Sharp Tech with Ben Thompson - What AI Is Today and What It Might Be Tomorrow

Episode Date: September 13, 2022

New AI image generators hint at a decentralized future, newspapers vs. the internet as an analogy for what’s next, potential implications for white collar workers, and how Steve Ballmer can help Apple.

Hello and welcome to Sharp Tech. I'm Andrew Sharp and on the other line, Ben Thompson. Ben, how you doing? I'm doing good. I'm excited. I'm ready to go. Hey, let's launch into it. We made it to launch week.

It's finally time to rip off the packaging. See how this CD sounds. Let the world hear it. It's going to be awesome. I'm excited as well. I like the illusion to CDs that sounds like foreshadowing. So, hey, let's talk about AI. Yeah, let's talk about tech from 1997 here on our tech podcast.

For part one today, I do want to talk about AI. If you want to be the expert on this podcast, it sounds like exactly the time there are we can go to. Exactly. Take it back to the Discman era when I'm most comfortable. You wrote about AI earlier this week. Can I tell you why I was particularly excited about doing this topic on the show? Absolutely. Because I feel like AI has been a buzzword over the past 10 years or so in the United States, especially over the past five years. And honestly, a huge portion of society still has no idea what AI actually is, what it could change, what kind of tradeoffs it presents. And I mean, we could put like 80% of U.S. Congress in that category. Maybe 80 is low, actually. Most of my friends. I was going to say five years feels generous.

Like they've been talking about AI for 30, 35, 40 years. And it's, I don't think we're still at the conceptions of what AI was thought to be in like 1980 or 1970 or whatever the term was coined. But there is at least to your point something substantial now to talk about, which really hasn't been the case for for all of techists. And hasn't stopped people from casting aspersions on AI and or like putting. unreasonable expectations on what might be achievable with AI, like going back to the 80s. I mean, virtual reality has been the future for like 50 years now. So let's start at the very beginning of your piece. You began by explaining that the impact of advanced AI might be similar to what we saw when the internet disrupted all the structural advantages that newspapers enjoyed.

Is that an accurate summary of your, the first, section of your piece? I think so. I mean, it's difficult to compare anything to the impact of the internet. Like, like the impact of the internet I've always compared to the impact of the printing press, which, you know, you know, just had a little small impact on Europe, completely transformed. You know, it was upstream of the Westphalia nation structure, sort of the downfall, the influence of the Catholic Church.

I mean, just small stuff like that. So, but I think that that speaks to the internet's impact. I honestly think we're, you know, you look around. Everyone's like, you know, what's going on with society, you know, things are crazy. Like, this is all downstream from the internet. And we're only starting to see the disruptions and changes sort of in the long run. So all that to say, I'm hesitant to to compare anything to that. Yeah.

Like that's a media basically. So I thought there was there was some useful sort of takeaways. And this gets to your point before about what is AI. And AI, the conception of it is what we call general AI. which is sort of like a thinking machine, right? That thinks like a human. It takes initiative like a human. It does these sorts of things.

And that's not where we're at. Where we're at, where we've been at for a while in broad strokes is probably better referred to as machine learning. And basically the idea is a, you know, traditional computer, you could argue like algorithms are in AI, right? Like the funny thing what AI is everything is AI until it's actually invented. and then it's no longer AI, it's just an algorithm. And, you know, that's sort of been the pattern over the last 30, 40 years.

And, like, imagine, imagine an AI that goes through and picks out the post that you want to see, just from your friends and family. And it's like, wow, that would be incredible. And then that's Facebook today. It's like, oh, that's just the Facebook algorithm. It's nothing special. Yeah. And, you know, or imagine, imagine you could just type something in a box. It would go across all the internet and find out exactly what you want.

Would that be amazing? And then out of that's just Google search, right? All these things do use and rest on determinative algorithms where you actually said, look at this, weigh this, do that, and then run a calculation and then surface something up. The difference with machine learning is you sort of have this corpus of data and you sort of tell the computer, look, I want you to solve this problem. Here's a whole bunch of data. You figure it out.

And so chess is like sort of an example where the first chess computers, you actually, we're actually all programming deterministically. Like if this happens, do this, if this happens, do that. Where chess got really good or where things like Go became possible. I was like, oh, go is much more harder than chess. There's an infinite number of sort of moves. Well, when you get in a situation where a computer can actually, like computers are very dumb. Like, they only understand ones and zeros.

Like every single bit of computing is only possible to the extent it's reducible to sort of ones and zeros. but they're unbelievably fast and they're getting faster. And the whole concept of Moore's law is about computers getting faster. So they can calculate those ones and zeros so much more quickly. And so any job that entails doing a lot of work, even if that work is relatively stupid, is lends itself to computers. So you present a computer with this massive problem space, all this data, and you give them a target and say, look, you figure it out.

And they will literally stupidly try out. everything, including sort of possibilities that no human would ever think about because it doesn't make any sense. Like humans, we sort of like, we use heuristics and we anticipate. Like, you know, you hear about things like your eyes can perceive things that aren't there because we fill it all the missing details. Like that's how the human mind works in general. Computers don't do that.

They just literally try everything and they sort of work through it and they iterate and they get closer and closer. And once they've done that, though, they've developed a sort of method for figuring out the answer and you can sort of apply that to different pieces. And that's called machine learning. And that is what AI is today. Now, the reason why you've heard about more over the last five years is there's been techniques and possibilities that have been huge leaps forward over that time period.

There's these concept, there's these, the technique called Transformers. That was a big deal. There's this, uh, AI's increasingly machine learning that can operate across dirty data, as opposed to like clean data where everything is perfectly categorized. You just throw a bunch of stuff at it and sort of figures it out. And so all those have made these applications of machine learning much more tangible and real and closer to usefulness from a normal person's perspective than before. But it's not like the AI we imagined. And you can imagine like we're in a very brief window where we can call stuff like Dolly or Mid Journey, this sort of image generation.

we can call it AI generated art because very soon it's going to be, oh, no, that's just algorithm generated art. AI is still coming in a few years. Yeah, well, let's ground people in the specific tech you were writing about in this most recent article. What is mid-jurney and why do you think it's significant in the evolution of the AI story here? There's two angles to this. The first one is sort of like big picture, how does this image generation stuff work?

And I think that's actually a really useful thing to understand. Because it actually, it gives you a good mental model for how machine warning works. So you take an image. You say this is an image of a person playing basketball. And there's fans in the stands and they're wearing, you get super detailed sort of explanation. Then you take that image and you layer noise on it, right? Like, like Gassian noise.

It is totally random, like just like dots and stuff like that all over it. And you keep and you layer on the noise. until the image is gone. It's just random noise. Now, because you know what that image used to be, you have a target for the computer to search for. So it's like, look, here's a piece of randomness. Here's what we want to be the end result.

Okay. Can you undo this randomness into something tangible, and this is what that tangible piece should look like? And what happens is the machine develops, like, its own sort of heuristics, its own sort of algorithms for getting from random noise to a description that that is, and the photo matches the description. Once you've done that, now you can tell the computer anything.

It doesn't have to be the image that it trained on. It could be, oh, hey, now generate a football player in the stands or generate a car driving down the road. And there's a second part, which I'm kind of skipping over, which is where you have this entire huge text-based model and all these associations with images. so it knows what things are. That's what I was going to ask about on behalf of our elected representatives who are confused by all of this.

So is the description of the stock photos that these tools are drawing from? Is that auto-generated as well? Like, is there something that has the ability to say this is a car? There's all these huge large language model things. And a big part of this is a lot of this part is really just going across the internet, finding images, seeing text around it, and building associations. And you just dump in massive amounts of data, and it starts to understand this description means this sort of thing.

And so it just needs to know. And so when you get to this image generation piece, you're not necessarily seeking out a specific image. In this case, you just want to generate an image of a football. So as long as it knows what a football is or a football player is, or it has a sense of like the style, I want a photorealistic one, I want a Pixar style one, I want whatever it might be. it can apply that understanding to these heuristics it's developed operating on this random noise and you give it a piece of random noise and out of that it pulls a picture that matches your description.

It's pretty wild stuff if you think about it. But there is, but you can, if you understand the pieces of where it learned that sort of thing, it's like, well, it's doing the same thing as pulling out the predefined picture from the images. it's just now doing it on sort of any arbitrary sort of input. So that's part one. That's sort of what's going on with this image generation stuff. And just to clarify for people, what's happening there is you plug in the image paperboy. You write paper boy into the tool.

And it generates an original image of a paper boy. It's not like Google images or something. And that's leading your article because you started out as a paper boy in Wisconsin. however many years ago. And so basically you can plug in any text and it will come back with an original image that the tool itself creates. I just want to make sure everyone gets that. Well, it original is a funny word here, right? Because all of these inputs of this scouring the internet and pulling in like just tons and like billions of images and data and building associations and all these sorts of things.

by definition, it is based on other content that already exists. And so is that original? I would argue absolutely yes. I mean, everything is a remix, right? Like, you know, all of society, all civilization is sort of based on what came before. Things are in the air. These new images look nothing like whatever went into this training set or data. There are people that try to quibble about this.

But to my mind, it's no difference. than, you know, me writing an essay. Yeah. And I've been reading and learning about tech for, you know, 30 years. Our entire body of law in the United States is a remix of the English common law like 400 years ago. So I'm for it. I think everybody's been remixing each other for generations. It's like the canonical example of a transformative work.

Right. Like this is, it's not like a variation on it. Like I'm going to deconstruct this image. It's literally a completely new image. Yeah. Yeah. And I guess the competing concern there would be that if you're drawing from the existing work of artists who are not being compensated, then I understand why people would be upset about it. But I also tend to be more sympathetic to the argument that, look, this is a totally transformative work.

And that sort of mitigates some of the copyright concerns. Well, and what artist is not drawing from other artists that came before? Like, to what extent is this different than an artist who comes up, you know, learning to copy different masters of the past, studies art history, and then comes out with a completely new approach and interpretation to art. Are they drawing, are they copying what came before? Like, no one would say so. Yeah. I mean, it's a total gray area because also music. It's not great.

I think it's great. I don't know, man. music, you copy like one piece of a song and suddenly Taylor Swift owes royalties to whoever she grabbed it from. So it's tricky. But again, I think this is the art that's in your most recent article is all pretty striking and unique. And so I don't know what they were drawing from, but I'm sure it was something completely different than what the final product was. Yeah. I mean, I think the music mixing or sampling is, I think, a pretty interesting consideration here. I think that's more akin to a collage in some respects because you're literally taking

like fully formed pieces of music. That said, and this is perhaps my personal bias here, I think the music sampling laws are ridiculous, right? Like the fact that like it's hard to understand how it's perfectly legal to take a snippet of text and use it in an article and, like, quote it. and refer to it. Yeah. I mean,

I guess you can't straight up plagiarize it. That is a sort of violation. But regardless, I think the point here is these aren't collages. And it's impossible to trace the input of what went into any image

from the, from whatever was the original, because the originals are literally like millions or billions of images. Yes. I agree with you on the specific point. And generally, if we're talking music,

I'm a big fan of the samples. And so I support anything. sort of regulation that creates room for more samples. We can take it back to 1997, Puff Daddy, the king of samples. I'm all about it. But to take it back to where your piece began, the newspaper comparison, why is that relevant to the evolution of this technology? So newspapers and newspaper editors and journalists, you know, thought and believed and

had themselves on the back for, you know, many, many years that we're doing a great service. People, you know, pay for our product that we work so hard to produce. Advertisers want to be associated with us because, you know, we deliver this great product. And I think the internet was a bit of a rude awakening in this regard. And the internet had a few different impacts. Number one was it exposed that actually the reason why newspapers made money is because they had geographic monopolies. It was the only place to get sort of that stuff in your area.

and, you know, that monopoly rested on, and this was the paper boy analogy, it rested on controlling infrastructure. Just like cable companies rely on having wires or telephone companies, newspapers had delivery boys. They had delivery trucks. They had, you know, printing presses. They had ad sales teams. They have all these pieces. And they were the only game in town. And because the market was relatively small, like, you know, say I grew up near Madison, Wisconsin, like who's going to come in?

and really, you know, go against that because it's not worth the reward. You're like every advertiser is just going to stick with what already works, what already has distribution. And so they control distribution. That's sort of the key point. The internet comes along and suddenly everyone has free distribution. And by distribution, I mean very narrowly the ability to deliver their product to people's eyeballs without paying any money.

That's the internet. And the implication, though, on one I think, oh, great, now we don't have to pay the paperboy, right? But the problem is the Wisconsin State Journal doesn't need to pay the paper boy, but also the Milwaukee Journal Sentinel doesn't need to pay the paper boy. And the New York Times doesn't need to pay the paper boy. Someone can be sitting in Madison and get access to every single newspaper in the world and they can pick and choose what they want to pay attention to. And so you have this decrease in distribution costs is directly correlated to this massive increase in competition. because now you're competing against everyone and everything. So that's that's problem number one. Problem number two is once people can look at and get access to sort of everything, then the question becomes like, why would you, like that applies to advertisers as well. Advertisers don't necessarily need to go to the local newspaper.

They can go to whoever, because they're not interested in, they don't care about the editorial. They actually want access to the end user. Like newspapers provided that. And now, well, maybe this ad now will give me an access to the end user. Or over time, maybe we can end up on a platform like Facebook that, you know, you know, understands exactly who we're looking at. The people are in there with an experience instead of ads being pasted along the side. Now the ad is native.

You're scrolling through this feed. Some of them are ads. Some of them are not ads. It's a way more immersive sort of experience. And suddenly that sort of, you know, that sort of moneymaker went away. And then the third thing is, you know, once news is out there, just broadly speaking, it has no more economic value. Like if you hear something happen, if there's a big scoop, like that scoop's super valuable up until the moment it's published. And then it's spread far and wide.

It's on Twitter. It's like it's almost like these drawings, right? Like it's in the air. Like there's no escaping it. You and I can talk about scoop X, Y, Z and the newspaper that invested all the money. to do this investigation, they can't come bust our podcast, say, hey, that's our idea.

You can't talk about that, right? And so you have all these confluence of events that really brought home the point that the way newspapers actually made money was by owning the means by which news was delivered. It wasn't by producing news. News didn't actually have economic value. It has value, but not economic value.

And I think the analogy, we talked about CDs at the beginning, it turned out you made money not by selling music, you made money by selling plastic discs. It was again by controlling distribution. And the music industry, Spotify dragged them kicking and screaming into this new reality where it used to be,

we had limited access to information, we had limited access to music. So in any market where there's scarcity, controlling distribution is the most powerful place to be where you can make the most money. The internet is a world of abundance. You have access to everything. And so what Spotify sells is convenience.

You can still go on like pirate sites and download whatever music you want. It's much, much harder these days. It's gotten so hard to pirate music. And this is me like five years ago. I had to make a call when I turned 30. Like I'm not going to do this anymore. And I'll just pay the fee because it's just not worth it to like trawl the dark web and try to steal music these days.

And Spotify does make it very convenient. convenient. I recently subscribed. It's very exciting. I will welcome. Yeah, welcome to 2012, but I think the term we use, the term of art is fell off the back of a truck, not I stole music. But so just to, I want to keep you out of legal trouble on this podcast. But, but that's what they're selling though is convenience. You can get the stuff elsewhere, but it's convenience. And it's leveraging and leaning into abundance. It's like, well, if the internet makes everything available, then let's make it super easy to access everything. available. And the real value on the internet, given this, we have access to so much stuff, are the companies that provide discovery that help you sort through this deluge of content. And those companies that sort of decide what you see, those are the companies that I call arrogators. The companies like Google, like Facebook, like Spotify, is trying to work to be this,

because consumers go to them because they're overwhelmed now. Now it's the opposite. It's not like I'm trying to find something. It's like there's so much stuff to listen to. Take Spotify as an example. How do you even start with all the music on there? Spotify, well, here's Discovery Playlists. Like this is, we've tuned this to your likes. We're going to use new songs. We think you're going to like it.

And now part of Spotify's evolving business model is they have playlists where it's like the old radio payola sort of thing. Like you pay and you get your song on the playlist. And that like that's super valuable because users trust Spotify or rely on Spotify. Well, it's true across the news market. place these days also. Like news is actually a great analogy here because like the way it's evolved is super interesting because it started out where it was a really good thing that there was less gatekeeping and like one of the byproducts of all these different newspapers enjoying a

monopoly is that they could all get pretty boring and unimaginative and didn't take very many risks. And like in the early aughts, it was like a boom. in sports writing and blogging and everything else. Like there was just a much wider selection. I say sports writing because that's what I was reading at 17 years old. But like it was a much better time to be a sports fan because there was so much selection that you could pick and choose what you wanted.

And then we hit a point in this last decade where there was just so much drek out there that it's like sorting through it became its own little task. And it all became pretty exhausting. And I started to have a little bit more money. And so I started to like freely subscribe because I just, if there's somebody I trust, I'll just go to that person and get my news there. And I now subscribe to like 10 different news sources because it's just easier than trying to navigate the like wild, wild west that we now all inhabit together in our

modern media marketplace. And, and I wonder. So is that way? I just want to jump out in there because I mentioned that that you're selling convenience, but you just added another piece which you're also selling trust and you're selling sort of a it's a lot of work

like to do your own research to you to use the totally what I believe we refer to as a term of art, right? And so you rely on someone because you trust their point of view. They've proven themselves. They have a reputation. And that's worth paying for.

And it really shows like what you sell, what is valuable sort of on the internet is really fundamentally different than what was valuable in the old world. It was like controlling distribution. You didn't have to, if you're Comcast, you don't need to be reliable. You didn't need to be trustworthy. You could be a total asshole. Because it didn't matter.

You own the wire in the ground. And this is where I do get frustrated at people that, especially sort of the anti-monopoly crusaders, that they just want to take the old frameworks and apply them to companies like Google or Facebook. I'm not saying these companies don't have massive power. They absolutely do. That power, though, it comes from consumers relying on them and choosing to use them. Google likes to say the competition is only a click away, and that makes people so angry. And I think it makes people so angry because it's true.

You really can go next door and use Bing. You can go next door and use DachGo. You can go to Yelp for restaurant reviews. And Yelp's mad. They love to complain because they used to get free rides on Google, and now Google has their own sort of Google local. It's like, well, hey, just get people to use the Elp app. Like, oh, but that's hard. That's expensive.

It's like, well, welcome to competition, like, which the internet is unbelievable amounts of competition where you win, not by restricting people's choices, but by having people willingly choose to follow you and trust you. And that's valuable. And you can charge for it. I've been charging for it for 10 years. Right. Well, so then as we look ahead here, how does all of this relate to what we might

C with AI over the next five to 10, a bit of a digression. That's all right. I love big picture thoughts about music sampling and or the evolution of the internet over the last 25 years. But how does AI fit into the story? Okay. So number one, why did I start out with newspapers? What's really fascinating about these models, first off, and I think we really saw this,

you know, Dali came out a couple years ago, then Dali 2 this year, I think was a real sort mind-bullying experience for folks is we've seen this sort of breaking apart of how do you convey ideas, right? Writing was a big deal, right? The printing press was a big deal. The internet is a big deal. But in all these cases, like they've undone a bottleneck, right? What you could write, you could undo the bottleneck of time.

Like the speaker and the listener used to have to be in the same place at the same time. Now they could be in different places or in different times. You know, the printing press sort of limited how far. that reach could be, now it made it infinite, which had an important effect of making it much more viable for more creators, more authors. And this is an important point. Every single time we remove one of these bottlenecks and make stuff more possible and more accessible, it does hurt the people who had exclusive rights or the exclusive ability to do that previously, right? You know, when only monks in castles could actually undertake the arduous work of copying by hand,

You know, then whoever controlled that had a lot of power. Once you could do a printing press, you could distribute stuff widely, suddenly the amount of people who could produce who could write stuff and not just write it, but also have it, you know, spread it around and get it consumed, increased a lot. And that's where you go back to my references to the fundamental changes in Europe and things on those lines that followed the printing press. So all these have big changes.

Where AI comes in is up till now, even with, the internet where you can spread whatever you write to anyone or you make a drawing put it online literally you know billions of people can see it for free uh is you still had to actually write you still had to actually draw and those are skills and if you think about it you're doing two distinct things when you make an original drawing number one you're coming up with the idea for the drawing and number two you're actually substantial that idea. You're actually putting it onto paper or onto pixels or whatever, whatever it might be. And what these AI models are doing is breaking that apart, where the ideation is a distinct process from the substantiation. And so I could never draw, you know, this sort of paperboy that you keep referring to. But I can have the idea. Well, it'd be cool to have an illustration of a paperboy, this particular style, you know, X, Y, Z. And then the. the AI, which again in a few months we're not going to call AI anymore, is then actually

substantiated it for me. And this is a, you can see how it fits in this process, even though it seems mind-blowing, there is this still going to be real value in actual ideation. Computers, despite having this amazing capability, they can't think. They're not, they're not creative. They're creative in the how is this substantiated sort of way, but the actual prompt and, you know, newness does still come from humans. But the possibilities are huge. Now, very bad if you're like a paid illustrator, particularly if your work is just mostly like churning stuff out, relatively mediocre. But hey, you have the ability to draw. So you have, you know, you have a job. Not going to be great. Just like it wasn't great for sort of relatively mediocre newspapers that

had a geographic monopoly. Right. If you're highly differentiated, this is like a superpower. Like, you know, I was working with an illustrator. Actually, we were working on this for the Sharp Tech branding. And one of our concepts that we didn't end up using, though, he actually started with Dolly.

And he did like hundreds of iterations of this sort of idea to give sort of like, to give him an inspiration about where to go. And then he actually drew the like what was the final version, which we almost used. But that's an idea of leveraging this. And it's still going to be messy. Yeah. And people need to realize how intricate some of the prompts have become already.

Like you entered, I don't know exactly what you entered to get the paperboy image, but it was pretty basic. Pretty basic. Right. Whereas like you go into the mid-journey discord. And it's cool the way they've set this up where it's all connected to a community. and people will have like several lines of instructions for the tool and it gets pretty intricate and it comes back with images that are in turn more precisely tailored to what the person is asking for. So the ability to meet that person's needs is already much further along than I would have imagined.

Right. There's already like a completely new skill set being born, which is like a prompt engineer, right? Like you know the exact sort of terms to put in to give what you. And the community is in there, like, giving each other tips and egging each other on. So as the AI gets better, the community is also getting better at using it. And it's kind of a cool little ecosystem. If anybody wants to go check out the mid-journey Discord. Well, this is the second point, which is, so opening eye comes out with Dolly 2 earlier this year, which I think was a big aha moment.

I wrote a piece about that at the time about how this ability to generate content is going to be important for like metaverses and sort of in the long run. But I had it in my head that AI by default is going to be super centralized just because this process I've been describing entails massive amounts of compute power. And this, you know, and who can actually afford that. And again, previously you really had to use super clean data. Like it was super well described. This is, you know, this picture with all these sort of inputs about what it is.

And in that world, it was sort of a. bit of a worry because like, well, Google's going to be even more dominant. Facebook's going to be even more dominant. Because they already have all that categorized, correct? Yeah, and they just have the capabilities and the resources to take care of that. And they already have massive amounts of computing power. And like Facebook has this entire corpus that's exclusive to itself, like Waldgarden paying off, you know. And so that was my assumption around AI.

The first sort of upset of that was Mid Journey, which came out this summer. And Mid Journey, they haven't disclosed exactly where they got all. other data. I think it's probably safe to assume they just scrape the internet. The internet's open. Go out there, grab stuff everywhere. Another legal gray area. Yeah. And then it's free to use, right? Now, it's not actually free. They're like, they need a lot of funding because this stuff actually costs, to actually generate the stuff still does cost money. But this idea that number one, it's a small team as far as I know. I think it's only like 10 people. They're actually doing this. Number two, this idea that as these models have gotten better,

the quantity of input has matters more than the quality. And you can actually generate something pretty compelling just by, again, scraping the internet, to going around just grabbing data everywhere. You don't need these super clean, neat inputs. And scraping the internet is just a compute job. You make a computer, again, computers are dumb, but they do stuff very fast. You say, go out there, go to every single website in the world, pull everything in, and pull in all the tax point, all the images, build your associations, do these sorts of things.

And it's like, whoa, suddenly this is much. more accessible and possible to build, you don't need an AI level, open AI level sort of infrastructure or Google sort of level infrastructure. Then the last shoot of drop, or the most previous year to drop was called stable diffusion. Stable diffusion is open source. Like someone did this work, built these models. You can download it to your computer and you can generate the images locally. You don't even need to go to like a cloud server with all these things. If you have like just a relatively eye and graphics card on your laptop.

I have a gaming laptop here. I can generate images. Now, stable diffuses images are the worst of the three. Open AIs are the best, or the most accurate. Mid-Journey has a cool aesthetic, but you can't quite get the realism

that you can get with Dolly. Like, Dolly can generate images that look like the real, like they're photographs. Yeah, the Dolly images are a little bit creepy, just for the record, whereas the Mid-Journey...

You do get to the Candy Valley sometimes. Exactly. The aesthetic from Mid-Journey is much cooler to be than what Dali is putting together. It's objectively impressive, but also creepy. Yeah, that's very fair. Stable diffusions, they're not that great. But you can run it on your own computer.

That was, for me, a mind-blowing moment. And it really changed some of my assumptions around what AIS and what it will be. I assumed it would be super centralized. And, you know, it's kind of like the Internet. Like, everyone thought the Internet would be super decentralized. Like, strategically kind of made. its bones by arguing that actually no, the internet leaves the centralization in the way we talked about earlier because discovery matters so much in a world of abundance.

So the discovery engines, the aggregators become super powerful. But what they don't do is close the internet, right? Because they're not controlling the pipes, it just consumers are going there. And this reality where the internet is still open, it's hard to make money, but there's tons of content and things everywhere. means that there's actually this opportunity for AI to be open to. Now, there's a good chance that there will be some dominant players in AI because they'll have the huge databases,

they'll have the huge resources, they'll have the huge compute. People, like, having cleaner data is still better than having dirtier data. You know, they'll have the ability and they'll have this iteration function where more people use their service, so they get more feedback on what works or what doesn't, which makes their models better. Like, that's one of the reasons why Google Search is sort of, you know,

hard to challenge. but there will also be the stable diffusions. There will also be the sort of startups that come along like a mid-journey and create these other possibilities. And to me, that's really exciting. Like, like, AI was, I didn't write a lot about it for a long time because, number one, it didn't seem applicable, you know, beyond things like search and Facebook, you know, feed algorithms and things on those lines.

Right. But two, this worry is like, yeah, it's going to be centralized. Like, what else is there to say? And my big wake-up is that maybe that's not the case. And I actually feel like we're on a paradigm shift. Like this is like smartphone level. Like the whole next generation of companies are going to be built on this stuff. And it's very exciting.

Yeah. Well, and you can see with stable diffusion or mid-jurney, there's a flicker of what a democratized ecosystem looks like right now. And even if stable diffusion sucks, I can see why it's a big deal. relative to like the dominant players that we expected to own this. It's going down the same road that Dolly. I mean, it's like it's behind, but it's headed in the same direction. Exactly. And that's pretty significant.

But as far as heading in any direction, and I feel bad because all of the congressmen have probably bailed by this point in the podcast. But you talk about this changing the world and being a potentially seismic event, maybe not on par with the printing press and the internet, but like a big deal. And I mean, what's hard for me is mid-Journey and Dali are both fun to play around with and cool for what they are. And like I said, like I'm kind of shocked at how far along the tech already is. But when we envision this technology actually changing the way normal people work and live, what kind of tasks are we talking about as far as disruption is concerned? No, it's a good question.

I mean, the image stuff is so compelling because it's very visual. it's easy to sort of like lock on to. Right. The tech stuff is actually probably going to be a bigger deal, at least initially to start. So one of the things, I think one of the really interesting applications that you're seeing is, is like coding with you, like programming. And like GitHub has something called co-pilot, a REPlet, which is an online cloud ID, just watched something I think is called Ghostwriter.

And this is actually a good example because people think, people who don't program think that coding is just knowing the coding language, right? Do you know Java? Do you know Swift? Do you know JavaScript? Do you know Python, whatever it might be? A lot of energy trains and working until like 6 a.m. randomly, like 36 hour sessions?

It's actually interesting to think about why do those 36 hour sessions happen? And the reason is because when your program, when you're writing a program, you're building the logical infrastructure. Again, these are all ones and zero. So computers are very logical. Programs have to be very logical. You build this up in your head about how all these pieces work. And you're almost in a trance because you have this entire infrastructure built up in your head.

And the act of coding is just getting that infrastructure in your head onto, into text. Because so the computers can understand the text. It's actually not dissimilar to writing, right? Like, I mean, I get in this translike state when I'm writing because I have the whole structure of the article in my head. I have all the different pieces, the digressions I want to go into, and I just have to walk in and get that down. And I actually, it's like, I'm a thing with my family. Like, they know what I'm writing. I'm like, like, maybe I'll come out to eat, but I'm like, I'm like comatose.

Like I can barely converse because if I lose that, if I fall out of that, it's like two hours to get back into it. Right? Because you have to like, you have to rebuild the structure in your head and then put it on a paper. This is, I'm talking about two pieces. there's the creation piece and there's a substantiation piece. And so you think about it like the actual creating the logic of a program is a distinct process from actually putting that down into text onto a thing. And so if you think about the idea, just like ideating an image is different than actually

drawing the image. Now, I get curious to be like, well, no, part of it is actually drawing it. I relate. Like I come up with ideas as I'm writing. I work on arguments as I'm right. There definitely is a connection. here, but it's not a complete overlap. Like there's a Venn diagram, but there's still two different circles.

And so you can think about something where if you, the real skill in programming is being able to construct that logic. And then why wouldn't the computer help you? So much the programming is mindless. It's just like boilerplate stuff that you have to put in. And you have to substantiate you on different pieces. You have to make these API calls and fill in these different bits. And why wouldn't you have a computer do that if you could?

You could be so much more productive. You could actually make fewer mistakes. And so I think this arena has massive potential to sort of fill these bits in. Another one, I think, applies to you. You went to law school. You were a lawyer. How much of law work is busy work. It's just like boilerplate.

It's filling in all the different sorts of pieces. I mean, look, it's not lost on lawyers that like 60% of the job could one day be performed by AI and render a lot of lawyers obsolete. I mean, like, I was joking about that when I was in law school. And even when I was practicing, after about a year, Westlaw introduced some new tools that did rely on machine learning and were really, really cool. Like, you can now upload a brief to Westlaw, which is a website that almost every lawyer uses. And Westlaw will check all the citations for you. But it will also look at the cases you're citing and then suggest other cases you can include in the brief that support the same legal principles. And all of that, it takes like the most annoying part of being a lawyer and makes it much, much easier.

And so I don't know that it's going to like replace law, replace like the attorneys altogether because I think you still need to know like what type of argument you want to make and be able to argue. it in open court, but I can see it transforming the profession. And like 20 years from now, I'm sure a lot of what lawyers are currently doing will be automated. Yeah, the only thing I disagree with you is 20 years from now. I think this is probably going to, like, this is where the societal changing stuff happens. Because like, you just made the, you just made the key point. Good lawyers are, are going to be not just as valuable than ever. They're going to be more valuable than ever.

because their capabilities and efficiencies are going to be supercharged by having this, by having this sort of extra abilities, their productivity can go up by a huge amount. But you still need to understand the case at hand, figure out what is applicable, and know how to direct the AI to do the boilerplate, to dig up all the case law, to do the sort of pieces that go into it. The folks that are going to suffer are the ones that aren't actually really good lawyers. They're just, like, good at researching stuff and, like, digging stuff. up and coming up with the arguments because that's all going to go away just like the sort of mid-tier graphical artists who doesn't really have any original ideas but can take a brief and you know dump something out you know right or and I'm not saying this is I don't want to

like dance on anyone's grave or anything on those lines but there is a degree of an inevitability to this we we've seen think about think about the internet like think about writing think about newspapers yeah like you can read unbelievably high quality stuff to a degree you couldn't previously. Now, there's a lot of, you said it before, there's a lot of Drek because Drek's cheap to produce, but because we have the discovery engines, you can actually just get the good stuff.

And that the absolute amount of that good stuff is way higher than it was previously. And so there's some folks that do very, very well because they have high quality stuff, they have good reputations, and they charge a fair price for it. You have a company at the New York Times that has an entire brand,

as a full experience, you know, is just killing it. And you have lots of other folks that in retrospect didn't make money by having stuff that was so good people would seek them out. They just had paperboys that would actually put their content directly on people's doorsteps. Yeah. Yeah. Well, and I think all of this alludes to the other implication and all this. And it doesn't sound like you're pro or against any of it.

It's so like you're not like putting value judgments on. it. It's just like as we look ahead, like, this is what's going to change. And an important change to consider is that a lot of white collar stuff could be done by AI generated programs. And so like if we're hollowing out the middle white collar jobs, not blue collar jobs. Right. We're hollowing out the middle of upper class, you know, workers. And that, I'm not sure what that looks like and whether that's good or bad for society. But, um, Well, I mean, we've already seen what happens when you haul out blue collars, right? Like that's what happened with globalization, with automation, where you have this entire category of job basically go away.

And it's been a bit of upheaval, I think, is a fair way to put it. Could have used some more regulation as we underwent that transformation. Sure, but what would the regulation be? I mean, I think this is where the challenge comes in. I mean, the, yes, you could say, oh, hey, we should have had. you know, not so much offshoring or something on those lines. That would have been my suggestion, yeah. Would we, but would we regulate automation?

Like automation is maybe a better analogy in this case, right? Because I think the analogy to globalization as far as manufacturing is sort of services like Upwork or whatever where you can hire an illustrator in the Philippines or in India or something on those lines who will do, and those exist, right? Like that service, those services are very popular. You can hire someone, you give them a brief and they'll generate something for you. And so that's maybe the, that's the global. globalization analogy, which the internet makes fairly trivial.

The analogy for this AI stuff is automation, where you have machines doing stuff that people used to do. And to the degree there is onshoreing of manufacturing, it's almost all driven by automation. Now, the advantage for the highly skilled worker is there is a limit to what machines can do. Like, you know, iPhones are mostly assembled by hand. People don't realize this because actually, once you get into the actual assembly part, there's a lot of details. I mean, there's massive amounts of automation to be clear, but some of the final pieces, right? Textiles are still almost all done by hand. But that's not exactly

like a comforting sort of disclosure. I know. And ironically, if we were to onshore the iPhone assembly, that's a job that most Americans just aren't very good at. So we wouldn't really be gaining anything in that scenario. But do you see what I'm saying is as this all gets better and and automation becomes even more advanced to where you actually don't need that many humans involved in the process, I could see us hitting a point where it makes sense to put limits on what we allow companies to do. And I know people who are fans of the free market will be pissed off about it, but like we don't want, you know, 70% of the American workforce to be rendered obsolete by some of this. Well, I mean, 70% of people used to farm or even higher. I mean,

I think that like there is some degree where technological progress is going to happen whether we want it or not. The the, the, so the question of regulation is maybe less about are we going to change what is inevitable? Or are we going to try to accelerate the upside in the future? And my concern about trying to lock what, lock in whatever is there is you, all you do is then get stuck in stasis. You're not actually benefiting from what's coming forward, but you're not preserving what is in the past because it's increasingly, it's increasingly pointless.

Like, like paying people to stay on farms doesn't actually, in the long run, preserve anything. Like, now, progress is painful. Like the, like, you want to go back to the Pertic Press and go to, you know, the Industrial Revolution and go through all history. Sure. Lots of people died, right?

Like, like, it was not great. Like, and, you know, there are, massive wars and like reorganizing the world's, you know, the way people are organized. And just for the record, we're currently experiencing globally like a similar sort of upheaval. And I think 100 years down the line, people are going to look back at the beginning of this century and think, oh yeah, that's when most of society just lost its mind as a result of the internet and not really understanding how to have a healthy relationship to this technology. Right.

But the problem is that this, this, and I am a. bit of a technological determinist here, I think. You're saying the term healthy relationship as if we sort of have a choice in the matter. I mean, you could say, you know what, I'm going to be, you know, this is maybe pertinent. You're an internet writer. You know, I'm going to, it's important to preserve the structure of society. I'm not going to go right online. I'm not going to have a blog because we need newspapers, our important pillars of society, we need them to continue to exist. And like what what would that have accomplished? Yeah.

You know, and you look at the music industry. No, we're going to walk in CDs. We're going to keep it. They had to have their rear end handed them by piracy because once the capability exists, people will find a way. And the default impulse is to inertia for people who are succeeding now. Like they're not going to change unless they absolutely have to. That's just a feature of human nature. nature. We change because of pain.

Right. So there's a danger, though, in putting that into regulation. Because now that inertia has the force of law. And, you know, you think about something like this automation, but this is where stable diffusion, I think, is the really important piece here. So the government passes a loss as you can't use AI image generation. Open AI goes out of business, you know, mid-jurney goes out of business. I have stable diffusion on my laptop. It's open source. Like, are you going to come from my laptop? Are you going to take it away? And that's going to iterate.

It's going to get better. Unless you're willing to pursue China style, we control the bits that go over the internet and we're going to chase you down and throw you in jail, it's not going to work. And this is just a critique I have of so much of Western approach to the internet. If you want to go the full China route, go the full China route.

If you want to do it halfway, you're going to get the worst of all worlds. And I think that would apply in this case as well. Sometimes I do want to go the full China route and get really restricted. But just one final point here, you end your article saying TikTok, which pulls content from across its network to keep users hooked, is the apotheosis of user-generated content. Metaverses may be the apotheosis of AI-generated content. And that links to a piece you wrote about six months ago on Dali and its implication for,

the Metaverse. All that aside, I just want to tell you, one of my goals for this podcast as we get started here, we don't have to do it on this episode. I do want to work toward a better word for the future of the internet than Metaverse. Something less stupid than Metaverse is like my 12-month goal. So I hope you join me in that mission.

Well, my philosophy is that I think Metaverse is really in the long run going to be like AI, wherein we're never going to get there because we're already there. like the real thing is like the the metaverse is just the internet right but it's just like a more sort of immersive experience that's possible or you can back out and you can just use text like your computer you can go to a command line you can actually operate your computer completely via text or you can use the gooey and like have have this sort of different experience and so I think I agree with you I think I think it's kind of a dumb term but it's still useful right like like

the thing about TikTok is you know professionally generated content you think would always win Like, we have the best, like, Quibi was like, like, you know, we're going to win mobile because we're going to actually have real directors and real actors. And we're going to deliver great stuff. And the problem is that, sure, you might deliver compelling stuff. It's not going to be that much. The problem with Quibi was they never pitched a regular person who would have told them, like, why the fuck does this exist? And do you really want me to subscribe to this? Absolutely not.

I think that whatever they were pitching sounded great in like a boardroom and just never made sense or, pass the smell test to any ordinary observer. Yeah, I mean, if we could do a whole podcast like, it would be fun. But I think like, like, if you're, if you're going to take the time to watch on these professionally produced and enjoyable, then I want to enjoy it on the big screen, right? Like I, I want to, I don't want to watch like House of the Dragon on my phone.

Like I want to like like it has to be a full sort of experience or like movies. I want to go watch it in a movie theater. But the, the advantage of TikTok or YouTube is anyone can do it, right? Now, if anyone can do it, you get a. a lot of drek. Like the vast majority of content on TikTok or on YouTube or any social media network is garbage. But the absolute amount of stuff is so large that even if the good stuff is only a small percentage, that absolute number is also large. And if you can find that and surface that in a consistent way, you end up with content that is so compelling. It's already

even more compelling than the professional stuff because people think outside the box. Like, there's stuff that people think of of doing that no one would have ever green light. No one would have ever given permission to. It's completely permissionless content creation that creates all these new sort of opportunities. And so TikTok's big innovation was we're not going to surface the best content from your friends. We're going to surface the best content across all of TikTok. Who you follow on TikTok is meaningless. It's just finding what you're interested in finding the best stuff that's associated with that.

That's why it's like the apotheosis of user sharing content. It's like the ultimate YouTube is really the other example here. Where this matters for metaverses is the challenge in creating immersive experiences is as expensive. All those AAA games, people have to draw all that stuff.

People are actually creating the textures. They're creating the characters. They're creating all these pieces. And as resolutions go up and performance goes up, the detail that these need to be created in goes up as well. That's expensive. So the problem is if you're going into a virtual world where you can look around, you want it to be open, you can explore. Someone has to do that well for it to resonate with anybody who's inhabiting that world.

It's impossible. It's impossible. You'll never, it's such an overwhelming challenge that it's not even viable unless. Got a lot of help. It can be generated. That's right. So computers are dumb, but you give them a goal. They can create stuff.

And so, again, we're not at a cost place as far as production where they can do it, but the relative cost of computer generating a world is already drastically smaller than a human generating it. And this is why this is going to be a key component of a more immersive internet, which I think is the better way to put a metaverse, is when content can be generated, not just by humans, but by sort of the computers themselves. Look at that. Immersive Internet. I think that's a great start to coming up with a better.

term than Metaverse because it really does capture what everyone is describing as far as metaverses are concerned. So I appreciate you getting us started here as we work toward a better future in web discourse. We're not going to get to part two tonight. I was enjoying part one and this circuitous discussion of AI and its implications for society. I think we embedded the part two in the middle by talking about completely unethical, related things. But yeah, our goal, our goal is to have a, a, a meaty part one and a touch on part two. Can I tell you, that's a good call. One take, we were going to talk about Apple and it's shifting priorities. I do have a take. And it's that Apple needs to recruit Steve Palmer to help them reinvigorate some of these launch events that they're having out there because the event that I watch. We don't want you involved in the business or the product. We just.

watch you do special consultant for events. Exactly. Just come in, crank up the energy a little bit, work the crowd before Tim Cook grows out there. Like, it's an easy win. And I was all psyched to watch my first ever Apple event. And I just couldn't believe how depressing it was. It's like Tim Cook giving a speech on a completely empty Apple campus. And you have all these other gorgeous buildings. All of them are empty. It's, it was all very. dystopian. And I think Balmer can help solve the problem for Apple. Right. That worked in the middle of pandemic.

It sort of fit the mood. But yeah, I mean, that was honestly my first immediate takeaway, too. We need to get back to live events because this is terrible. There you go. Well, we will cover all that and more in the weeks to come. I'm very excited

to get rolling here. And listeners can email us at email at sharptech. from fm because we're going to be putting together two episodes per week. The first episode is going to be Ben and I with a two-part show that will be public. The second episode will be for Straterex subscribers or if you subscribe to this podcast individually and we're going to be answering feedback on this conversation we just had, some of the conversations we had in our back catalog, and any other questions you might have. It's going to be a little bit of an

adventure or the second show each week. Yeah, it definitely an adventure. I mean, I think it'd be interesting to talk about things that I don't write about, you like, maybe like, you know, whether it be products or things on those lines. It could also be very terrifying. So we're going to we're going to play it by year. Yes. I have no previous interactions with your listeners or reader base. So it's going to be an adventure for me as I'm going to be the one reviewing the emails. but I look forward to all of it. And for now, Ben, we will come back later this week with our first ever Q&A, our first ever mailbag episode.

So get your emails in because we're very excited to check them out. I'll talk to you soon.