Intersection of AI and Music Technology with Trevor Hinesley, founder and CTO at Soundstripe

(Intro Narrator at 00:00:00) Today, Trevor from Soundstripe joins us in the studio to talk about generative AI and the music industry. You're listening to Joel Beasley, Modern CTO.

(Joel Beasley at 00:00:14) So you can't tell me all your intellectual property and private stuff.

(Trevor at 00:00:18) Well, I could tell, yeah, right, exactly. I can tell you that we have about 10,000 songs now that we own. So that's, like, that we spent years building.

(Trevor at 00:00:27) I think probably the last time we talked was that, like, two years ago?

(Joel Beasley at 00:00:31) Yeah, I think it was in 2021, right?

(Trevor at 00:00:33) Okay, then yeah, we were probably at, like, half that. Yeah, so done a lot of investment in music catalogs, doing some really cool stuff on, really what our bread and butter is, is we've created this, like, on the business side. I hate to use, like, the most overused flywheel, but really that's exactly what it is. Like, we have this whole thing that when each piece spins, the whole thing moves faster, and I like geeking out on this stuff too, because the business side itself, like, we really focused on supply chain.

(Trevor at 00:01:03) That was, like, a big part of our, and it's enabled so many other things now. So, yeah.

(Joel Beasley at 00:01:06) And what's the elevator pitch for what you guys do?

(Trevor at 00:01:09) Yeah, we basically, Soundstripe, you know, we sell unlimited music to creators. So everybody from fledgling YouTube creators all the way to Fortune 100s. We have, you know, persona-based pricing and plans for pretty much any need. So if you're trying to put music in content, whether it's a podcast or a video or some other medium we haven't even thought of yet, we license that out for use.

(Joel Beasley at 00:01:32) Nice. Can you buy unique rights to a specific thing?

(Trevor at 00:01:35) We don't do, like, basically sell the rights outright in most cases. We have done, actually, for some of our larger clients, we have done that a few times. So we do offer, like, a bespoke thing if you're a larger enterprise and you're looking for a really efficient way to get music that is branded for you and that kind of stuff. We've, like, in some instances, we've actually done, like, 24-hour turnarounds on fully produced music.

(Joel Beasley at 00:02:07) That's awesome. So shout out to Josh. He did a 24-hour turnaround on the intro music for the show.

(Trevor at 00:02:11) Yeah, that's awesome.

(Joel Beasley at 00:02:13) I called him one day. I was like, hey, I was listening to the show. You need to do something, please. And he's like, okay, I'll put it on the calendar for next week. I was like, I need you to do it

(Trevor at 00:02:22) like now.

(Joel Beasley at 00:02:22) Now. Next episode. He's like, okay. So he goes away and comes back, like, that night or the next morning, and he's just like, here's the first version of it. And I heard it, and I was like, that's it.

(Joel Beasley at 00:02:33) Done.

(Trevor at 00:02:33) That is awesome. Phenomenal. That is awesome. Yeah.

(Joel Beasley at 00:02:36) He's great with music.

(Trevor at 00:02:36) I love that. That's great.

(Joel Beasley at 00:02:38) Most of the content you guys have is created by humans.

(Trevor at 00:02:41) Yes, it is. Yeah, I mean, the only thing right now that is technically machine is stuff like, you know, using samples and things like that. But, yeah, we work with human artists all over the globe. So we have, you know, we have a few different models that we work with artists, but we make it really easy. Like, basically, we have a set of turn-in standards when we get the music from them, but they don't have to mix or master it or whatever. We have post-production we have on staff, and it goes through this whole, we do all the admin for it. It's basically, like, you just create and turn it over, and we handle everything else.

(Joel Beasley at 00:03:17) Well, that's pretty cool. Yeah, we have a lot of experience with that. So, you know, we do our show, but, you know, we do about 20 other shows. Okay.

(Trevor at 00:03:23) I did not know. Is this, that's like a, is it a newer thing for you guys?

(Joel Beasley at 00:03:26) Like, about 16 months ago.

(Trevor at 00:03:28) So, I

(Joel Beasley at 00:03:28) don't think we were doing it when we talked.

(Trevor at 00:03:30) I don't think you were either. That's awesome.

(Joel Beasley at 00:03:31) Some of our customers asked for it, and so we started doing it.

(Trevor at 00:03:34) I love that.

(Joel Beasley at 00:03:35) And we've gotten, I think if we actually looked at both of our business processes operationally, there's a lot of similarities, probably some overlap, because we were designed to take in lots of content and then, sure, polish it up and then push it back out to the world.

(Trevor at 00:03:51) How much of that have you automated?

(Joel Beasley at 00:03:53) Well, we've recently, the producers right now were battling this concept of it's a craft

(Trevor at 00:04:02) sure,

(Joel Beasley at 00:04:02) to do the editing

(Trevor at 00:04:03) yeah,

(Joel Beasley at 00:04:03) to where an AI is now able to do it better. And there, but there's still some things because it's not perfect.

(Trevor at 00:04:10) Yeah, for sure.

(Joel Beasley at 00:04:10) So you can't just completely give it 100% over, but you can give 98% of it over.

(Trevor at 00:04:16) Sure, which is like most things at this point. It's crazy. Yeah, I mean, we use, yeah, ChatGPT, clearly, like, there's a lot of use for that. But, you know, we have team members using GitHub Copilot and, like, all the, there's so many, like, a friend of mine kind of, it's kind of like a generative adversarial network, which is like an older, you know, ML thing. I think I don't remember when the first of those came out, but it was one of the, you know, bigger AI advancements over the last few decades. And when the whole concept being you've got essentially two models playing against each other. And a friend of mine sort of, like, he's not a developer, like, in that space or whatever. He's a marketer. But he kind of created his own means of doing that where he basically has, like, ChatGPT spit out, you know, article headlines or copy or, like, ad titles or whatever it may be.

(Trevor at 00:05:07) He'll have it spit out a few variations of it and then pump all that into, like, you know, Facebook ad AI, right? Determine whichever one won based on engagement and then put that right back into his next prompt in ChatGPT and say, from the last set, this. So he has this, like, long-running chat, yeah, with ChatGPT where it's just, so he's basically using each model to play off each other and it's like his ad performance is crushing.

(Joel Beasley at 00:05:34) Oh, I know. There's a couple out there. Usually they're fairly expensive, yeah, and they're bespoke or whatnot. Yep. But I had a guy on probably about five years ago on the show, and that was his startup. What they do is, because these models existed for a while now, sure, right? They've just gotten significantly better and hit public consciousness.

(Trevor at 00:05:52) Exactly, yeah.

(Joel Beasley at 00:05:53) So he would just, it would create, like, 100 ads, push them through the API, take the best performing ones, shut them down, and it would just cyclically do that. Yeah. And then it would just come back with, okay, this is what works best. And you could monitor what it was generating and, sure, and contour it along the way.

(Joel Beasley at 00:06:11) Yeah, I want more of this, less of that. But there's so many dollars in marketing, and once you have the API to an ad platform, why would you not?

(Trevor at 00:06:21) Yes, exactly.

(Joel Beasley at 00:06:22) If you have a machine that can just create thousands of variations and tell you the most effective one,

(Trevor at 00:06:27) just hit the button. Exactly. I mean, that's not a, at least for the, you know, as soon as the ad market became kind of, like, the online auction system that it is, I mean, there's really no reason to try to get, quote unquote, creative there because the whole point is just to reach people with your message, right? So as long as it's on brand and within your, your voice, ethically, like, let a machine automate that, you know?

(Joel Beasley at 00:06:54) But that's the rub, the artist, totally, having them back off for the efficiencies. And previously I'd had built some software that would make an accounting team much smaller. And so we would, this is 15 years ago, sure, we'd go in, 20 people in the accounting team, we'd put this in and come down to two, right, right? And the people who were curious about what we were doing and wanted to learn were the ones that had jobs after.

(Joel Beasley at 00:07:19) Right. And the people who were dismissive of it were the ones that were there, because they didn't know, not because we, no one liked them, because, sure, they just didn't know how to use the tools that now run that department, right. And so when I see this comment fiasco on LinkedIn and stuff of people arguing over the morality of this generated AI, right, while they're arguing and philosophizing, if that's a word, right, right, right.

(Joel Beasley at 00:07:43) I don't know, we'll make

(Trevor at 00:07:44) it up. We'll go with it. It's working out.

(Joel Beasley at 00:07:46) He said it's cool. Yeah, exactly. Everyone else is generating and pushing the tools forward.

(Trevor at 00:07:50) Yeah, I mean, it's still moving forward whether you kind of like it or not,

(Joel Beasley at 00:07:53) yeah.

(Trevor at 00:07:53) Where it's at, yeah, for sure. And, I mean, there are legitimate conversations happening too, around especially in music. I mean, that's like, yeah, I had a good conversation the other day with kind of a new connection of mine that owns a company that is in the generative space. Essentially what we were talking about is, like, you know, there are some, there are definitely some interesting tools out there where, like, Boomy. If you've heard of Boomy? Okay. Or Bandlab.com.

(Joel Beasley at 00:08:21) There's a

(Trevor at 00:08:21) couple others, but they're basically tools where Boomy, in particular, you can, like, you just specify some parameters and click a button,

(Joel Beasley at 00:08:29) oh, I've seen those,

(Trevor at 00:08:29) and it pumps a song

(Joel Beasley at 00:08:30) out, right?

(Trevor at 00:08:31) But Boomy's whole thing was, when you create music, they'll just distribute it to Spotify for you too. So it's, like, fully automated. But then a couple weeks ago, Spotify basically determined that, like, 13% of the world's catalog was Boomy as of two weeks ago.

(Joel Beasley at 00:08:49) Oh, wow.

(Trevor at 00:08:49) So they've put out, like, millions of tracks on Spotify, and Spotify shut it all down.

(Joel Beasley at 00:08:53) I'd imagine.

(Trevor at 00:08:54) So, yeah, because it was, you know, just basically flooding the system. And when I was talking to my new friend about this, we were, based where we landed is that, like, you know, there is value, part of the value you get from things like, I love Midjourney. Yeah. I use Midjourney for all sorts of stuff.

(Trevor at 00:09:14) We've even had some very particular situations. Like, one was, you know, my girlfriend had mentioned, she, we have, like, a gallery wall in our dining room, and she was like, she had this idea for this crazy kind of, it was just like a wacky UFO meets Santa theme. That was what she wanted to do for last year's, like, gallery wall.

(Joel Beasley at 00:09:32) And I

(Trevor at 00:09:32) was like, okay, cool. It was really hard to find cool artwork for that. So I just put it in Midjourney, and then we, like, all we have to do is print it out and put it on canvas or whatever, right? So, but the cool thing was, is like, depending on my prompt and all these other things, like, I felt like I had agency, mhmm, over that, right? I was able to dictate what was happening. The difference between something like Boomy, you know, and it's fascinating if nothing else, the fact that that tech has come this far.

(Trevor at 00:10:00) But, like, just clicking a button, like, I don't have an emotional attachment to anything that's making, right? And so what, where the interesting rub I've seen that I think is legitimate is when it gets to the point of, like, like, right now, even with systems like ChatGPT or Midjourney or whatever, there is still that kind of guidance process from a human, right? But, like, what happens, you know, and this is more of the speculative realm, but what happens when we go further down that train to where, like, it has all the context of everything I do online and everything I, all my emails and XYZ, and it knows my interests and my own voice and everything else.

(Trevor at 00:10:41) And, like, I don't really have to prompt or dictate anything. It can just kind of figure out, you know, maybe I, like, for instance, like, if I'm saying, like, I need, um, trying to think of something that's more, like, general. But in that situation where I created that, like, UFO Santa art or whatever, I had really specific parameters I wanted on that, right? But, again, this is, you know, many years down the road, maybe.

(Trevor at 00:11:09) But if we get to a place where, like, all I'm saying is we have a gallery wall help, and it spits out a UFO Santa thing because it uses all this probabilistic data to determine, exactly, like, what happens? Like, does the world become super boring or homogenous? I don't know.

(Joel Beasley at 00:11:28) I think it stays about the same.

(Trevor at 00:11:30) It probably does.

(Joel Beasley at 00:11:32) Because here's the thing. Right now, there's still the people that won't put the effort into going and doing anything, and they're just, they end up with the same Target wall picture that everyone else has.

(Trevor at 00:11:43) This is okay.

(Joel Beasley at 00:11:44) So there's, what I have been excited about is the easier this becomes, the more you'll see a separation of people that have discipline and don't. Love that, man. Because the discipline is what's going to shine. It's going to be the differentiation between most people and the exceptional people. Because I can, I've always argued, if I gave you a button that was a money button, you could press it and money would go into your account.

(Joel Beasley at 00:12:11) But only, like, a thousand bucks, right? You press it, every time there's a thousand bucks. Eventually you'd, like, stop pressing it.

(Trevor at 00:12:17) Right.

(Joel Beasley at 00:12:18) But it's kind of hard to imagine that you would stop pressing it.

(Trevor at 00:12:21) Sure, but you

(Joel Beasley at 00:12:21) just be like, how much time do I, I still, oh, I spent, like, 10 minutes today doing it.

(Trevor at 00:12:25) Yeah, right, right.

(Joel Beasley at 00:12:26) Like, you could have, like, infinite money, but you don't, right, right? And so it's a discipline thing. And I think so I think discipline is going to shine, and I think we're going to see a huge growth in in-person events and activity.

(Trevor at 00:12:40) Oh, yeah.

(Joel Beasley at 00:12:41) Because I want to do business with you, right. It's great our robots are running around negotiating. You know, maybe your lawyer robot's suing another person. Your money robot's negotiating with your finance guy, right? Like, you got all your things happening. Yeah. But at the end of the day, you still have to wake up and do something, sure.

(Joel Beasley at 00:12:57) Well, some people are okay with waking up and, like, just Judge Judy-ing it the whole day with bonbons. I'm not that person.

(Trevor at 00:13:05) Yeah, I know, me either.

(Joel Beasley at 00:13:06) I would, I have to do something.

(Trevor at 00:13:08) Yes, for sure.

(Joel Beasley at 00:13:08) And so what is that something? And it'll likely be doing projects with people you enjoy spending time with in order to take the world to that next level, sure, yep. And that's what it's going to be.

(Joel Beasley at 00:13:20) And so relationships and discipline and in-person activities, I think those three things are going to be hugely important over the next decade.

(Trevor at 00:13:28) Yeah, no, it's a great point too, because I think that the, as far as, like, the creative sphere is concerned, there is a ton of, it's a good example. Take, like, vinyl records, right? So, like, I have all the music I could ever need for $10, $15 a month, period.

(Joel Beasley at 00:13:44) Yep.

(Trevor at 00:13:45) I still buy vinyl records because there is, I like the tangibility. There is whatever. And at the same time, I also, like, you know, I appreciate that, okay, so Midjourney exists. I can get anything I want through that.

(Trevor at 00:13:59) And, again, you know, fast forward a few years and when it can truly spit out exactly what you're thinking with whatever quality, right? I still value human art too. Yeah. And if anything, it has a premium on it now because

(Joel Beasley at 00:14:12) a

(Trevor at 00:14:12) person put time and blood and sweat into this, you know? And I think the same is true for, you know, there's a lot of advancements happening. Did you see the Unreal Engine 5.2 demo? The

(Joel Beasley at 00:14:25) No. How recent was it?

(Trevor at 00:14:26) This was a few weeks ago. It wasn't that bad.

(Joel Beasley at 00:14:29) I saw the one from a year and a half ago.

(Trevor at 00:14:31) Okay. So it was bonkers. Bonkers. Okay, so this one was recent.

(Trevor at 00:14:34) And what's funny about it is it's just 5.2. It's not even a major version update, and it had the most mind-boggling—they added some deterministic procedural generation. Where in the example they gave, they had—well, this was cool too—they had a 3D-scanned Rivian truck driving through this thing, and it looked real. I mean, it looked real real.

(Trevor at 00:14:57) And they're driving it through this kind of rocky canyon or whatever. They get to a Y fork in the canyon, right? And in front of them, the thing that is forking it is this huge—it's just a canyon wall. You know? There's a Y and there's a big canyon wall. And out of the canyon wall, you can see a lot of this organic—there's foliage, you know, in certain spots. Little rocks and things that look like they, you know, took place over thousands of years. Branches sticking out of the walls and—right. Well, while they're playing the game, so it's still run time. The designer of the game—you know, it's a demo, of course—but they're on stage, and the designer's like, "Hey, we changed our mind. We don't want this fork here in this canyon or whatever." So they just move it to the side. And when they do, all of the procedural evolution redoes all of the branches and rocks and whatever. And there's even a new little stream pops up because that's a new element or whatever.

(Trevor at 00:16:00) And what was interesting about it, though, is—it was things that, you know, that's not my world. I'm not working in Unreal, so I wouldn't have thought of this. But I just thought it was so cool that they had—because it was deterministic, you could have an artist that's still guiding this thing. Right?

(Joel Beasley at 00:16:19) Yes.

(Trevor at 00:16:19) And they're moving it around. They're trying a few things. And then if they're like, "I don't like where it's at now," move it back two slots or whatever. When you do, it looks exactly like it did then because it was deterministic.

(Trevor at 00:16:30) So you're basically—you think of it like a Google Doc, right? You just Command-Z and it looks like it did, but it's this super hyper-complex thing. And every time they moved this, I mean, that would have been days of work for a person, and it would have looked exactly the same. And what's fascinating to me about—again, talking about the creative sphere is—what new types of things are going to be created or come out when the time to doing so is shrunk? And how fast are things going to keep moving to where you've got—or even going on the discipline train you were talking about. Right? People that have discipline but don't have that skill set. Right? What happens when they start jumping into stuff like this?

(Joel Beasley at 00:17:15) They're doing it.

(Trevor at 00:17:16) Exactly. It's fascinating. And so I love the empowerment that comes with that.

(Joel Beasley at 00:17:21) You just opened it up too. Because here's the thing. We've all had calculators. We've had stuff automated for a little bit, right? For a minute.

(Joel Beasley at 00:17:29) Now they're getting the creativity side automated, and they're going to have a field day with it.

(Trevor at 00:17:33) Exactly. Yeah. Well, and it's, you know, for a long time—I can't remember the quote I saw, but it was basically, we thought that creative jobs were the only ones safe from robot overlords.

(Joel Beasley at 00:17:46) Yeah.

(Trevor at 00:17:46) You know, and that all the factory jobs would go away. Turns out it is much more difficult to make an automated assembly line of something than it is to teach something what creativity looks like. Yeah. Yeah.

(Trevor at 00:18:00) Anyway, it's a fascinating topic because it's still very much the Wild West. And my hot button opinion is that—sort of this new—just like you said, these generative models are not new. Right? They just—when ChatGPT came into beta a few months ago, it just flipped everything. Yeah.

(Trevor at 00:18:17) The public sphere—I think it hit a new level of, "Oh, wow. This is much more real." You know?

(Joel Beasley at 00:18:24) I texted my—so the venture fund that invested in me five years ago, Florida Funders. Yeah. I was the third investment. They've grown to be the largest VC in Florida. Wow.

(Joel Beasley at 00:18:33) They have hundreds of investments. Wow. So I know the original couple people there, and they're very nice, and they've been working with me for years. And about a year ago, I had seen the—I tracked the progress of a handful of projects. Sure.

(Joel Beasley at 00:18:46) Right? And I had seen where it was going to be, and I was like, "Oh, no. This is going to be crazy." So I texted Mark. I said, "Mark, watch out for any investments that are with this generative text modeling and everything." And I gave him a couple examples of how I would implement it or just general ideas that flew through my mind. Sure. And he was like, "Cool. Cool." And then I talked to him a couple weeks ago, and we were on a meeting, and that came up, and I was like, "Mark, remember that text?" He goes, "Oh, my gosh. I do remember the text. He goes, "Wow."

(Joel Beasley at 00:19:15) Oh, I love that.

(Trevor at 00:19:15) Yeah. Yeah. You need a gold plaque on your wall a little bit.

(Trevor at 00:19:18) "I predicted ChatGPT" is what it is.

(Joel Beasley at 00:19:21) Dude, I predicted the—we had to cut it from the show at the time, but I predicted an acquisition ZoomInfo was going to make.

(Trevor at 00:19:30) No way.

(Joel Beasley at 00:19:31) They were in the middle of it, and he looked at me, and we had to cut it out, and it was this whole big thing.

(Trevor at 00:19:36) That is—yeah.

(Joel Beasley at 00:19:37) And then seven months later, he made this big—one of the—I think it's the CTO of ZoomInfo, right? The data company or whatever. Yeah. Yeah. And he made this big post on LinkedIn.

(Joel Beasley at 00:19:47) He said, "Joel, a.k.a. Nostradamus." I was like, "Well, first, let me Google who Nostradamus is." But I was like, "That's awesome."

(Trevor at 00:19:56) That is amazing. Also, Nostradamus had your beard, I think, so it's way too accurate. Pretty sure he did. Yeah.

(Joel Beasley at 00:20:01) You know Duck Dynasty guy? Will Robertson? Yeah. I got to meet him this last weekend.

(Trevor at 00:20:06) No way. And—

(Joel Beasley at 00:20:06) He's like, "Look at that beard. It's awesome." You know—

(Trevor at 00:20:09) You know you've made it when he says your beard is great.

(Joel Beasley at 00:20:11) You know?

(Trevor at 00:20:12) It's like, "I'm in the beard camp now." I know.

(Joel Beasley at 00:20:14) I told him. I was like, "I snuck backstage. I told everyone I was your nephew." And he looked at me seriously. He's like, "Oh, my God. Really? What did they say?"

(Joel Beasley at 00:20:22) I was like, "No, bro. I paid. Paid to come back."

(Trevor at 00:20:26) That is amazing. Um, yeah. In that—I—the other thing in particular with, you know, when it showed its head, ChatGPT, towards the end of last year, the technology itself is—I truly believe, and this was my hot take, is that this is sort of, if I had to define Web3, this is it. Mhmm. Not that I don't see value in blockchain technology or crypto in general, but I truly believe that what we're seeing right now is the next step change of the internet.

(Trevor at 00:21:07) How we think about search is changing actively.

(Joel Beasley at 00:21:11) Right? Mhmm.

(Trevor at 00:21:11) Clearly. How content is being generated and what content is being generated and who's generating it—robot—is changing. How we're doing our jobs, what jobs exist. That's what happened with Web 2.0 and Web 1.0. So it's just—this to me is—you know, I don't claim to be able to predict it as well as you can, Nostradamus, but—but I do.

(Trevor at 00:21:36) I think that there's a lot of—over the next few years, we're going to see stuff that was literally in Star Trek. Again, just like with the iPad and—Oh, yeah. You know? I mean, it's—that's what we're on the train for now.

(Joel Beasley at 00:21:51) Yeah. I don't think it was anything too special, predicting it. I just typically—I actually am surprised that I find people are surprised when you just connect basic dots and they're like, "How'd you know that?" And I was like, "Well, first of all, I'm a customer of ZoomInfo. Secondly, intent data at the time was booming as this idea, and I suggested to him that he purchase an intent data company because I wanted him to—because I wanted it ingrained into the product. Sure.

(Joel Beasley at 00:22:16) Because I was a customer.

(Trevor at 00:22:17) Sure.

(Joel Beasley at 00:22:18) And so that's how that got started. But for you, what I would think is going to happen for your company—

(Trevor at 00:22:25) Right.

(Joel Beasley at 00:22:25) Is I think you would take how you prompt Midjourney and use that for music creation.

(Trevor at 00:22:33) Yeah. So there's—and there is—have you seen MusicLM, which is Google's—okay. So MusicLM is that. It's Google's thing. They just put out a—they did a white paper for it beginning of the year.

(Trevor at 00:22:45) But about a week or two ago, they put a beta out of a UI that you can actually go and prompt. Yeah. Yeah. The only thing is—the technology is amazing. No doubt.

(Trevor at 00:22:57) But it is surprisingly similar to generative music that has already existed, like the Boomys and the Amper Music of the world. Amper Music got acquired by Shutterstock, I think, five or six years ago. But it's—and the reason I say that is because it's actually pretty dang good at certain types of EDM. Right? Which makes sense.

(Trevor at 00:23:20) That's already a computer-oriented genre. Right? Yeah. But, I mean, if you try to do authentic jazz or blues or whatever, I mean, what is even happening? You know?

(Trevor at 00:23:29) It just doesn't do it—or metal or whatever. But the concept is certainly interesting, especially again, I know how much value I get out of something like a Midjourney or whatever. And as we—we're always looking at ways even outside of the generative sphere in terms of content creation because, you know, generative can also mean insight derivation. Right? Pulling insights out of things too.

(Trevor at 00:23:54) So there's a lot of interesting stuff happening that—and frankly, it's a matter of prioritizing because there's so many cool things we could or are doing, some of which I can or can't talk about. But the things that we're actively working on are—what would our customers care about? You know? Because the—going back to even the crypto thing, that was probably my biggest head-scratcher with a lot of what was happening in Web3 of the last couple years is there were so many things happening that were solution in search of a problem where it was Web3 was the hammer, so everything became a nail. Right?

(Trevor at 00:24:34) I remember listening to a podcast where a very large game company that has their own web store online was—I think it was the president or someone that was talking about—they had created a—they were creating a secondary market through their digital marketplace. Right? And they were using blockchain because X, Y, Z, blah, blah, blah. It was a centralized marketplace. You didn't need blockchain for that.

(Trevor at 00:24:57) They did it for marketing, and I understand that. But—

(Joel Beasley at 00:24:58) But—

(Trevor at 00:24:59) It was also—that's a prime example of—and then there was another company that put—and this was a big one, marquee name. I can't remember who it was, but all they did was they changed the name of one of their products. They put "bit" in front of the name, and their stock went up by 20% that morning. Just—what? They didn't do anything else.

(Trevor at 00:25:17) Um, but—and all that stuff happens in any bubble, you know, and kind of—so—and I'm certainly not faulting the technology for that. It's just that the difference with what I think we're seeing with AI is—it's truly going to the customer-facing thing. Customers didn't care about half that stuff with Web3. Right? The value was typically on the side of the creator or whoever's benefiting from the secondary market or whatever.

(Trevor at 00:25:43) In the case with AI, the value is there for everybody. It has already been shown to be the case for consumers, businesses, et cetera. And there is so much—um—to your point about the flame wars that are happening online. We spent as a society, I feel like we spent so much time pushing against the grain or trying to put a wall up to change. It is going to happen. We as a species just build things.

(Trevor at 00:26:15) That is literally since the beginning of time. That is what we do. We're going to keep building and do it, blah, blah, blah. No matter what the legislation winds up being, it's going to—I mean, look at Uber, taxi industry on its head. They were built with our dollars.

(Trevor at 00:26:33) Exactly. And I—I'm just frankly, at least for the generative side of music, what I am excited to see are the ways in which this can enable a new economy for musicians because I truly believe that people that aren't even musicians will become that, just like we're talking about with the discipline thing and being able to use tools that help you do that. Right?

(Joel Beasley at 00:26:52) Like, I'm not a great musician. I'm a medium musician. Right? And the idea that I can then use tools to augment that. Exactly. Without having to spend the ten years—

(Trevor at 00:27:04) Right.

(Joel Beasley at 00:27:04) To get there because I can—my problem is I can hear it in my head. Sure. And I'm just like, "How do I make that happen?" Sure. And so with the assistive tools, what that'd do is that it would empower me that I have a physical time block issue with being able to acquire those skills.

(Joel Beasley at 00:27:20) Yeah. And now that's gone. Yep. Because I don't really need to. Sure.

(Joel Beasley at 00:27:24) Because now I have this assistive tool that can allow me to do it. Exactly. It's like you arguing, "I'm going to walk to Las Vegas instead of taking a plane." Right. It's like you use whatever the modern technology is around you.

(Trevor at 00:27:35) Exactly. 100%. And, you know, I think a lot of the argument there in particular—in what I would argue is probably more of an elitist stance—is like, "Oh, well, you'd—it undermines all the time people have put into their craft." Right? But at the same time, that only exists because we haven't had this.

(Trevor at 00:27:55) Yeah. Like, that is literally the only reason that is the case. Right? Or, by that same token, should anyone be allowed to use a keyboard with that argument? It's using a computer to process samples that they may or may not have created.

(Trevor at 00:28:10) You know? So I totally agree. I think it's honestly—the—there will be hurdles that have to get navigated. It's happening with every—I mean, whether it be litigation or whatever. All that stuff.

(Trevor at 00:28:24) But it always winds up getting figured out. I don't think there has ever been a technological innovation that was a wave that changed things like this that just all of a sudden stopped because we just wanted it to.

(Joel Beasley at 00:28:36) They won't. It won't. You can't. You can't. What we experienced because, you know, it took me a while to figure out—so I'm 35, but it took me a while to figure out, why people at each juncture in the evolution of the economy and humanity always are freaking out about the job loss.

(Joel Beasley at 00:28:56) It's like, we've been dealing with this since the beginning of time. The only variable is that has been speeding up. Right. And so it's speeding. And that might—that's definitely something to consider.

(Joel Beasley at 00:29:03) If we—Yeah. If we decimate an industry overnight, that's not something we have efficiency with in the fifties. It would take a long time for it to play out. So we will have to figure out, what do we do then? Sure.

(Joel Beasley at 00:29:15) Yeah. But that's going to be a super, super tough problem. I did—and speaking of lawyers, I did talk with the lawyer of the—I think it was David Guetta or Guetta. The guy that did the Eminem thing.

(Trevor at 00:29:28) It was David Guetta.

(Joel Beasley at 00:29:29) Yeah, he did the Eminem thing live.

(Trevor at 00:29:31) Yep.

(Joel Beasley at 00:29:32) And his lawyer came on the show.

(Trevor at 00:29:34) Oh, cool.

(Joel Beasley at 00:29:35) Yeah, and we were talking to him about it. And the take—and Josh, correct me if I'm wrong—but the takeaway from it was that it's just the Wild West right now.

(Trevor at 00:29:43) Yes, 100%.

(Joel Beasley at 00:29:44) There's just gonna be tons of lawsuits, and then case law will happen. But it's just—he looked very happy. I imagined he was making a lot of money.

(Trevor at 00:29:52) I would imagine so too.

(Joel Beasley at 00:29:53) Yeah. If you're the foremost expert in the world on dealing with this—I mean, I'm a creator. I definitely agree that if you're gonna be using Eminem's likeness, he should get some money for it.

(Trevor at 00:30:05) Sure.

(Joel Beasley at 00:30:05) Because he crafted it.

(Trevor at 00:30:07) Sure.

(Joel Beasley at 00:30:07) And then if somebody's gonna use your likeness, you should get some money for it. I have no idea how to price it. I have no idea how to enforce it.

(Trevor at 00:30:14) Sure.

(Joel Beasley at 00:30:14) And I can tell you one thing for certain: people are gonna do it either way.

(Trevor at 00:30:17) Either way, right.

(Joel Beasley at 00:30:18) It's the Napster, it's the—yeah, LimeWire.

(Trevor at 00:30:21) 100%. Well, and the other thing that's tricky is that it's in this weird gray area between—so likeness is typically handled with, if I understand this right, I think trademark, right? Whereas, you know, your melodies and your compositions and stuff—that's handled with copyright. But where it straddles this line is, this is sort of—because it is, I think it leans towards being a trademark thing, right? Because it is Eminem's likeness. But also there is a—they're creating new work from it, derivative, rather than, you know, they're not just printing his image a bunch. It's just, like you said, it's a new world. And frankly, I don't know—once case law happens, I don't know if it will ultimately wind up changing copyright law, or if this is truly a form of replication that we don't have a name for yet. You know? Because it isn't—it's not Eminem's—like, we use likeness for image, name, that kind of stuff. Those are all static properties, right? But is your voice your likeness? Because, for instance, you could probably find two people in the world at some point in history whose voice sounded the same.

(Joel Beasley at 00:31:45) Oh, there's—some—people that look alike are real. If you have not—if you're 30 and you have never had someone send you a Google image or something of a person that looks exactly like you, your doppelganger.

(Trevor at 00:31:56) Exactly.

(Joel Beasley at 00:31:56) And you see that photo and you're like—you can tell the slight variance, but you're like, wow, that's pretty spot on.

(Trevor at 00:32:03) It's pretty crazy.

(Joel Beasley at 00:32:04) Look at this—the doubles. You ever see in the action movies when they put the people next to each other? It's like, oh my goodness.

(Trevor at 00:32:09) Exactly. Yeah.

(Joel Beasley at 00:32:10) They look exactly the same.

(Trevor at 00:32:11) It's true. And I also don't know—you know, and if anybody deals with likeness law, they probably already have all this figured out, so I certainly don't know any of that. But what I do think is interesting is what happens with—because how different is it from, like, if I do a cover song, right? I can find someone that sounds a lot like a singer, and I'm not violating anything if I make that a cover song. And then I own the recording.

(Joel Beasley at 00:32:37) Yeah.

(Trevor at 00:32:38) Because—and, well, music copyright law is super complicated, but there's the recording and then there's the composition. So whoever wrote the composition still owns that. But if I do a cover, I own the recording. But I'm also not violating anything if it's not actually that person's voice. So is that different, though, than using a machine that sounds like someone's—this is where it's—

(Joel Beasley at 00:32:58) And then how would you—what sort of tools would come about that could determine—it's horrible. Your OpenAI's hit rate's 20% on detecting its own type of text.

(Trevor at 00:33:09) Yes.

(Joel Beasley at 00:33:09) It's just so imperfect. They can't even do it. So how would you differentiate between me just telling you, oh, that's my voice, and my inflection—and I just tweaked the settings of the filter my voice is going through, and it just happens to sound a lot like Eminem. And I just happen to be talking about topics that Eminem talks about. So it's gonna be super messy as we go through it. But it's a lot of fun for me because my business is not that close to the music. Like, you see Ed Sheeran—I don't know, that's how I say his last name. If that's not how you say his last name, I don't care.

(Trevor at 00:33:45) I think it's Sheeran, but I'll let you have that one today.

(Joel Beasley at 00:33:46) Yeah, it's the fancy version. When he's in a suit, it's Sheeran.

(Trevor at 00:33:52) Yeah, that's amazing. That's a new one.

(Joel Beasley at 00:33:56) Did he win his thing?

(Trevor at 00:33:57) He did. He did. He did, and he deserved to win that. Yeah, that case was crazy, in my opinion. But I think it was—

(Joel Beasley at 00:34:04) So far-fetched. I thought it was a joke when I heard about it.

(Trevor at 00:34:08) Yeah. I'm like, what? Same. No, he totally won. And he even, after he won, they did an interview, and he's like—I can't remember exactly what he said, but he was basically like, I can't believe we made it to court with that. That's basically what I'm—

(Joel Beasley at 00:34:22) If that artist was alive, this would never have happened. It was Marvin Gaye's estate?

(Joel Beasley at 00:34:26) Yeah. It's because it's his great-great-grandkids, nephews, cousins—they're just doing it for money. 100%. Yeah, and I get it. And I would say that they should 100% do it if it was a legit—

(Trevor at 00:34:38) Yes. Infringement. Infringement, absolutely.

(Joel Beasley at 00:34:40) This wasn't even close.

(Trevor at 00:34:41) No.

(Joel Beasley at 00:34:42) That's like me saying, you know, G-D-E minor is my thing.

(Trevor at 00:34:47) Exactly. That Wonderwall couldn't use that chord progression because every song 30 years prior used it, you know? Right. Yeah. So, you know, I do—I actually love talking about the Ed Sheeran thing in particular. I wonder what happens with—so you referenced earlier, robot lawyer, right? Where it gets interesting is, you know, the—I think GPT-4 is performing in the top 10% on the bar exam or something.

(Joel Beasley at 00:35:13) Yeah.

(Trevor at 00:35:14) Oh, yeah, right? Okay. What happens, though, in super gray area cases where there's determinations being made on, you know, how—and this is the classic morality of can a robot determine what's moral or whatever. But maybe not Ed Sheeran's because that one was, again, bogus. But there was a lawsuit a few years ago with a Katy Perry song and a Christian rapper that said that she copied the beat or whatever. And through some adjustments—like, if you change the pitch of the beat, basically it sounded very similar. So they went to court over it, and I don't even remember what the outcome was, but those determinations are still made based on a judgment, right? So do we need new laws around—is there an arbiter that decides for a robot? Do we let them make those decisions?

(Joel Beasley at 00:36:06) It's assistive tools for the judge, maybe.

(Trevor at 00:36:09) Yeah, right.

(Joel Beasley at 00:36:09) Exactly. The attorneys will use all their tools.

(Trevor at 00:36:12) Yep.

(Joel Beasley at 00:36:12) By the way, there's crazy AI tools in law now.

(Trevor at 00:36:16) Are there? Oh, man.

(Joel Beasley at 00:36:16) I believe it. I did a law project several—about ten years ago, so I still have some friends over there. And I've sort of followed the progression of what was there ten years ago when we were looking at—

(Trevor at 00:36:25) Yeah.

(Joel Beasley at 00:36:26) —concepts and what's there now. These AI can read and understand the document, point out issues, suggest different cross-examined case laws. And basically, the way it works is it's like a sidebar in the app. It's like this intelligence sidebar. It can give you—you can ask it questions. You can do things. You can sort of control it over there. And it's suggesting things to you. It's pointing out different parts of text. You can take people's—in discovery, I can just take all of your emails and push them in there and then have it—

(Trevor at 00:37:02) Whoa.

(Joel Beasley at 00:37:03) —go through all of your emails—

(Trevor at 00:37:04) 30,000 emails.

(Joel Beasley at 00:37:05) —to find the thing, because that's one of the hard things for discovery. If you get 30,000 emails, you have to have a team—

(Trevor at 00:37:10) Reviewing all of them.

(Joel Beasley at 00:37:11) —reviewing all of them, hoping you don't miss something. Now, 30 seconds later, you can—and you can even tell it, hey, go see if there's any incriminating evidence. You can talk to it like a person. You say, go find any incriminating evidence that is inside of this. And these things are smart, dude. How much have you played with GPT-4?

(Trevor at 00:37:29) Oh, a lot. I mean, I've been paying subscribers soon as they made it available. Yeah, absolutely. I use it all the time.

(Joel Beasley at 00:37:34) Yeah, we use it constantly.

(Trevor at 00:37:36) It's amazing. And if you get—this is the other thing. This is now starting to happen on LinkedIn. People are pointing this out, but people have only been using it to—people haven't really taken advantage of how you can prompt these things.

(Joel Beasley at 00:37:47) Oh, I know.

(Trevor at 00:37:48) It's crazy. Anyway, go ahead with what you're saying.

(Joel Beasley at 00:37:50) Well, no, I don't remember what I was saying. But to your point, every time I meet someone in public and they're not—they're typically nontechnical—I'll give you an example. So I go to church, and at my church there was a secondary pastor, not the main guy. And he had to write this paper. And at the Wednesday night group, I was like, oh, GPT. And he's like, okay. And so he played with it, and then Sunday he saw me. He's like, dude, this is garbage. And I was like, all right, well, tell me, what did you say? How did you—? He's like, oh, I said write me a 14-page paper on Matthew and whatever.

(Trevor at 00:38:22) Oh, okay.

(Joel Beasley at 00:38:22) And he went off, and I said, oh, yeah, that's not how you do it.

(Trevor at 00:38:25) Yeah, exactly.

(Joel Beasley at 00:38:27) And so I started to give him some examples of—you know, don't just say write my paper. Figure out different areas and ask it, prompt it in different ways, and give examples. You know?

(Trevor at 00:38:38) Even put your own past content in there to put your voice in.

(Joel Beasley at 00:38:42) Oh, yeah. Yes. It does really good saying "act like." Yes. Yeah, that's one of my favorites.

(Trevor at 00:38:46) The—you mentioned earlier the 20% detection, right? Yeah, yeah. You can thwart it 100% of the time if you just ask it to change its voice. It cannot work with that. Literally, basically, if you ask it to have an empathic or empathetic tone, if you ask it to use your voice and give it, you know, here's a paragraph I've written or whatever, detection is impossible, basically. It goes away.

(Joel Beasley at 00:39:09) Yeah, I've put my own transcripts through it.

(Trevor at 00:39:11) Yeah.

(Joel Beasley at 00:39:12) I'm like, oh, you evil Joel. Then I break my MacBook, and then I realize that it's hosted somewhere else.

(Trevor at 00:39:20) Exactly.

(Joel Beasley at 00:39:21) And I feel bad. They pick up the scraps from my MacBook. Call up Wozniak. Hey, bro, we got problems. The world's falling.

(Trevor at 00:39:30) Oh, I love that. Yeah, it's—yeah, it's super exciting. It's also—I just—people not really knowing how to prompt and stuff, I think we just don't even know what all is gonna come out of this.

(Joel Beasley at 00:39:43) Yeah, yeah, yeah. Wild. Prompting is the skill. It's gonna be a hugely important skill. It's gonna be the equivalent of being able to write code.

(Trevor at 00:39:52) Yes.

(Joel Beasley at 00:39:52) You're gonna need to prompt.

(Trevor at 00:39:54) Yeah. Well, even, you know, like, using GitHub Copilot, which is their assistive thing—they have a new version coming out that has GPT-4 under the hood. I mean, for example, right now, it can only—or at least a year ago—the last—honestly, I used ChatGPT for most of my coding assistance now.

(Joel Beasley at 00:40:15) Yeah.

(Trevor at 00:40:16) And because Copilot—access is not yet. But how Copilot works before is it would either—it would take context from the file you're working in or whatever. And maybe this is the ChatGPT plugin. I can't remember. I've tried a hundred of these things. But either way, a lot of them will take context from the file you're working in and can make suggestions or, you know, you can prompt it if you need some boilerplate written or whatever. But the other thing, too, is most of the programming problems or the things that take the longest typically are stuff like debugging—like discovery with emails, right? Because there's so much review that has to happen and trying things or whatever, or solving really hard problems that have been solved before, but you have a particular use case, right? For example, I was playing around with writing a browser-based video editor about a year ago.

(Joel Beasley at 00:41:02) Okay.

(Trevor at 00:41:02) And I wanted to do, you know, if you drop in, I don't know, a video clip into an editor, and I wanted to do an effect where it's stacked images. Like, it looks like photos being dropped on each other, right? Well, I need to cock each photo—like, rotate it, right? Yeah. Yeah. Okay. Well, to do a rotation handle, like you would expect with an image editor, there's trigonometry involved with that. I haven't done trig in years. I haven't needed to.

(Joel Beasley at 00:41:27) Right?

(Trevor at 00:41:28) I've never done it. Well, trig is basically what you need for that and matrix multiplication and stuff that I just don't use. I haven't had any need for it. And it took me—I was spending days having to relearn stuff I haven't done since tenth grade. You know? And literally, in an instant, I could have had that whole slew of time given back to me, and I'm never going to use that again. Why would I need to spend all that time on that?

(Joel Beasley at 00:41:58) You don't.

(Trevor at 00:41:59) Exactly. And so going again to the discovery piece and stuff like that, the more context these tools are able to handle, right—like, if I can pipe our entire organization's code repo into it and it can understand how our database schema works and all these files are talking to each other, et cetera—how much more effective can our small engineering team be, right? And how much cost saving can we get out of that? How much more agile can we move compared to larger competitors and things like that? That's the stuff I'm super excited about with it.

(Joel Beasley at 00:42:26) Well, I'm actually curious because I started looking into this in the past couple weeks. Moving past the idea of I can tell it within the context of a conversation and find value there, use it for sales follow-up emails by pushing stuff in there and saying, hey, how would I respond to this? All sorts of different ways of using it. And what I found is, as I wanted to use it more, I needed it to have a larger understanding of the context.

(Trevor at 00:42:55) Yep.

(Joel Beasley at 00:42:56) And then I quickly found that you're limited.

(Trevor at 00:42:58) Yes.

(Joel Beasley at 00:42:59) So I wanted it—I took a transcript of a sales call, and it was fairly long. And I put it—I tried to put it in. It's like, okay, you're limited to 2,000-something tokens. And then I was like, all right.

(Joel Beasley at 00:43:11) Well, I started talking to it and saying, like, what type of systems allow you to get around this?

(Trevor at 00:43:17) That's awesome.

(Joel Beasley at 00:43:17) And it started explaining to me the different types of models that you can use and the different ways, and I got all the way to the point where it's giving me, like, code samples. That's awesome. Right? Because I think that's what a lot of people want to do.

(Joel Beasley at 00:43:29) I want to drop my, you know, SQL or whatever, database and, uh, not SQL, Postgres. Drop my Postgres. Sorry. Sorry, guys. I was unclassy for a minute.

(Trevor at 00:43:38) It's all good.

(Joel Beasley at 00:43:39) It's funny. That elephant action. Dropping Postgres in there. Yeah. And I want and then I just want to talk to it. Yep. You know? And I want it to understand. It'll pick up that it's a real estate application based off of, like, the table names. Yep. And then I also would love to throw the code in there. Do you know currently, because, again, this is actually the conversation I was having yesterday with it when I was trying to do this, so this is well timed.

(Trevor at 00:44:02) Yep.

(Joel Beasley at 00:44:03) Do you know of any tools where I can just throw a bunch of unstructured data at it and it could hold all of it in context?

(Trevor at 00:44:11) I don't know of anything, like, off the top of my head. I couldn't point you to anything. But what I do know is that based on some blurbs that, like, because GitHub's Copilot X that's built on GPT-4, if I remember right, some of their marketing around that was that it would essentially, since your repo's in GitHub, right, it would be able to have context for your organization's repos, which that to me is huge. That's gangbusters. However, the problem you're talking about extends outside of coding even because there are, like, good example, um, there are marketing teams who are struggling to get their entire, like, they may have, like, a, essentially, you know, a spreadsheet type database thing of their brand voice. Right?

(Joel Beasley at 00:44:53) Mhmm.

(Trevor at 00:44:53) Like, that have examples of here's how we would answer or, like, on these particular subjects, here's the way we approach them. And then, you know, there may be, like, boolean type things too. Like, yes or no, do we cuss in our posts?

(Joel Beasley at 00:45:08) Mhmm.

(Trevor at 00:45:08) You know, stuff like that. But the problem is that, like, depending on what you're prompting or what you need, right, like, maybe I need an article written and I need it to have this database of what our brand voice is, but also I want to paste in another article that I'm taking inspiration from and, like, I've maxed out the limit, right?

(Joel Beasley at 00:45:26) Mhmm.

(Trevor at 00:45:27) And so, uh, this just came up in a conversation that I was having with a marketer friend of mine about two weeks ago, where there are tools that are, like, sort of trying to help. And I don't know if they are, you know, trying to essentially summarize your brand voice and pump it in from your table or, like, how exactly it's getting that context into the token limit. There are things that are trying that, but there is no singular tool that I know of now outside of building your own model, because, you know, you can create a model where you have enough resources that you could dump context in and fine tune it for what you need. But there's also, like, at that point, clearly it's not GPT-4, right?

(Joel Beasley at 00:46:09) Yeah.

(Trevor at 00:46:09) And you're also, like, when you're getting to the level of needing to fine tune it, you're back into R&D land just doing what OpenAI is doing. Yeah. So it's, there's a trade-off where, like, unless you have that kind of resources, you could make something amazing, but you could also wind up running into the same stumbling blocks that they're running into, I'm sure, as they build things. You know? Yeah. So, yeah, I don't know of anything right now. And, you know, some listeners might have some suggestions, but I don't know of anything that takes, honestly, more than the prompts even, because now I think there's a, I don't remember, they extended the token limit for Plus users or whatever. It's like 8,000 or something now. I don't know. I know that there are other tools that do similar stuff, like Replit, if you're familiar with that. It's basically like, Replit is like a cloud-hosted code environment. They were, like, one of the earliest to that, but they have their own assistive AI coding technology. Mhmm. And if I remember right, they do some sort of context awareness since all your code's in the cloud, right? Yeah. But, yeah, I don't know.

(Joel Beasley at 00:47:13) I had some success with, I took the conversation and I started breaking it up by, like, five-minute chunks or, yeah, fifteen-minute chunks or whatever. But I told it. I said, okay, I have this conversation that's larger than the token limit, so now I'm going to paste it in sequentially and so on. And I did it, and it, like, it definitely was annoying because it was trying to respond to every thing.

(Trevor at 00:47:39) Yep.

(Joel Beasley at 00:47:40) But after I got it all in there, probably, like, six pastes, it really understood. I went back and looked at the conversations, and I intentionally formed a question where it would have to understand, yep, from more than one paste something. And it did. Yeah. And I was like, alright, well, then now I'm just slowly learning the limits. I looked at the API. They said that the fine-tuning API can hold that across conversations.

(Trevor at 00:48:07) Yes. And that's essentially when, when you're in a, to my understanding, when you're in a, like, we open a new chat, right, in ChatGPT, every back and forth is essentially you fine-tuning it for that chat.

(Joel Beasley at 00:48:18) For that chat.

(Trevor at 00:48:19) Exactly. And I think that's basically what the API is replicating is allowing you to do that in, yeah, and you can use it that way. I don't know what, I don't know, and I've never hit this, but I don't know if there's a limit to that, you know? Like, if there's a, and do they specify that on the API? If there's a limit to how much you can fine tune?

(Joel Beasley at 00:48:36) Um, I only got as far as trying to understand the right words to even talk about this stuff. So I had an idea.

(Trevor at 00:48:44) Sure. Yeah. Absolutely.

(Joel Beasley at 00:48:44) So because of this idea that I had, I talked to my buddy and we're like, hey, let's go just kind of kick the tires and see if, see if there's anything even here.

(Trevor at 00:48:52) Right.

(Joel Beasley at 00:48:53) And so I had to take from my conversational talking with different people from that level to an actual, got it, like, almost practitioner level to push it beyond what I knew how to make it do. And that was all, like, a lot of this week. Yeah. And me just exploring and trying to figure out. And it's been really helpful because, as you know, when you achieve mastery of something, you have this mental model in your head, right? And you need that in order to talk about it and reason about it and figure out what's possible. Yep. Right?

(Joel Beasley at 00:49:25) So I was just working on building that model of, okay, exactly what do these models look like? Sure. How do you make them? Like, where do they live?

(Trevor at 00:49:36) Yep.

(Joel Beasley at 00:49:36) Uh, how would you boot one? I'm a, you know, software engineer, so, sure, I've never done much with this.

(Trevor at 00:49:41) Right.

(Joel Beasley at 00:49:41) But I was like, what would that look like? You know? Yep. How does it work exactly? And so I just asked a ton of questions and learned a whole bunch about it. But the idea that I have had before, which isn't what I was specifically looking at this week, is, like, I just want to give it, like, my parents own a medical company, right? They, yep, like, doctors go see them, right? Got it. That type of medical company. And they do, like, 3 to $5,000,000 a year, right? So it's a small business, and they sell all sorts of things, like 40 products from vitamins to the patient visit and all of that stuff. And I wanted, and they keep their stuff in QuickBooks, okay, right? And so a couple years ago, they were asking me to take a look at the P&L and the stuff because they have seen me in my interest in business, right, and asked me what I was thinking about it. And I quickly found out that there is so much amazing data that they have on all of these people and products and transactions.

(Trevor at 00:50:43) Interesting.

(Joel Beasley at 00:50:44) That they don't use at all, right. So for example, when we wanted to do reports of, like, okay, well, what's the best-selling product across this time frame, or we had all of these questions, and the labor to build the reports was so significant.

(Trevor at 00:50:59) Yeah. I bet.

(Joel Beasley at 00:51:00) And, you know, the cleaning up of the data because the reports are fairly rigid and, yep, all of that. And so I said, why can't I just take this QuickBooks database and shove it into, sure, an AI model and then just talk to it, right? Like, why is that not something we can do today? Why can't I, and you own a business, right. Why can't we take our P&Ls and our, sure, business transactions and our costs and shove it into these systems and say, okay, act like, and pick our favorite business person.

(Trevor at 00:51:30) Right.

(Joel Beasley at 00:51:31) Who's written X books, yeah, on the topic, yep, and use their principles and advise me, something I'd normally have to stroke a six-figure check for.

(Trevor at 00:51:39) Exactly.

(Joel Beasley at 00:51:40) Advise me on what I'm not seeing in my business, yeah, or the next couple moves I should be considering. That is when I'm going to get real excited, when we get to that point.

(Trevor at 00:51:50) I agree. I agree. Because right now, there's a lot of, just like you said, there's the initial lift, like, the actual getting all that stuff together, figuring out a way to where it's even analyzable by you. Like, how do you even, you know, like, for instance, if you have data that's not in a structured format, like, like a QuickBooks database or whatever, maybe it's spread across a whole bunch of spreadsheets with different formats and whatever, uh, maybe it's from disparate sources, QuickBooks and a spreadsheet and whatever else and cash and whatever, you know, like you have all these records of all these things in different places. And honestly, the financial piece itself, like financial analysis, is that same problem. It pretty much exists in most, you know, we have tools that, like, you've got Mixpanels of the world that help you actually track behavioral data and stuff, right? But analyzing that stuff is still a manual process. Building your dashboards, picking out insights and things. Like, I think about this all the time of, like, why isn't there more? And there are some tools that surface some of this stuff, but, like, basically, we, you know, we have, there are teams all over the world now typically called product, right, that a large portion of their job is they spend time, just like finance spends time digging through spreadsheets, where they are having to dig through all this data to pull out the insights when, like, well, why can't we just have a model surface that and then they can go do things with it?

(Joel Beasley at 00:53:20) Who's the guy, who's the Broadcom, Tony from Broadcom? Yep. What's the phrase that they're doing now?

(Intro Narrator at 00:53:26) Meaningful observability.

(Joel Beasley at 00:53:29) They're building, because they work with, like, Knox and Socks and, yeah, yeah, pants and, yep, yeah. Okay. All the stuff. K. And he said, you know, Broadcom's huge, right?

(Trevor at 00:53:39) Yep.

(Joel Beasley at 00:53:40) And so their customers are huge. And they say, you know, I think his phrase was, my customers are telling me, I don't need more metrics.

(Trevor at 00:53:47) Right. I don't need more insights on it. Yeah.

(Joel Beasley at 00:53:50) Like, I need to understand what's meaningful about it.

(Trevor at 00:53:52) Yes. Bingo.

(Joel Beasley at 00:53:53) And so now they're building these types of tools. It was actually pretty cool because I had him on, like, two years, a year or two ago, and he was talking about it as, um, so he was a practitioner, and then Broadcom brought him in as, like, the token customer to work with the teams and work with the customers. And he talked about it in theory about a year and a half, two years ago, and then he came on again a couple weeks ago. He's just over in Memphis. He's over in Memphis. Okay. So he came on a couple weeks ago, and it's, they actually have a tool now in the wild.

(Trevor at 00:54:26) Oh, cool. And they built it? Proprietary?

(Joel Beasley at 00:54:30) Yes. Yeah. Cool. Yeah. So they have this tool now in the wild. I don't know if it's, like, publicly available, but I know it's being used.

(Trevor at 00:54:36) Love that.

(Joel Beasley at 00:54:37) And he came up with this phrase, meaningful observability.

(Trevor at 00:54:40) That's interesting. That's exactly it.

(Joel Beasley at 00:54:42) Yeah. He, like, pioneered that stuff at, um, like, FedEx in the early eighties or nineties when they were doing all of their logistics and moving to the, you know, the computer systems and all of that, figuring out how to handle what was he doing? Uh, what Gremlin does? Site reliability engineering.

(Trevor at 00:54:59) Yeah.

(Joel Beasley at 00:54:59) He's like the godfather of SRE.

(Trevor at 00:55:01) Yeah. Wow. What a champion. Dude, that, so logistics in general, like shipping logistics, freight, all that stuff, like, to me, that's a ripe area for this kind of thing because it's, I mean, it's like, while there are always, like, unknowns, you know, like, I mean, at this point, I don't know if there's a single shipping or freight problem that has never happened. Like, there's historical data that can be used for probabilistic purposes. So, like, it seems like, and I know a lot of, there's companies like Flexport and some of those others that are dealing with things like that, like shipping logistics. Uh, they do, I think all their stuff is, um, they deal with, like, harbors and ports.

(Joel Beasley at 00:55:43) Okay.

(Trevor at 00:55:44) Like, ships and stuff. But, like, it's the same deal with what we're talking about. Like, logistics itself seems to be like a problem that is perfect for this kind of thing because it's all, like, up until now, we've either had, like, you know, these siloed, what I'm kind of thinking of is, like, last gen AI systems, right, that are narrowly focused, good at particular, you know, like, particular slices of their thing or whatever. But we're at a stage now, it's the same reason that GPT can be used for all these different things that, like, the narrowness of text models has widened far enough to where, like, if there's possibility for enough context, like what we're talking about, I mean, anything that is not, humans shouldn't have to do all that stuff. That should be basically, and it also shouldn't require a whole bunch of, you know, the timeline to general artificial intelligence is a whole other conversation, but I do think it's far enough now at this point where most of these problems are just different flavors of the same thing for the models we're talking about.

(Joel Beasley at 00:56:53) I think we're there.

(Trevor at 00:56:54) You think, oh, come on.

(Joel Beasley at 00:56:55) I think it's already happened. I have a good argument for it today.

(Trevor at 00:56:57) Okay. Go ahead.

(Joel Beasley at 00:56:58) I wouldn't just, like, throw that out there.

(Trevor at 00:57:00) Oh, hit me. Yeah.

(Joel Beasley at 00:57:00) Very simple argument. Uh, ChatGPT is smarter than a lot of people I know.

(Trevor at 00:57:05) Touché. Then it's, like, definition of what is, what is GAI, you know?

(Joel Beasley at 00:57:09) Yeah. I can talk to it or I can. Look, I can talk to people. I'm not sure if they're all there, you know?

(Trevor at 00:57:14) That is so true, unfortunately. Yeah.

(Joel Beasley at 00:57:17) But, yeah, I've gotten better in my age to, like, distance, you know, and find the right people to spend time with.

(Trevor at 00:57:23) Man, it's true because, okay, the term AI, I mean, a computer is an AI. Like, we've had it for sixty, seventy years, right? Yeah. Since the first computers. So then what is the, and in that case, you know, like, there's a, if the, like, the way I think about it, I guess, is, like, what IQ level are we saying is general intelligence? Because to your point, we have robots that have some level of IQ now.

(Joel Beasley at 00:57:57) Yeah. Well, they can pass the bar exam. Exactly. How many people on the planet can pass the bar exam? Exactly.

(Joel Beasley at 00:58:04) It's not an 80% thing.

(Trevor at 00:58:06) Yeah. Exactly.

(Joel Beasley at 00:58:07) It just isn't.

(Trevor at 00:58:08) So I wonder, I mean, you know, and I think part of the definition, though, is that it can essentially, like, ChatGPT would be able to do all of these things to turn itself into a sentient being somehow if it was truly—

(Joel Beasley at 00:58:20) Oh, okay. I thought it was just the Turing test where you couldn't tell.

(Trevor at 00:58:23) No. It is. I mean, it is. But I think that's the, at least that's the rhetoric, you know, the whole Skynet thing. What happens when it's self-aware and that, like—

(Joel Beasley at 00:58:30) Yeah. There's people that aren't self-aware.

(Trevor at 00:58:32) Right. That is also a good point.

(Joel Beasley at 00:58:35) That is also a good point.

(Trevor at 00:58:35) Also a good point.

(Joel Beasley at 00:58:36) I heard people arguing in LinkedIn. Oh, it was crazy. I got dog-piled by these five girl PhDs because I commented on one of the girl's PhDs, and then she had her whole crew come at me. Holy crap. It was, I was like, I don't know if you're all robots or if it's real or whatever.

(Joel Beasley at 00:58:50) Right. It's all changing. But it was at least kind of fun for me, so I screenshot it and show it to my wife. But they were complaining and saying something—somebody had said, you know, it is AI, and then they're like, it's not. It's just tokens, and they're just consuming information and figuring out what the next best thing to say is.

(Joel Beasley at 00:59:06) And I'm like, dot dot dot. That's what humans do. I mean, that's what we do. And they're like, it's not creativity. They just take in things around them and past experiences and then use that to create something new.

(Joel Beasley at 00:59:16) And I'm like, human consciousness. Yeah. Crickets.

(Trevor at 00:59:19) Yeah. Yeah.

(Joel Beasley at 00:59:19) Yeah. Hello. That's what we do every day.

(Trevor at 00:59:21) Yep. I mean, that is learning, and that is also the argument with, like, a lot, even the legislation that's happening around, you know, like, oh, well, there's been a couple images and stuff that have popped up that have had morphed versions of a stock image library's logo in it because it obviously used that as training data and tried to recreate it or whatever. But, like, I don't have to pay to go look at Google Images and learn from those, or I can trace them. I can, you know, do whatever. So is it just because it's faster that we're having these conversations around having to charge for training data and that kind of stuff? And, like, I don't have it.

(Trevor at 01:00:10) Like, I understand the value in that. I mean, we even have, you know, we, for instance, at Soundstripe, we have an API where we basically can provide music content for models that are training around that. Right? But we're charging for numerous reasons. Right?

(Trevor at 01:00:26) Like, we paid to create all this music. We also are providing a programmatic interface to it that we're supporting server infrastructure and all of that. There's a lot of white glove. There's a whole other thing. Right?

(Trevor at 01:00:36) But as far as something just going to Google and learning from that as a training set, it's publicly available data. So I can go learn from that for free, anyone that has access to the internet. So if a robot has access to the internet, it's kind of to the thing you're talking about. Right? Like, yes.

(Trevor at 01:00:54) That is technically what it's doing. But the whole, how much different is that than how humans—my example of the Google thing and where it's learning from, that's how we learn. Yeah. And it's literally the same thing. So I don't, you know?

(Trevor at 01:01:09) And then but the argument there is to what resolution. Right? Because we can learn and then apply it to other things, but that's the idea behind multimodal. It will be able to do that even if it can't.

(Joel Beasley at 01:01:20) So yeah. When it comes to the complainers, which is what I'm calling them, I had heard, I think Jim Rohn said something. He's like, happy people are happy people. Expect them to be happy people. Complainers are complainers.

(Joel Beasley at 01:01:33) If you expect them not to be a complainer, you're the idiot. Right? Wow. It's who they are. Right.

(Joel Beasley at 01:01:38) And so to understand that these people will always exist as we make our progress forward is important. Also, with the AGI, because I get to do the show and talk and I, you know, read a lot about it, I kind of try to get this feeling, this understanding of what are people really saying? Like, when you said, we're not an agent. And I think what people are really saying is, and then I'll tell you exactly why I think there's a parallel to VR. But I think what people are really saying is we're not at singularity yet.

(Trevor at 01:02:10) Yeah. Exactly. Yeah.

(Joel Beasley at 01:02:11) And that's what it is. Like, yeah. We're not there yet. We are not at the singularity point.

(Joel Beasley at 01:02:17) It's not completely all-knowing, imperfect, and whatnot, and it's not bipedal walking around with human skin. Right. It's not there yet.

(Trevor at 01:02:23) Right.

(Joel Beasley at 01:02:23) It's not there. But as far as if you take the word intelligent, the ability to learn—I mean, trees can learn if you block their sunlight.

(Trevor at 01:02:26) Right. Right. Right. We're not we're not uploading our consciousness into the silicon yet.

(Joel Beasley at 01:02:30) We're not there. That's a form of intelligence. Artificial intelligence, just silicon-based intelligence has been around, like you said, since the sixties, if not sooner.

(Joel Beasley at 01:02:44) Yep. But the parallel over—and one of the reasons why I think it's been so successful in culture is because the VR had been so disappointing for so long and had so many false starts.

(Trevor at 01:02:59) Right.

(Joel Beasley at 01:02:59) Magic Leap, you know, every—there's these constant, like, oh, you're gonna put it on, and you're—it's—and everybody had this big image of what it was going to be. And then you put it on, and there's just nothing but disappointment. Yeah. Right? It's just sad disappointment.

(Joel Beasley at 01:03:13) And it's like, what is this? And the moment I put on Magic Leap for the first time, I was like, oh, feel bad for those investors.

(Trevor at 01:03:19) Oh, man.

(Joel Beasley at 01:03:21) Glad I don't have any money in Magic Leap. Oh, man. Five years later. Five years later.

(Trevor at 01:03:25) Exactly. Yeah.

(Joel Beasley at 01:03:26) But yeah. So that is the opposite of what ChatGPT did. It kind of just showed up one day.

(Trevor at 01:03:33) Exactly. Publicly. Yeah.

(Joel Beasley at 01:03:34) And then there wasn't this big reveal and this, you know, I don't think that there was publicly tons of money invested into it, and it was this big unicorn chasey thing. Like, you know, the ride services are, like Uber or whatever.

(Trevor at 01:03:46) Right. Right.

(Joel Beasley at 01:03:47) There wasn't this big buildup to it. It was just one day OpenAI was like, here you go. Yeah. And then, and then it was so good.

(Trevor at 01:03:54) So good.

(Joel Beasley at 01:03:55) It was so good that it didn't create the disappointment. It met or exceeded the expectation. Yeah. And that's why I think it was the fastest-growing technology in the history of humans.

(Trevor at 01:04:07) That if you've seen the breakdown of that, how much faster it is than everything else that occurs? Like, Facebook and all these other—

(Joel Beasley at 01:04:16) I didn't validate it, but I saw it and I believed it.

(Trevor at 01:04:19) Yeah. Just like ChatGPT is what it sees.

(Joel Beasley at 01:04:21) It probably does a little more work than me.

(Trevor at 01:04:25) Okay. So the other thing that's interesting, I'm curious your thought on this. The idea of, like—I can't remember the term for it, but it's basically where it's ingesting its own content. Right? So, you know, like, if I go have it write a blog and then I put it online and then they keep training it on online data.

(Joel Beasley at 01:04:45) Okay.

(Trevor at 01:04:45) Right? One example that I saw, someone who's—I can't even remember exactly who it was. I just remember it was someone that is certainly a pro in the field. Like, they've been around for 30 years doing AI stuff. And they were talking about is the same reason, well, one of the reasons that OpenAI has built that kind of detection model, because they're trying to weed out that stuff so that it doesn't become a race to zero.

(Trevor at 01:05:07) Because, basically, the way this was described was, okay. If I ask, let's take Midjourney for example. It's like, if I ask it to create an image of a cat. Right? It's going to—unfortunate.

(Trevor at 01:05:21) Let's go with dog. I ask it to create an image of a dog. It's gonna do just like what it's learned. It's gonna try to create an image of a dog. Right?

(Trevor at 01:05:28) And it may look exactly like the real thing. Right? But it's not actually a real image of a dog. Right? So to some degree, even if it's 99.999999% close to a real image, it's still some degree off.

(Joel Beasley at 01:05:40) Right? Well, you mean it's not a photo. It's not a photo. Because if you draw a dog, that's an—

(Trevor at 01:05:47) Totally. Yeah. Right. Exactly. Right?

(Trevor at 01:05:48) But it's not a perfect, like—

(Joel Beasley at 01:05:50) It's not an actual dog you can walk up to in existence and touch.

(Trevor at 01:05:53) Exactly. And today's, today's Midjourney. Right? We'll just talk about today's. It's not perfect.

(Trevor at 01:05:58) Like, there may be a finger that's slightly off or—

(Joel Beasley at 01:05:59) Yeah.

(Trevor at 01:06:00) Right? Okay. The issue is, and the way I heard this described was if that image makes it back to Google, right, I post it on my site and then it gets indexed and whatever else, and then it learns from that. It has now taken that as an image of what a dog looks like.

(Joel Beasley at 01:06:13) And so—

(Trevor at 01:06:13) The next time I go ask it for a dog, it's gonna make a slightly worse option. And then eventually, it's going to get to—

(Joel Beasley at 01:06:18) Work in one pixel. Yeah. I know how this is gonna end up rolling out. Yeah. You're gonna have these data packs, and you're gonna be able to say, okay.

(Joel Beasley at 01:06:25) GPT, use—you know, it's gonna have its core knowledge. Right. Right? And then you're gonna be able to install data packs, like plugins.

(Trevor at 01:06:34) Which are you from data sets, companies. Yeah. Data—

(Joel Beasley at 01:06:36) Data sets. Yeah. Yeah. Yeah. Basically, like, I want the Josh, you know, understanding of this.

(Joel Beasley at 01:06:40) Sure. And I can plug that in, and I can, you know, give it weight to some degree or whatnot.

(Trevor at 01:06:45) Right.

(Joel Beasley at 01:06:45) And we're gonna go around, and we're gonna essentially be picking our sources of our knowledge banks, if you will.

(Trevor at 01:06:52) Yeah. Totally. But what—how do you, okay. So take ChatGPT where they've—I don't remember how many billions of parameters it has at this point, but it's a lot. All of those data points, how do they then sort through—how do they make sure to 100% accuracy that they're not feeding its own stuff back into it?

(Trevor at 01:07:13) Right?

(Joel Beasley at 01:07:13) We can't.

(Trevor at 01:07:14) And that's the thing is that, like, does it by definition, it will degrade it. Right? But will it degrade it to—what I don't know because I'm not in that part of the field is, does that actually degrade it any different than a blurry image of a dog? Yeah. That we took. Right?

(Trevor at 01:07:30) Like, I don't know.

(Joel Beasley at 01:07:30) I don't know. Well, here's a big asterisk. I'm so far out of my depth, but this is fun. But I just don't want people taking it too seriously, the things that I'm saying.

(Trevor at 01:07:38) Oh, same. I'm not an AI engineer. I'm—

(Joel Beasley at 01:07:41) Yeah. I'm far from that. If you guys need some business logic or enterprise applications, what up? But yeah.

(Trevor at 01:07:46) Yeah. There you go. Yeah.

(Joel Beasley at 01:07:48) Oh, it's great. I think we—there's a couple things to consider. The first thing to consider is that I believe it was largely train—it's got different versions of it. So you can arguably go back—the current version is trained up to 2021. Yeah.

(Joel Beasley at 01:08:03) There weren't a ton of generative models out there producing content that it was trained on. So you at least have sort of that, and then all the models behind it that you can say with a degree of confidence that there's not a large amount of this AI-generated stuff. True. Second, I think it would definitely reinforce it, and I definitely see the problem of eating its own stuff.

(Joel Beasley at 01:08:28) So I see that. And I think what will happen is we talk about this problem, but where it becomes important is when we're actually trying to solve something and then the dog has a finger.

(Trevor at 01:08:40) Yeah. Right.

(Joel Beasley at 01:08:41) So when we get down to that level, you could—you can fill it up. It can think a dog has a finger or whatever.

(Trevor at 01:08:47) Sure.

(Joel Beasley at 01:08:47) Great. The general consciousness of the GPT thinks the dog has a finger because it's self-referenced and Midjourneyed up and everything. Right? Yeah.

(Joel Beasley at 01:08:56) And but then I'm actually needing this dog entity. So I'm gonna go get my dog data pack.

(Trevor at 01:09:02) Right.

(Joel Beasley at 01:09:03) Yeah. Yeah. From Duck Dynasty or whatever.

(Trevor at 01:09:04) Right? And basically fine tunes it.

(Joel Beasley at 01:09:06) Right. It—that's exactly it. Yeah. And then you're gonna have these creators that, this is my—similar to how you see it in the financial services space, where you can, like, eToro. You can copy a trader.

(Joel Beasley at 01:09:18) Right. They find the trader. You can copy them. And then you could have your data pack, and you're known for your data pack. And then people—you will then license that interesting.

(Joel Beasley at 01:09:27) As the training set. I think that's the future of where we're gonna be headed.

(Trevor at 01:09:33) I think that's fascinating because, I mean, again, that's kind of like even with our API that we provide to companies that are training models, it's a similar thing for sure.

(Joel Beasley at 01:09:40) But we gotta do a shout out. You gotta go sign up for Soundstripe. What's the call to action?

(Trevor at 01:09:44) Yeah. Yeah. Check it out. It's soundstripe.com. Once you go to the website, we have music for literally any creative use.

(Trevor at 01:09:51) So if you're making videos, podcasts, whatever, we have stuff for you. So check it out.

(Joel Beasley at 01:09:56) Thank you so much for listening. And if you found this episode useful, please share it with a friend or colleague who you think would get value from it. And if you have topics that you would like to hear discussed on the podcast, either add me on LinkedIn, or send me an email [email protected].

(Joel Beasley at 01:10:14) Every time I get an email or LinkedIn message, it absolutely makes my day and inspires me to keep going.