Valentino Stoll (00:00)
All right. We are live. Alexa, turn on the odd air sign. She's not going to do it. Welcome back to another episode of the Ruby AI podcast. We are joined by a very special guest today, Scott Werner and another cohost, Joe Leo and I, Valentino Sol. We are all together here today to talk about Ruby and AI.

Joe Leo (00:02)
Hey everyone.

She never does it.

Valentino Stoll (00:24)
Joe you want to say hello?

Joe Leo (00:26)
Yeah, I want to say hi. I

I thought the topic today was about how I can up my nerd t-shirt game. Because I've been showing up in the half-dressed, short-sleeved, worker man shirts. And I've just got old Garuko t-shirts that I'm too embarrassed to wear out anymore, so I just wear them to bed.

Valentino Stoll (00:33)
I'm sure.

Yeah, you know, I'm surprised there hasn't been a hyper specialized, you know, engineering related t-shirt maker that just like uses AI to like generate very specific like niche things of like, you know, in the Ruby world, there's so many of them. Like I would buy tons of them, you know, you know, per team, certain recurring. Right.

Joe Leo (00:51)
Mm-hmm.

there are, yeah. You could even, I would buy them all the time, you could even do it for like, you know, per team, like certain recurring bugs.

You know, it's like, you know, name error on that, you know, repos controller. That got us again.

Scott (01:11)
Ha ha ha.

Valentino Stoll (01:15)
That would be really funny, you know, like the swear jar or like another

company that used to have like, if you like made master, you know, red, you had to put like a dollar in the jar. And at the end, like when it filled up, they had a party or something. It's a good rule. You know, like get a t-shirt, like you have to pay, you know, make your own the t-shirt for the event, you know, like, yeah.

Joe Leo (01:23)
Mm-hmm.

It's a good rule. We had a, ⁓

Scott (01:28)
That's fine.

Joe Leo (01:36)
Yeah, I

like that. Back in the old days when everybody used to like come into the office together, I used to consult at Boeing in New Jersey and they had this system where...

The CI report was in these glowing LED lights across the office. And if you broke main, which back then was called master, your GitHub handle would be showing up in red. And so everybody would know that you broke it, which was effective.

Valentino Stoll (02:01)
so everybody would know that you...

That great.

wall of shame.

Joe Leo (02:07)
Yeah.

So Scott, we wanted you on here for a number of reasons. It's interesting as I was preparing for this interview, nominally it's like, yeah, we want to talk sub layer.

Valentino Stoll (02:09)
So Scott, we want to draw here.

Joe Leo (02:20)
But there's so much else that you are involved in. You've got the Works on My Machine blog, everybody, and Substack, which everybody should subscribe to. You've got this whole community on Discord, ⁓ and you've got all these different experiments that you throw out there to say nothing about artificial Ruby, which is how you and I met the Meetup. there's a lot to cover.

Valentino Stoll (02:29)
You've got this whole community on Discord.

Scott (02:30)
thank you.

Valentino Stoll (02:38)
to say nothing about.

you know.

Scott (02:46)
Yeah.

Joe Leo (02:46)
So why don't you just start by telling us a little bit about how you came to to ⁓ AI and Ruby and made it your main professional goal.

Valentino Stoll (02:47)
So why don't you just start by telling us a little bit about how you came to ⁓ AI and Ruby.

Scott (02:57)
Sure, sure. Yeah, I mean, to start off, I want to say thank you for having me on. I'm so excited that you guys got this going. When we started Artificial Ruby, it was basically, I had really only found the Ruby AI Builder Discord and there were only a few active people in there. And to see you guys connect and then start this podcast, it's just incredible. ⁓

Valentino Stoll (02:57)
Yeah, let's start off, I wanna say thank you for having me on. I'm so excited that you guys got this going.

Basically, I had really only found the...

connect and then start this podcast. It's just incredible.

Scott (03:20)
I'm super, I'm

super excited to see this and, and subscribe and continue listening to more and more episodes. But yeah, how did I get to make Ruby and AI my professional focus? So I've been a Rubyist for a long time. I, you know, started, I guess, somewhere around 2008 and was kind of like right around the time that

Valentino Stoll (03:24)
and subscribe and continue listening.

⁓ But yeah, how did I get to Ruby and

Joe Leo (03:35)
Mm-hmm.

Valentino Stoll (03:36)
So I've been a Rubyist for a long time.

somewhere around 2008 and kind of like right around the time.

Scott (03:46)
Why the Lucky Stiff, it was like right before Why the Lucky Stiff disappeared. And it was, he was super inspirational. It was very cool to see like him leading the kind of, know, having fun with programming and like showing that it can be an out, like a creative medium, which, you know, I think, I think a lot of us get into software because like we're excited about creating things. And then we go to school and it's like, you know, whatever.

Valentino Stoll (03:47)
It was like right before Why the Lucky Stiff disappeared and it was super inspirational. It was very cool to see like him leading the...

having fun with programming and showing that it can be.

Thanks

Scott (04:13)
optimize this algorithm. like, I know that fun gets kind of, you know, you lose the fun along the way. And ⁓ to see somebody like him and then see the Ruby community back then, you know, show that it's not like office space, which kind of dating myself, but you know, there was that there was that thread that like, the software and programming is, you know, is going to be like office space. And, you know, Ruby definitely wasn't at the time. So on my way there,

Valentino Stoll (04:15)
And like, kind of that fun gets kind of, you lose the fun along the way. And to see somebody like him and then see the Ruby community back then, ⁓ show that it's not like office space, which kind of dating myself. But there was that thread that like, the software and programming is, know, like, is gonna be like.

Joe Leo (04:15)
you

Valentino Stoll (04:40)
So I'm on my way there

and I've been through startups, through big companies like Groupon and Adobe.

Scott (04:42)
and have been through startups to big companies like Groupon and Adobe. My

last startup was a Ruby startup called SaySpring, which was kind of like right at the start of Alexa and Google Home. We built a prototyping platform for voice apps. So giving designers, product people the tools to build a voice experience and then test it out on device. We got acquired by Adobe, spent...

Valentino Stoll (04:51)
which was kind of like right at the start of Alexa and Google Home.

build a voice experience and then test it out on device. We got acquired by Adobe.

Scott (05:08)
four years at Adobe building out the team. That's now the product called Adobe Podcast. We had some different internal names before when I was still there. But Ruby has been the common thread throughout my career and seeing its impact on the industry kind of at the beginning of Web2 and how, it was kind of was kind of Ruby and Rails were kind of the language and the framework of

Valentino Stoll (05:16)
But Ruby has been the common thread throughout my career and seeing its impact on the industry kind of at the beginning.

and

Scott (05:35)
of IPOs throughout the 2010s. And, you know, Ruby was kind of the default when I started Sublayer. But along the way, like one of the things that I realized, you know, as you look at the, as you look at the industry, right, like every time there's a big shift, the more flexible, more ⁓ dynamic languages usually lead the charge at like figuring out how to interact with this.

Joe Leo (05:37)
Yeah, you're right.

Valentino Stoll (05:40)
kind of difficult when I started something.

that I realized as you look at the industry.

flexible, more dynamic languages usually lead the charge at like, bigger...

Scott (06:00)
right? And come up with the new ideas. So you had, you know, at the start of desktop, had, you had small talk and a lot of the ideas that came out of Xerox PARC, which then were kind of like formalized in C++ and Java. And then the web came along and then Ruby and JavaScript kind of led the charge. So we were trying to figure out what, you know, what these database back web applications were capable of. And, you know, the idea of putting together a big Java team to build something like.

Valentino Stoll (06:01)
come up with a.

the start of desktop, had a small talk and a lot of the ideas that came out of Xerox PARC, which then were kind of like formalized in C++ and Java, and then the web came along and then Ruby and JavaScript.

these database back web applications were capable of.

Scott (06:28)
Twitter or Groupon probably wouldn't have happened, right? And so when I looked at AI, this big platform shift, know, I skipped over something, Ruby and JavaScript, and now we're in this age of, you know, Rust and TypeScript and Go for a lot of the web 2.0 applications, mostly because like we're not coming up with new patterns, right? Like we've figured out how to stabilize, we know what we want to build.

Valentino Stoll (06:31)
And so when I looked at AI, it's big platform.

over something. The Ruby and JavaScript and now we're in this age of Rust and TypeScript and Go.

education.

figured out how to stabilize what we want to build.

But now we're going into AI.

Scott (06:55)
But now we're going into AI where we don't

really exactly know what we're going to build yet. Right? And so the dynamic, more flexible, easier to use languages offer the ability to kind of whip things together. Have an idea, whip it together really quick and show it to somebody. And I could show the code to you guys and it'll be, you know, under a hundred lines. And you'll be able to quickly understand it because of Ruby's, you know, the style of Ruby of like writing it very easy to read and,

Joe Leo (07:00)
Yes.

Valentino Stoll (07:01)
or flexible, easier to use languages offer the ability to kind of whip things together. Have an idea, whip it together really

show

style ruby of like writing and very easy.

Joe Leo (07:21)
Mm-hmm.

Scott (07:22)
heavy abstraction. that was my idea behind using Ruby is that it would make it possible to get a lot done. And also, as we're exploring these new possibilities, I don't have to spend a couple weeks to see if my idea actually makes sense.

Valentino Stoll (07:24)
That was my idea behind using Groovy is that it would make it possible to get a lot done and also...

Joe Leo (07:38)
Yeah, that's absolutely true. Yeah, I mean, I've always loved it for how expressive it is. I think that you hit on something when you spoke at Artificial Ruby a couple of months ago that I've carried with me, which is this concept that in ways big and small, the Ruby community of those mid-aughts through the next decade or so. ⁓

Valentino Stoll (07:44)
think that you hit on something when you spoke in artificial movement a couple of months ago that I've carried with me, which is this concept that...

in ways big and small, the Ruby community of those mid-aughts through the next decade or so ⁓

Joe Leo (08:04)
really shaped the next couple of decades of software engineering.

Valentino Stoll (08:04)
really shaped the next couple of decades of software engineering.

Joe Leo (08:07)
And they made large contributions to how we write software today. And so why not, as this paradigm shifts to AI, why not Ruby for making those contributions big and small again?

Valentino Stoll (08:08)
Amen.

Yeah, I totally agree. mean, I remember going to my first Rails comp, which was funny. Like, I feel like that's many people's intro to Ruby is through Rails, which is kind of funny because ⁓ Rails is so complicated of a thing, right? Like, you have to know all of the underlying principles that guided it to where it is today, right? Like, REST and, you know, obviously things have changed, but there's a bunch of building blocks that it was built for.

Joe Leo (08:30)
Yeah, it was mine.

Hmm.

Valentino Stoll (08:46)
You can't just jump into rails necessarily because you'll hit a bunch of edge cases and then just be like, well, I don't know what that means. We're more experienced, so it's really great because we hit those roadblocks and we just jump over them and we can move to the next thing. I feel like it's often skipped over Ruby.

for like just jumping into rails and Ruby is so much easier to dive into like I remember watching why the lucky stiffs, you know intro You know the the comic books and just being like this makes total sense and like doing a shoes app and you know, you could just like Literally have fun in very simple code style and that made much more sense to me personally and I feel like a lot of people miss that like Ruby entry point ⁓ and and I feel like

Scott (09:18)
Yeah.

Right.

Valentino Stoll (09:30)
You know, AI is kind of... I remember a talk from the last artificial Ruby I caught up where they... I forgot his name.

but he like intro'd like a, you know, learn Vim, right? Like through AI. Which is like, I use AI all the time for learning, right? Like it is, it's a great resource. Yeah, it's like, it'll throw a ton of information at you and like maybe some of it's not true, but like it gives you enough links and stuff. You could like go to the actual content. And so like you do end up exploring like, you know, things you don't know more because of that, right? Cause you don't trust it inherently, right? So it like makes you research, right? Like, which is like,

Scott (09:41)
Yeah.

Joe Leo (09:42)
Mm-hmm.

Scott (10:03)
Yeah.

Valentino Stoll (10:07)
Incredible like I would never research on my own right like I would be like Google search and like the top ten results and then be like well one of these has to have the answer that I want and I'll miss like a ton of content right because It wasn't surface and so with AI It'll just like glob together a bunch of stuff and I'll be like I have to start poking around and like finding the things like that interests me which I Prefer that learning style personally, but maybe it's also leading to missing other

Scott (10:08)
Thank

Thank

Joe Leo (10:32)
Mm-hmm.

Valentino Stoll (10:35)
other

stuff though. I don't know. But I feel like that like, you know, that like you said Scott, like that creative aspect.

where Ruby really shines. That's what it was made for, was just enjoyment, really. Matz was like, I'm tired of writing C, which is funny because he's still writing C. But he wanted to make something where you didn't have to. And it could be more enjoyable. And you could let that creative spirit fly. And definitely does it.

Scott (10:51)
You

Joe Leo (10:53)
I

know. ⁓

Valentino Stoll (11:03)
I love your monkey's paw. If anybody wants to have some fun, I definitely recommend checking that out. Do want to just like briefly describe like what monkey's paw is? Yeah. So, monkey's paw...

Joe Leo (11:11)
Yeah.

Scott (11:15)
Yeah, sure. So, so monkeys.pa

is a prompt based web framework. And the way it works is you write your prompts in markdown files in a wishes folder. those are it's kind of actually similar routing to like next.js where like the file name is the path. And it sends it to send it to an LLM and generates a page. ⁓

Valentino Stoll (11:21)
web framework and the way it works is you write your prompts in markdown files in a wishes folder and those are kind of actually similar routing to like next.js where like the file name is I love it the wishes. And it sends it to an LLM and generates a page ⁓

Scott (11:41)
And so the idea is that you can kind of focus on more of the

Valentino Stoll (11:41)
and so the idea is that you can you can kind of focus on more

Scott (11:46)
content and let the LLM just kind of hallucinate what the page might look like given styles or page content. I actually, used it at ⁓ artificial Ruby when I kind of released it to ⁓ actually do my talk and do my presentation.

Valentino Stoll (11:52)
Actually I used it at artificial Ruby when I kind of released it to actually do my talk and do my presentation.

That's awesome. I then used it, I have a that I was preparing for like a couple of...

Joe Leo (12:01)
I then used it, I had a presentation that I was preparing for like a couple of weeks later

Scott (12:04)
Or did you?

Joe Leo (12:09)
and I used it to generate the first version and it was bananas, it was so fun. But that's where I started. I started with the Monkey's Paw generation. I didn't know what I was gonna talk about, was about AI, it was about some of the latest trends, whatever it for kind of a general audience, not just engineers.

Scott (12:13)
Hahaha

Thank

Valentino Stoll (12:23)
you know, some of the latest trends.

Joe Leo (12:28)
I was like, You know, just let this thing rip. Yeah.

Valentino Stoll (12:30)
This is the exact spirit of Ruby right like, you know

something like that seems like fun and like it has a purpose but like, you know, it's very like whimsical in nature and like it just spawns like actual practical things like practical results like Emerge from it and I feel like that is where Ruby like really shines and why like I feel like Joe and I wanted to like start this off right like cuz This spirit is like gonna be the driving force

Joe Leo (12:45)
Mm-hmm.

Valentino Stoll (12:59)
of where AI ultimately leads, right? Because it's like a butterfly effect, right? Like, AI will just generate massive amounts of stuff, and massive amounts of stuff will generate from that, right? It's like a cacophony of generation at this point, right? Like, we're just like, like you said, you make a bunch of wishes, and a bunch of stuff happens, right? And people just keep repeating it, right? We're almost at this point of, I don't know,

Joe Leo (13:12)
Mm-hmm.

Valentino Stoll (13:24)
Like the AWS effect, right? Where like, you just spin up a new machine and then you cut it over, right? Like, you know, who cares like what happened before that, right? Like, it's almost like what AI is happening now where you're just like, all right, I want a new presentation. Like, I don't care what you generated the last presentation, like just generate a new one. Like, what could it cost, right? Like.

Joe Leo (13:34)
you

Scott (13:46)
Right, sorry.

Joe Leo (13:46)
Right. Yep. go ahead, Scott.

Valentino Stoll (13:48)
It definitely helped solve the blind page problem, right? You can get yourself to a starting point to start editing.

Scott (13:48)
It definitely helped solve the blank page problem, right? You can get yourself to a starting point to start editing and moving much faster, right? Everybody hates looking at the blank page.

Joe Leo (14:01)
Yeah, I know I do. I can certainly understand why it's helping me pass that, right? Whether it's like research or it's code or ⁓ saying what I want and then getting something back, even if it is totally imperfect, it's a good starting point. What I think is interesting, and I want to dive into this with you, Scott, is how much...

Valentino Stoll (14:02)
Yeah, I know I do and I can certainly understand why it's helping me pass that.

or it's code, or know, saying what I want and then getting something.

What I think is interesting, and I want to dive into this with you Scott, is how much

Joe Leo (14:25)
control do we give up when we say, okay, give us the boilerplate, you know, AI, give me the first draft. Because what I've noticed is that that first draft, sometimes it stays largely intact and sometimes it's changed entirely. But software engineering, as you've already pinpointed, is such a creative medium. So what is it that we, what is it that we,

Valentino Stoll (14:28)
Okay.

give us the boilerplate, know, AANA give me the first draft. Because what I've noticed is that that first draft, sometimes it stays large.

software engineering is such a creative medium. So what is it that we

Joe Leo (14:53)
give up and what is it that we gain when

Valentino Stoll (14:53)
give up and what is it that we gain?

Joe Leo (14:56)
we collaborate with artificial intelligence to get a result in software.

Scott (15:00)
Getting deep.

Valentino Stoll (15:00)
Indeed.

Scott (15:01)
You know, I think I look at it as I think collaborator is a good word. you know, I think you lose a lot if you just like throw a prompt out, get the output and then paste the paste the output without editing into whatever you're doing into an email or do a blog post. But I don't know, did either of you watch the Beatles documentary that came out a few years ago? The Get Back?

Valentino Stoll (15:01)
You know, I think look at it as, I think collaborator is a good word. I think you lose a lot if you just throw a prompt out, get the output, then paste the output without editing into whatever.

Venus documentary.

Joe Leo (15:29)
Uh-uh.

Valentino Stoll (15:30)
get back.

I'm gonna go a little out there with this. So there was a scene where Paul McCartney was sitting there just just noodling around and it showed like his creative process of like, know, it sounded nothing like it sounded a little bit like get back. But like each time he went through it, he like changed something he started like, you

Joe Leo (15:32)
Nah, I'm good.

Scott (15:33)
So now I'm go a little out there with this. So there was a scene where Paul McCartney was sitting there just just noodling around and it showed like his creative process of like, you know, it sounded nothing like, it sounded a little bit like Get Back, but like each time he went through it, he like changed something, he started like, you know,

just letting...

Valentino Stoll (16:00)
just letting,

Scott (16:02)
letting words

Valentino Stoll (16:02)
letting.

Scott (16:02)
come out of his mouth, not even caring about like what they were at first and just like kind of humming to a melody. And then it eventually like George came in and Ringo came in and John came in and they eventually kind of get back over a bunch of iterations. But it started with something that like wasn't as polished or wasn't as clean. And so I think, I think like, you know, if we think about programming as a creative medium and creative.

Valentino Stoll (16:08)
And then it eventually, like, George came in and...

and they eventually kind of built it back over a bunch of iterations. But it started with something that like wasn't as polished or wasn't as clean.

Scott (16:29)
collaboration with AI as a little bit of like that of just like, just give me some raw material to start molding around and, then see what we come up with. Like even I think I mentioned when I demoed Monkey's Paw, ⁓ it's, it actually kind of enables a new relationship with the software that you build because like I built it, but as I was using it, I was learning how to use it and just like trying things out. Like, huh, I wonder, I wonder if I could say.

Valentino Stoll (16:33)
some raw material to start molding around and then see what we come up with. Like even, I think I mentioned when I demoed Monkey's Paw, it actually enables a new relationship with the software that you build. Because like I built it, but as I using it, I was learning how to use it and just trying things

wonder if I could say

Scott (16:57)
hide this text and when you click this button have the text slide out and it wrote the JavaScript just perfectly and I didn't actually like write any specific code for that to happen. And so I see it more as like if you approach software from that, ⁓ you know, this isn't something solidified or something set in stone, which a flexible language that's easy to write and quick to get stuff done with enables you to kind of like

Valentino Stoll (16:57)
hide this text and when you click this button have the text slide out and it wrote the JavaScript just perfectly and I didn't actually write any specific code for that to happen. And so I see it more as like if you approach software from that, this isn't something solidified or something set in stone, which a flexible language that's easy to write and quick to get stuff done with enabled you to kind of like

Joe Leo (17:08)
Mm-hmm.

Scott (17:24)
those iterations and riff and mold it more like kind of clay rather than, you know, what else do you build? Steel beams and cement or something.

Valentino Stoll (17:31)
It's

Joe Leo (17:32)
Right,

right.

Valentino Stoll (17:34)
funny it sounds a lot like a baby learning how to talk

Right? Like, you know, I feel like as a parent, you'll like hear your kid like and understand them, even though like somebody else will come and be like, what the heck did they say? Right? but like you'll understand that. they like, you know, slowly like the gurgled and like slurred words and like mispronunciation like starts to transform. I wonder if we're just like learning how to like understand and talk.

the language of these AI systems, right? Like, it sounds very similar in nature. It's kind of like a wild, but like we've built this thing and now we're like learning how to communicate with it, right? Like.

Joe Leo (18:09)
Yeah.

Yeah, I feel that way as well. Scott, I wanted to ask you this. I read something on Works on My Machine that I thought was really smart and not necessarily what anyone else is saying about the future of AI and software development.

Valentino Stoll (18:24)
smart and not necessarily

and software development.

Joe Leo (18:31)
And for people who don't know Scott, I would define him as an optimistic

Valentino Stoll (18:32)
And for people who don't know Scott, I would define him as an

Joe Leo (18:36)
futurist. There's not a lot of AI ⁓ folks that really know what they're talking about that are talking in concrete terms with such an optimistic viewpoint. And I find it really refreshing. And I also think that it's grounded in a lot of reality. So when you released Blueprint,

Valentino Stoll (18:38)
There's not a lot of AI folks that really know what they're talking about.

So.

Joe Leo (18:56)
from Sublayer, right, because it's built off of or integrating with Sublayer, correct? So when you released it, you had a post where you talked about, like everyone else, the end of software development, right? But you had an interesting take on it where you said it may be the end of software development as we know it. In other words, simply writing for...

Valentino Stoll (19:05)
like everyone else, the end of development. But you had an interesting take on it when you said it may be the end of software development as we know it.

simply.

Joe Leo (19:19)
simply creating templated code or simply building off of abstractions. However, if you look at it as a higher order of software engineering whereby we are creating abstractions, it's only the beginning. And that's why I wanted to get deep with you on the last question because I kind of want to understand this concept better. So what can I do, you know, the average software developer with, you know, 10, let's say 10 years of experience?

Valentino Stoll (19:19)
simply creating.

distractions.

That's why I wanted to get deep into the last question, because I kind of want to understand how that's said.

Scott (19:39)
Mm.

Valentino Stoll (19:41)
than I do the average.

Let's say 10 years.

Joe Leo (19:49)
I certainly know abstractions, I certainly use abstractions, but does that really mean I can create them and how do I learn how to do that?

Valentino Stoll (19:49)
I certainly know abstractions, I certainly use abstractions.

Scott (19:56)
Whew! Well, so yeah, think the follow-up to kind of what I was, my answer to your last question was that like, know, Paul McCartney starting with, you know, like a rough, you know, rough music. And, you know, not everybody can do that. You know, it took his taste and like the Beatles taste to be able to come up with, get back out of, you know, that raw material that they had.

Valentino Stoll (19:56)
⁓ Well, so yeah, think the follow up to kind of what I was, my answer to your

Joe Leo (19:57)
Ha!

Right.

Valentino Stoll (20:09)
music and you know not everybody can do that you know it took his taste and like the Beatles taste to be able to come up with get back out of.

Joe Leo (20:18)
Yeah,

right. I've been doing this for 20 years. I'm not Paul McCartney of Ruby, you know? so, yeah.

Valentino Stoll (20:20)
doing this for 20 years, I'm not Paul McCartney.

Scott (20:22)
I back there, I can barely play what Paul McCartney plays.

Valentino Stoll (20:31)
But yeah, I think that's another example. I think that we are going to be forced to ⁓ develop that taste and develop that skill of building these abstractions.

Scott (20:33)
Another example, think that we are going to be forced to ⁓ develop that taste and develop that skill of building these abstractions. But I think...

You know, one of the things that AI enabled is a lot more, a lot more opportunity to do that quicker, which I think, you know, the, the 10 years in the industry is probably going to be really tough. think somebody new in the industry is going to be very, very set up to be able to kind of get to where everybody else is a lot quicker.

Valentino Stoll (20:55)
I think the ten years in the industry is probably...

Are you set up to be able

Scott (21:09)
Because they're not starting with any preconceived notions, no preconceived opinions. ⁓ The people that have gotten far enough in their career to start experimenting with building libraries and building their own abstractions are going to get superpowers. But it's like that middle is really going to be really tough, right? Because you're kind of stuck. You haven't had the chance to really get... ⁓

Valentino Stoll (21:10)
because they're not starting with any preconceived notions, no preconceived opinions. ⁓ The people that have gotten far enough in their career to start experimenting with it'll...

Joe Leo (21:10)
It's interesting.

Valentino Stoll (21:32)
chance to really ⁓

Scott (21:34)
to get much exposure to building an interface, building some abstractions, building some libraries, and then seeing the effects. ⁓

Valentino Stoll (21:34)
get much exposure to building an interface, building some abstracts, building some libraries, and then seeing the effects.

Scott (21:44)
And so you have all these opinions and these thoughts build up, but you're really focused on where you are in your career. Which actually reminds me of something similar happened with poker. You guys remember back in the early 2000s how everybody was playing online poker all of sudden.

Valentino Stoll (21:45)
And so you have all these opinions and thoughts filled out, you're...

You guys remember back in the early 2000s how everybody was playing.

Joe Leo (22:01)
Yeah.

Scott (22:02)
I was one of those people. But you actually saw this split in professional poker where you had the people that had been playing their entire lives playing in person and they were at the top of the game, but they were also much older. And then you had this group of people that were mostly young people. so the tables were kind of split with a lot older people that had been playing for decades and then people that had been playing for a few years.

Valentino Stoll (22:02)
I was one of those people. But you actually saw this split in professional poker where you had the people that had been playing their entire...

Joe Leo (22:03)
Yeah.

Scott (22:27)
Because those people that had been playing for a few years were playing like seven tables online at once, being able to play like all hours, the hands were dealt so much faster. And so they saw, they got the reps in so much faster. Whereas like if you were, if you had spent like 20 years playing in person and weren't like a computer person, like you were stuck learning slower than the people that had just started like a couple of years ago because they could have.

Joe Leo (22:34)
you

Valentino Stoll (22:37)
saw they got the reps in so much faster whereas like if you were you had spent like 20 years playing in person

learning slower.

Scott (22:53)
They could play, you know, hundreds of thousands of hands in the time that took you to play 5,000. And then also like poker tracker and be able to analyze and see the effect. Whereas people new in the industry are going to be able to put together designs, see how, see how they like them, see how the AI can interact with them and get that feedback so much faster than you might've gotten in the, in the industry the past, however long, decades.

Valentino Stoll (22:59)
and then also like Hover Tracker to be able to analyze.

together, design, see how they like them, see how the AI can interact with them, and get that feedback so much faster than...

So I'm curious what your thoughts are in here, because there's kind of like two things that could potentially happen, right? We're like, because our current bottlenecks, as you mentioned, on works on my machine are, you know, the meaning, what is it? Meaning making work, right? Like trying to make sense out of like what things mean.

Scott (23:37)
Any minute. Yeah.

Valentino Stoll (23:42)
which is really interesting. But it's almost like the user feedback loop, right, of our current, like, what we're building things for is like, that is the ultimate bottleneck of like what you're creating things for, right? Like if...

If they don't give you any feedback, like it'll just keep making stuff, right? Like, and you'll just keep making stuff. And it's the same problem, like in a business, right? Where you're creating a whole bunch of things that you don't know if it's going to work or not. And you just hope people keep paying the bills, right? And so I wonder if like, we'll see the opportunity for like failure early, right? Like is this thing worth building? Right.

Joe Leo (24:10)
Yeah.

Valentino Stoll (24:20)
Will we see more of that or will we see more of better home products? It's almost like two pathways where you either build something that's useful or you realize really quickly that it's not gonna be useful.

right? Because you can generate things so quickly. And I just think like back when I was building things on my own, like I would have loved to know early if it was like not going to be something fruitful. Right. And like, I feel like I've known so many people that are just like they've gone and they're just like, this is my idea. I'm like, I need to see this through. Right. And it's like a lot of times it's just like not a product market fit and like it will just never happen. And so I'm curious, like we're kind of getting like the best of both things happening at the same

Scott (24:59)
time.

Valentino Stoll (25:04)
time where people get to see the product market fit way earlier because it's so quick to build or if it's just going to lead to a bunch of stuff where you kind of use here and there and you just have so many products that you're bought into. What are your thoughts on that aspect of the AI as a product building nature?

Scott (25:25)
think, I guess Joe kind of alluded to it in the beginning, we're kind of doing a lot of different things at Sublayer. ⁓ I think that's what it's going to start looking like, because if you follow the traditional two-week sprints or whatever, you would build for two weeks, you would get that in customers' hands, and then you would learn from customers interacting with what you built.

Valentino Stoll (25:28)
guess Joe kind of...

You would build for two weeks, you would get that in customers' hands.

Scott (25:49)
When you compress that building time down to two days or two hours, you still want that two weeks of customers interacting with the product and learning. And so you're not going to do like another two hours and another two hours, another two hours, like kind of back that up because like you're not going to actually learn from what you're doing. And so like, I think what, I think what we're going to start to see is a lot more, is doing a lot more things in parallel.

Valentino Stoll (25:50)
that building time down to two days or two hours, you still want that two weeks of customers interacting with the product.

That was about it.

Joe Leo (26:02)
Hmm.

Valentino Stoll (26:08)
think what we're going to start to see is a lot more.

Scott (26:15)
which has traditionally not been what people tell you you should do. But like, what else are you going to do? Right? Like you're just going to, you're going to do the, you know, do the week of features in two hours and then sit there the rest of the week. ⁓ You know, maybe you're, you know, maybe you're doing an event, a newsletter, a podcast, you know, building another product.

Valentino Stoll (26:15)
which has traditionally not been what people tell you you should do. But, like, what else are you going to do? Right? Like, are just going to do the...

Joe Leo (26:31)
Mm.

Valentino Stoll (26:32)
Maybe you're doing...

Joe Leo (26:35)
Hmm.

Valentino Stoll (26:39)
Are you telling me we have to revisit Y Combinator mathematics again?

Joe Leo (26:40)
Just for example. That's right.

Scott (26:42)
you

So I think that's one of the things that I think is going to change where we're going to do a lot more. I think to where your latter point, I think there are going to be a lot more apps to switch between on the customer end. I think there's going to be a lot more stuff to...

Valentino Stoll (26:45)
So I think that's one of the things that I think is going on.

So where do you focus your efforts in relation to that user feedback turn aspect of things? Where do you find AI works well in that chain?

Scott (27:15)
as in like supporting... ⁓

Valentino Stoll (27:17)
supporting

your products from the, you know, the usage.

Scott (27:21)
We haven't really done much. I guess it helps us kind of categorize or organize a lot of the feedback and what am I trying to say here?

Valentino Stoll (27:22)
guess it helps us kind of categorize or organize a lot of

Scott (27:31)
You know, we, guess we've mostly been focused on using it to getting, get to that point of like, get something, get it out. And then we're going to learn, we're going to like, kind of learn from this while we're doing other things. because there's so much going on that like that's that I think has been the thing of just like always, always doing more things because like AI is moving so fast and you need to kind of be able to keep up. And so it's been less about, ⁓ less about focusing AI on the customer feedback side.

Valentino Stoll (27:31)
You know, we, guess we've mostly been focused on using it to get to that point of like, get something, get it out. And then we're gonna learn, we're gonna like.

going on, that I think has been the thing, just like always, always doing.

been less about focusing AI on.

Scott (28:00)
and more about getting us to the point where we're going to get customer feedback.

Joe Leo (28:05)
Yeah, I think you nailed it with the split, meaning that yes, we can build very fast, but AI has not sped up the customer feedback process. A person still has to go try out this thing and then give you something, hopefully, that you can use to build on the next iteration.

Scott (28:28)
Right. And so imagine, imagine that, that experience from the customer side, like, you know, we're only seeing the beginnings of this change in like most software where people are going to be able to release a lot more features, release a lot more products, new products coming out all the time. And, you know, how are customers going to handle that? Where, you know, you.

Valentino Stoll (28:29)
And so imagine that experience from the customer side. We're only seeing the beginnings of this change in most software.

release a lot more features, release a lot more products, new products coming out all the time, and how are customers going to handle that? ⁓

Scott (28:52)
You know, look at how many people get outraged when like a favorite site changes their design, like Facebook changes their design and so many people are frustrated where, where did this button go? Like, or, like, how do I do this thing now? Uh, that's going to start happening to all their products a lot faster. Um, and you know, we're already kind of experiencing a little bit in AI, like, um, like look at, think about how much work we spend keeping up with what.

Valentino Stoll (28:54)
many people get outraged when like a favorite site changes their design.

frustrated.

to start happening to all the...

Joe Leo (29:09)
Hmm.

Valentino Stoll (29:11)
and we're already kind of experiencing a little bit.

Think about how much work we spend keeping up with what

OpeningEye is doing, and Anthropix is doing, and Jim and Google is doing, and Mitchell and all of the other platforms and all the other...

Scott (29:20)
OpeningEye is doing, and Anthrombic is doing, and Google is doing, and Mistral, and all of the other platforms, and all the other labs.

Joe Leo (29:24)
Mm-hmm.

Scott (29:27)
And that's a full-time job, right? And they're all kind of focusing those tools and those services internally. And whatever is the best for a particular thing changes week by week or day by day, right? I have a feeling customers are going to have a really

Valentino Stoll (29:36)
whatever is the best for a particular thing changes week by week or day by

Joe Leo (29:43)
Yeah, absolutely.

Valentino Stoll (29:43)
I have

a feeling like customers are...

Scott (29:48)
us as product builders and then customers are going to have really tough time dealing with this like, ⁓ just over supply and over like changing in velocity of like the way that all their products have been or that they're used to their products delivering.

Valentino Stoll (29:49)
us as product builders and then customers are going to a really tough time dealing with this like, ⁓ just oversupply and over like changing.

Yeah, mean, circling back to the meaning making work, Like there's the processing of user feedback meaning making, right? Like, you know, even in a Claude prompt session or like coding assistant session, like you give it feedback and it's still like, doesn't really know, right? Like it'll make some guesses and like try, it's more helpful, but like still doesn't really solve things all the time, right? And so we're still like.

Scott (30:21)
right?

Joe Leo (30:24)
Mm-hmm.

Valentino Stoll (30:30)
that were critical at that point too, right? So like, you know, that meaning making aspect of the user feedback is like more of like where the work is gonna land, right? And so it like makes me think like...

Scott (30:39)
Yeah.

Valentino Stoll (30:42)
I've been using all these coding systems to just knock out side projects left and right of things, just ideas that I've had, but I'm not maintaining any of those. And so I worry there's going to be a lot of that out there. hopefully somebody makes a maintainer AI assistant or something. But even then, I don't know. I worry about the quality.

Joe Leo (30:52)
Yeah

Scott (31:03)
you

Valentino Stoll (31:12)
of products like dissolving in the near term.

Scott (31:16)
Yeah.

Yeah, I mean, I I was just going to say to Valentino, to your point, like, you know, I think that that that part around understanding the feedback is the is a lot of the taste of like, what is this person saying? What is what pain point are they expressing that, you know, is going to be what are they expressing? And then do we actually want to do anything about it? Right. Like, is this the way that we're taking this product? And that's something that

Joe Leo (31:17)
So Scott, sorry, you continue.

Valentino Stoll (31:17)
Yeah, I mean, I think. I was sorry, you continue. I was just going say to Valentino, to your point, like, you know, I think that's that part around understanding.

that is going to be, what are they expressing and then do we actually want to do anything about it? Is this the way that we're taking this product? And that's something

Scott (31:43)
⁓ that it's going to be, I'm skeptical that LLMs are

Valentino Stoll (31:44)
that's going to be unskeptical that that one.

Scott (31:46)
going to be able to like have that answer of like, is this what you want to do?

Valentino Stoll (31:51)
Cause I think even back to my time of like diving in a sub layer and be like, I found this stupid bug. Like, let me go to the discord channel and say it. And then I'd be like, I found it. And then it's like, well, I, you know, wasted your time, like bringing this bug that wasn't really a bug. And it was really just like something in my, you know, bash configuration or something, right? Like, and like, I saw, I wonder like, cause I would love like a tool at the point of use where I could just be like, okay, submit this feedback to the project. And like, you don't necessarily have to see it and maybe like,

Scott (32:09)
Hmm.

Mm-hmm. ⁓

Valentino Stoll (32:20)
and LLM or AI or something is like processing that, but like, you know, where does that end up to you so that you can make more sense of that too, right? Because like, the AIs are great at like filtering and like helping with some things, but like it becomes that critical point of like, all right, you need that human somewhere in the loop, right? Like, so like where does it fit, right? Like where, excuse me, not it, but where do we fit?

Scott (32:40)
you

Joe Leo (32:43)
A little slip.

Scott (32:44)
Yeah.

Valentino Stoll (32:44)
Hahaha

I don't know. So I have a hard time finding that like human placement, right? Like where does the human best fit in the artificial process, right? Like, do you have like any guidance there on like where you look specifically? Like, are there any tells where you're like, like, obviously this is something like we need somebody to like intercept.

Scott (33:05)
Huh, that's a really good one. I...

think I've come up with any kind of like good heuristics or anything for that. I've definitely, you know, I guess I haven't really examined, that's a really good question, I haven't really examined, I've kind of just been like using...

Valentino Stoll (33:18)
I haven't...

the

standard that we've set in software.

Scott (33:24)
the standard that we've set in software

for a while. If there's going to be code committed, I want to make sure that there's a PR review from the AI. If there are going to be external impacts beyond just me, definitely have a human set of eyes on it. But yeah, that's a good... And I mean, I guess I...

Valentino Stoll (33:28)
there's going to be code committed.

If they're going to be external impacts beyond...

And I mean, guess I,

Scott (33:48)
It's mostly been

Valentino Stoll (33:48)
it's mostly been.

Scott (33:50)
just relying on intuition. Cause like, think, ⁓ like one of the latest works on my machine posts, I, ⁓ set up where we're changing our team photos every day based on a historical fact. And I commented, I commented to somebody that like, I actually don't even check whether this historical fact is actually real because it doesn't actually matter. All it is, is just like feed, you know, feed, feeding into the image generation. like.

Valentino Stoll (33:52)
I think ⁓ one of the latest works on my machine posts. I set up where we're changing our team photos every day based on a historical fact. I commented to somebody that I actually don't even check whether this historical.

Joe Leo (34:01)
Right, I read that. It's good.

Yeah.

Valentino Stoll (34:12)
is just like feeding, you know, feeding.

Scott (34:17)
I could be lying that like, you know, something about the Beach Boys happened today in history.

Valentino Stoll (34:17)
could be lying.

Joe Leo (34:22)
Right. Yeah, when the stakes are low, it's easier to let things, you know, let the LLM kind of do its thing. You know, and I think, just taking it back to Monkey's Paw, there's a little bit of that too, right? When I generate the first draft of something, it's for me, right? I can let the LLM kind of hallucinate and do some things and maybe it'll be interesting.

Valentino Stoll (34:23)
When stakes are low, it's easier to let things...

Scott (34:39)
Yeah, yeah, and actually I think that's something thinking about like that question. That's something that's been on my mind a little bit too around, you know, getting out of your comfort zone and really questioning, do you actually need a person to review this? Which, you know, I was having a conversation a while back around kind of like as Ruby and Rails were getting big.

Valentino Stoll (34:40)
Yeah, and actually I think that's something I'm about.

There's something that's been on my mind.

coming good, getting out of your comfort zone and really questioning.

Scott (35:00)
the sentiment, I don't know if this is widely in the industry, but the sentiment that the person shared was that it looked like from the Java standpoint, that like, we had solved types and the Rails community came around and were like, yeah, but let's do none of that. it's like, can still makes people uncomfortable, right? But I see AI as the opportunity to kind of like really push on, like, do we actually need to be uncomfortable about?

Valentino Stoll (35:01)
sentiment that is widely in the industry but the sentiment that the person shared was ⁓ that it looked like from the Java standpoint that like we had solved types and the Rails community came around like let's do none of that and it's like I think Rails still makes people uncomfortable right?

Joe Leo (35:16)
Right.

Valentino Stoll (35:21)
But I see AI as the opportunity to really push on my...

Scott (35:27)
this in the same way that like, do we really actually need to know that this is a string or whatever type this particular thing is, or can we work with it for a while and get something useful out of it? I think AI gives us that opportunity to really like put that into question and like, Joe, to your point around getting that first version done, we've been using Bolt for, you know, prototypes and like...

Joe Leo (35:28)
Mm.

Valentino Stoll (35:35)
it.

Joe Leo (35:50)
Yeah, me too.

Scott (35:53)
a few minutes, like I can share a prototype with my team of like this is the idea and everybody can interact with it.

Valentino Stoll (35:56)
This is the idea.

Joe Leo (35:59)
Yeah, of course, you know, I, I joke that I built, so I built phoenix.love, our landing page using Bolt, but I didn't release it using Bolt. had one of my, one of my engineers had to go back and, you know, fix all the vibe code and stuff, you know, but yes, it got, it got the point across really quickly to the marketing team. They were look at it and say, yes, this, no, this, change that.

Scott (36:11)
Right.

Valentino Stoll (36:19)
this, no this,

Joe Leo (36:22)
It's fascinating. And I think you've hit the nail on the head. I said this in the last episode. I'm far from ready to answer this question, but the fact that a PR comes up and something that has been written entirely by an AI and LLM, that introduces a concept into the dynamics of working teams that has never been there before. And it happens over and over again. And how we, as a...

Valentino Stoll (36:23)
I you hit the nail on the head. I said this in last episode. I'm far from ready to answer this one.

you

This is a concept.

Joe Leo (36:45)
well, individually and then collectively as a team and then collectively as an industry, answer that question, I think, is really fascinating. I don't have that answer. I think we'll learn it over time.

Valentino Stoll (36:51)
really fast. don't have.

Yeah, that's a really interesting point too. It's almost like roles are dissolving a bit, right? Because anybody can put together one of these prototypes, right? If a business...

executive wants to explore what a new interface would look like, he could go to Bolt New 2 and like, what would this look like? But also at the same time, some customer support person could be like, well, what would a new tool to help me do this thing look like? That would help my day job and help other team members. Everybody is almost experiencing the same thing and the rules are almost like re-arranging.

Scott (37:30)
Yeah.

Valentino Stoll (37:36)
themselves right where like you know even I was the other day somebody on my team at gusto they were using their designer and they're using rep lit to like generate a full-blown prototype of the front end of a new thing that didn't even exist and be like hey what do you think of this new like way of organizing and restructuring the data and like interface and does this work and I was just like the feedback loop was like so tight

Joe Leo (37:37)
Mm-hmm.

Scott (37:50)
Hmph.

Joe Leo (37:56)
Mm.

Valentino Stoll (38:01)
You that like you could just get through a ton of meetings You didn't even have to have meetings at that point, right? Because you could just be like here it is like what's swath, right? And like everybody can play with it and be like leave comments on it like just like the you're right like everything you know Scott is like changing to like You know of how we do things and so it'll be interesting to see kind of where that leads ⁓ and

Joe Leo (38:09)
Yeah, so everybody loves that. Yeah.

Scott (38:26)
Yeah.

Valentino Stoll (38:28)
I don't know. I'm curious to see how it affects the Ruby language itself too. Because, you know, like there's a bunch of things that have just been roadblocked over time.

for just like utility purposes, but also like the whole namespace thing for those aren't familiar, like, know, namespacing in Ruby is like, has been a challenge for years, right? Of like conflicting of where it's nested, where it's required, and so there's a new proposal to try and like jump past that and like,

Joe Leo (38:49)
Mm-hmm.

Valentino Stoll (38:59)
help lock things down better in isolation. And it's really exciting. And I have a suspicion that AI is somehow involved in pushing this thing forward, right? That's interesting. Yeah, I read that.

Joe Leo (39:03)
Mm-hmm.

That's interesting. Yeah, I read that and I didn't

even think that, but yeah, you're probably right. Like that's, it's kind of a big thing, right? Because you don't want to have collisions from gems with the, with your library code across different gems. And yeah.

Valentino Stoll (39:23)
And even

when you go to generate Ruby code from AI too, it'll have these conflicts. It'll be like, you've defined this class in another module, but it's also defined here because I generated it. Because it doesn't know the history of everything that's been required, which maybe it could. But is that worth it at the point of use? I don't know. But also to just.

Joe Leo (39:30)
Mm-hmm.

Right.

Valentino Stoll (39:47)
Seeing like how an AI engine would work inside the language itself is really interesting. I know there's some people working on that too. So I don't know. Scott, we're 50 minutes in and we haven't let you tell us what Sublayer Tech is. We talked a little bit about Sublayer in the company.

Joe Leo (39:54)
Yeah.

Scott (39:55)
yeah.

Joe Leo (39:56)
Scott, we're 50 minutes in and we haven't let you tell us what Sublayer the Tech is. You talked a little bit about Sublayer the company. ⁓

So why don't you tell us about it? you know, we've got a couple of questions.

Scott (40:09)
Sure, yeah, so we're kind of doing a bunch of different things. Primarily, think the agent framework that we've built is the core and it's been driving a lot of what we've been building, a lot of the things that I share on works on my machine. And I think the two core ideas of the framework are, one, I...

Valentino Stoll (40:09)
Yeah, we're kind of doing...

Scott (40:29)
we kind of started off like really rejecting the conversational abstraction. And I kind of see that as a, it's useful. It's useful for certain things, but like from the engineering perspective, if we think of LLMs as this black box of, a black box function of string to string, we can do a lot with it. We can do a lot more than just have a conversation. And so,

Valentino Stoll (40:30)
started off like really rejecting the conversational abstraction and I kind of see that as it's useful it's useful for certain things but like from the engineering perspective

bugs function of string to string. We can do a lot with it.

Scott (40:58)
One of the things we're doing with the framework is we've kind of built these things we're calling generators where you basically, you set it up as a prompt, it takes some data and then you set up an output structure that you want and you get an object back. And you can do whatever you want with it. You can build a conversation, but a lot of what we do is get it to generate.

Valentino Stoll (40:58)
One of the things we're doing with the framework is

things we're calling generators where you basically you set it up as a prom, take some data, and then you set up

output structure that you want and you get an object. And you can do whatever you want with it. You can build a conversation. But a lot of what we do is get

Joe Leo (41:14)
you

Valentino Stoll (41:23)
get it to generate.

Scott (41:25)
generate code, generate user stories, generate tests, generate all different types of things that you can then add into pipelines or add into another application. And so we found that very powerful.

Valentino Stoll (41:31)
and then add in the...

found that very powerful.

Scott (41:38)
And let's see, the other big thing that we've kind of tried to do with the framework is this idea that I tried to make work really early on was never caught on, we're calling it promptable architecture. And the idea is that if you build things in a really like extremely loosely coupled way, you can have AI just generate more and more code and use it immediately. A couple other things like, you

Valentino Stoll (41:38)
Yeah.

⁓ that I tried to make word really early.

and the idea is that if you build things...

couple.

Scott (42:07)
really clean interfaces at the time it was much smaller code. But the idea behind the framework is we're setting the conventions and setting the foundations for code generation. So we've got generators which are an interface to an LLM and then actions which do things to the outside world. And then you build software with those concepts and combining them where you're

Valentino Stoll (42:08)
interfaces at the time.

my code.

Scott (42:37)
Maybe pulling data from an API, that data goes into a generator, generator does something with it, and maybe it goes to some other place. Like if you're pulling, maybe you pull the data from a JIRA ticket, then that goes to a generator to generate cucumber steps. And then the output of that generator goes to another action that stores it in a file. And so you can kind of piece those things together.

Valentino Stoll (42:51)
data from a year ago.

Scott (43:05)
But the nice thing is that since those things are so small, you can rely on a few-shot examples and in-context learning to generate new pieces, new actions, or new generators. And it didn't really take off the way I thought it would. ⁓ I've been able to use it, but I haven't been able to explain it very well. And the way that the framework is, it's very minimal. It kind of relies on you to build those things.

Valentino Stoll (43:06)
thing is that since those things are so small...

on a few shot examples and in contact.

didn't really take off the way I thought it would. ⁓ I've been able to use it, but I haven't been able to kind of explain it very well.

And the way that the framework is, it's very minimal.

It's kind of funny because I see it as like, you know, a coding agent.

Right? it's like almost one of the best use cases is like getting it to like, you know, take a bunch of context and like generate use, generate its own code to do something. Right. And like, it has all the fundamentals for you to organize how to do that. And then just goes and does that. you don't necessarily have to get it to generate code that it executes. right. But like, that is kind of like something you can do with it. And I always saw it as like, you know, when I saw

all of these coding agents like surface I thought well like you could do that with sub layer in my head and like you know it opened up some things I was experimenting with but it's super cool like everything is like so like modular and

Scott (44:12)
cool.

Valentino Stoll (44:17)
I know you only touched briefly on blueprints, but that is really incredible where you just give it reference code snippets and it learns and can regenerate those aspects of things. Especially as you get to a big giant project, you say, here's how we do things for this. You have all these cases and you end up with all this documentation, which is honestly great for the LLM, but you wanted to do something with that.

Scott (44:23)
Thanks.

You

Good.

Valentino Stoll (44:45)
not just like go read

it and then generate it yourself, right? Like you want it to do it and you just tell it, right? Like, and that's where like the libraries you've made like awesome. Like, ⁓ yeah, totally. Especially like in stronger typed like structures like GraphQL or something like that, where you have like very specifically defined things, you know, generation is like a hundred percent accurate. Like never had an issue.

Scott (44:55)
Thank you.

Valentino Stoll (45:10)
with that style of coding within the sub layer and blueprints. Really awesome. So I'm curious, where has blueprints led? How has that evolved internally? Do you still use blueprints? Yeah, so I mean, I think one of the things that we learned was launching blueprints and getting it out there and getting it people's hands was that it did kind

Scott (45:15)
Cool.

Yeah. So, I mean, I think one of the things that we learned was launching blueprints and getting it out there and getting in people's hands was that it did kind of require

a little bit of a shift in the way of thinking about like how you're building software, which was a a tough sell. And so we've actually done two things with the concept. One, we actually brought it into the sub layer framework itself. And so we have generators where you can, a command line generator where you can put a description of a

Valentino Stoll (45:42)
So we've actually...

Scott (45:55)
generator or an action and then we've got a bunch of examples of different generators and actions that you can just write a description and generate one right there in your project which is kind of like what we're doing on the docs where you can the same idea where it's blueprints we were thinking that could be an interactive docs kind of thing and then the other you know kind of came out of a works on my machine experiment where I have that ⁓

Valentino Stoll (46:15)
the other kind of came out of a works on my machine experiment where I have that

overnight it looks at this sub layer action repository, looks at all the examples in there. have Gemini,

Scott (46:21)
Overnight it looks at this sub layer actions repository, looks at all the examples in there. have Gemini, Claude

and GPT-4 come up with ideas for new actions that don't exist and then make PRs for new actions. that repo could keep growing, but I think there are probably 700 something PRs to review that I just don't have time to do anymore.

Valentino Stoll (46:33)
do action.

Joe Leo (46:44)
It's awesome. And as an aside, I love the generators idea because I am constantly forgetting migration generators and I'm constantly Googling them or asking chat GPT and I don't want to do that anymore. So I'm just not going to.

Scott (46:59)
Yeah.

Valentino Stoll (47:01)
Yeah, just so many things you can make.

Joe Leo (47:04)
I'm curious to know, you

you build tech, I don't want to say on top of LLMs, but you're definitely utilizing them. And we face this as well with Phoenix. What, like, I guess, do you have any sort of existential dread that the LLM, you know, the next release of the LLM is just going to do all that, right?

Valentino Stoll (47:10)
I'll talk.

existential dread that the next release of the LLM is just going to do all

Joe Leo (47:30)
of it and you know make make a portion of what is unique and compelling about Sublayer Obsolete.

Scott (47:37)
We've gone through that a couple of times actually. And so I think, you know, if you look, I know if I look back at like all the things that we've done over the last few years, it has definitely been a case of like, you know, trying to skate to where the puck is going, not skating far enough, maybe skating way too far out there. And, and then trying to find that sweet spot where, you know, the models improve.

Valentino Stoll (47:37)
gone through that a couple times actually. And so I think, you you look...

Joe Leo (47:39)
Yeah.

Valentino Stoll (47:45)
that we've done over the last few years, it has definitely been a case of like, you know, trying to skate to where the puck is going, not skating far enough, maybe skating way too far out there, and then trying to find that sweet spot where, you know, the models improve

and they actually improve.

Scott (48:05)
and they actually improve ⁓

what you're working on and not like subsdume it.

Valentino Stoll (48:10)
since we met.

and

Scott (48:14)
That's tough. That's a tough one.

Joe Leo (48:15)
No, we've

Valentino Stoll (48:17)
We've gone through it once, right? So we're not as old as you with Phoenix. We've gone through it one time and I feel like I need to learn from somebody who's gone through it a couple of times.

Joe Leo (48:17)
gone through it once, right? So we're not as old as you with Phoenix, and so we've gone through it one time, and I feel like I need to learn from somebody who's gone through it a couple of times.

Scott (48:25)
Yeah, I mean, I think it's like, I don't have any good answers. I think it's really like, you know, even another product that we released, augmentations, think, you know, that was less of model

Valentino Stoll (48:25)
Yeah, I mean, think it's... I don't have the answer. I think it's really like, know, even another product that we released, augmentations, I think, you know, that was less of a model.

Joe Leo (48:29)
Yeah.

Scott (48:41)
capabilities and really just like a lot of the big labs are also trying to figure out how to productize the AI that they're providing and...

Valentino Stoll (48:41)
capabilities and really just like a lot of the big labs are

Scott (48:53)
PR reviews was a very obvious step. so like we've done with augmentations is this, we call it semantic linting where you can write plain text, how you want your code reviewed, which we were thinking more of like, if all of us were working together and we have a discussion about like that class, we're gonna need to refactor that. Let's not touch it. Let's not make it any worse unless we absolutely have to.

Valentino Stoll (49:04)
which we were thinking more of like, you all of us were working together and we had a discussion about like that class, we're gonna need to refactor that. Let's not touch it, let's not make it any worse unless we're...

Scott (49:16)
you can actually set that up as a review. so like, hey, this model has too many responsibilities, let's not add a new one. Or here's a different way to do what you just tried to do. And then now, like a lot of the big labs that have connected to GitHub have the PR.

Valentino Stoll (49:16)
you can actually set that up in the room.

Here's a different.

that now.

Joe Leo (49:30)
Yeah, I can imagine

that being frustrating. The thing that I've found just as my own experience over the last six months is that generally speaking, the LLMs and the labs behind them, want to be product companies at the end of the day. We have always made our

Valentino Stoll (49:37)
Generally speaking, yellow labs.

Joe Leo (49:46)
as a consulting firm and I think the more consultative the product can be, in other words, the more that it leaves open exactly like Sublayer does, the idea, like the opportunity for exploration and the opportunity to increase the entire, increase the value across your entire project, the less likely it is to be eaten up by the next LLM generation. That said, you know, we've had

Valentino Stoll (49:49)
So.

the player does.

Peace.

as likely it is to be.

That said, we had.

Joe Leo (50:12)
You know, we started out as like just generating tests and now it seems like everybody just generates tests, right? And so it can't just be that. And we've also found that, we've got to be able to shift really quickly because if the LLM can do this thing, then great, then we should, A, we need to use it and B, we need to be able to, you know, offer some kind of differentiator.

Valentino Stoll (50:14)
now.

So can't.

immediately.

Scott (50:30)
Right, right. I mean, I've had similar conversations. mean, you know, going to everybody moving to the function calling syntax, the model's really getting good at the function calling syntax, which that was the other thing I was going to mention about sublayers. We've kind of, we're kind of doing function calling in reverse. Like we're using function calling to guarantee the structured output, which I haven't, I think I've only seen instructor, instructor use.

Valentino Stoll (50:31)
I mean, I had similar conversations going to everybody.

We're kind of doing functions.

Scott (50:54)
But yeah, the function calling, have Gemini, the recent PDF data extraction, something like 6,000 PDFs for a dollar. Like imagine if you had built this complex OCR system for analyzing PDFs, all of that is legacy code. All of that's tech deck now, right? That's like a, you know, whatever.

Valentino Stoll (50:59)
extraction something like 6,000 PDFs for a dollar. Like imagine if you had like built this like complex OCR system for like analyzing PDFs.

In

a way though, like the...

people pay for like the buttoning up that you've done around that, right? Like I think of like, uh, even like, Hey, from like, you know, base camp, right? Like, or not base camp 37 signals, whatever they're calling themselves. Uh, you know, like it's an email client, like there's so many email clients, right? Like, people pay for that, like, you know, buttoning up like smooth, like nice design, like even Apple, right? Like why would somebody pay so much for like, it's a phone, right? Like there's a bunch of quality.

like wrapped around all of the basics that are ubiquitous across all phones and smartphones, right? So like I feel like the same is true of like the software products, right? Like Phoenix as an example, right? Like you're laser focused on tests, right? Like and test generation and like making sure that the tests align with the quality of software that's being created for it, right? Like and making sure that it doesn't degrade and like, you know, somebody like GitHub, they aren't focused

on tests right like they're they're focused on you know git and github right like and so like you know those kind of players they have such a wider net like they're never going to be able to compete with you focusing on the one thing and that's going to be true of any product that you make right like and i feel like there's always going to be the fear of like some big thing coming and like replacing your pdf thing but really like just pay gemini to like do that processing save all that

Joe Leo (52:18)
Yeah.

Valentino Stoll (52:41)
Like, you know, maintenance burden of the thing you generated and and like get those quality pieces around that, you know, PDF aspect of it like bolted on to your your customer and it would survive right like That's my thoughts. I I'm not a business owner. So I don't

Joe Leo (52:50)
Yeah.

Well, first, yeah, so first balance, you know.

Scott and I both hope that you're right.

then second, think that's what I'm getting at. So you find out that this part of your application has been made obsolete. How quickly can you stop using it, move to something else, right? And you know, is supposed to help us move fast, right? And so how quickly can we say, okay, this part of the process is dead. We need to leverage what's here, know, scrap the rest, re-architect what's, you know, what's left. And...

Valentino Stoll (53:03)
I think that's what I'm getting at. So you find out that this thing is part of your application could have made absolutely how quickly can you stop using it?

Joe Leo (53:26)
And we can't miss a beat. We've to move quickly on it. That can be challenging.

Valentino Stoll (53:27)
Yeah.

Scott (53:30)
Yeah.

And I think that goes to the kind of the conversation we were having earlier around abstraction interface design, where there was a post I read recently of start with the interface and doing something like that, where this is a good interface I like to work with. this function call might be a bunch of services or tons of thousands of lines of code that AI generated, or it's the Tencent API call. It doesn't matter because your system is designed for

Valentino Stoll (53:46)
call.

system.

Scott (53:58)
interface that you have rather than the implementation.

Valentino Stoll (54:01)
Great.

Yeah, hard problem to solve. I trust that, you know, collectively we'll be able to solve it. Hopefully, hopefully you'll be able to survive it. I believe in you guys.

Joe Leo (54:06)
Mm-hmm.

Yeah, I

do too. think Scott's approach and what you talked about before about, well, keep.

keep a bunch of irons in the fire, right? You have sublayers, this basis, you've been spinning off these different potential products that are either open-sourced or not. But it's like, okay, well, you've got multiple bites of the apple. You can take multiple shots to see what really is gonna take off.

Valentino Stoll (54:40)
Yeah, yeah the agility

is on your side.

Scott (54:42)
Supported by AI. Because yeah, I mean, think that's the thing too, is the one that we noticed is that like there might be, one, there's, I don't know if you guys have found a way to deal with this, but like a lot of times you see somebody that does something successful. A week or two weeks later, there are a bunch of copies. A couple of weeks after that, there are open source versions.

Joe Leo (54:44)
Right.

Valentino Stoll (54:58)
a or two weeks later, there are a bunch of copies.

Scott (55:03)
A week or two after that, a lot of the big AI platforms have it as one-click templates for building that thing. And so the opportunity of you have a good idea, you get it out there, a lot of other people are able to catch up really quickly.

Valentino Stoll (55:09)
So the.

Yeah, I don't disagree with that. at the same time, it's the same of like software frameworks and things like that, right? Like one will stick and it'll get used, right? And even though like you can go and duplicate it yourself, it's like, there's not really much value in that. Cause like there be everybody else has like gone around it and like there's like this community consensus on like, okay, well like we all know how to do this and it just makes things easier. And it'll be interesting to see like if you can

Joe Leo (55:33)
Hey.

Valentino Stoll (55:47)
focus, right, like what you're building around that idea. I feel like that's probably the most valuable thing you can do is just focusing on how, you know, the community consensus can converge on your products, right? Like, which I think you're doing a great job at, to be honest. Yeah.

Scott (55:52)
in.

Thank you. Yeah,

thanks. Yeah, think one of the things I've noticed too, though, is like we are definitely in this age where like a lot of people...

A lot of people who are here and building are also experimenting too. so a lot of people, know, a lot of times you see, you know, you'll release something, you'll see somebody release something and then there will be like a flood. You'll see like a dozen of like, yeah, I made something like this two weeks ago. Like the, turn your repo into a single text file for, for use in context.

Joe Leo (56:32)
Yeah.

Scott (56:33)
There was like a two week period where like everybody, somebody got popular as one. And then like every couple, like every day there were like a couple of new ones. Like, I made something like that the last week. I actually needed this. Here's my command line version. Here's my web version. I think we're in this like really cool, this must've been what it was like in the, you know, like the early Unix days.

Valentino Stoll (56:33)
There was like a two week period where somebody got popular as one and then like every couple...

Joe Leo (56:35)
you

Valentino Stoll (56:45)
Really?

Joe Leo (56:46)
Mm-hmm.

Scott (56:55)
I wasn't there. You've got this new platform and everybody has the capabilities to mold it and change it. So you had all these new utilities bringing up that solved problems that you find a lot of people actually have.

Valentino Stoll (56:56)
You've got this new platform and everybody has the capabilities to...

Joe Leo (57:09)
Mm-hmm.

Valentino Stoll (57:09)
It's

funny, there's an episode of Co-Recursive Podcast, which is pretty great. They had somebody on that made the whatever the MS Paint version on the Mac was. I forget what it's called. But he saw a classified ad.

that was like a computer forum, like in paper form. he wrote in and was like, oh, this sounds really interesting. They started a thing. And he wrote in a cabin in the woods the first paint version on the Mac and submitted it and sent it in and got it reviewed and it circulated. And Apple ended up buying it from him, right? And he ended up working for Apple.

Joe Leo (57:51)
Yeah.

Valentino Stoll (57:52)
And I feel like, you know, there's a similar thing happening now where like, you know, people are like coming together and just being like, hey, look, I did this thing. And like, you know, a lot maybe just like be like, yeah, that's interesting. Like, you know, mess around with it and use it a bit. And some of it is like blowing up and like turning into things, right? ⁓ And.

Joe Leo (58:11)
Mm-hmm.

Valentino Stoll (58:13)
It's just exciting to be able to just participate in that if you have even if you have nothing right like if you're in a cabin in the woods and you don't have electricity right like You could still like Participate in this like, know huge thing that's happening. It's just wild Yeah, so we've been talking a while Scott ⁓

Scott (58:17)
Right?

Yeah.

Joe Leo (58:32)
Yeah, I totally agree.

Scott (58:36)
Thank you.

Valentino Stoll (58:38)
I feel like we could just keep it going all day to be honest. Is there any imparting wisdom here? Last notes you want to talk about?

Joe Leo (58:40)
I know.

Scott (58:46)
Well, I guess I had one question. like this is, you know, you guys are both works on my machine reader. So I did have this question for Joe as, you know, dovetails into something that we've like started noticing ourselves and building. When you have Phoenix running,

Valentino Stoll (58:51)
I did have this question.

Joe Leo (58:58)
Sure.

Valentino Stoll (58:58)
When you have Phoenix running.

Scott (59:01)
What do ⁓ find you and your team do? Like, do you sit there and watch it? Do you do something else? Like, what is that experience like?

Valentino Stoll (59:01)
do you find you and your team do? Like do you sit there and watch it? Do you do something else? Like what is that experience?

Joe Leo (59:08)
It's a little bit like it's running in the background. Where we are right now is that we can switch out models fairly quickly and we can rerun commands, but that's not fully automated. So a lot of times we kick it off and we continue working. We check in on it. We've got some pretty extensive reporting on it. And then...

Valentino Stoll (59:09)
It's a little bit like it's running in the background.

where we are.

switch out models fairly quickly.

and we continue working.

Joe Leo (59:29)
When it's finished, we'll get a tally. say, OK, it's generated 3,500 tests, and 200 are failing, the rest are passing, whatever it might be. So we send it to do it again. And it might be that we switch the model. switch. We don't really switch the context. We might switch the command ⁓ and run it again. So it's mostly on its own. we probably could fully automate it. But right now, at this point, we kind of want

Valentino Stoll (59:29)
it's finished, we'll get a tally and say,

you send it to do it again, and it might be that we switch the model.

We don't.

and run it again. it's mostly on its own.

Joe Leo (59:55)
check-in points.

Scott (59:56)
How often do you do you have do you have like alerting or anything like notifications for like it's you know it's waiting for your input or it's like at one of those checkpoints or is it a little bit like you know like we should check on it

Valentino Stoll (59:59)
you have alerting or anything, like notifications for like, it's waiting for your input?

Joe Leo (1:00:11)
I believe this is a lead engineer question, but I think we were notified when it's complete. ⁓ we've got. Yeah. And most of this, so I shared on the last podcast, we started this shamefully. started this in Python, but we have so much of it instrumented or, or the workflow built out in Ruby now. And so there's a job that is completing. And so at the very least we've got, you know, an active job.

Valentino Stoll (1:00:11)
I believe...

Scott (1:00:16)
Okay.

Valentino Stoll (1:00:18)
Yeah, and most of this, I could share on the last podcast we started, it's shameful we started this podcast. But we have so much of it instrumented or recorded.

So there's a job that is completing. So at the very least, we've got it.

Joe Leo (1:00:35)
see what's running and know when it's done. I'm not sure if we built an alert in yet. Probably we're just checking on it.

Scott (1:00:41)
Gotcha. Yeah, so the reason I ask is like one of the things that we've, you know, we've been building a lot of stuff and like reflecting on, you know, what is, what's missing here and like, you know, we'll, we'll have the AI agents running in some tab or some terminal or doing a deep research and

Valentino Stoll (1:00:42)
Yeah, so the reason I ask is like one of the things that we've been building on.

missing here and like we'll have the...

Joe Leo (1:00:55)
Mm-hmm.

Valentino Stoll (1:00:57)
research and uploaded

Scott (1:00:59)
or, you know, upload it,

Valentino Stoll (1:01:00)
it.

Scott (1:01:00)
analyze this video I just uploaded and you'll get, you'll have an error or it'll be sitting there waiting and you've forgotten about it. You've like, you know, moved over to your email and you're like spent two hours, you know, responding to email or whatever. And this thing has been sitting there waiting for, you know, an answer. Um, and so that's why I asked, like, is that the, you know, the experience or like, I don't know, my, I'm still, still trying to break the habit of like, I put a prompt in the bolt and like, watch it build the app. I just like sit there.

Valentino Stoll (1:01:06)
getting there waiting you forgot about it you like you don't read email you like spent two hours

Joe Leo (1:01:08)
Yeah.

Valentino Stoll (1:01:13)
things sitting there waiting for.

So that's why I asked, is that the big experience? like, I know, I'm still trying to break the habit of like, I put a...

Joe Leo (1:01:29)
Yeah, but it is amazing when it happens. mean, I understand that inclination.

Scott (1:01:31)
Yeah.

Valentino Stoll (1:01:33)
It's funny you almost need like an idle time upsell right like built into every every one of these LLM things You know while you wait Do you want to create a new bolt like

Scott (1:01:37)
There.

Joe Leo (1:01:37)
Yeah.

Yeah, we could do something else, right?

Scott (1:01:41)
Yeah. Right. Yeah. I

Joe Leo (1:01:47)
Yeah.

Scott (1:01:48)
had a friend tweeted something like, coding, more like vibe waiting.

Joe Leo (1:01:53)
Yeah, it's true, but the way it is more fun than, you know, compiling something.

Scott (1:01:56)
Right.

Valentino Stoll (1:01:56)
Yeah.

Scott (1:01:59)
But yeah, we're kind of working on trying to solve that aspect of like, you've got all these powerful agents doing all these things that are all demanding your attention. How do you organize that? How do you categorize that? So, you know, really, really curious to hear and maybe we'll chat sometime when we have something to demo, see if it would be useful for what you guys are doing in Phoenix.

Valentino Stoll (1:01:59)
Yeah, we're kind of working on trying to solve.

Joe Leo (1:02:11)
that's interesting.

Yeah, I knew you were setting me up for something. But I would, yeah, I would like to

Valentino Stoll (1:02:20)
Yeah, I knew you were setting me up for something. Yeah, I would

Joe Leo (1:02:23)
check

Valentino Stoll (1:02:24)
like to check it out, you

Joe Leo (1:02:24)
it out. You know, I mean, right now we onboard, you know, we onboard one or two companies a week, but that's going to, you know, that's going to increase when we kind of take the restrictor plate off. And then, yeah, there's going to be more than, more than a couple of jobs to be monitoring. And I want not the smartest engineer I've ever met monitoring them. want somebody else to do it, you know? I want to, you know, delegate that task.

Valentino Stoll (1:02:41)
Well, it's somebody else who did it.

Yeah, I know we are like kind of at time here, you've got me going here. But I feel like that is definitely something like that is a black hole of like, know, orchestration observability is like such a like missing thing because it's not it's custom to every single.

Joe Leo (1:02:49)
Yeah.

Scott (1:02:49)
you

Joe Leo (1:02:59)
Mm-hmm.

Valentino Stoll (1:03:05)
workflow, right? So like, it's not something you can even like offer as a product because it becomes too complicated and coupled with all the business logic and stuff that like you can't really explain for everything, right? Yeah, I don't know.

Joe Leo (1:03:17)
Or can you? Scott's gonna let us know.

Scott (1:03:18)
think we might be able to. Well, not exactly what you're saying, but some solution, some

get us further along on that.

Eventually we're gonna have to probably talk about it and then like the open source clones will come shortly after

Valentino Stoll (1:03:26)
Eventually we're going have to publicly talk about it and then the open source phones will be happy to

release something like in secret so that like you have to like be a member in order to like know that it even exists. Right. But then like everybody wants to get in, you know, almost like a, you know, a secret club style release. Right. Like I wonder if that's even possible. Right. Because everybody just loves to like whistle blow. Like, you know, like in like, you know, nobody should be able to whistle blow. Maybe that's the wrong word. But like, you know, like leak the like

Scott (1:03:47)
Alright.

Joe Leo (1:03:59)
Yeah, they want to be the ones

with the insider info, which you can't be proved that, right, unless you spill it.

Valentino Stoll (1:04:00)
the secret right right and so like I wonder if that's even yeah I wonder if it's possible yeah like to even say like maybe like I don't know you could have something yeah I have no idea yeah

Scott (1:04:14)
You know, that actually reminds me of Ben Thompson in Strutechary. had this post recently where the end he talked about, it was like kind of right after deep research came out and he had this post, the end of the post he finished it with like, value is going to accrue to the secrets that you can keep from the AI.

that once it's out there and the AI knows it, everybody has access to it. And so there will be value in the things you can keep from the AI. Which goes right into what you were saying.

Joe Leo (1:04:41)
Yeah, I read that too. I'm a big, yeah, I'm a big

Valentino Stoll (1:04:43)
That's really interesting.

Joe Leo (1:04:44)
Stratocary fan and yeah, that's, it's cool to think about. But on a serious note, look for the Ruby AI sub stack coming soon where we do secret releases that only subscribers can see.

Valentino Stoll (1:04:55)
Yeah, totally

Scott (1:04:55)
You

Valentino Stoll (1:04:58)
Yeah, I'm on board for that Awesome well, you know, think we're at time here Scott was great to have you on

Joe Leo (1:04:59)
Yeah.

Valentino Stoll (1:05:06)
Maybe as like a quick departing thing, like share one thing that you found super cool or interesting, want other people to know about and expose it to the AI world here.

Scott (1:05:17)
Expose it to the AI world. Well, I guess the other tab I have open here right now is TLDraw. TLDraw's computer. It's kind of like a visual interface for... I actually don't know how to describe it. But I use it to generate the image back here every day.

Valentino Stoll (1:05:19)
The other tab I have open here right now is TL.

Scott (1:05:34)
But you can basically create these like visual boxes that have instructions, text, image, and create these kind of like reusable pipelines for AI. Mine's definitely simple. There's way more complicated things, but I'm still in awe of what they've done with TL Draw.

Valentino Stoll (1:05:35)
Thank

Yeah, that's super fun.

Joe Leo (1:05:50)
⁓ I can go next. Well, what I have to share is actually a series of blog posts. It was started last year. It's a demystifying Ruby thread. I'll post the first one here.

Valentino Stoll (1:05:51)
Yeah, what about you, Joe?

Joe Leo (1:06:04)
And it goes into threads, processes, and ractors. And I just found that it really explains ractors in a way that is very easy to digest. I remember the first time somebody tried to explain it to me, I was like, I don't know what the heck you're talking about. So it explains it with simple examples and gets it across threads, fibers, ractors, processes really simply. And then it goes on and starts to talk about

Valentino Stoll (1:06:05)
into threads, processes, and racks.

explain.

So it explains it with simple examples and gets it across threads, fibers, wrappers, processes really simply. And then it goes on and starts talking about.

Joe Leo (1:06:31)
and a future blog post, garbage collection, and then the last one was just published a couple of weeks ago about memory, the heap of the stack, et cetera. So it's good.

Valentino Stoll (1:06:33)
Garbage collection and then the last one.

Very cool, I'm gonna check these out. Yeah, I love a good concurrency read.

Joe Leo (1:06:44)
Yeah,

me too.

Valentino Stoll (1:06:46)
Yeah, I've got one thing. I just released it this morning actually. It's kind of fun. ⁓ It's called the AI Software Architect. It's basically a series of markdown files specifically laid out in a folder structure that lets you...

Joe Leo (1:06:50)
Go.

Valentino Stoll (1:07:00)
basically implant a bunch of members that you can define that are software architects of various domains. And you can get it to review your code base. You can get it to draft up ADRs for new features that you want developed or technology you want introduced. And you can help it like even review the code changes that you have like in place and just ask it, does this align with our architecture system wide? Like what are the pitfalls?

been using it a lot and just some side projects I'm bringing into some things I'm doing at Gusto as well to just sanity check a lot of my work as well as I start to do so. And it's a very interesting idea of just like markdown files.

Scott (1:07:35)
Thanks.

Joe Leo (1:07:42)
have gotten

sort of like a new lease on life at Deaf Method especially because well maybe they were always popular but we're definitely doing them a lot more because we can have AI write them so we don't have to spend time writing them ourselves.

Valentino Stoll (1:07:53)
Yeah, you know I...

Scott (1:07:52)
Thank

Valentino Stoll (1:07:54)
I followed this approach for a side project called a Gentic I'm working on. It's like not really anything worthy of looking at yet, but I basically just set it up in a workflow where I'm just saying, you know, this is what I want to accomplish and this is something that I want to do. like, you know, coding assistants are great at planning and then like making to-do lists and stuff. But instead of like just diving right into it, starting with, okay, like let's think about this purely from an architecture standpoint.

point using all these references and people that I know that like are, you know, historically known for well-designed systems like Sandy Metz and Sarah May and all these people that like happen to fall into the language that I'm working in, right? well, you can do anybody, and just have them like, you know, have the system like analyze itself and where this new thing fits into it. And then from there, start a plan like aligning with those new goals and working through

like a churn like that just like really fun and like I know to me personally like I feel like I'm like hand waving a lot of knowledge that I have of like how things work and I definitely like shape things you know tell it to do things and what not to do

Joe Leo (1:08:53)
Yeah.

Valentino Stoll (1:09:06)
But I'm hopeful that like over time this can become more of like a, you know, consistent, like just call and response and help people walk through building more quality things. I don't know. It may just end up on the pile, you know.

Scott (1:09:19)
This

is very cool.

Joe Leo (1:09:20)
Yeah, and I like that you've started the tradition of launches on podcast recording day.

Valentino Stoll (1:09:27)
That's right.

I've heard it here first.

Joe Leo (1:09:29)
Ha ha ha ha ha

Scott (1:09:29)
Yeah.

Valentino Stoll (1:09:30)
All right, well thanks again, Scott, for coming on. It's always awesome talking to you. For those that are listening, come to one of the artificial Ruby events in New York City now, maybe a new city in the future. We'll see. I would love to see more of these pop up, because that's how Joe and I met. Honestly, Scott, you and I met there. I know we were on Discord for a while, but meeting in person is just so much better.

Scott (1:09:39)
Yes.

Joe Leo (1:09:39)
Absolutely.

Yeah.

Mm-hmm.

Scott (1:09:51)
Yep.

Yeah.

Joe Leo (1:09:57)
Yeah, I have loved artificial Ruby and I talk about it all the time about how it's just given me so much just hope for the future in Ruby. It's been a life-affirming experience going to those meetups where there's just 100 people and everybody's just having a good time and talking about Ruby. I haven't had that experience in years and this has been really fantastic.

Valentino Stoll (1:10:06)
Yeah.

Scott (1:10:09)
Thanks.

Thank

Valentino Stoll (1:10:19)
Yeah.

Scott (1:10:20)
Awesome. I really appreciate that guys. I really appreciate you having me on here. And yeah, like I said in the beginning, that you guys connected through it and set up this podcast is beyond my wildest dreams of what I was hoping when we started. It's actually gonna be one year in July. So beyond my wildest dreams when we started it as a small happy hour. But yeah, that's incredible.

Valentino Stoll (1:10:27)
guys connected through it.

That's awesome. ⁓

Joe Leo (1:10:37)
Excellent.

Valentino Stoll (1:10:41)
Yeah, totally. That was fun too.

Joe Leo (1:10:44)
Yeah.

Scott (1:10:44)
were there since the

first one. Yeah, it's evolved a lot since then, but the energy was there since day one, which is cool.

Valentino Stoll (1:10:49)
Yeah, totally.

Alright, well thanks again. Thank you for joining if you're listening. I hope people are listening. I'm sure we'll ramp this up because there's just so many fun, interesting things to talk about.

Joe Leo (1:11:03)
All right, take care everybody. Thanks, Scott.

Scott (1:11:05)
Yeah, thank you guys. Talk to you later.