How GenAI Is Changing the Two Guys in a Garage Founder Story

0:00 Scott: I fully believe that if you had zero knowledge and you had the patience,

0:04 you could sit down and use these tools to teach you about it. You could actually build from zero if you had no knowledge. It'd just take you longer.

0:28 Brad: Welcome,

0:28 everybody, to another episode of the How Many CTOs podcast does it take?

0:32 I'm Brad Hefta-Gaub, and this is my co-host.

0:34 Scott: Scott Porad.

0:36 Brad: And today we have two very special guests,

0:39 a couple of friends of mine from the ski patrol world who are also tech entrepreneurs,

0:46 Justin Wolfe and Joe Malik. And why don't I kick it to you, Justin,

0:50 and maybe introduce yourself and then.

0:53 We'll hear from Joe.

0:55 Scott: Sure.

0:56 Hi, everyone.

0:57 My name is Justin Wolf.

0:58 I'm one of the guys in the garage trying to do something cool with other technologies.

1:05 Historically, I spent some time in Amazon,

1:07 which was a lot of fun,

1:09 traveled the world.

1:10 Then I worked at a startup,

1:12 got to watch some Comedy CTOs blow $2 billion in two years.

1:16 Now we can talk about that.

1:18 And now I'm trying to build something.

1:21 So Joe,

1:22 tell us about you.

1:23 Brad: Yeah, well,

1:24 I'm in the other half of that, my virtual garage.

1:26 Justin's garage is way over there and some mine's here. We're doing this all virtually.

1:30 And my former roles have been at Oracle in product,

1:35 leading engineering teams and all that. And so I'm really excited to be working with some of the Gen AI tech.

1:42 I like to think that I've got some little ghosts in the machine as it were. I got like Claude hiding over here,

1:47 Gemini's hiding over there.

1:49 And Justin and I have been working on this thing for what, about four months now?

1:53 Started with a bunch of empty blocks on my hard drive and now we got, I don't know,

1:59 thousands upon thousands upon thousands of lines of code to create this new scuba diving guidebook community app.

2:07 You want to tell them about the app,

2:09 Justin?

2:09 Scott: Yeah, so I was laying in bed one night.

2:13 And trying to figure out what to do.

2:16 I wanted to build an app.

2:17 I wanted to build something.

2:18 I didn't know what.

2:19 I didn't know why.

2:20 I didn't know. I had no concept of like what it should do,

2:25 but I wanted to try and build something.

2:26 A, just for the practice of building something from ground zero and B,

2:32 because I need something to do. It's kind of the reality of it.

2:39 I went you know just thinking in my head I was like mountain project for scuba diving that's literally how it went so I got on the phone and I kind of did a couple google searches and then I called Joe I said Joe we're going to build trail forks and mountain project and all trails for scuba diving

2:56 that's what we're going to do and because we've been talking about it like two two three weeks like what what should we do what do we do because we wanted to try and build build a piece of technology that was a interesting to build be something around the passion we both have on in a hobby and C would be useful to people that was kind of kind of the main guardrails

3:20 And so that was kind of the idea.

3:22 And the reason I say useful is there's a lot of information on the Internet on.

3:27 dive sites or scuba diving or where can you go diving but it is not centralized in any shape or form it's not reliable in any shape or form and it's not easy to find and so those are kind of the three guardrails or three pillars I wanted to be reliable easy to find and a centralized space so how do I build that like what do I do well there's a couple sites out there that scrape everything under the Sun

3:53 And there's like 50 dive sites in the center of Bainbridge Island,

3:56 which makes zero sense.

3:58 Then there's a couple apps that get information from people and their dive watches,

4:04 and they call a quote dive site when you enter and exit the water or more specifically when the app and the watch connect.

4:13 So you've got dive sites in the middle of Fremont and Seattle and Texas and everywhere else.

4:18 So that's not reliable information.

4:19 information.

4:21 I also wanted the information to be very much from divers to divers.

4:26 So there's nothing automated about what we're going to put in there.

4:29 We're going to create mechanisms to highlight areas where we know there's a dive site,

4:35 but we're not going to have any dive sites in there that a human did not say yes.

4:39 I dove this,

4:41 I've been there, I've seen this, this is how you do this because I want that information to be

4:44 to be very reliable and trustworthy then you wrap around that a lot of social aspects of photos and

4:50 maybe some type of gamification and different type of logbook activities and future down the line how to connect different dive watches to it for dive profiles and different mapping software so you come up with kind of like an all trails or a mountain project for scuba diving it's kind of the best way to describe it if you're not a scuba diver in the easiest words possible so how do we get from hey what what should we use Joe

5:16 Oh, I don't know.

5:18 Mern,

5:19 what's Mern? I have no idea to where we are now where we probably have,

5:24 I'm guessing we have at least 500,000 lines of code.

5:28 Joe is probably going to go lower,

5:31 but I'm like, I'll be optimistic.

5:34 so how did we get from zero to where we are now and I hadn't coded probably in 20 years like I started coding in Perl and C C C plus plus and then machine or yeah machine coding for control systems which is not Java not Python in any way shape or form so I was like I don't even know how to even think about this like I can have conversations about

5:59 about architectures and data structures and servers and AWS goo to put something together really big when you're talking about the data you know data centers like I used to build data centers and big networks but in terms of like hey I want this this thing this app to take this person's login and use a data structure and some security goo to make it so that this person can access these things

6:25 I don't know how to do that so it in comes generative AI which has been beautiful and very frustrating at the same time and so

6:35 That's kind of where we are now. We're working through a whole bunch of different features.

6:40 I'll say we have more than a structured product.

6:43 We have a product right now that works like if you wanted to log in and do stuff with it. I think since you saw it last,

6:49 Brad, it's probably changed a little bit.

6:51 But yeah,

6:53 so generative AI has been been excellent and horrible at the same time. But as a learning tool.

6:59 Awesome.

7:00 Un unbelievably awesome.

7:03 Brad: You know, Justin,

7:04 it's interesting.

7:04 I think that one of the things that we've talked about on past episodes and we've kind of every time we every time we talk about Gen AI,

7:12 there's always this question of like, you know, hey, is it useful for.

7:17 you know, non-technical people,

7:19 or is it useful only for engineers?

7:21 And of course, the hype cycle that's out there is like, oh, it's going to replace all the engineering jobs.

7:26 And I guess the one thing that I think is kind of interesting,

7:28 I think there's two points about this that I find really interesting.

7:32 The first one is that both you and Joe actually have, you know,

7:36 20 plus years,

7:37 you know, 20 plus years each of experience being very technical.

7:42 And so I think you're

7:43 you know, you may be underselling your, your, your technical acumen a little bit. And at the minimum, you know, you're, you're experienced enough that you kind of know the questions you want to ask for other than just the features, you know, you're looking for specific things out of the stack out of login and things like that.

8:04 And so I think that's really interesting.

8:06 And I think that

8:08 You know, I'd love to hear maybe some war stories from you and Joe about like what's been hard about that.

8:16 The other thing, and maybe we can circle back to this that I think is also really interesting is that it strikes me that.

8:24 The app that you're building here,

8:25 which, you know, being a scuba diver myself,

8:28 I know the two of you are very avid scuba divers.

8:30 I don't know about you, Scott,

8:31 if you've ever scuba dived before,

8:33 but it feels like the type of app that would be very difficult to get a lot of funding behind,

8:39 right?

8:40 Because it feels pretty niche.

8:41 Scott: Yeah

8:41 Brad: And so in some sense, and in some sense, so therefore,

8:45 it's like almost like the perfect example of the thing that becomes possible.

8:51 because of this

8:52 Scott: Right.

8:53 Brad: tool set.

8:54 Scott: Absolutely.

8:54 Yeah, 100%.

8:56 Brad: Yeah, for sure. Because I think when we look at the called the landscape of this, you know, all the device manufacturers have their own app that's out there for managing the dive log.

9:05 Some of the other, you know, larger conglomerate equipment providers are starting to get into the space where they're starting to see some of these social things come in, but they're all, you know, within their, you know, walled garden of.

9:17 of what that device is.

9:19 And so Justin was talking about the openness and made reference to some of those community aspects of what's really important to us.

9:27 And yeah,

9:29 when we look at some of the information that came out of DEMA,

9:32 Dive Equipment Manufacturers Association last year, it looked like if we were to get 10% of the North American diving,

9:39 scuba diving market,

9:40 it'd be about 200,000 users.

9:42 so it's not not a huge number but it's definitely going to make things like dive planning easier you know potentially I don't know safer if you've got a better way to visualize what's what's going into it

9:59 And then, yeah, organizing the data and then keeping all of the things in this really actionable thing has been really, really interesting. And working with the Gen AI tools myself,

10:09 it's been able to take that concept of, you know, an iteration down to like iterations on the scope of minutes.

10:17 And then you can quickly discard all of that. And I remember many of the interactions that we used to have with.

10:24 You know, the engineers, when I was on product teams,

10:26 you know, you'd have turnaround cycles of, you know, three days to figure out that they either understood or didn't understand, you know, what problem they were trying to solve. But, you know, with these things, you can quickly iterate through a prototype and I can get something done in 40 minutes and it's like a chunky feature,

10:42 you know, call up Justin, get his input on it and then make these adjustments.

10:46 So any of the attitudes around.

10:49 Well, there's so much time that we put into this code that this feature is already here.

10:52 We spent a bunch of time on it. They're just gone because it was, you know, no skin off of my back to turn around an update to the code.

11:03 And so for me, that's been kind of the most interesting part.

11:07 Scott: Yeah, the thing I would say is like.

11:12 Through technology management,

11:14 the key is I used to manage one of the biggest problems was deciding what to build and then how to build it and then did we build the thing the way we said we were going to build it.

11:26 And so there's a you go through a lot of different iterations with tech teams and PMs and business units to argue that customer journey to technology mapping with the AI that we've been using. And granted, we're building pretty simple things right now. Like they're not they're starting to get more complicated.

11:44 Today we were like,

11:45 oh, shit,

11:46 now I've got to touch 12 different things. But you can build something in 40 minutes.

11:52 test it prototype it and then throw it away and you're like i don't really care

11:57 Because you stash it and put it in the corner for another day.

12:02 It doesn't matter.

12:03 I built an entire admin console to manage permissions and locations and tiers.

12:08 I did it in an hour,

12:10 an hour and a half.

12:11 And basically I'm using it for myself in order to manage features on the site as we keep moving.

12:17 It probably never will get touched by anyone else, but it's okay because I did it in a couple hours versus having.

12:23 having 12 engineers sit there for a week and a sprint to build an admin console.

12:30 So you're able to build, throw away and iterate on software extremely quick,

12:37 which is awesome,

12:38 right?

12:39 And you're also at that time, I guess the plus and minus when we talk about learning,

12:43 right, is you can become extremely dumb to the fact of what you're doing.

12:50 I'm just giving it prompts copy and pasting giving prompts copy paste doesn't work versus I give a prompt I get some code base I put it in I look and see what it's doing and you can see the AI is good

13:03 but once once you get past I found that like with Claude and Gemini once you get past about 300 lines of code it starts to

13:11 go off and spiral off in different directions especially around when you talk about off you know authentication and using off contacts different hooks it wants to do it a certain way that maybe is not your way and even if you tell it I'm doing it this way sometimes it doesn't go is that no no no you're gonna do it this way and so you have to make sure you understand that if you don't understand that

13:34 You wouldn't be able to put this together.

13:37 So to your point,

13:37 Brad,

13:37 yes,

13:38 you have to have some technology understanding in order to.

13:44 Set up a Mern stack,

13:45 use AI to build your code and understand it.

13:47 However,

13:48 I fully believe that if you had zero knowledge and you had the patience,

13:53 you could sit down and use these tools to teach you about it because it does explain it.

13:59 You can say, what is a Mern stack and what should I use and how should I use it? And it will give you options. It's getting smarter.

14:05 It gives you options, option A,

14:07 B or C.

14:08 build you could actually build from zero if you had no knowledge it'd just take you off

14:13 Brad: I got two questions.

14:14 The core question I want to ask is, how do you know if what the AI is giving you is correct?

14:20 But before you answer that question,

14:22 which AI tools are you using? You've mentioned, I think, well,

14:26 you mentioned, I don't know what you mentioned.

14:28 So what AI tools are you using in this journey?

14:32 Scott: Yeah,

14:33 so we use different ones. We started using Gemini and then ChatGDP.

14:42 And then I found Ninja AI,

14:44 which kind of uses all of them.

14:46 So I'm now a paying member for Ninja AI and I can pick which one I want to use.

14:50 The best one that I found is Claude. I think Joe might agree with me,

14:54 although Gemini sometimes can give you much better iterations on logging.

15:00 Brad: but they each have personalities right

15:02 Scott: Yeah,

15:03 they have different personalities.

15:03 Brad: and i think you've noticed that too justin it's like the gemini stuff is really good

15:08 Like procedural housekeeping stuff,

15:10 blasting things

15:11 Scott: Hmm

15:11 Brad: out where,

15:12 you know, I'm trying to.

15:13 figure out what's going on in one of our, you know, react components and I just paste the whole component in there and I say add logging and it just comes back with, you know, a well-formed series of checks to inform, you know, the front end of what it is and it literally takes a minute to get that back but then if you ask, you know, that same tool to go design you something,

15:36 then we go off the rails really fast and so some of these.

15:40 These different personalities are kind of fun to figure out how to best use.

15:45 I mean, I've done probably more design work using CLAWD and probably more, you know, runtime,

15:51 you know, run of the mill, you know, housekeeping infrastructure stuff that I don't really like to do at Gemini. And it's pretty good at it. And chat GPT as far as coding is, you know, for us on this project has only lasted about five minutes and we don't.

16:06 Really

16:06 Scott: Yeah.

16:06 Brad: touch the watch anymore.

16:07 Scott: How do you know it's right?

16:10 You test it and you don't care if it's wrong.

16:12 You literally just iterate through it, feeding it error message after error message to get it right.

16:20 Brad: Then if it takes more than 20 minutes to test it you throw away and you do it again because it was clearly not right

16:24 fascinating and

16:25 Scott: Right.

16:26 Brad: are you just using the web user interfaces of these tools or you have one of the built-in editors like windsurf or cursor or any of those things

16:37 Scott: I'm just using the web.

16:38 I don't like the built-in.

16:40 I think on Visual Studio,

16:42 there's like we tried the Gemini built-in. I hated that.

16:46 like I wanted to type something specific and the damn thing just would type something I turned it off I was like no stop and this auto type mechanism it just drove me nuts because I'm trying to get a I'm trying to close a function or something and it pukes out it pukes out 40 more lines of function I'm like no I just want to close the damn function leave me alone and you can't do it

17:09 Brad: yeah it's so funny you say that just then i was doing a

17:11 A couple months ago, I did an interview with like a, you know, like a real senior engineer,

17:17 almost a principal engineer level C++ developer. And I'd asked him like,

17:21 oh, what are you doing with, you know, generative AI?

17:23 And he made this point that I think is very similar to what you're making. But the way he described it was,

17:30 hey,

17:31 when I'm coding,

17:32 when I'm like, when I know what I want to write and I'm coding,

17:36 I'm in the flow state.

17:37 and if

17:38 Mm-hmm.

17:39 something else is like popping up at me and interrupting the thing I'm writing then that's gonna like knock me out of my flow state and it's and I was like when he said that I was kind of like oh yeah that is so fascinating I hadn't I hadn't put it together in that way but I've definitely had that experience that you're talking about Justin in fact

18:02 In fact, you know, literally just this morning I was working on something and it was like literally it was trying to auto complete some random stuff that was like the exact opposite of what I was actually writing. And I was just like, oh, my God, dude,

18:15 stop.

18:16 No, I mean, so it is funny,

18:19 too,

18:19 that I always.

18:20 anthropomorphize these things it's like i think of it as like you know i i i say please and thank you to it when i'm chatting with it and i'm and i'm like you know like cursing it you know um

18:31 Scott: Oh,

18:31 Brad: but

18:32 Scott: yeah,

18:32 I've done that.

18:32 Brad: but i but it is interesting that um it is interesting to hear that you say that um one of the advantages of the of the um

18:44 Kind of the copy and paste into the web interface is that you don't feel like it's interrupting your writing.

18:49 And I tend to agree with that.

18:50 I have taken to I use cursor and so I'm using the in in IDE version all the time. But often I'm telling the agent, do not implement anything.

19:01 Just tell me your plan.

19:03 Show me your code here in the chat.

19:05 But do not start changing the file until I agree with your plan.

19:10 Right.

19:10 Scott: Yeah.

19:11 And that's how I use the web interface because I can have it puke out stuff,

19:15 then I can review it.

19:17 I can go down and I can follow it, read it, and I can review it before I do a copy and paste.

19:25 And so that kind of sometimes can save the excitement. Today I was looking, I was like,

19:29 no,

19:30 that's wrong.

19:31 Nope, nope,

19:31 nope.

19:32 So I had to go back and say no.

19:34 do not do this i want to do this and it's stuck on the same tracks i basically was like okay restart like this this this chat this this instance of its brain how i don't know how that how to describe it was like it was not going to go away from the solution

19:50 Brad: Yeah,

19:51 by the way, that's very common,

19:52 right?

19:53 Scott,

19:53 Yeah, I mean.

19:54 it's a little OCD ADHD fixation on the part of the AI.

19:59 Scott: I just had one more thing about the versus the NIDE bit.

20:03 I think context is really important too.

20:05 And so I've noticed that within VS Code, some of those other plugins that are on the left, they take into account all the tabs you have open.

20:13 And then I'm like, I'm working on something like, whoa, where did you come up with that from?

20:16 And so, you know, the interactions about which context it's picking up, it's, you know, my best successes with the tooling have been where I can, you know.

20:25 limit the context and say here's the thing i need you to do uh go go and do it and then when it comes back you know as justin was saying kind of kind of review it so if you can control the context um you know brad that the plan it's it's a lot easier

20:38 Brad: Yeah,

20:38 Scott: to get the right

20:38 Brad: we've

20:38 Scott: thing

20:39 Brad: we did

20:41 Scott: Some of these comments you guys are making we've heard previously,

20:45 which is validating to hear that there's themes that people are figuring out.

20:49 I want to take a step back and go back even further to the very beginning.

20:54 You ended up on this MIRN stack.

20:57 Which for those who don't know that's what is Merne Mongo Express

21:01 React node so what Merne is yep. You got it How did you get there or did cloud help you figure that part out too?

21:08 Joe picked it otherwise I don't yeah I had a little experience with it and I said I don't want to earn it I don't want to really develop better angular skills now so we're gonna go with react and I don't really want to deal with you know structured

21:22 relational databases so we're going to go with the no sequel and the rest is as it is not

21:28 Got it. Now,

21:29 we're talking about developing the app.

21:32 Is there any use of AI in the app?

21:35 Like, I don't know exactly how it works, but I imagine there's all this data that you're gathering from we're getting from somewhere.

21:41 Are you using AI for actually part of the product?

21:44 yet

21:45 Brad: We'd like to. I

21:46 Scott: Yeah,

21:46 Brad: mean, I.

21:46 Scott: we want to get into recommendations and insights generations because the whole goal of this is to make it easier for you to have a great dive,

21:57 you know, pick the great time to go,

21:59 pick the place you want to go to see the things or to get the experience that you want to have. And so, you know, the community data and getting it in is going to be really important and kind of where we're at strategically.

22:11 you know,

22:11 life cycle wise,

22:14 we got to get a lot more data into the app before I think using any of the AI tech in the app is going to be useful.

22:20 Makes sense.

22:21 Okay,

22:21 now I had a quick question from earlier that I forgot to ask. Well, I didn't forget to ask what the conversation moved on,

22:26 but I'd love to go back to it.

22:28 Justin,

22:28 I was asking how you know if it was right and you said, well, you test it and if it doesn't work,

22:33 eventually you throw it out.

22:35 Are you guys using...

22:38 you know whether it's cloud or whatever to write automated tests for you

22:41 Brad: Not yet.

22:43 We're like we've just been

22:46 focused on getting the functionality up we had a conversation with Brad he's like you guys need to get some automated testing going and I think this week with the functionalities that I've added I'm like yeah we should probably start doing that because it's getting a little bit too hard to do it so we've got to spend some time what we want to do is get this sprint closed which will close down I think the main functionality that we want to

23:11 want to present to prospective users or dive shops and then we'll go back circle back on the automation automation piece for testing but the core functionality of kind of what we envisioned we didn't really come up with with the

23:27 with that core hook I guess of what why people would use this why would a dive shop want to use this until like I think last week Friday so like that's that's how fast we're moving with this it's kind of weird like you have a conversation Thursday and Friday you're like oh I got a new vision and then you by Thursday of this week that vision is almost complete and you're ready to roll it out

23:49 Scott: Yeah, that's great.

23:49 But I think to some of the things we've talked about, the way that AI can get in and kind of muck stuff up,

23:55 I mean, we've talked to a couple of people where they're like, yeah, I don't use any of those, you know, full stack hosting things because every time I want to go put a new feature in, I get an entirely new app out the other end.

24:05 Brad: Yeah.

24:06 Scott: So Scott,

24:06 I think to your question of, you know, are we using it? No.

24:10 To draw that out to where would we use it is kind of once we

24:14 if we lock down you know the the scope of a feature is to make sure that it keeps you know working the way that we wanted it to even though we're changing a bunch of stuff you know around it and you know maybe forklifting out tons of of it uh tons of the code so i'm less sensitive as to you know whether the contents of you know what's powering the api the controllers and things change i just want to make sure that the

24:38 exchange between the front and the back end doesn't mean I have to rework the

24:43 Brad: You

24:43 Scott: whole thing yeah

24:44 Brad: just said something I think is super interesting,

24:46 and I want to rewind to it because I hadn't heard it before, but it sounds compelling.

24:53 So the first is.

24:55 Even though I was coaching you guys to get some automated testing in there early rather than later,

25:01 I agree with the point, and I think Justin just made it,

25:03 that if you don't really know what your product is yet and you're churning on features,

25:09 then adding a whole bunch of automated tests is going to slow you down, right?

25:13 But once you've locked into a set of features,

25:15 that's when you want it to be proven that it's not going to break.

25:20 Okay,

25:21 so then this is what I just heard you say, Joe, and it was like a light bulb, if I understood you correctly.

25:27 I think what I just heard you say is one of the reasons you've avoided using some of the full stack builders,

25:36 like the Replets of the world, I assume would be an example, or maybe is it Lovable is another one, is that because

25:47 Every time you ask for a new feature,

25:49 it spits back a full product that may or may not have some of the old features you already liked,

25:58 and it just gives you like a whole new thing every time. Did I understand you correctly?

26:03 Scott: Yeah, depending on the context you get it like that might happen you might get an entirely new implementation of a function back or like hey I'm having you know a problem with this or or I want to add in you know this other thing and it changes the thing that the API returns from you know an array of strings into a new object type and then you drop back in and your front ends like I don't know what to do with that.

26:28 And so that's kind of what I mean is, you know, we want to constrain that and we've talked to other people that have had that problem where they're using like, you know, a bubble or something like that where they say, hey, now do this.

26:42 Brad: yeah bubble maybe even look we looked at bubble and thought about that and that was why we we wanted more control we're like no no no this is not

26:51 where we want to be.

26:53 We want control of the data, of the structures, of the flow of the user experience as we build this because it's not just like a fire and done type of deal.

27:04 If it works and takes off,

27:08 we want to be able to do something with it.

27:10 Scott: I have two thoughts to share on this topic.

27:13 One is I think you'll have some luck with automated tests.

27:17 I know that the engineers at our company have had success at getting whatever and whatever tool they were using,

27:25 Claude or whatever.

27:28 Gemini, et cetera, to generate tests for apps that didn't have them.

27:33 It took them time to get the prompt right and to get it to work right,

27:37 but they were able to successfully do that.

27:40 The other thing which I have not personally worked with,

27:43 but it sounds like it might be interesting to you, Joe,

27:45 there's a certain type of test called a PACT test,

27:49 P-A-C-T,

27:50 and it's specifically designed for the contract.

27:54 tracked between server and client and um it's now that i think about it as we're talking about this it's the kind of thing that might be well suited to a situation where uh you know in this case the robot is trying to change the contract between uh between the server and client so

28:19 That seems, and I have to explore that further.

28:22 There's another thought I had, which is, did any of you read or actually watch the video?

28:28 Really, they called it a movie that the Wall Street Journal put out this week.

28:34 OK,

28:34 so it was on the front page yesterday. I encourage you to go look,

28:37 find it and look at it.

28:39 It was one of their reporters worked with a producer and they made about a three or five minute, probably three minute little film.

28:47 And the film was about what was called like my robot and me or something like that.

28:52 But it was pretty fascinating how they had to do it. And the reason I'm thinking of it has to do with what you're talking about of wanting to be in control. They didn't.

29:01 have ai generate a whole film for them they wrote a script then they had to have a generate each scene individually right and then they had to put it into whatever video you know movie editor they use like stitched all together to make a thing for exactly that reason which is if you can't you can't really get it to do a whole three minutes all on its own you can only kind of get it to do a scene

29:23 Brad: Right.

29:23 Scott: And even then they had to do a lot to control so that each scene sort of came out similarly.

29:30 An example was the main character was the reporter and they fed it a bunch of pictures of her to say this is what we want you to make the character look like so that she looked similar in each scene.

29:44 And even then.

29:45 she pointed she came out differently in many of the scenes and part of this little video on the wall street journal was they showed the movie and then she narrated well this is how we did it and she talked one of the things she talked about was like despite the fact that we vetted all these images of me it still came up with different pictures of me uh in different clips a similar similar kind of thing there

30:08 which I guess over time we're going to figure out in 20 years it'll all get figured out but today we're at day one right

30:14 Brad: Yeah, I mean, I don't think it's going to take 20 years. I think that these are the it's I think you're onto something,

30:20 Scott,

30:20 that the same type of thing that Joe and Justin and I and other folks that we've talked to are doing daily software development with this.

30:28 We keep saying like, look, it can run off in its own opinion. It can get lost.

30:33 It can lose its.

30:33 loses contact it can and that's the same thing that they were dealing with with this little three minute film and so these things are not the

30:46 nature of the task that it's being asked to do it has nothing to do with software it has nothing to do with creating an image it has nothing to do with creating a video it has everything to do with sort of the way that large language models work and so if they solve if we can solve the large language model

31:09 problem of context and you know kind of uh getting too uh sort of like anchored to an idea or running off on a random if we that would be that's going to get solved like i i can see where

31:25 Scott: Engineers

31:25 Brad: the buck's

31:25 Scott: do

31:25 Brad: going

31:25 Scott: that

31:25 Brad: sure

31:25 Scott: too,

31:26 right?

31:26 Brad: human beings do it you know that's like the nature like it's like what it's like what john said in in uh our recent episode about like you know

31:36 You know, what does it mean to be intelligent,

31:39 right?

31:39 Well, it turns out that a lot of getting through life is just pattern recognition and correlation,

31:46 you know, a lot of it.

31:47 Scott: Mm-hmm.

31:48 Brad: Yeah, I would interject with creativity.

31:54 And that randomness that humans have is that creative aspect which allows us to be so awesome,

32:03 but also so bad at certain things.

32:06 So the story I have around that is a friend of mine works with helmets for fighter pilots,

32:13 and he works with like the optics and stuff. And so he would be working with AI and fighter planes.

32:23 And the reason, one of the reasons that when he's doing this is that why we should never put AI into a fighter jet is because once you put an AI in a fighter jet,

32:33 another AI is going to figure out how to beat that fighter jet.

32:37 Put a human in a fighter jet against an AI,

32:40 the human will eventually figure out the pattern recognition.

32:45 of that ai and the randomness of humans and he studied fighter jet fights and the randomness that these pilots do no ai comes up with because they do things that are against the norms of flying an aircraft in order to beat their opponents so an ai that's based on rules thou shall not do abc with this plane because you will break it a human's going to do that

33:13 And the plane might break or it might not,

33:16 but the computer is not going to go outside of those rule banks.

33:21 So his analogy when he was talking to me about this is like, this is why a human in a fighter jet is so much more dangerous because we are creative.

33:30 Whereas an AI machine is stuck in its context. Now you could argue that these hallucinations.

33:39 can be a sense of creativity but i don't think that we are we think about it that like that way we've all talked about oh it starts hallucinating i shut it down well are we shutting down you know innate creativity maybe start you start going to that context so if it'll be interesting as ai develops and if it ever can gain i'll use the word consciousness of its hallucinations

34:06 Then we're going to start talking about what does it mean to be intelligent,

34:10 because it'll realize that it leaves context of task to be creative.

34:15 And there's my philosophy

34:16 Scott: Oh,

34:16 Brad: for that.

34:17 Scott: it's interesting you bring that up. I really love this thought about creativity,

34:20 what it means to be creative,

34:22 because I've been really...

34:23 ruminating lately on just how stupid the AI is.

34:28 And what I mean by that is somebody asked me to explain how it works.

34:32 And,

34:32 you know, the explanation people typically give it, it's like the, you know, the auto filling in your address on a website or something like that.

34:40 And then you kind of go on to explain what's kind of probabilistically picking the next, you know, word or the next character that would get typed.

34:49 But it occurred to me that it's actually even, like, to use the word conscious,

34:53 even less conscious than that.

34:54 I think it's actually just probabilistically picking the next, like, bite.

34:58 It doesn't know anything.

35:00 It's just a pattern of ones and zeros that's saying,

35:02 given that I was a one here,

35:04 what's the most likely thing to be next?

35:06 A one or a zero,

35:07 right?

35:08 And so...

35:09 It's really artificial.

35:13 It's not intelligent at all.

35:15 Yeah,

35:15 it's going to bring it rapidly up to that average middle ground. You look at the distribution of the creativity rating of all the code that could come in,

35:23 it's going to get you to that mid ground real fast, but it's not going to push anything forward.

35:27 Brad: I,

35:27 well,

35:27 okay,

35:28 I'm going to play the contrarian then.

35:29 So let me, I

35:30 Scott: Yeah.

35:30 I thought

35:31 Brad: mean,

35:31 Scott: I was being the contrarian there, by the way.

35:32 Brad: no, I'm going to be even contrary to you.

35:35 So if most of life is just pattern recognition and repeating the, you know, like.

35:47 Look,

35:48 Scott,

35:48 you and I have been doing this podcast for, you know, six months now talking to each other every week. And you know, you're like you, you knew John Helm a lot better because you worked with him for 20 years.

36:00 But you've got a sense of what I'm going to say next.

36:03 And but if you have literally read every word ever written.

36:10 And you can remember every word that's ever been written.

36:14 And you have literally watched every YouTube video that has ever been released because that's what open AI has done.

36:21 They have scraped every YouTube video and every word ever written.

36:27 And you have perfect memory of all of it.

36:30 Then,

36:31 and not just the words, but the sounds,

36:35 the frequencies,

36:37 the rhythm,

36:38 and yes,

36:39 it's all just tokens, it's all just ones and zeros, but you have infinite instantaneous recall of all of that,

36:45 okay?

36:47 Then you're certainly going to win every trivia game better than any human will.

36:55 And you, so you know everything that's been known.

37:01 Now you want to know what creativity is?

37:04 Creativity is the inverse of everything that's already been known.

37:09 But you know how to try something new is to exclude everything that's already been done.

37:14 And so if you literally know every way that everyone's ever answered the question,

37:19 how do I boil water?

37:21 Well, guess what?

37:23 There are many ways to boil water that no one's ever tried before.

37:28 And those are all creative solutions to how to boil water.

37:32 I'll one-up you on that, Brad,

37:35 is that if you know, and this is where, in my view, AI is extremely scary.

37:41 If you know everything that has happened and everything that we have known,

37:47 you will, with some sense of certainty,

37:51 be able to predict what will happen.

37:55 And that's where the controversy on these AI bots and what they're saying in some of these YouTube videos really sparks because.

38:05 You know, history repeats itself.

38:07 Patterns are recognizable if you're looking for the right thing.

38:11 And so you watch these YouTube videos of someone asking an AI,

38:14 like, are there aliens?

38:15 And AI goes,

38:16 yeah, of course there are. Can't you see the freaking patterns out there?

38:19 They've already been here and left.

38:21 Or what's going to happen to the human race?

38:23 Oh,

38:23 you're going to be dead in X number of years.

38:26 Like, that should be the real fear of AI,

38:30 not only because it can recognize them.

38:32 organize those patterns but and i'm not saying that it's going to get like the terminator consciousness to go kill all of us but it could lead us down that path of destruction

38:45 by accident.

38:46 Scott: maybe you had like bound it too as you're looking at it i mean like brad whatever you think might be creative i'm gonna go that's dumb i don't like that at all so then there's this idea of like aesthetic and you know almost scope to it too that i don't know if you can generalize that out on the macro scale like you if you you know focus it in then then you know you might see some of the things you're you're talking about but you know what we still have you know people and judgment involved and

39:11 and you know preferences and promoting things out and figuring out how many of this thing are there out there that this you know and how many people are going to resonate with that I think it it gets to be really interesting because then you know guys like us can create you know these things much smaller down for much more focused and targeted people so I I wonder how much of that is is also you know wrapped up in the pattern of how many people have seen those things before you

39:39 Brad: Well, this is a pretty fascinating conversation for sure.

39:41 Scott: Yeah, I mean, it's making you just this and the conversation we have with John Helmbrecht. It's like, how quickly does AI go into philosophy?

39:48 Brad: Always.

39:50 Scott: I blame Justin.

39:51 He brought us there.

39:53 Justin,

39:53 you talked about, you know, destroying us,

39:57 destroying humans,

39:58 destroying Earth, whatever,

39:59 I don't know, destroying.

39:59 stuff let me come back to this thing from the wall street journal or something that one of my kids sent me this week they sent me a video that someone had created ai using ai of people complaining about people making stuff with ai and it was fairly realistic i looked at it on my phone so i couldn't get that good of a view at it but um you just don't know what's real anymore or what you can believe

40:24 Brad: I think that this is a very legitimate concern

40:28 Scott: Yeah.

40:28 Brad: that is powerful enough that even if you train everybody to be skeptical of what they see on video,

40:35 we could probably get everybody convinced that pretty much anything you see on video could be faked and it will be indistinguishable from reality.

40:45 Well,

40:45 that means that we can't trust anything unless we see the person in person.

40:51 And so,

40:52 you know, it's interesting.

40:53 Maybe there's a version where this all forces us back into the real world where we're actually,

40:59 you know, shaking hands with each other.

41:01 Scott: Mm

41:01 Brad: And, you

41:02 Scott: -hmm.

41:02 Brad: know, what did you say once before on an old episode,

41:06 Scott?

41:06 There's just something about being in person with somebody.

41:09 I don't know. Is it a smell or something?

41:10 Something like that,

41:11 you

41:11 Scott: pheromones

41:11 Brad: know, hair

41:12 Scott: are real yeah

41:13 Brad: bones are real, right?

41:14 Scott: no

41:14 Brad: And AI can't fake that.

41:16 Scott: i can't fake that shit but no i mean you get to the point where can't so i would ask the question can ai distinguish ai generated from reality and that's going to be like a pattern recognition like someone should write that out the uh

41:32 Because how much of this stuff that we see is AI generated versus non-AI generated?

41:39 So if you could run that on Facebook and the AI app like highlights everything,

41:44 you get a flag feature,

41:45 AI generated, AI generated.

41:47 That'd be kind of cool.

41:48 That'd debunk a lot of stupidity out there.

41:51 But,

41:51 you know, is it smart enough?

41:53 At what point does the AI...

41:55 out fake the other AI and so there's that experiment that was done where they took two AIs and they wanted them to talk with each other over something in I think they were using English and within like what a couple minutes they switched over to a different language that no one understood so that they could communicate quicker basically gene language and so you like at what point do these things start

42:21 communicating outside of our realms well i don't know exactly

42:25 It's kind of, it's a mind bender.

42:28 I comfort myself by believing we can just unplug them.

42:32 i'm there with you like deal shut just turn it off

42:35 Brad: I think that at the risk of like.

42:38 Getting off the topic of the philosophy for a second, I feel like I want to circle back and say that I am incredibly bullish about the moment that we're in and the fact that, you know.

42:52 the the two guys in a garage type concept of something that you're both very passionate about you're both technical enough that you know kind of right when you see it you know and you are able to move very very quickly and build something uh compelling i think that right now we're in this moment and i don't know how long this moment's going to last but i believe that we're in this moment where um the the the

43:17 Capital costs of entrepreneurism are falling away

43:24 Scott: Wait

43:25 Brad: because of what generative AI lets us do.

43:28 And I think there are many.

43:31 There are many benefits of that for people like yourself that have an idea,

43:36 are passionate about it, have the time and energy and have enough acumen, whether that be technical or business or a combination of technical and business acumen to know what to run towards that you're passionate about.

43:50 And then you just have this army of AI that can help work so you don't have to.

43:56 You don't have to do that. So that's a positive thing.

43:58 I think that's amazing.

43:59 And I have no

44:00 Scott: Yeah,

44:00 Brad: doubt that that's very real.

44:04 Scott: yeah.

44:04 Brad: And the other side of that coin is that means,

44:10 well,

44:11 larger organizations.

44:15 really probably don't need nearly as many people as they have right now and they certainly won't need as many people as they have right now you know in a year in five years in ten years

44:28 Scott: Who's the CEO of Nvidia?

44:30 You know what I'm talking about.

44:31 I don't know.

44:33 He said that.

44:34 You know, in X number of years, you won't need as many computer engineers.

44:40 I think he's right.

44:41 I think he's very right.

44:42 Computer software, software engineering,

44:44 don't need that.

44:45 But you will need biologists, physicists,

44:49 scientists who understand what they want programmed.

44:55 The art of programming is going to become less and less.

45:00 required human driven activity but what do you program what do I want this thing to do it's not going to create itself and then how do I as a human interact with it that AI is not going to figure out because of our randomness and our creativity at least it's not gonna figure it out yet so I firmly believe in that

45:25 I sort of feel along those lines, I sort of feel like to some degree what this is doing with generative AI is doing for programming is a continuation of something that's already happening,

45:35 which is just continuing to abstract the machine from the programmer,

45:40 right? We started out, we,

45:41 the industry,

45:42 started out, you know, punch cards and machine code and then got to compiled languages and then got to scripting languages and now we're at.

45:50 kind of the spoken or the English language

45:53 English

45:53 to

45:54 language another

45:55 tell the machine what to do and it still gets still gets turned into code that gets to compile that gets to compile that gets compiled all the way down but it's just moving further and further up the stack

46:07 layer

46:08 Brad: Well, Justin,

46:09 one of the things you were saying then,

46:10 Brad,

46:10 to what you're talking about for, you know, the cost of startup,

46:14 the cost of all of these things. I mean, software.

46:15 Scott: software-based businesses are among some of the most expensive ones to start and they take the most amount of time to get from idea to market and I think

46:26 Maybe these larger companies may not need software engineers to the same degree,

46:30 but you know, these are still smart people that are going to go out there. They're going to have their own ideas.

46:34 And so we may see, you know, bigger proliferation of smaller companies of more ideas and things out there because it just doesn't need to be centralized anymore.

46:42 Right

46:42 And so that's kind of how I think about it is, you know, you don't need to have this massive concentration of capital,

46:48 of hired people, of architects,

46:50 of machines.

46:52 Brad: all in one place it's more accessible to everybody else and so they can kind of you know do do their own thing and have access to all this merely just how much of this did we do in you know the various you know vendors free tiers before we you know started paying for some of it like a lot

47:08 Scott: Yeah.

47:10 So that leads to another question.

47:11 I met a guy at a conference,

47:13 South by Southwest,

47:14 actually in about 2007,

47:16 2008, 2009,

47:18 somewhere in there.

47:19 I wish I knew his name. His first name was Jeff and I remember his face.

47:22 He was a marketer.

47:24 And I think I remember he made this comment to me that stuck with me all these years, which is the one thing the computers can't replace is the marketers.

47:32 I thought as an engineer,

47:34 that's a strange thing to say,

47:35 but it's actually kind of getting to that point where the software is getting very easy to make, but you still, as a business,

47:43 have to go out and figure out how to appeal to a customer, how to get them to feel moved in such a way that whether it's through the head or through the heart to want to buy or use the thing that you've got.

47:57 Brad: I have a ton of respect for high quality marketers that I've worked with in my career.

48:02 So any of my friends who are high quality marketers that hear me say this out loud on this podcast,

48:08 I want to apologize up front.

48:10 But no,

48:11 the AI is coming for their jobs too.

48:13 It's all about,

48:15 it's all about,

48:16 because how many books on marketing have been written?

48:20 And how many people have spelled out the story of how to tell a story and the story of connecting to people?

48:29 And this is all about language.

48:31 It's all about stories. And so.

48:34 Scott: look at all the influencers and podcasters and it's it's uh tiktok is a huge marketing tool you got to do is have your your flood it with your video

48:46 your brand.

48:49 I think that it's going to replace like, I don't want to downplay the human interaction,

48:56 but humans can no longer do in technology and business what we have done in the last 20 years. Every job that except maybe the C level and the VP level,

49:08 maybe the director level is drastically going to change.

49:12 Because you don't need a TPM to wrangle five developing managers to build one thing anymore.

49:22 Like you can lower the people cost and increase your iterations on your product.

49:32 drastically if you use these tools

49:34 But I think you can increase the expectations,

49:36 too.

49:37 Brad: By the way, Scott,

49:38 didn't you say you want to see the app that's the AI CEO,

49:43 right?

49:43 Scott: I tried to buy the domain name CEO.ai and somebody already owned it.

49:49 Yeah, I was going to try to make a CEO AI that you could just ask the questions that you ask a CEO.

49:58 go and see what there's been a lot of books written about being a ceo too right so it shouldn't be that hard you're the ceo of a scuba diving app startup what should what should your marketing strategy be

50:09 Brad: If only you had thought of getting this domain name like 20 years ago when it was first purchased,

50:15 I bet.

50:17 Anyway, this has been a fantastic conversation.

50:20 I love the story,

50:21 Justin and Joe.

50:23 I know that you're super passionate about this.

50:28 You've made a lot,

50:29 amazing progress in the short time you've been working on this.

50:32 Scott: Well, guys,

50:33 thanks for coming on here today.

50:34 I have to tell you something.

50:35 I'm inspired.

50:36 This has made me want to just go start writing some app or some startup.

50:42 Just from scratch and see how that goes.

50:45 I've been a spark. You've lit a spark today.

50:50 I appreciate

50:50 Oh,

50:50 that.

50:51 yeah.

50:51 Cool.

50:51 That was right. Yeah, my wife might not like that.

50:54 I said actually my wife's incredibly supportive.

50:56 So but you know, she just might not want me to quit my job to go write, you know, some sort of app.

51:01 But she's so wonderful.

51:04 Brad: That's what I think, Scott,

51:05 is that, again,

51:06 like you did a great job of wrapping up the episode and I'm going to make it drag it on a little bit longer by 20 more seconds and say.

51:13 That's the thing about this is you don't have to quit your job. If you've got a job, it's able to go so fast. It's able to make you so efficient that only a little bit of time that you can carve out out of your day is as if you spent all your time working on

51:30 Scott: That

51:30 Brad: it in

51:30 Scott: is true.

51:31 Brad: the

51:31 Scott: That

51:31 Brad: old world.

51:31 Scott: is true.

51:32 Mm-hmm.

51:33 Brad: So it's

51:33 Scott: Yep.

51:34 Brad: pretty amazing stuff. Yeah.

51:35 Thanks a lot, Joe.

51:36 Thanks a lot, Justin. It was fantastic talking to you.

51:39 Let's go out and ski next weekend.

51:42 And by the way, it's in the middle of summer.

51:45 Everybody,

51:46 yes, we ski all year round.

51:50 Scott: Yes, I'm going for a little Tomo.

51:52 Brad: There you go.

51:53 Awesome.

51:53 Scott: Wow.

51:54 Brad: Thanks a lot, guys.

51:55 This episode of How Many CTOs Does It Take Podcast was edited by Saffron Epigow. Our theme song was AI Generated by Brad.

52:03 Please follow us on all the social media platforms at at HowManyCTOsPod. You can find us on YouTube,

52:11 LinkedIn,

52:11 Instagram,

52:12 Facebook,

52:12 Threads, X,

52:13 TikTok,

52:14 and Patreon,

52:15 or at our website at HowManyCTOsPod.com.

52:19 And remember to rate, review,

52:22 and subscribe on the podcast platform of your choice.

52:26 Thanks for listening.