Immad and Raj have a wide-ranging conversation on how AI will impact creative roles with the CEO of rapidly growing Jasper.AI, Dave Rogenmoser

And I even remember I was in this Slack group kind of early days with like everybody else that had access to GPT-3 and everyone was doing very trivial things with it. And I'm sitting there thinking, guys, we're sitting on a gold mine here. Like this is incredibly powerful and everyone is using it to like translate the Declaration of Independence into Elvish.

Hi everyone, welcome to the Curiosity podcast where we go deep on a wide variety of technical topics with the smartest leaders in the world. I'm Immad Akhund the co-founder and CEO of Mercury. Also, I'm an active investor and I've invested in more than 300 companies.

I'm Raj Suri, I'm co-founder of Lyft and Presto Automation

And Dave Rogan Moer who's the CEO and co-founder of Jasper. Dave, do you wanna to do a very short intro on like what Jasper does?

Yeah, Jasper is a company we started about two years ago. We help marketing teams write high quality, great content and do it in about half the time with the help of ai. So been riding this generative AI wave and being one of the first companies to help really productize that and take it to market to end users.

Just curious, like what uh, got you excited about this field and what will get you motivated to do what you do?

Yeah, I mean it was kind of just perfect time, perfect place, perfect team. It was about two and a half years ago we were spinning down our company proof that Ahau graciously invested in back before we really knew what it was five years ago. Liked the team a lot, but it laid off some people just cause we were trying to get profitable. Were kind of saying, Hey, we don't really want to keep pursuing this, you know, vision. And we were, it was a marketing tool that helped increase conversion rates on websites and we were trying to get to the personalization space and for a variety of reasons we decided this wasn't gonna work out. And then I'd seen G P T three, I'd seen all the like Twitter threads and just fun little demos and thought, oh that would be really great for writing marketing content.

I was teaching this course on how to do Facebook ads for B2B SaaS companies and one of the lessons, I think it was like week three was how to write a great ad and how to write great ad copy. And I'd been doing Facebook ads for years and years and you know, I taught this framework of how I wrote our ads and it was always challenging to get the people in this course. There's maybe like 12 other founders in there or CMOs and it was always challenging to get them to really follow my formula and and do it as well as me. My first thought was, oh I could just build a little tool that could like help them write their ads for this course. And like, maybe it's more than that someday, but like at the very least like that would be pretty cool and I could like sell more courses And so we just built this really quick MVP to solve that problem and it was super sketchy.

You know, you couldn't do hardly anything. I mean there's no settings you couldn't even cancel. There wasn't a cancellation button, which was like our way to minimize churn and keep it at zero , there's nowhere to go. And yeah, launched it and you know, I did some like calls with some of our coaching clients like as it was kind of just an mvp, they couldn't even like use it, it was just a screen share and people were just like cussing on the phone and just saying like, oh my gosh, how do I get this today? And for like all the years I'd ever been in business with my co-founders the previous seven years, like I'd never gotten that reaction. And so I knew pretty quickly like, this is different than anything I've ever done. And kind of quickly learned that this had the potential to be a really great company.

Dave, I'm invested in like more than 300 companies now and your story both pulling off a pivot and also your kind of growth shots after you did it. I'm not gonna give away any stats, I don't know what's public, but it was just astronomical. It was like a crazy story to see. I remember every, every month you'd send your investor update and I'd be like, you think it's gonna happen next month too? And he'd be like, , do you remember

That right. Oh I look, I I just dreaded sending those monthly investor updates and they, they they, well they frequently got less, you know, they weren't monthly because it was just like nothing is even happening here. And, and then I just think it got to the point there was probably like a six month window between like the last proof one. I kind of go dark for six months. We build Jasper and then I'm, it's the funniest like follow up one, it's like hey, we're still alive and it shows this like flat chart for three years and then just this like massive uptick on the right. So

That explosive growth is like very hard to come by. Like Roger's a co-founder of Lyft so maybe he's seen it before mm-hmm . But do you think it was something about the AI and the fact that it was like being utilized that like got people really excited about sharing it with people or was it just that it was solving this pain point that was like really painful to people?

Yeah, it was just solving a real pain point in a dramatically better and magical way. And so it worked so well in a way that was like so surprising that like everyone just told everyone about it.

It seems such rapid growth. Was there like a moment where you, like you flicked a switch, you just saw like a explosion of usage? Was there like that, you know, we've all been there as part of um, startup that we've worked at or was it just kind of like building over um, over several months?

I mean it was very quick. We had, we had this goal to reach $50,000 R by like mid-summer, it's like six months in we would reach like $50,000 r and I think we reached it two weeks in and then you know, we're like wow, this is really great. I'm like oh man, let's try to reach 200 k you know, by then. And then you know, a few weeks later we like hit 200 K r and then we did this webinar about three months in like at the end of March we were just selling one plan, so it's like a $29 a month plan. Everybody was on this $29 a month plan. There was really no expansion or anything which hadn't built any of that in. And so we built out this pro plan and hosted this webinar and I've done a bunch of webinars in the past over the years, like probably over a hundred and usually we'd have 20 people show up or you have 60 people or three people. And this one we had like 2,500 registrants and then the room was totally full with like a thousand people and I pitched this pro plan and I think we added $400,000 of R in like three days. Mm-hmm

, what was the cost of the propane?

It was like $109 a month I think. And so I had like all this crazy expansion and just like people were going nuts and that was a pretty magical moment when it was just like holy cow, this thing is like left the station and it's working and yeah just felt really good after I feel like seven or eight years of grinding away to finally hit something that you feel like you were working towards the whole time.

You mentioned I think earlier at some point that you're not AI researcher but you'd like to understand how these different technologies can be applied to different use cases. Where do you think this technology is gonna go in terms of like what use cases do you think are gonna become more and more prevalent that maybe people aren't even thinking about today? I think what you're doing now is kind of obvious, this is obviously the way of the future, but what use cases do you think are not obvious?

Yeah, I think obviously you see a lot of the generation of text and images and then you'll see you know, video become better and better and it'll kind of continue, you know, to generate great stuff. What large language models are really great at it is just understanding and so, you know, it's kind of like what we've always wanted Siri to do or Alexa and like I can just like talk to it and like regular language and get it to do stuff and I think that's the big like next leap is that these models will start being able to do more things for us outside of just generate text and images. You know, you could connect multiple apps together. You could say, you know, hey go start my car. It could actually go and do that. You could say hey, you know, write me 10 Facebook ads and like upload them to Facebook and it'll be able to do that.

And so I just think like we're gonna get to this place where they're just doing more things in the digital world. I mean at some point it'll be the physical world too that aren't confined to like a very like narrow use case or a very narrow tool and that actually like connect different tools together. It'll be across the different tools. So I still think like we're super early and people are just kind of learning like what to do with it. I think most of the conversations are still more on the research side and there's kind of just now coming online a new wave of entrepreneurs that are saying, Hey, how do we build really great products and tools with this And that's gonna be incredible.

Your use case is very fault resistant. It's like copy for marketing, which it doesn't matter if it's a little incorrect and you have a human normally like look at it afterwards. I worried that the, you know, a lot of other use cases are not so tolerant. Actually the one that you just said like switch on my cars not so bad but like people are like, oh why don't we apply this to like doctors, right? That's the kind of thing where like if you make a mistake someone could die . So uh, do you think we'll get to a position where like we can apply this to use cases that aren't so fault tolerant or is it mostly focused on like these kind of fault tolerant use cases?

I think so. Humans are humans, even doctors and all that. So even there's some level of okay, like we we're comfortable with some risk here and doctors make mistakes and there's liability insurance. It's a little bit worse if it's a computer makes a mistake and and that's what you die of as opposed to a PhD doctor but or an md. But yeah, I think like the hallucination is a big problem and that these large language models output things that aren't true and they do it so authoritatively and convincingly and you could read a whole paper and it kind of lulls you to sleep because everything you read is true and then it kind of slips in you know, one or two things that aren't true but they look pretty true. And so I've seen research and we've even experimented some things that decrease a hallucinations and definitely make a more accurate, I think at some point we might need to rethink kind of how we train the models and actually the process of even getting to where we are now in order to like really solve that. But I think there's likely gonna be, you know, human in the loop for the really high risk use cases for a long, long time. I think it's solvable.

Yeah but, and one thing you said talked about was like digital to physical interfaces, those are gonna become more lucrative to develop now that you have a really powerful digital intelligence that could drive physical use cases. Do you see a lot more investment going into that as well? I mean this is something that I work with a lot in Presto is digital to physical thinking about how digital can impact physical.

I have not seen it that much and that's somewhat surprising to me. There's definitely a lack of that. I assume that you've seen this if that's kind of what you're focusing on. It's like the physical world is like so hard to change and it's not really keeping up with technology and it's like, you know, my computer's like wholly different every single day and everything is like coming alive and it's like, you know, I go down and my toaster's the same and my microwave's the same. You know, everything's like normal life outside of your computer. So yeah, so I'm really excited but I have not seen a lot of investment kind of flowing in or like even a lot of really cool companies coming out that make that leap using generative AI to go impact physical products.

Yeah, I think it's gonna happen. There's a lot of companies and investment out of MIT in particular, you know, around these types of problems. But you're right that it's not enough. There's disproportionate. I mean that's one of the downsides of like the startup economy is that like startups like to work on things that are cheap and working in the physical world is not cheap. It's very expensive to invest in the physical world and many companies fail but at some point that becomes the bottleneck and you have to undo that bottleneck. It was a hardware innovation with the smartphone that got us to where we were. Right. And being able to deliver software to everyone's pockets. So

I mean VCs keep getting burned when they touch the physical world, right?

Yeah. Is it just not lucrative enough to really put money there?

It's hard to pull it off. I mean if you look at the top tech companies, you know, both Amazon and Apple are very heavy in the physical world so it's definitely lucrative if you can pull it off.

It's like the VCs can't do it but the big companies can, right? Like it takes an apple to figure it out for the smartphones, it took an Amazon to figure it out for cloud, right? And the VCs can't stomach it, there's too much investment. You get a magic leap type company, right? Yeah. Where you know it's very difficult for them to sustain that type of loss over many, many years and they need to keep generating hype. It's not easy

Is the way for startups to do it more to keep bringing down the cost of interacting with the physical world and you see like breakthroughs happening there.

I mean I think one of the interesting things in AI is like maybe yeah a lot of like the innovations that I've invested in and are interesting AI in robotics. So does AI reduce the cost of robotics because a lot of robotics like you're doing like kind of very point-based machine learning or well I guess we all call AI now solutions whereas like maybe with generative AI you could have like a generic AI layer that you kind of just trained to like move the robot a little bit and you don't have to build that understanding from scratch, which maybe helps.

Yeah, I've seen talk and research just around like the ability to do all these simulations. Mm-hmm virtually now, you know, I used to have to actually build out all the prototypes and test it all and now you can do you know, hundreds of thousands of simulations and iterations before you even go to the physical world and can really prove out a lot of concepts there, which would certainly lower the cost in a big way.

One thing I've wondered about is whether it's G P T or one of the other kind of APIs, like if the L L M space in general is like very accessible so you can build a startup on it very easily, does that mean that you don't really have a moat Like you know there'll be like a thousand Jaspers or a thousand other applications and and it'll be very hard to build like a very kinda sustainable large company as a application there or do you think that's not the case?

I think there's something there. I wouldn't be spending my life building on that but you know there's certainly like some risk of that And it's funny, you know I talk to, we kinda have one foot in the AI research world and like talk to a lot of folks there all the time and then I've kind of got one foot. I mean our customers don't care about that. They're just trying to solve this problem. And so I've always gotta simplify and spend a lot of time just with customers and you know, people that aren't AI experts and the AI experts always think that the model is a commodity and they're like, yeah this is commodity. There's nothing here. It's gonna be you know, all go to zero here very soon and you're gonna see, you know like they all talk like that and they're like the hard part is the product and distribution and go to market and the ui like the, that's where the value is.

And then I talk to the other people that aren't their experts in and they are like, oh yeah the products, that's easy. That's a commodity. The value is in the model and the data and all of that And and these are really smart people on both sides. And so I think it, you end up thinking the thing that you're not as good at or not an expert in is where all the value is. But it's really funny cuz both feel that very dogmatically. I think that there's lots of moats, it's very rare for a B2B software company to have really strong moats. Yeah. It doesn't have the network effects like a Facebook or Lyft or things like that. And so obviously like we've gotta innovate, we've got to to build differentiation in moats where we can find them. I don't think there's anything that we're gonna do that is going to be long-term defensible and I could just kind of rest on my laurels and like not keep executing.

You see that and you know you said there might be a thousand Jaspers, like there's probably a thousand Jaspers that launched today that we've gotta go beat and now, now fortunately like again we're really good at a lot of this and it's more to a company than just kind of building a ui. We've got great distribution, we've got a great community, we've got really great insights from talking to customers and obsessing over customers every single day where some people are not doing that. And so I think there's like a lot to build in a company that is hard to do and various people will do that to different degrees but I don't think that the models are also like these defensible places that once you build a large language model it can just last for a long time. You're seeing them come out so quickly and copy each other and yeah you're just seeing that the innovation moves so fast. It's like I think what we talk about a lot is find defensibility where we can but more than that we're trying to build a team and an engine to continue to innovate regularly so that we can always have speed to market and speed to innovation there and keep that lead.

Yeah you know I think the dirty secret amongst SaaS companies, especially amongst technical founders that we don't really talk about that much cuz we get enamored with the technology is that like a lot of the mode ends up being like brand or like your customer network or just like your referenceability among customers and word of mouth and people don't talk about that, especially amongst technical folks, it's kind of anathema in some ways. But you know I talked to um, senior person at a very large SaaS company recently and I asked them what their MO was cuz it was like a very simple product. It was literally like a type of CRM type product and they're like our motives, we get our customers together once a month with new prospects and people love to come to these dinners or events and that's what gets us into more customers and it's a simple thing and it is simple execution thing that had nothing to do with the technology that was pretty instructive as well. This is like billions of dollars in revenue a company

That's amazing. Yeah, people don't like to talk about that and and again like can not think about like my week, most of the really hard things that I'm doing are around building a team building organization like building great partnerships. You know it's like all these things that are not only product related endeavors, like these are really hard, they're really stretching me and like I'm gifted in those things and they're still incredibly hard and it's so it's like I think with any company or founding team, like there's gonna be things that are really, really hard to do that are necessary for building a great company and you've kind of gotta do it all but you certainly have to nail the product and build differentiation where you can.

I think one thing that's good about Jasper and makes it more sustainable is this kind of business use case. I think it's much harder if you're boating a consumer company where I do struggle with most of the generative AI consumer companies that feel like a little fatty. Yeah. And it's hard to see how like they charge these, like some of these profile pick ones were like charging like these insane values, which I get like it's kind of fun to do it but like it just doesn't seem like a very sustainable business on the consumer side.

I know you'd have to really build a machine that can just keep cranking out winners all the time and Yeah. You know get it blows up and you get the number one in the app store and you make a bunch of money and then before you know it, all the apps do what you do 24 hours later and you know, you've gotta kind of keep doing that. We definitely love selling the businesses. We have like a lot of consumers and like prosumers and you know, single seat users that use Jasper and that's definitely like a harder place to compete and a harder group of people to continually please over and over when the switching costs are lower and they don't mind kind of going somewhere else. But we do see, we get into teams and we get into you know, larger companies and there's a lot less of the noise and a lot more of just you building value for your customers and it's just an easier segment I think to go and penetrate.

What's the longest form content or like most complex to have a content you guys help enable Today

Our number one use case is blog posts and there's probably a lot of different examples of different kinds of blog posts and people call, you know, white papers, blog posts, things like that. But you know, I mean several thousand words, maybe four to 6,000 words would be kind of the high end window of, of the kind of content people are creating. People do use Jasper to write books and probably two weeks after we launched Jasper we had no like long form stuff we had like this like Facebook ad tool. Some guy wrote a whole book in a weekend kind of using this Facebook ad tool and he kind of figured out how to hack it, get like say what he wanted and it must have been just an absolutely miserable experience kind of using this like one sentence at a time generator and wrote this entire book and that really blew my mind and we quickly saw that people were hacking our product to write really long form stuff and finding good results with it. And so we kind of made this shift from more of the short form ad copy landing page copy type product to really long form product and yeah, people seem to really like that.

I mean this question might seem a little basic to you but I don't people actually understand it. Like how, what does it take to go from I can go to chat g PT and say write me a blog post about x versus I can go to Jasper and say write me a blog post about X. What makes it different to go to Jasper and do it versus like chat g PT directly?

Yeah, so when we started out they had an an API and NOBO was like really building on it and I even remember I was in this slack group kind of early days with like everybody else had access to G P T three and everyone was doing very trivial things with it. And I'm sitting there thinking guys we're sitting on a gold mine here. Like this is incredibly powerful and everyone is using it to like translate the Declaration of Independence into El Elvis and just like these like weird like, like trivial things and all thought that was super cool and they're all complaining about latency and you know, all these things. I was just like forget all that businesses want this and they're willing to work through the kinks to go and do that. So like early days it was a lot of like prompt engineering and it was just figuring out how do you prompt this, you know, behind the scenes how do you give it like really high quality examples and like a lot of the work was there, you know, plus there's like different kind of settings and tweaking that you can do to the API to improve the results, you know.

But it was like, I would say straightforward, not necessarily easy, you had to have like really like strong knowledge of the use case in order to get better and better results. But it was that I was like the models have gotten better, the prompting has like diminished the value of like the prompting itself has diminished cuz they're just so much better out of the box and they might not need examples to really do well. And so like we're kind of seeing, I think that layer diminish, but what you do have opportunity to do is fine tune is to take, you know, as we have more and more data, we've just seen more and more great examples of great blog posts of ad copy that people like of paragraphs that are high quality and at least for us, like outputs that our users prefer. And so now we've been able to take that and we use a variety of different models.

OpenAI is a great partner, certainly use like a ton of their stuff. Like the majority of our outputs kind of come from them. But then we also have, you know, other partners and other models that we're tweaking that we're able to co and fine tune these models. Like for Jasper specific use cases, I think this is where there's a lot of value for companies to build is open AI and chat t is gonna be inherently more generalized and it's gonna work for a broad number of use cases for a broad number of people doing all sorts of stuff. But like people come to Jasper for like a job to do and like Jasper's where people go to work. And so going back to the question about chat G P T, I mean Jasper has Jasper chat so if you kind of wanted to use a chat like interface to do that, you can certainly do the same things in Jasper as you couldn't chat G P T where we would add different tools around blog post creation.

So we also have templates that make it easier and more structured to go and create blog post headlines and outlines and things like that. We've got workflow tools where you could generate each step of the blog post kind of piece by piece and you know, intro topic and then it creates an outline and then it creates paragraphs for each outline and then it cleans 'em up and oh you can click a button, it'll, it'll create art and images for each one of the blog posts like sections in there. And so like I think there's a lot of value in like what is the job to be done of the user and how do we just work backwards from there to pull together like a variety of tools in a very like opinionated way around that use case. I think we've found that when many users have more structure, they do better when you're kind of, it's very unstructured and they look at it, it's unfamiliar interface, you know, they don't really know kind of how to get from point A to point B but many users, you know, do like that structure and so we build tooling that helps you do that faster.

Do you think personalization is important? Like you, I assume you fine tune on like a generic thing. Like you can't say like say things in the voice of MR very easily. I mean I guess I have enough tweets out there that maybe people can, but in general it's like personalizing it is still pretty tricky, right?

Yeah, it is tricky. We're making some really great progress there. I think as an industry and as Jasper, you know, it still doesn't really make sense to fine tune a model per customer in the backend of Jasper. I mean it's not like there's just like one model,

You mean like it technically doesn't make sense or it's

Yes, it technically doesn't make sense. Yeah, cost time, all of that. Like we could do it and for a super big contract we'd be like, all right, we're gonna go fine tune the whole app and every model for that use case there. But

Like how much would it cost?

It wouldn't cost that much in compute and training, it would be more collecting all of the data, our team kind of synthesizing that, cleaning that, making sure we have enough of it, making sure we have the right quality of data. Like from a customer, yeah we'd go grab all your tweets, we'd go grab, you know, whatever we kind of could, but there's like you don't actually have the right data to go and like fine tune off of for like some use cases. Mm-hmm . And I think it's more like operationally and just like the complexity of now kind of having 90 separate models for this one account that would make it like very complex to like the operationalize and the backend there. But I think we'll get there and what we are doing more, we just launched Jasper brand voice and we're like bedding that with some different companies where you can upload information about your company, about your style, about your tone, about your brand voice, about your product descriptions and all of that. And then as you're writing we can go retrieve that. And again, it's not a fine-tuned model but it is retrieving the right information at the right time about your company that does provide a very unique experience and it functionally feels like very personalized to you and it is personalized to you but it's, I think it's in a different way than like most people think that it would happen where they would think it would come from like custom models.

We were saying earlier about people creating like books from your system. It was pretty fascinating. I mean do you think like three to five years from now, like basically every newspaper, every book is made in a very different way where the news is is made with just a couple prompts, here's some updates or maybe it comes from like, you know, a Reuters feed or something like that and it gets adapted to the New York Times by AI and then a New York Times editor just reviews it and it goes into goes into the paper. I mean do you think that's the way like books and all like articles are will be, will be published in the not too distant future?

I don't even think it's years out. I think it's probably way closer than that. And I think maybe the natural questions like well what happens to creativity if like if kind of that's how writing happens or books happen. What we have found with our customers is that the like spectrum of creativity doesn't go away and it's not like oh everyone gets s smushed into if you use Jasper. Like you know, you're all kind of the same amount of creativity and all that but like our power users are like coming up with really remarkable, clever creative ideas and ways to use this and ways to tie in great stories with image generation through Jasper Art and like all these things and then like less creative people like come up with just far worse stuff. And so I don't think people will be sitting around looking at blank pages typing out every sentence from scratch.

I don't think it'll take years and years to write a book. I think you could write a great book very quickly, but you'd still have to understand like people understand storytelling, understand how to combine these tools in order to like create a really compelling experience for the reader, for the user. And the bar will be raised, right? Like there'll be more books that kind of pass the bar, you know, but I still think you're gonna see like exceptional books, exceptional movies, exceptional media sites that have a unique insight into a group of people and figure out how to apply this in a really unique way. Like I don't think that's going away at all. But I do think a lot of this like just write, I think people will move from writers to editors in a very big way.

I actually think that most of the current stage of AI doesn't solve the creativity part actually removes the grudgey part.

Yep,

Yep. I don't know if you know this Raj, but father used to write for the next web and you know she had to write like five articles a day. So like it was like a machine. So there was very little actual like creativity, it was like much more like assembling facts together as quickly as possible. And actually I think it will enable people in those kind of roles to actually be creative because if you can like have the facts assembled really quickly, then you actually spend most of your time on the creativity side rather than like just writing down the facts and accepting everything.

And it can eliminate like language barriers, you know like uh, there are people in the US coming from like all these different countries. My wife for example is native Chinese and she's creative but she can't write well in English as well as us. So you know, she can write that in Chinese and can get great translation to English or good interpretation due to

That. Yeah,

There's just be so many people that can participate in the creative process in a variety of mediums. One of our like maybe most interesting customers, his name is Jim, he, you know, was like an academic and copywriter and he outside of a grocery store, it was about five years ago, he got attacked and suffered a really traumatic head injury and that injury turned into aphasia where he had a really hard time recalling words and communicating and he went from this very smart guy that could communicate his thoughts and write and all these things to a very smart guy that now could not do that and couldn't write the things he was writing before. And so when AI came onto the scene he was like an early adopter of it and for him it really unlocked his mind and his ability to communicate again in a very big way.

And he's become probably like our most creative user. He creates really amazing things that always blow our minds. He's very creative with his prompting, he'd actually applied to design school, you know, he is not a designer but he applied to design school as their first AI designer, which again sounds kind of crazy. You'd think that like, oh like design schools are gonna hate this. Like this school's leaning into it and they're like hey this is just like a new medium. Mm-hmm and he is like wonderfully creative and so I think he asked like are we more or less creative as this culture with this? It's like I would argue more because now you've got people like him that could never even participate in that process that are pushing the bounds forward and like creating really unique things out there. So you're gonna see just a ton of that where people that could have been adding and creative and creating a lot of value there like just couldn't and now they can and that's really powerful.

Yeah, I felt a little bit that LLMs are the most interesting thing and the generative kind of images and even the video things are like a little bit more, I dunno like they're cool but I don't, I find it hard to come up with like actual good applications for them that are not just like kind of like oh this was cool and I feel like the LMS maybe have like a ability to apply to a lot of use cases that it's very broad. Like it's just basically talking to something in gang knowledge. Like do you agree with that or do you think there is actually like kind of the art or maybe video side of it could actually have like this kind of broad applications?

Well I think they will have very broad applications. They're may be a little bit behind and I think with images what they have that makes 'em so viral and powerful is that you can generate something, you instantly see it and it's like wow that just came out of a sentence. And when people generate a paragraph it's like kind of hard. Like you gotta sit down and like read it to understand it was this good? Is this just junk, is it just nothing? And so it's like an instant wow factor with images and something that like makes me go viral. I think the challenge is editing them, making them actually exactly what you want. Where if I generate a paragraph, it's not quite what I want, I can just go delete a few words, copy and paste this thing. In other words, but like how do you like fix this image is something I like but it's got seven fingers you have to go into Photoshop and like be an expert there.

So it's like, it's hard to like get stuff into the final state that you wanted in with images, but like that's just a short term bug and I think it very quickly, you know, I mean I don't know, just kinda look around anywhere, any book cover, any images for ad copy, anything you post on social media, any billboard, like all of these things will just be created at least first pass by generative models. I think it'll be equally big. I don't know what kind of like the total restful market is for one versus the other and like which one has like more application. But I think we're gonna be very quickly seen real practical use cases with the image and video side.

Yeah, I would agree. I mean I think the applications for images and video seem tremendous to me. I mean marketing via video could be really much cheaper to do right now it's expensive, that's why we don't see a lot of it. But when it's cheaper it's gonna become a lot more common. And of course there's a huge creator industry, right? And TikTok and these types of things, people creating video, I mean that's becoming a massive industry, although that's becoming people's careers that's gonna become um, largely AI driven I think going forward. So yeah, I'm very excited about images and videos. Even tv, I mean like TV and movies to some degree animation. That whole industry is gonna be AI powered for sure. Of course there's gonna be still a lot of humans there, you know, but I'm very excited for that evolution.

Have you guys seen the movie Everything Everywhere All at Once? Oh

Yeah. Great movie

Perhaps like the most creative movie in a while and visually very interesting. I haven't seen it. Can you

Describe what's special about it?

I've always seen like half of it. It's just like very unique and creative and it's like, not like any movie that you'd kind of like normally see pop up.

Yeah, you hard to describe. You gotta see it. Oh

Really? It's just kind of hard to describe. I don't know, it's just kind of like a visual like smorgasborg of storytelling and visual effects. And I was talking with the CEO of Runway ml, which like a AI generative AI like video editing platform. And they said he was like looking at I think maybe the credits of the movie and it was like a seven person visual team and he's like, that's really strange. Like that should be a, a massive team. Like I wonder like what they did, like how did they pull this off? Are they using contractors that they just been put up there or whatever? So he like got ahold of them, he was like, hey like how'd you create this insane movie with just such a small team and and actually they said oh we used runway for a lot of this and I think that's what you're gonna see. Like that team would've before had to go get way more funding and get the powers at be sign off on the movie and all that stuff to create this like really creative film and now we get to see it because a few people wanted to make it and it turns out to be like really great. So I'm really excited about that.

That's funny that you didn't even know his own software was used for

It. Oh I know. And I was like dude that should just be your, like that's your headline on your website, everything Everywhere All at once and all at once. Really powerful.

This is somewhat unrelated but somewhat related. It's like do you think actors get like that whole industry of acting and do you think they get somewhat replaced by AI? Because I mean that's one you, it's hard to get a human in the loop on that and the technology is good enough that you should be able to, I

Don't know if we'll need actors in the future. Like I could see you having like jet, like AI generated actors where it's a team of people. I think you see this and you know, I think it's happening you know more and more in like Asia where there's some of these like AI generated like stars and celebrities and, and they actually like it's still important to tie the story in and you still have to be really good to build that character. But now it wouldn't be limited to people that look a certain way on screen and have the right connections. Like you could have everybody's like building actors that are unique and have a story that you connect to and deliver things in a certain way and like you'll just kind of like license out these characters.

I mean what if people really really care about seeing like that specific actor and the bond with them. Like maybe it'll just be that like a given actor can now do 10 x more movies.

Yeah, so I think in the shorter term, obviously we've got our batch of actors now that are real people and they'll continue to have like poll and yeah I think they'll probably start licensing out their face and you know, it's just way cheaper. You could either have Tom Hanks for you know, 20 million or you could have his likeness for like 800,000 and you know, you just put it on the body double or something like that. Like I, I think that'll happen in the short term. But yeah, I just don't know if you're gonna need human actors, you know, in the future but you still will need people that can create stories and create Yeah, you know, I guess the persona that is the actor,

You need directors, you'll need writers, you'll need editors for sure, right? You'll need to some degree designers, but their job is gonna be editing what the AI comes up with. It's gonna be a different job altogether. But yeah, I think actor is one of those things that could be, you know, significantly disrupted. To Matt's point, people are gonna start bonding with the AI characters soon enough and then they won't care that much.

Take like Leonardo DiCaprio, like he knows something about his skill that like people can't replicate. They don't know. Like again, I'm sure he is like hyper intelligent and emotionally aware about like what he's doing. And again, I think there'll still be people like that and there's still a value there, but like maybe in the future he can create and run seven different actors, you know, taking that unique insight that he has and they're gonna be the seven best actors in the world because he knows something that we don't. And even if you had all the same tooling, you couldn't quite capture what he's doing there.

The best actors participate really intensively in the creative process of building the character, right? And so they're, they're not just like being a face on the screen, they're actually, you know, doing more than that. So yeah, someone like Leonardo DiCaprio probably is a good example of that who's like very involved in the creative process and there's probably a number of those folks out there, whereas there's some people who are just faces on the screen, those people will get disrupted.

I think one interesting downside to this is I think there's a world where the stories that we like share as a culture, like we lose those and they get fragmented and everyone's watching their own unique movie, you know? And so you'll be able to just generate a movie that really fits your taste and your style and aligns with themes that you believe about the world and you know, is super entertaining for you. But nobody else has watched that movie in the world and you can't talk about it. You can't be challenged in, you know, some of these ideas that you have. You can't like you lose that shared experience

To some extent that kind of happened with streaming already, right? Yeah. Like it used to be the case that everyone would watch this episode and everyone would talk about it the day after. Now it's just like the whole season's out on Netflix and like sometimes people talk about seasons as a whole, but I do feel like the collective we are waiting for this episode, we just watched it together and it was like a big thing

And we could talk about it the next day and yeah, you know, what'd you think? And obviously there's this kind of like entertainment factor, but I think there's value in like a culture like having, we all have saving Private Ryan as a movie that you're like man, it teaches us something that we value as a culture and we can kind of share that and we kind of like build something on that and like, you know what if all of that goes away and like we're not reading the same books and they're all just again kind of maybe more and more like echo chamber, there's like a risk to that maybe we end up not caring or maybe that is so powerful that we kind of see it start to like come back together instead of being fragmented cuz people actually prefer that.

Yeah, I feel like we go back and forth to some degree because like as a society we do crave that. We do crave a common connection and common media to some degree and that's why things like squid games or like White Lotus or things like that, which we do still emerge even in today's age of streaming because people do want to have a shared experience and they talk about things with their friends. Mm-hmm and they want other people to see the same thing that they've seen. It happens probably a lot more rarely now than it did before, but it still happens. And yeah, you know, it's also just some of the criticism behind social media. I mean you can customize your Twitter feed to be only things that that you look at, right? Versus things that everyone's looking at. So there is that aspect that's already happening today and that's always a risk, but I think the natural human urge to be part of a collective that's counteracts that as well. So there's gonna be this back and forth that continues and the question is how far we'll get into this world where it's just like everything is personalized.

Yeah and I think that kind of goes into like a bigger conversation just as like a culture, like I think we need to be intentional about what we want these tools to do in our lives. Like I don't think there's any stopping it or putting the cat back in the bag, but like we've gotta decide do we care about artists rights and getting them rewarded for providing training data? And again, it's not clear, I think it's kind of up to the culture to decide like what do we kind of want to do and what do we want to value there? Do we kind of value the like original artistic expression or do we kind of value high quality training data and great models and all of that. And so I dunno, I think there's like a lot of the conversations and decisions that need to happen here over the next five, 10 years of just like what role do we want this to play in our lives? And you know, there's no stopping it but I think there's shaping it.

Isn't there a problem that the government can kind of square it up if you take lms if like someone's like, oh actually like if you ever use this training data you gotta pay X off your revenue for it And like I don't it, it's hard to implement. I would feel like most ways that a government would end up implementing it would probably be worse off than where we are right now.

Yep. I mean the pace at which this is moving is just so different than the pace at which governments evolve and move and you know, I'm not saying there's not like some place for like appropriate involvement there, but I look more to just society and culture and just kind of us collectively deciding what do we think is important here and, and how are we gonna kind of reward behavior that's in line with how we want to be as a culture.

Yeah, I mean just looking at like recent history, right? With a few things. So you know, obviously the music industry and you know, the emergence of Napster and all these things like highly disrupted a very lucrative business for music producers, right? And there was no government to step in and and try to take care of the, the artists there, right? They had to find a new business model, they were significantly disrupted. I think the same thing with Uber and Lyft, right? Like coming in and disrupting the taxi industry, there was no government to step in. You know, the taxi industry just got disrupted. That's quite likely to happen here. I think there's gonna be a set of people who are gonna have to change the way they work because of this technology. And we talked about actors, you know, there's other, I dunno, copywriters, there's things like that whose job are gonna change dramatically because of this and it's just gonna happen. I mean that's just the history of technology. Yeah. And uh, you're right that society at some point is gonna have to figure this out, but I don't really think society's gonna have a vote. I think it's gonna be relatively small group of people who are gonna try to figure things out for themselves and they'll raise a fuss about it, but ultimately they're gonna have to change.

No, I agree. And I think, like you said, this has always happened. I mean we have far less farmers than we've ever had in previous points in history and I think people kind of think, well does that mean everyone's just gonna be homeless? It's like you don't see like all the old farmers like homeless now and poor. Like they just like developed other skills.

Okay. Did you guys see this post by Mark Andreessen about why AI can't possibly take all your jobs?

I saw it but didn't read it. Uh,

his thesis, which I thought was really amusing and I think it's probably true is I don't know if you've seen this graph where it says like basically anything touched by technology and innovation has had massive deflation. Whether it's like yes TVs or consumer goods or whatever. And then anything touched by like government and to some extent like people and services costs like health education. And there's a third one though, it's always quoted, it's got like massively inflated costs and his point was that like AI is gonna have a difficulty going after those kind of regulatory captured industries and basically the cost of everything else is gonna go near zero and all the jobs will will be in these kinda industries with regulatory capture and all the GDP basically.

Interesting. Do you think

That's true though? I mean, don't you think like healthcare and education for example, seem like ripe for disruption by ai? You know, like first round of diagnosis can easily be done by an AI probably better than a primary care doctor.

They seem ripe for disruption. But you don't not think that regulatory capture will stop the disruption. Like, are we really gonna let an AI without a doctor give people a diagnosis?

Yeah. Like will that even be legal will insurance? You know, it's like, yeah, I could see that getting so tangled up or like it's pretty good and it could be really good for like on par for a long time before you're even able to like really use that.

Yeah. Maybe in other countries where there's just not enough doctors, like there's actually like a pretty big global doctor shortage , so maybe this will have more of an impact and places where like they're just lacking these basic services.

I can see a government's getting a little bit upset about it, but I, I'll tell you, there is a lot of flexibility right now with like things like telehealth. It's very easy to get like a doctor to like give you anything like a prescription or whatever, you know, and have insurance pay for it. Like there's, yes, there's some certifications and stuff that are captured in the regulatory scheme, but getting AI maybe illegitimately at first using aieg legitimately, but then it slowly starts to become more obvious that it's AI driven. I mean, telehealth and AI are bosom buddies, right? They're gonna be good friends. Everyone using telehealth, both the doctors on the other side as well as the consumers themselves will start just using AI naturally I think to help themselves out. And right now when you have a problem, you probably Google search, right? So you're like, oh, I have this symptom, what do I do? I mean that's gonna all go to an AI and the governments can't stop that. People are gonna go to a website and do that. The doctors are gonna start using AD because they don't wanna remember what they read in a textbook like 10 years ago. So they're, they're gonna start using AI as well. So it's gonna be hard to stop, I think.

Yeah, when I looked at that chart also appeared like a lot of the prices that it inflated, I think it was like college tuition, childcare, even like food and beverage. It was like stuff that's hard to globalize, you know, like you can't outsource your childcare to somebody overseas. And some of it's like, almost like this like physical world stuff that, you know, I think it'd be interesting to figure out if we can kind of reduce the cost of labor and augment more of the physical world. I'm not saying we need like robot childcare yet. I've got three little boys and sometimes that would be nice. But uh, it's kind of like these are things that like naturally have to be people driven and can't be offshore and maybe we'll see technology improve those things.

I don't know Dave, if you've seen like G P T four, I've heard from a few people who've kind of tried it out and they seem like think it's a massive improvement. What do you see as like the short term, like next kind of six to 12 months improvements that are happening in AI as yeah, as far as like you probably have a view on it that other people don't. What do you think's happening and what will they change?

We're working with pretty much any big model provider and kind of getting early-ish access to the models and seeing what's out there. And I don't fully know what all you know OpenAI calls G P T four. I mean it's not like a single thing, it's like on the spectrum. And so yeah, we're testing a lot of them. I think 3.5 was a pretty big leap over GBT three. I think four is, you know, kind of continuing on that spectrum. And so I dunno, the models I've played around with, which I'm not just being coy, I don't actually know kind of, is this what they're called G PT four are better. But I think it's like overhyped on Twitter and thinking that, oh it's just you click one button and you get you know, a 4,000 page novel just like that. It's not that. But I do think you're gonna see better zero shot prompting where it can do a wider range of things out of the box.

You'll see these models doing better with some more like abstract thinking like poetry, rhyming, connecting the dots between two things. Like write me a poem in the style of Beyonce about this topic that, you know, it's just like kind of connecting these like random things I think they'll like do better at. I think you're gonna see context windows expand dramatically where right now most of the models are limited to like two to 4,000 tokens and so you just can't fit a lot of content in there either kind of in the input or the output, which like really limits the kind of work that you can do or the kind of like context that it can hold. So I think you're gonna see like those windows expand dramatically to where maybe you have a 50,000 token context window or a hundred thousand or a million, which just opens up a lot more possibilities.

Like what would be the most interesting possibility from like a huge context window.

Well then you could click a button and generate 4,000 words, you know in one. Now would it, would it be good or not? Who knows? There'd probably be like a lot of drift. Yeah. Yeah. But I think the big thing is you could store more company or personal context in that. And so if I say hey, write a blog post about mercury or you know about whatever the importance of banking, I could send in five previous blog posts as examples. Mm-hmm in that one prompt and it would be able to feed off of those.

I see. So it's like kind of like quicker fine tuning.

Yeah, you'd have to fine tune for that right now, which is just like so much more work.

Do you think, like I always find one thing that like makes the AI is like less human is they don't seem to like keep their own context. Like they don't have a life obviously , but they also just like don't remember their previous things. Whereas the humans always like, okay, you know, this is my story. Like they don't have a story. I wonder if they can have like morph a story if they keep like a continuous context.

And even you see that some with, I know there's some examples of, you know, when Bing chat came at it, it kind of appeared like after five interactions back and forth, the like original context and maybe even like the original prompting like fell out of the context window and it was start to kind of go off the rails because like it's a context problem, it just can't fit all of it. And so you've gotta be selective and opinionated as a company and again like users don't understand this and so it's, it's really challenging to figure out how to include the appropriate context in a way that does what the people want it to do without them really even having any idea how to do that. And it's not always clear cut that just, oh just send in the last bits of what they were writing about. Sometimes you know, for writing a blog post, you know, maybe you always wanna send in the title and the description, even if it's a really long blog post. The beginning of the blog post might not fit into context when you've gotta cut that off, but you still wanna keep the title in the description in order to kind of keep it on track there. So there's just a lot of engineering that has to happen around this that would be nice to not have to do.

Speaking of this idea of like AI is becoming a little bit more human, I know there's this whole like fear factor driven by sci-fi about sentient ai, but do you think these language models should become a little bit more human in terms of like how they dealing with things like empathy or dealing with different types of soft touch, you know, like if it helps their humor to be a little bit more relevant and on point. Mm-hmm , do you think that's one part of the evolution of these language models is to make them more human deliberately even though you know, there might have some, you know, and you could maybe cap the downside by some protection so it doesn't have all the negative attributes of a human.

Yeah, no I certainly think that could make 'em, you know, maybe a little bit more palatable or adoptable. I mean they do a really great job at imitating those things and some might say well humans are kind of imitating empathy, you know, a lot of times too. And I think for us, like I always come back to what's the job to be done, what's the ideal output or model or set of models to accomplish that? And so like for marketing, you know, empathy matters and mean human matters and that's a really powerful way to write copy and to write blog posts and write stories. And so like in that regard, yeah we've gotta help our models be more human and more empathetic and more casual. But I think other use cases may not need that at all and not actually detract by making it more empathetic or more human-Like

Do you worry about kinda AGI or AI alignment or do you think like that's overblown?

I wouldn't say I worry about it. I can understand kind of both sides and kind of see a world where that is not a net positive for people. I'm an optimist and so I'm kind of just like, let's go create and build and you know, let's be thoughtful about these things as they come up, but I don't really worry about it. None of the futures where we get destroyed by turned into paperclips, you know, by the paperclip baking machine . I could see where that could happen, but it seems like lower probability to me, but it's also extreme downside so maybe even if it's incredibly low probability, it uh needs to be taken very, very seriously and I can see that for sure too. What do you think?

I think actually like the level of AI we have is really interesting and there's tons of applications from it, but I don't think it's anywhere near AGI and human level intelligence. Like I just don't think we understand it well enough. Like I think we are far away from it. So that's my first blocker. I think my second blocker is, I think the more intelligence something is the more empathetic it is And I mean that's what's happened with humanity to most of the extent, right? Like I think humans were like real assholes like a thousand years ago, like we had like rampant slavery and there's just a lot of bad things and uh, humans have improved I think the smarter the societies of God. So there's no reason to think that a really smart AI would not also head in that direction.

Yeah, there's this like window, you know, let's say it gets smarter and more powerful really fast. There's like a pretty small window in which it would like care to destroy us, but before it gets too smart and powerful where it's like add doesn't even bother me.

I love these things that like people debate and technology and scientific circles are really actually like religious debates to some extent. Uh, yeah, you know, I grew up in like a religious household and one of the things my parents could understand is like how someone non-religious can still be moral. They were like, oh, if they don't have God to like guide their morality, obviously they're all evil , right? Like that was like the thing in reality actually people can be very moral and make good ethical decisions without this like kind of force forcing them to

Yeah.

Yeah. I, I think there's also this point about the desire to like live and the desire to actually even destroy other entities that you can live is very biological. It's actually a defining characteristic of life. Our biological life is that our life wants to live,

But all of these LMS are being trained on humans, right? So like when they have human tendencies

They'll talk about human tendencies but I don't know if they'll have an innate human, like I don't think an l LM is gonna become sentient. I think, you know, we're talking about agi, that's a different kettle fish. So is it gonna be, is an AGI gonna want to survive? Is it gonna have that deep si? I mean it could be programmed I guess to do that and probably you could program an AI to be destructive, a hostile country you know, could maybe do that right deliberately. Maybe that's the thing we should care more about is like if someone's gonna do it deliberately as a way to like hurt another country.

What a positive note to end this on . Yeah, we have another five minutes. Uh, Dave, I think one thing that I always try to learn about this stuff is like the AI field and applications of AI is changing so fast. Like where do you kind of go to, you know, make sure you are on top of it, what sources have you found like the most kinda informative or communities you're in?

I mean certainly Twitter is kind of where I'll see breaking stuff and different demos and and interesting things there. I'll interact some with like the research community and just kind of see like what's kind of out there further out a little bit more nebulous, you know, a little bit not practical today. And for that I'm usually just like texting a couple different friends and kind of seeing what you know they're seeing there. I think it is important to have like a foundation, you know, again, I'm not a AI researcher but I've taken a few courses and I've watched some YouTube videos just on kind of like understanding like the core concepts and like, you know, what's even kind of happening into the hood here. Like I think that's really helpful even if you know you're not gonna be an AI expert. But yeah, honestly Twitter is probably where I kind of see the most interesting and like up to date things. By the time you write a book about it, it's like not even relevant anymore, you know? So it's gotta be a pretty quick to stay on top of things.

Dave, was there anything else you wanted to cover today? Any other topics or something that you wanted to chat about that you were interested in or any questions you have for us?

A lot of people ask me about kind of how to build products with this and what's made Jasper work when others haven't worked as well or whatever. And I think there's still so much opportunity today, even as it gets so crowded and competitive for people that are willing to start with the customer and work backwards and instead of starting with the technology and like going and like working towards a customer there and obviously like the technology sparks it so it's gonna kind of start there, but I don't see a lot of people like talking to enough customers and like developing a point of view or getting a unique insight. And in a world where kind of like a lot of it is becoming a commodity, it's ip like what you gather from talking to 30 customers is like your own IP. And what you build out of that is like a unique insight that you have on the world and you have on a specific segment.

And so people will ask, well you know, what can I build? And for me, the way I would approach if I starting over new is I, I would pick a segment, pick a customer base, let's say it's doctors, let's say it's lawyers. And I would just start to like really talk to them and I would start to really understand their workflows and what do they use to do this thing and then what do you use before that? What do you use after that? And just like going really deep there and then turning around and being like, okay, cool. Like, like I understand the pain points, I understand what they're saying that is so painful and takes up so much time and is so tedious. Like is there something en generative AI that can solve this extremely well or something Not in generative ai though again, I would probably recommend people are building generative ai.

You were not just like talking to marketers, you were a marketer, right? Like you'd written a course about it. So like you had this like super du deep insight.

Yes. In my case I was that and I had spent seven years Yeah, yeah. Fully engaged in that community. And so that's the other shortcut too. It's like if you just are the customer Yeah. You spent seven years ago, it's like you can do . Yeah. Yeah. But if you, if you could solve your own problem, obviously that's where like so many great companies have like come from. But um, you know, like there's just, this is gonna be so prolific across so many different industries that I think it just starts with like talking to the customer and like working backwards from there. And I think there's real differentiation and power in that that's not as sexy as like starting with the, the model and you know, working out, but it's far more powerful. So that's always my encouragement to people is like, go talk to customers for a while and then go and look at the tech and I think you'll come out with some really interesting solutions that it's very likely other people are not thinking about because they haven't gone done the legwork there.

Yeah. And you're more likely to be able to commercialize them because you're solving someone's actual problem.

Yeah. So that's just what we always talk to our team about is just, hey, like we wanna be customer obsessed. We're always starting with the customer working our way out as widespread as we all think all of this stuff is and G P T three and all that. Like most people in the world don't know, don't care, will never care what a G P T four is. And they've got pain points in their day that they're looking to have somebody solve in a delightful, easy, elegant way. And if you can go and do that, they will love you for a long, long time. They'll pay for it. And that's a repeatable model that a company can take to keep building great products. Yeah,

What a great point to, in this great podcast, really appreciate you taking this time, Dave. Super interesting. Yeah, this

Was fantastic. Yeah. Thanks guys.