Copy.ai CEO Paul Yacoubian on keeping up with GPT’s rapid evolution & AI’s potential impact | E1719

0:00Hey, everybody. Hey, everybody. We all know the AI space is moving at a blistering pace right now. So we decided to do a little ad hoc miniseries on it. We're calling it Innovators in AI. Very creative. Every week, I'm going to interview a founder in the AI space here on This Week in Startups. And every founder that joins is going to be building something awesome. That is the criteria. We don't want just pontification. We want to actually see real products. And today, we have the CEO of Copy AI, Paul Yakubian. We have a great conversation about the insane pace of AI, as I mentioned, and building products for a specific niche, in this case, writing, copy AI, right, they do writing on top of various LLMs.

0:42And he explains how he's dealing with the fact that, hey, he started, I think, with GPT two or three, then they went 3.5, four, five is coming out, BARD is coming out from Google, so many different LLMs out there. How do you actually build a product that people are willing to pay for when these core platforms are moving so fast that they're going to absorb the innovations of people building on top of them almost in real time. This is a major issue for founders who are building in GPT or in AI with OpenAI's products. So it's going to be a great and insightful episode. Please stick with us. This Week in Startups is brought to you by Mercury, where innovation meets peace of mind.

1:27Now more than ever, startups need a safe place to put their cash. Mercury offers a simple way to manage bank risk and protect every dollar with up to$5 million in FDIC insurance and a money market fund. Visit mercury.com to apply in minutes. The Microsoft for Startups Founders Hub helps all founders build a better startup at a lower cost from day one. Startups get up to$150 ,000 in Azure credits, access to OpenAI APIs, free dev tools like GitHub, technical advisory, access to mentors and experts, and so much more. There is no funding requirement and it only takes minutes to join. Sign up today at aka.ms slash thisweekinstartups.

2:15And Mayfair helps venture-backed companies protect and grow their cash automatically. Diversify where you hold your cash, increase your FDIC insurance coverage, and earn up to 4.35 % yield on deposits. Go to getmayfair.com slash twist to get started today. All right, everybody. Next up on the program is a company that's been doing generative AI, specifically in copy and words, for a couple of years. I met Paul a couple of years ago when he was raising money for copy.ai. My bestie, David Sachs from Kraft Ventures, is an investor in the company. and i remember meeting you paul and paul's last name is yukubian yukubian right that's right that's right there you go yeah um you i remember were working on a couple of experiments when we met one of them was like doing taglines for using that's right i don't know if you were using chat gpt at that point but when did you first become aware of chat gpt and or generative ai for a copy Good question.

3:24So I started using GPT-2 in the end of 2019. And we were using it to create startup ideas. So it was insanely creative, but it was really bad at telling the truth. Because it wasn't trained on as much data as GPT-3 was. And it was immediately clear that the creative power was where the value driver would come from with the generative AI models. And I thought it'd be seven years away from being commercially viable. because it would just make stuff up and you had to really help constrain the outputs of it to drive value from it. So I was creating science fictional products and just had this explosion of ideas.

4:04What was the interface like for that chat GPT-2 at the time? Was it all like a dev tool? Because I remember it was a lot of startups were trying to get access to it. Yeah, GPT-2. And it wasn't really available. It was actually open source. So somebody had to set up the model, host it, and then build a UI for it. And the simplest UI was really just a text box where you type something in and then just click a button and auto complete it for you. And so I was using that, but I wanted a lot of ideas. And so I kept hitting the button and then deleting what it printed out and then hit it again. And I was like, a search interface would make way more sense, like a UX interface for brainstorming.

4:45And so I thought it would be about seven years before the model would actually be good enough because only about 10 % of the results even made sense at that time with GPT-2. And then fast forward until July of 2020, which is only like six months later, and they launched GPT-3, which is 100 times the size of GPT-2, and all the numbers flipped. So all of a sudden, most of the time, it was making sense what was coming out. Ah, so when you were doing 2.0, this is really interesting to look at the history of it. I'm sure people are going to study this rapid pace that this has gone on. There was no web-based service.

5:25There's no cloud-based service. You had to download it. And then did you have to provide a model or was there already some training in it based on some model? Yeah, they had pre-trained it and they had open sourced the model. So somebody else ended up hosting it and providing that interface. Got it. And so that's what I was using. Were they upfront about what was in the model or did they just say, Hey, we trained it on some stuff. Or were they explicit? Hey, we downloaded the Wikipedia. Hey, we did Quora. Hey, we did Reddit. Were they at that time? Because that's something that I've been trying to get my head around.

5:56And hopefully you can educate us in the audience as to what exactly. How do you know what's been it's been trained on? Yeah, they I think in the paper listed out a number of the sources. One is a common crawl data set. It's just a huge, huge file that's open source. You can download it and you can use that to train the model. Wikipedia was a big, a big component of it. Common crawl as in a common crawl of open web. So just a bunch of web pages that are in an open source search engine of sorts. That's right. They got Reddit and I don't think they got Quora at that time. I think. Yeah, probably had a paywall or something or a registration wall.

6:32So they couldn't get in there. Right. I think since then they have, they have probably gotten access somehow to Quora. it's data set so you see 3.0 and it goes from one in ten times to getting it right to one in ten times getting it wrong you said it another way nine in ten times at least making sense like it wasn't completely gobbledygook gibberish exactly yeah so the first the first thing that came to my mind i was like this is the next big wave in tech because all that you have all these new use cases that are unlocked and we would i thought we would see you know immediately i've tweeted this out i I thought it was a phase change of the internet.

7:09So this is the first time the internet can be trained and synthesized into a model and then talk back to you directly. Right. So that's a profound change in the amount of information that you have access to at any moment in time. And given that you're kind of able to talk to the internet, I thought that conversational interfaces would be the most relevant way to access that information that's inside the models. so we built built a slack bot and we asked uh open ai if we could launch it and their safety team said no you can't launch that it's too dangerous and so we had to go back to the trying board and we built um built the an app that could just summarize whatever you threw at it so we launched that we got a mention in the uh wall street journal um for that and then we realize, you know, people, that's like a sometimes use case.

8:06I mean, that is a component of a skill that they models have, but that's not something that people will subscribe for and pay us money every month. Listen, as a founder, ensuring your cash is safe is priority number one. It's been a bit crazy out there. I don't want to put caps locks on right now, but I will if you need me to. The FDIC$250 ,000 limit is just not enough for most businesses. So let me tell you about Mercury Through its partner banks and sweep networks, Mercury customers can access up to$5 million in FDIC insurance. That's 20 times the per bank limit. Sounds a lot safer, doesn't it?

8:45Well, these sweep networks protect your deposits by spreading them across multiple banks, which limits the risk of any single point of failure. And with Mercury Vault, any funds above the FDIC limit can easily be invested in a money market fund, mostly comprised of U.S. government-backed securities. So it's easy to get started with opening an account. You can apply in minutes, and many customers are approved and onboarded in less than two hours. Mercury also offers great resources for founders, including my favorite, Mercury Raise. This program connects founders with investors to help them raise funds.

9:20Head to mercury.com to join more than 100 ,000 startups that trust Mercury for banking. And if you're interested in Mercury Rays, applications for both Seed and Series A are open through April 20, right? So you got to get it in there by April 20th. Mercury is a financial technology company, not a bank. Banking services are provided by Choice Financial Group and Evolve Bank and Trust, members FDIC. when it's summarizing something explain to a lay person what it's actually doing so you have a paragraph let's say you have a i don't know a movie review from the new york times it's a thousand words about some you know really detailed movie like tar right academy award uh uh oscar nominated best picture but what does it do what is the what is the um what is the model actually doing when it is summarizing?

10:14So there are different techniques for summarization for these models based on neural networks. The neural net is learning skills. So it has learned skills that you would have otherwise needed specialty natural language processing algorithms to do. And that's really what sets it apart. So when you had 176 billion parameters in the model and they're structured into layers, these layers get really good at doing certain things. And for the summarization tasks, it can learn what summarization means. What does the word mean to summarize something? It can pick that up from the corpus that it's trained on, so all the text that it's trained on.

10:58And then that is a task it can do pretty well. Which to me, I was like, okay, now this one model we can use to do all of these natural language processing tasks without needing a bunch of NLP engineers on the team. It's literally one API call, and I can just prompt it in natural language and tell it what I want, and it'll give that back to me. So the first, that was a major unlock because it suddenly made the process of building an MVP, a software MVP that uses natural language technology, made that really easy. So you could build a prototype and launch it. And so when we built our next prototype, it was called taglines.ai.

11:42It was just a tagline generator. That took 48 hours from beginning to end to build the entire thing. And we launched it 48 hours after saying, okay, maybe people would buy, maybe marketers would buy words if we can generate them. And so within the first week, we had validated that not only would you have professional copywriters using these tools, but you'd have marketers and freelancers and small business owners, students. So that was a huge validation of the market size and the TAM potential because it was such a wide user base. And then we also validated that people would use a tagline generator to do email subject lines, headers for websites, just all kinds of PowerPoint slide titles and really validated that the number of use cases were really wide for that as well.

12:35And then throw up like a Stripe paywall. And then validated people would sign up for a subscription. So we started charging$3 a month. We got some subscribers, raise it to six, and then raise it to 10. And we continue to get subscribers coming in. So at that point, we validated one that it was commercially viable, that you had a wide market and a wide set of use cases. Right. And that makes sense. There are copy editors in the world. You discovered this is working nine out of 10 times. and if you can make them more efficient for$10 a month, if they, the average copywriter, I think freelance probably gets paid 50 bucks an hour, 40 bucks an hour, 60 bucks an hour, something that range.

13:14I mean, it's not even one hour of their time. So that's always one of those great tools for founders is to say, what does this person get paid an hour? Okay, this attorney gets paid$1 ,000 an hour, save them one hour a month, they'll pay you$1 ,000 or they'll be okay with it, right? Exactly. So when we went through that exercise, it was really clear that from all the way from a student to professional would find value in these ai models and then the thing that we learned from going from two to three was that if you throw more data at it it's going to get even better and better performing these tasks at an expert level or beyond so now when we get to gpt4 to get back to this sort of these jumps there was three i think there was 3.5 and four in terms of major releases you're working directly with the open ai team you have some insider access to it, I understand.

14:03So you were kind of in touch with them and seeing what was coming, what was coming between 3.5 and 3.5 and 4. Yeah, so for 3.5, they ended up tuning it to human feedback. And so what that would do is it would allow you to kind of describe the intent that you had, and you wouldn't have to prompt it with examples as much. So when we first built our tools, each tool, he had to give an examples of what good looks like, and he had to make sure that even the examples are very diverse. When they ended up fine tuning it on human feedback, they kind of program that in across a wide variety of use cases.

14:40And so that enabled the chat GPT experience where you can just tell it what you want, and it'll figure out and actually create that content that you want. And so when you saw three, 3.54, maybe you could describe for the audience, the step functions there, you said before, hey, you know, it's right one out of 10 times then it's right nine out of 10 times now let's go to 3.5 and 4. 3.5 yeah it was awesome it was like a big a big leap and a lot of the complexity and prompting it kind of went away a lot of um even the use cases for more fine tuning of models so um you'll see a lot of startups talk about how valuable their their fine tune models are like oh we have proprietary models well at the end of the day not there aren't that many models that are really relevant for for you to have fine-tuned what a model is again to a layperson and what's an incredible example of that where you're restricting it and building a model that maybe results in better output yeah so the these are foundational models so these are built to be general purpose that That means they can do a lot of things at a pretty solid rate.

15:55What they figured out was that the general purpose model could actually outperform specialized models if they kind of made them bigger and trained it on more data. And that is nuts. That's nuts. So I don't know if you've seen some of the GPT-4 metrics, but they said, oh, well, it passed the bar exam. It passed all these exams. right it's getting better across every single category so let's pause on that for a second yeah you used to build in the industry what was called narrow ai yes narrow learning so hey we're going to learn how to beat um you know humans at poker chess go you know pick a video game fortnight and we'll really just work in that narrow subject area now this more general model i'm a little confusing to use the word general because it's loaded because there's general ai as a concept which is it's thinking like a human but their average or their default model maybe is a better word for it their default model is better at doing the l stat or the sats or the bar exam than a model that was trained just to do the bar exam that's what we're they're learning with four that's right and you seem to say hey that's insane why is that insane why is that insane uh it really does level the playing field if you have access to those the gpt4 model so you don't need a whole team of ai engineers you don't need a whole team of machine learning engineers you can actually build you can just go straight into building the application and yeah keep going i was going to say something that i think larger companies are running into issues around is they've kind of delegated the like hey ai team go figure out this generative ai stuff but the ai team you know if the ai team comes back and says you know what we can't beat gpt three and a half we can't beat gpt four they'll get fired right right that's fascinating so if i was working inside of i don't know amazon let's just say walmart and i said hey our internal team is a great example amazon's a great example yeah i'm in amazon i want to optimize amazon prime members to show them better reviews and show them better q a so i take the q a section on an amazon page and the review section and we'll summarize it and make it better for users if gpt4 can do that better than the internal team why is there even an internal team yeah that's pretty wild depends on the use case so if you looked at like the iphone voicemail system you know try to transcribe a voicemail message they had i don't know how many thousand ai engineers at apple and then open ai had a team of i think two or three people that figured out how to how to complete the um voice to text task and they ended up launching a model called Whisper, and they open sourced it.

19:03They literally gave away the technology, and it dramatically outperformed all of the existing technologies, all of it. All right, everybody, our friends from Microsoft are here. Tom Davis, a senior director at Microsoft for startups, and you're a former founder. You are here today to talk to us about the giant leaps that Microsoft has made in the AI space. What does this mean for startups? I see a ton of different tools. I've been playing with ChatTPT for, I have a paid account, but I'm also seeing things happen with GitHub. Absolutely. So the work that we've been doing with OpenAI over the last few years has really set ourselves up with a foundation around, we've built this sort of AI supercomputer from the ground up and we've been looked at everything from gpu configurations to networking and things and really what we're now able to do is sort of allow startups to access all of this innovation through our founders hub so we've been building this to do something at scale so startups can now build their own ai applications and and build out and train llms as well.

20:15And this has really helped us to become a far better cloud for AI broadly. And being able to drive that down to the startup ecosystem is fantastic. It's open to everybody. There's no funding requirement. You don't have to be anointed by a VC. Five minutes to sign up. You get six figures of benefits, Azure credits, GitHub, open APIs, which everybody's really having fun playing with and so much more. So go ahead and sign up right now, aka.ms slash thisweekinstartups, aka.ms slash this week in startups thanks so much tom thank you so as we're looking at this yeah this is almost like the snake is eating its own tail kind of situation we're all going you know uh copywriters you can name your company's copy ai you got 12 million dollars in reoccurring revenue and your premise which turned out to be absolutely correct is hey this thing is advancing fast copywriters are going to use this tool to be willing to pay for is going to make them bionic make them superhuman um won't replace them but might replace the bottom third that are terrible at their jobs or make the bottom third great at their jobs or good at their jobs was that your central tenant um about this because the replacing of a job or elimination of the job in other words i'm the developer or i'm the ceo i'll just ask chat gpt for copy and i don't need a copy editor uh what what is your belief now let's say copy editing your wheelhouse copy editing going to go away and the CEO or the sales team just does the cop ask Chappie T to do it or will there be bionic copywriters who use this as a starting point i think before you know before we launched copy i had no idea what copywriters even did so it was it was never about copywriting it was that was basically the first use case that i thought would be relatively translatable against the models but the thing that really motivates me and my co-founder is the idea that people have great like a high degree of creativity especially kids yeah they're super creative i've got two daughters that are extremely creative and then they we pre-train them through school right just like these models are now pre-trained on the internet and all that pre-training kind of stamps out the creativity and by the time they graduate, they're very tracked in the things that they want to do.

22:38So it's like, oh, I'm going to be an attorney or I'm going to be an accountant or I'm going to be a software engineer. And I think that that whole process is going to come to an end here because GPT-4 can do all of the liberal arts better than any expert could. And now that's a tool that's in your pocket, which means we don't need to train people on a lot of these very technical tasks that we used to because we've now trained computers to do that repeatedly so there will be a very um you know a challenging transition for a lot of industries a lot of institutions like higher education education in general and we're gonna have to figure out okay what what's the best use of of a person's time well that that usually comes down to the individual and what we what i'm seeing you know the whole my whole hope here is that people can be way more creative because each time they are creative they have an outlet to actually build something they can actually get things done so a non-copywriter a non-writer somebody's just not good at words can now go in there and be good at words so you take the exactly 80 of people who are not good at that now they can be good at it i may not be good at drawing or illustrating but i can pop up stable diffusion or any number of tools and i could be a good illustrator so now you've got the 99 of people who have no illustration skill can't draw for not uh and they can be creative so it's not that that everybody's job goes away it's that everybody gets good at everybody's job yes that's an interesting way to look at it and that's that's incredibly powerful especially if when you're looking at entrepreneurship so they're If you've ever tried to create your own website, for example, that takes a lot of time and it can be very stressful for somebody that's trying to get something off the ground.

24:33That is now going to be pretty much instantaneous. Like, oh, I need a website. Boom. I've got one. I need a presentation. Boom. I can just generate a whole deck. And then work from there. So you're starting on Firebase. Yeah. And so... Which happens now. You can go on Squarespace, pick a template. And now you'd be able to, a future version of Squarespace will be, Hey, make me a website and show me five different designs. I need one that's funkier. I need one that's a little more avant-garde. Now the challenge, the challenge is what kind of business are you going to create? That's going to actually be able to compete against what the AI models can do.

25:17Yeah. Well, and to your point, the AI developers who are making this, are literally working themselves out of a job the better they do the less they're needed because the general model can do better than the specific model so at least in that case the narrow specific models are being subsumed into the general models yes yeah you're seeing that and then you're also seeing it um in the software itself so i'll give you kind of a run through here so yeah software is why is software valuable jason uh because it makes people more efficient which means it reduces the cost of any good or service right and and what like why software is valuable is because it's reusable so your investment to build it part of it yeah right it's a reusability so we're now heading into a world where software can be generated on the fly right and it's almost the cost of creating new software is going to zero right that means that the returns on investing in that software should also go to zero, which causes some problems in the venture capital community.

26:32I'm an angel investor. You're an angel investor. So we really, this whole thing is going to get transformed from a technology standpoint. Fantastic. I mean, if you can build a company with two people instead of 10 or 10 people instead of 30, well, that means you need less investment and the earliest investors are the only ones who get to seat at the table i'm fine with that i'm fine with people who do series b and c is getting blocked from having an opportunity to invest because the company's too profitable is that sort of your premise it takes less money to run this just like cloud computing drop the cost of it yeah so one it should expand the number of people that can build things which should be good overall for entrepreneurship uh two like you said on the growth stages that that's going to increasingly make i think a little bit less sense and you've seen you saw yc shut down the continuity fund which is their growth stage fund uh this week and and so in that and then when we look at the ai landscape a lot of the repeatable tasks are in the sales and marketing space so like personalization now can be done on a one-to-one basis.

27:44At scale, even things like querying databases, you know, that's something that people had to do on their own. So you'd have to figure out, well, how do I write the right search query to get the right results? Maybe on LinkedIn. So what, what you're going to see happen and this, you know, chat is a great example of this. We'll express intent, right? Which is what our query is when we type into chat. And then the AI will figure out what it needs to go do to get us back what we're looking for. And so that is a world of AI agents where we're letting the AI go and do things for us. And that world looks very different as well.

28:28So AI doesn't need an interface to go grab data, right? It can just query APIs directly and go get the information it needs. If it's allowed to, if it has access to those. Exactly. So I had a tweet about this. I said, by the end of the year, as my prediction, the biggest search engine on the planet won't be Google, it'll be Bing. And it won't be because people are using it, it's because the bots are using it. So so many bots are doing queries that it's just going to increase the size of the corpus and the learning model. the corpus the corpus uh kind of grows when you publish things and so there are going to be you know agent flows and bot flows that go and create more content and new content um but for the on the retrieval side that a lot of those queries are going to get sucked in to you know powering some result right and then that result gets re-indexed it gets re-indexed and now people are speculating is going to pollute the uh search indexes to the point of absurdity somebody like yourself or some group of people offshore uh just like they created content farms they could literally just start creating they could create an ai right now with chat gpt that registers 100 new domains a day makes 100 different recipes and then articles about those recipes and then does the same thing every day until there are so many recipe sites that the internet is flooded with garbage and nobody knows which one actually works.

29:57That's the fear. That's the negative thing. Yeah, Google kind of knows. How? Well, they have the original index. Got it. Yeah. And they can detect AI-generated content. Overall, they're... How do they do that? How do they generate... How do they know it's chat GPT-generated? There are signatures in the outputs. that allow you to calculate. It's a perplexity score sometimes, and that's an actual measurement of how likely the next word is to follow the previous one. Got it. So it's too perfect. It can't be too perfect, and you can make it not be as perfect just by giving it certain instructions and say, hey, mix it up a little bit.

30:41I think this is how they caught the chess cheater kid, was he was doing the book move. yes he was and he was accurate yeah how often does somebody do the book move and i play in the chess.com thing and it will tell you like this is the book move this is the you know game theory optimal move basically and if everybody's doing game theory optimal so now you have to put something in there randomized to not do game theory optimal so maybe at some point it figures that out you know it's interesting though they said when the digital camera came out uh that this would change hollywood forever that digital filmmakers and wayne wang did a digital film um and bennett miller had done a digital film on this is all on the vx 1000 i believe was the name of the sony camera the sony one yeah it became super popular in the late 90s in new york and all these and skateboarding and skateboarding yes well skateboarders use it right yeah and it was like this thing is so good that when you blow it up and you put it on a projector you can't kind of tell or most audiences can't tell a cinematographer obviously and then there were filters in you know different pieces of software that would make it look even better anyway the point was because you don't have to worry about film stock developing it the cost just got sucked out of this you could record forever even a bad actor could just keep take doing different takes you would get there and it didn't actually change the world i mean it did create youtube it did create a lot of other content so maybe a nuts extent it did but it didn't change film so what's different about that argument than your argument for um you know ai replacing everything well the the iphone won right iphone won the camera world it destroyed it it also won music so my my like base case is that apple wins probably this too to a large degree one because the phones are powerful enough and the models are going to shrink small enough that they can fit on a phone and so even some of these image generation models they can fit now on the phone text gen and same thing on a long enough time frame they will fit on the phone and um i think that you know that's always crazy but then people just get used to it like oh yeah iphone 20 will have whatever it is native well just like laptops now have wi-fi built in there was a time you had to add bluetooth with a dongle or wi-fi with a dongle or a pcm cia card whatever that was called you know you have different card slots on your laptop you put in different peripherals now it's all on a chip somewhere and you don't have to do anything for it exactly i have one more point here so my kids are my kids are eight and five and they just think computers can generate images just totally natural to them you know and in the um i think it's adults and people that have struggled to use computers their adult lives it's like this is these are the people that have you know a challenge like accepting that ai is going to power computers to do these tasks for us but kids and kids are like yeah of course of course yeah of course it's going to yeah i mean it's just like well it's so obvious that a computer could do it to them that it's not even special or interesting you know if you're of the age where you remember the internet uh happening in broadband um and there was a moment in time where i was on a college campus in the late 80s and i had seen higher speed internet speeds bitnet and arpanet and people only had dial-ups so i was using dial-up at home but i was like you know this this can get a lot faster there are faster systems in this and i had installed token ring and banyan vines and then eventually ethernet in local area networks and like you know this thing's going to be able to carry images and pictures and sound and eventually video and so it's kind of the same thing like you your limitation right now that you're like oh it can produce interesting copy that's sometimes good sometimes bad okay images and once in a while a three second looped video i mean you're gonna be able to talk to this thing and just make your own star wars story and it'll be indistinguishable than i don't know the original series or whatever 10 years ago series without any polish on it i mean that basically is doesn't take much of a jump actually only takes a huge jump if you haven't watched rapid change right well i think in the next year you'll see it probably just take over business in general.

35:05So we know it can generate business ideas. We know it can create websites from an idea. We know it can figure out who the core persona is for that service. We know it can access APIs that would retrieve who the target contacts are. We know it can write personalized emails to them. Right? So it's going to come up with an idea? You make your SaaS startup and then it starts selling it for you. It's all in the same sequence of prompts. Yeah, that's all it is. So that's all that is here today. It's just a matter of infrastructure. And that's something we're working on right now. So what I mean, what you've described means your own business commodified.

35:52So are you saying that your own business is going to or your or are you saying your business will be commodified every 24 months and you'll just have to extend it? no i'd say like most businesses are not commoditized by that loop you know most businesses are up some kind of like real world thing happening let's say like jp morgan right that's not you can't really replicate jp morgan with that kind of system but there are a lot of one services that can be replicated like specialists consulting a lot of that stuff can be replicated a lot of software companies can be replicated yeah i mean you're talking about regulation i mean you could i mean somebody's going to put ai and web three point web three together in crypto and be like make me a crypto that does x y and z they're trying to i mean they're trying to shut you know shut all that down with um regulation regulation the the crypto choke point yes if you f around too much you will find out what the government is capable of When you're a founder, priority number one is hitting your payroll.

37:00This means you need fast access to cash no matter what goes down. So startups need to hold their money in more than just one FDIC and short account. We all know that, right? Been pretty choppy out there of late. So let me tell you about Mayfair. Mayfair is a cash vault powered by Stripe. And it's built for startups looking to diversify and grow their funds automatically. Takes 10 minutes to set up. It works alongside your existing bank accounts. and most importantly, FDIC in short. Here's how it works. You sign up for a new Mayfair account. You keep operating cash at your existing bank and deposit the rest at Mayfair for safekeeping.

37:33Then move your money back whenever you need it or let Mayfair manage your balances automatically. Mayfair can help you earn yield with no investment risk or lockups. Earn 4.35 % APY on daily balances paid for by the bank. Your cash stays in cash and is never invested into money market funds, bonds or other instruments withdraw your funds at any time just like a checking account mayfair is designed to keep your money safe they partner with stripe to move your money securely and your funds are held in the custody of evolved bank and trust which powers companies like mercury and wise your funds receive automatic pass through fdic insurance so here's your call to action twist listeners that sign up today earn 4.35 percent yield on their daily balances that's more than 10 times the national average savings account go to get mayfair.com slash twist to start managing your cash like a fortune 500 company today that's get mayfair.com slash twist okay so show me copy ai and what it's doing today because you and i were talking at south by uh and i was like huh a lot of the stuff you showed me is now built into chat gpt4 what does that do for your business yeah so i'll give you that same question which is hey you know the stuff you were talking about two years ago to your credit i literally put everybody in my company i said everybody has homework this weekend we're a finance company we all work six seven days a week if you don't like it don't work at a finance company um and i said just spend an hour or two playing with uh chat gpt and um if you tell me on monday how it helps you do your job faster better or what did they come back with uh the sales team started doing prospecting in it and other people started doing marketing plans in it and i haven't gone through everybody's responses but uh the my team on this podcast has been doing podcast notes with it and that works really well which we thought was a goof 60 days ago when chat gpt 3.5 came out i was like just ask you what the revenue of this you know netflix was the last five years and it was wrong and then chat gpt 4 comes out it's like okay it's right so it's starting to things that we even noticed three months ago four months ago were wrong so you could and then we asked it to take you know um we will take a zoom recording of a meeting with a founder transcribe it in this past week we took the transcripts which now zoom is doing and then we had chat gpt4 summarize it and summarize it in three different lengths so now in our notion instance not only do we have the video with the interview with the founder recorded we have the transcript and then i guess two or three versions of it so i could read the short version of it and i was like it's pretty accurate i mean i wouldn't rely on it to make an investment decision but certainly would be interesting to read before you meet with the founder what about gpt5 it could do it you have to assume gpt5 will be extremely less error prone and you will just trust it just like gps the first gps remember people ran off the side of roads be like turn left and somebody would wind up in a cornfield like it said turn left and it's like exactly a dummy you still have to look just also like autopilot it's like it's gonna stop at this red light unless like there's a christmas tree at the red light you know confusing it um but it feels like you can trust it i think trust is going to be the main issue is can i trust this by default yes i trust it more than a human and i trust self-driving on a highway more than i trust a human today yes you know yeah and we're we went through the that valley where the quality was lower than people can produce.

41:03And that was like the GPT-3 Valley. And that's where a lot of people were saying, oh, well, the content's terrible and it's awful quality and it's just going to pollute the internet. That's a very short duration until you basically can replicate all the steps that a person would go through to write a high-quality article. Just replicate that at scale. And I think one of the key points that even chat GPT-3 is not... It's kind of missing right now is the scalability component. So in order for you to get value out of it, you have to manually query it. So your marketing team has to say, okay, we're going to go do this or your sales team is manually figuring it.

41:46The area of opportunity for... In our viewpoint of the future is really scaled, fully automated workflows. So you give the AI, you kind of tell it what you want to do, or over time, just give it access to your systems, and then it will go and configure itself. And it will try to improve your business for you autonomously. So you can't really get there from chat. And so, but part of the backend infrastructure is very similar because in the world of autonomous business process agents, that is like an agent building platform. And so you do need to say, okay, I want to do this. Here's the list of tools you have access to.

42:33Go do it. now instead of reporting back to me report back to yourself and then figure out how you can improve from there right so what you want to do is is build these um you know extend the data flows that you have coming off of your current software stack so you have a newsletter right yep what do you do with the information uh the data that comes out of the newsletter performance nothing don't care right so is there value there sure right and it's and it usually comes down to like the team just doesn't have enough time to do everything right it's an afterthought to think about and to coach yourself there's no person on there so you could say every time the newsletter comes out at the end of the week tell us which stories got the most clicks and then what are stories like those that we can do more of boom and that's just set every week and it just sends them an agent so it's like google alerts or magic leap had this idea that you'd have agents out there doing things for you but if you take a google alert which is on your last name or your company name this is a google alert that can do more advanced things yeah exactly and and even the newsletter maybe you have like very high value contacts in your in your newsletter database Sure.

43:56Yeah. Right. And you're like, Hey, I actually want to personalize theirs. So I want to take this content and just literally personalize it down like one to one to them because I want every single message they get from us to be extremely high quality. Interesting. Right. So those are systems that are going to be you're going to be able to build. We're building out with pilot enterprise partners that we're working with now that are deploying some of these workflows into their existing business processes. fascinating the there the thing that compounds is once we set up a workflow um let's call it like the newsletter use case most newsletters would need it and so that is the thing that we're going to copy that template in your library like notion has templates or any other platform has templates and yeah that gets super interesting in a creative we did uh just as an example here's a couple of you know when we do when the when our researchers are going to do a google search like i'm doing a live show and they're like hey when did netflix launch when did it pass 100 million subs when did disney plus launch when it allows do 100 subs it's now getting that right it wasn't getting that right previously but here's examples of chat gpt doing that and it does seem like bard from google which i've been playing with has more recent information so in your mind how far behind is bard do you think in reality um probably the the measure that matters is the user base for the chat programs i think that would that would be the the difference maker there that ultimately will determine who wins so then google does win because even if open ai has tens of millions of users google has billions across chrome android youtube etc so google wins it's i wouldn't count anybody out i wouldn't count google out i mean google has they have all the data sets that matter yep so they should be able to go faster and have the users and have the browsers they have to really get their vision together about what they're building and for them you know their goal was to make um you know all human knowledge accessible and useful.