of any good thing that I contributed some mistake that I made with the text or whatever is what's just out there forever.
So I don't know if I wanted to train on me because I think I'm more afraid of my mistakes being exponentially amplified than I am excited about having my good stuff being exponentially amplified. Do you have thoughts on that? Yeah, it's interesting because you're talking about, you're talking about doctrine. And so it's important that the right stuff gets out there.
There's been several people come along in the past and say, oh, look at this new revelation I have of God and it's true and it's totally right and you should totally believe it. I don't believe it. But what I find interesting is there's a little bit of retcon stuff going on where you try to make retroactive continuity and you try to like go back and change the record so that things make, it looks like they make sense, but it doesn't. I really don't like that.
And so what you're trying to do is you're trying to get back to the original source material and you're trying to inject that into what may be the foundation for whatever the future of the search engine is. I find that to be quite interesting. The things that I make videos about aren't near as controversial or I guess important for kind of controversy. It's not as important as what you do.
It's quite interesting. Do you assume that somebody's AI somewhere is trading on your phone calls or your conversations that are happening within your shot of a smartphone somewhere? Crap dude, thanks for that. I've never thought about that.
I mean, you know what is interesting? It most certainly is going to be training on this podcast episode. That's fascinating because we're going to publish it. Which is why we should say, I can't, I'm not going to do this.
Just some wrong stuff? Like you want to just say about what happens? You can see what happens. No, I think it's strange.
We're in unprecedented times. One thing I do want to talk about and to relate all this stuff together is how AI is changing the workplace, specifically for software engineers. I know many software engineers that say, holy cow, this is changing so rapidly. I've got a friend.
His name's Carson. He's really, really smart. He's a very good software engineer. I have a chance to work with him on a project way back in the day.
Would you be down to have him on and we'll just ask him some questions? Yes, of course. Because I feel like I'm shooting in the dark. And just on the receiving end of this, I don't understand what's going on behind it.
I would love to pick his brain. Let me reach out and see if we can get Carson on. Alright Matt, this is Carson. How's it going?
Nice to meet you, man. Yeah, you too. Why did I say Carson? I don't know either.
First means they're always awkward. It didn't have to be. Yeah, but I made it. I'm sorry.
So Matt Carson is one of the smartest computer scientists I've ever met just in practice. Like I've asked him a ton of questions and he always has a really good answer. He's got his finger on the pulse of new technology. He's really good.
We had a conversation the other night, Carson. And we were talking about how AI has changed the computer science field specifically when it comes to coding, right? Yes. Can you just kind of give Matt a primer just so Matt has no knowledge of this at all?
I'm just kind of used to when I went to school and learned Fortran. I know we had to learn syntax and we'll tell you. Don't worry about that, man. So we had to type in the code and we had to run our code.
We had to compile it, all that kind of stuff. And you've done that, right? That's what you do. Yeah.
So, let's take a step back. So with your software job, you're typically working in a team of people, right? So you're needing to start with, you know, even five years ago, you'd start with your requirements, what you're going to need to build, what are the inputs and outputs, what's the architecture going to use. And then you hand that off to a coder who actually codes it up.
And you needed multiple people to do this in the past. You needed someone who could write documentation or at least someone who had the skills to write documentation and a code. You needed someone who could think about the architectures. And about two years ago, when chat GPT came out, it kind of blew everyone's mind because you could tell chat GPT, hey, I want a little function that can, you know, print the word dog and it would write it for you like the computer would just write you code.
And that blew everybody's minds. I forget it was three or three years ago. And it's only gotten better from there. So today we've gone from a place where you used to have a single process where you send in your text, it gives you back a little bit of text and only has a small context window thing under, oh, you know, say 4,000 words or so, that would have been three years ago.
So now you can go back and forth and it can at one moment understand 200,000 words. And it has the power to write code for you, not just like talk to you, but write code on your computer. It can synthesize your requirements. So I guess to give an example of my process now, I can start with an idea.
You know, last night I had this idea that I wanted to make some audio books. I collect these old books and they don't have audio books. They're too old, right? I had these books from like the 1750s.
And I mean, fat chance you're going to find some dude reading an audio book from the 1750s. But you know, there's these text to speech things, but they're not just like easy to use for a whole book. You know, they only work with like 30 second increments. So if you wanted to find a text to speech, start feeding a whole book and end up with an audio book at the end.
I mean, you have to be able to code to do that. And since I use all these elements as a part of my pipeline, I used elements to help me write this app that would give me a book read to me by text to speech. I think I had the idea that I wanted to do this at 1030 PM last night. And I started working and I basically outlined, okay, here are my requirements.
And I fed that to an LM. And I told it I want you to write me an official requirements documents. So I word vomited. I didn't do anything fancy.
I gave like a really, really honestly a pretty crappy paragraph. And it turned it into a professional well written document saying here are the exact requirements that you need to make this piece of software. And I was like, okay, now you've got your requirements. Use those requirements to tell me what the architecture is going to be.
So it took them and it was like, hey, you need these libraries. You know, you're going to want to use these plugins. You're going to want to think about organizing your code in this way. And I was like, oh, thanks, that's great.
And I didn't even think about that. And I was like, hey, can you write the code for me? Because you know, it used to be you had to talk online with the code, right? You had to go to www.chatcpt.com and you had to talk to it and you had to copy paste your code and you had to copy paste it back into your browser.
And then you change your code, you had to put it back in the browser. No, that's not how it works anymore. There are code editors that have the AI built into them. We can ask to talk about these terms later, but they have what are called agent workflows and train of thought.
And they will type for you. So I was like, hey, that architecture looks good. Will you make it? And it went into my terminal.
It made the folders. It made the files. At this point, I haven't coded anything, man. I had literally a paragraph of Word vomit and an idea and it has coded this for me.
And I'm looking over it and I'm thinking, okay, maybe I want to make this decision. I go back and forth for a little bit with it. I make some changes. But yet wrote the whole thing and by 1130, I had a book and a pipeline that was done.
And now if I want to, I can go listen to my copy of Buffon's 1753, a natural history volume for it. And I can put it on my phone and listen to it. Wow. Okay.
I tried that this week too, except the difference between you and me is I don't know how to code at all. I don't know anything about any of this, right? So I was like, well, okay, here people talk about Python. Let's go with that.
So I was playing with chat, gpt, and I was made up an app. I was like, I make me an app that the tracks bowel movements regularly, and I was like, Oh, nice. I'm making a poop cracker app. I'm like, I'm not speaking to mine, right?
And I just want to see what the interface was like. That's all I want to know. Like, can you just code me a sample app for this? And I was amazed at how simple that was.
And I put it in front of a buddy who does this kind of thing for a living. And he was like, well, I mean, I do this and I do this, but yeah, we're off and running here. I was like, okay, well, obviously I need you want me to know, no, no, no, I'm just, I'm just learning the interface. I'm just playing with this.
But that was mind boggling to me that someone who knows absolutely nothing. And of course I just followed up with a, okay, well, how do I make this actually something I can interface with? And it was happy to give me three, four steps. I'm like, yeah, I can do all that if I really wanted to.
It'll be ugly, but I can do it. What an unbelievable opening of this secret knowledge. I mean, you all have had access to this oracle, this place where dreams can become some sort of interfaceable reality. And now an idiot can do it with, again, one paragraph of word vomit.
That's done. I mean, the whole Python, I don't know if this is why they named it that, but I assume it's named for the guardian to the oracle at Delphi. That's super knowledge as hell. There is this dragon serpent that lived there and you had to slay or get past the Python to access this level of knowledge.
Now an idiot can just easily walk right past with a sleeping Python, I guess. I'll have got this. LLM's get a lot of hype and they get a lot of hype about whose job they're going to replace and what they're capable of. And a lot of the resolves around them doing an average general job.
What one must understand about an average general job and about the human language in general is the possibilities are vast and endless. The ways you can put together a sentence, the ways you can put together stuff to do their work. Coding is not vast and endless. Coding is a limited space where there is a very constrained set of commands that exist in coding and they are very well documented.
This set of commands does this exact thing. Coding in general is uniquely positioned to be taken over by LLM's because it can be so powerful in that domain. Does that make sense? Yeah.
Yeah, there's rules and you have to play by the rules. It's unlike, you know, if you wanted to write a book for you, that's a harder question. That they might do better or worse at. But coding, there's a right answer to coding.
It's like, you know, it's like a very complicated math problem. So what are people saying in your community? I know that you adapt. So are there luddites that are like, no, I'm not going to use AI in your field?
Are there people like, no, no, that's just, are they just angry? Are they adapting quickly or getting cut out? What's happening? I work in an area that's very excited about it's machine learning in general.
You know, I work with teams that are very excited about language models and about ML and we work on the applications of large language models and on ML in general to science. So I'm a little biased in my worldview in that regard. But it is simultaneously very exciting and very scary because on the one hand, it seems like every week a new amazing tool comes out, you know, it was just yesterday if you asked you to do something, it would give you some horrible answer. And now I feel comfortable letting it run terminal commands on my computer.
Let me run back to your question. So you want to know, you want to know how I want to hear the scary part before we go back to the question you said. Yeah. Why is it scary?
It's scared us to juice. I want to know why you're troubling. Yeah. So the fear is who is the fear is that you can code, man.
That's the fear, right? That's because I have a, you know, I have a whole life of coding and thinking about coding and architectures. And when I go like, let's say I wanted a new job, I don't, let's say I did. You know, if I was going to go and get hired, I'm bringing this, you know, 10 years of experience or whatever, do I need that three years from now?
Or does chat GPT have all that experience and the project manager can write the code from themselves? Because I walked you through my pipeline, right? What I started with was the requirements and at the end I got coding product. Where is the coder in that workflow?
Right now, I think you to really make good stuff, you still need a coder in the workflow. You need a coder for two reasons. One is for a genuine understanding of modern architectures. And one is just to, you know, troubleshoot things when they really are hard and when they go wrong.
But we've seen a lot of growth in the past two years. And there's a lot of people just like you who are realizing that they can code things for the first time. And I don't think we've reached a plateau of advancement. We don't seem to have plateau.
We seem to still be increasing. And you know, today you might still need to need a coder, but tomorrow you might just need your project manager. So the scary part is what it does to the job market. Yeah.
The first most observable way that it could threaten to upset the system right now. Yeah, I need to start learning how to plumb and machine and, you know, do some stuff that AI can't take over yet. Are you being serious? This year?
No. Five years? Oh, yeah. Tell me more about that.
Yeah. So right now there's various factors, right? If you had looked back at the 1970s and you'd ask yourself what would happen if computers got faster and more people learned how to code, there weren't a lot of coding jobs. And so you have this idea of what the coding job market looks like in the 1970s.
So obviously today there are hundreds of millions, potentially more coding jobs than existed then. So on the one hand, it's possible that demand for coders increases to an extent to where it's not, you know, the demand for new software applications increases to the point where despite the productivity increases, we can continue to supply them. But there's always that point where what happens if the demand for new software does not keep pace with the average productivity of a single coder. So it depends on how those lines converge, right?
What about for all of humanity? I mean, the immediate job market concerns are pretty evident. Just my experience this week, bumbling through what I bumbled through. I immediately went to folks who are in your line of work across my mind.
But what do you see and what do people you talk to in your line of work see in terms of the implications for human knowledge and progress and all of those bigger existential questions? Yeah. Well, let's talk about that in two ways. First, staying on specifically the job market.
So okay, do we all know what an agent to work club is? I do not cover that. No. Yeah.
Okay. So there's this idea. Okay. So you have this one paradigm where you have an L limb.
It's basically a statistical next word predictor. You tell it a chunk of words. It puts them into its little machine learning model and it predicts the next word and that prediction is the text it puts on the screen. So that is your standard non-agentic L limb.
Well, what an agentic L limb does is it basically says, okay, let us give this model access to an external tool. So for example, maybe you give it a calculator. So maybe the AI itself is bad at math, but it is capable of like sending inputs to a calculator and the calculator is great at math. So what if you're doing something more complicated than a calculator?
If you're giving access to Google and what if it gets good at Google or what if you're giving access to a calendar and you start broadening what the access it has to in terms of agentic tools. So agentic workflows are potentially the thing that is disruptive to the general job market. I wouldn't say they're in their infancy. They're not in their infancy.
They've been used pretty successfully for, I mean, at least a year, like Chachi BT has had Google searches for, I think, a year now and it runs Python in the back end, right? So like if you ask a complicated math program, it will secretly open a Python window. It will write Python code. It will calculate the answer to the, to the, your question and then it will display that answer as if it was part of the conversation and it will tab where you can click it and you can open it and see the code if you want, but that is hidden from the user.
I mean, not, not in a variously hidden, but you know, that's just, it's agentic on the back end, right? What? Yeah. And so what I was talking about earlier, so I use this, um, I use a lot of different AI tools.
I try to just stay up to date, but one of them is called Windsurf and it's the AI coding editor. And it literally updates your own files. So that's agentic. So imagine you're someone like a, uh, imagine you're someone like a project coordinator who is responsible for making meetings and you take meeting notes and you do, you know, emails and you get requirements from people and that's your job and you're really good at it.
I don't think I is here yet. But you can imagine all of these individual components are kind of figured out, like AI summarizations are pretty dang great. I don't know if you've ever just like word vomited onto a page and said, Hey, please organize this carefully for me. They do a pretty acceptable job and like making a counter meeting, not that hard, right?
I don't know if you've looked at like the perplexity integrations on Android for the replacing the Gemini, but they're pretty good. They've been recently, they're pretty good. So once you start combining these technologies, that's where you start hitting, is it going to replace a human worker or is it going to make one worker be able to do the work of three, which is kind of like replacing a human worker. Yeah.
That's an important point. People say like, Hey, it will never totally replace a human. Maybe not. But if it makes a human three times more productive, that's the same thing.
Making one human be able to do three times as much work does replace two humans. Have you heard of this thing called Jevons paradox? Have you heard of this? No, no, I have to ask you things like this because you're so smart.
You probably, I just recently read about it, but basically back in the 1860s, there was this dude, he was a British guy and he was talking about the big conversation was coal mining back in the day because steam engines were becoming more and more efficient and they required less coal. And so there was this big fear that you wouldn't need coal miners anymore because everything's going to be so efficient that you don't you don't need coal. However, what happened is because the steam engines became more efficient and you could get more productivity out of them, everyone started using them. Therefore, it actually caused an increase in coal.
Yeah. And so it was a paradox. It's called Jevons paradox. William Stanja Jevon is the guy I'm reading here now.
But anyway, there is a group of people that I don't know, they theorize that there will be paradoxically more demand for, I guess, people managing these agentic. What did you call it? Agentic workflows. Yeah.
Okay. What's your response to that concept? That is the question. So when I mentioned earlier that there were not very many coding jobs in the 70s and now there's hundreds and millions of coding jobs because there's more demand for it, there's that question of does the growth and demand match the amount of the supply of work that you have?