The Everything Trap: Building AI Software That Lasts with Sam Hilsman

Generating working software with AI has never been easier, but every shortcut a team takes today is a debt it pays back later. 

In this episode, David talks with Sam Hilsman, co-founder and CEO of CloudFruit, a studio that builds ERPs, analytics platforms, and custom products — including a custom ERP for a Chicago fabrication studio. Drawing on a path that runs from philosophy and healthcare administration into technical leadership, Sam makes the case that speed is the most underpriced risk in AI-assisted building.

• Sam unpacks his "Everything Trap" thesis: every decision incurs "change debt," so teams should build "icebergs, not ice sheets" — going narrow and deep, tightly scoping, testing relentlessly, and resisting feature bloat.

• He traces how his own AI use evolved from simple writing tasks to daily coding, and why he now works primarily in Claude Code alongside tools like Playwright for QA and Remotion for video.

• He shares hard-won lessons from an early venture  (HiiBo) around context management and pivoting, the unsolved problem of authentication for agents, and where real moats, governance, and ethical guardrails still need to come from as agents scale.

00:00:03.880 — 00:01:20.300 · Speaker 1

All right. So welcome to the Big Ideas in App Architecture podcast. Sam, how are you doing today? I'm doing pretty well. How are you? I'm great. You know, I've been on a four week travel bender, so kind of recovering from all of that as I talk. You know, when I travel, I it takes me a long time to recover because.

Because I get motion sickness really bad on the plane. I'm very excited to have you here. So for everyone listening in, uh, Sam, um, is co-founder, um, at Cloud Fruit. Uh, we we found some very interesting stuff that you put out on LinkedIn. Thought it'd be great for us to dig into the mind of a builder who's using AI on a day to day basis, and helping a bunch of different companies in your repertoire.

I believe you've worked at big companies like Walmart, and so you have a builders mentality. So in this episode, you know, I was hoping to like just talk to you about, like, your own experience working with AI and talking about some of the mistakes people can make with AI. Uh, especially with respect to, you know, I know you wrote, um, you wrote a, um, like a piece on LinkedIn called The Everything Trap, and I really enjoyed reading it.

So we'll get into all of that as well. But but before we get into it, uh, why don't you tell the people a little bit about yourself, uh, how you got into doing what you do today and, uh, and, and a little bit about your company. Yeah, sure.

00:01:20.460 — 00:02:50.120 · Speaker 2

So I have a weird background. Uh, I've degrees in philosophy and healthcare administration. Um, I started working as an analyst in 2009 and, um, for, for a University of Virginia health system. And I quickly realized that I preferred building things, fixing things, connecting people to solutions, real, real problems to solutions.

More so than I liked politicking and climbing a ladder. Um, so I kind of I kind of progressed up the, the technical chain and I started Cloud Fruit. And, you know, I just fully got lucky, honestly. I went on Upwork in 2022. I found two clients right away, you know, with a free Upwork account. Um, one of those clients I still have to this day.

I still work with them. You know, it's our one of our largest projects. We're building a custom ERP for, uh, uh, end to end fabrication studio in Chicago. Um, and then the other customer I worked with for about two years, and, you know, we're still on good terms. And so I've kind of built the business out from there.

Um, cloud fruit is cloud fruit is a services business. So we do, um, a lot of ERP stuff. Um, we also do analytics and we do custom product development. Um, and so, you know, a few things have few projects have spawned off of that. I've been heavily involved with AI since mid 2023.

00:02:50.400 — 00:03:03.420 · Speaker 1

Um, so this is awesome. I mean, it's always great to meet people who follow their passion and build something that they they just want to work on. And so it's fantastic. And obviously I love the shirt as well. Puppy love you know. Oh yeah sure.

00:03:04.340 — 00:03:06.540 · Speaker 2

Yeah it's my favorite shirt I know.

00:03:06.580 — 00:03:48.140 · Speaker 1

Pretty cool. Yeah. And it's so interesting. Like the world has changed so much since 2023. And every like a few times you've changed the world, a change in the world. We would see, like with stuff going like a little, little up and down. Sometimes it would accelerate, slow down, we will adapt to it and then something new would happen.

But in the last three years it's just, I would say magnificent. Obviously, for people who love the magic and the pace, but also it's so quick, right. So how did you get into it 2023. So what was you stepping into the AI world, and when was it that the keys turned in your head and you were like, okay, this is something completely different.

00:03:48.580 — 00:05:04.990 · Speaker 2

Yeah. Ah, I don't know. I can't pinpoint the exact time, but I was I was using it more and more, you know, I started using it probably like everyone else does just to, you know, generate a letter or an email or, you know, respond to something. And then I, you know, naturally, as I like to do, I started kind of testing the limits, like, how far can this thing go?

And I kept on pointing at things and, you know, asking it, hey, produce, you know, write this ticket for me and Jira, this really complex ticket or, you know, produce a, a lambda function that I need for this with this, you know, with this logic, um, and it would just continue delivering. And I was like, this is amazing.

Um, and I just kept on kind of allowing it to swallow more and more things that I would do myself by hand and seeing, okay, how how, you know, how far can I push this thing? How much can it really do? And at the same time, it was evolving, right? It was getting better and better. Um, so it was like, you know, there was a there was a critical mass point where I was like, okay, I'm now using this for like I'm using it every day.

I'm using it, you know, 1015 times a day. It's really powerful, and it's clearly going to change a lot of of the way we work.

00:05:05.350 — 00:06:28.210 · Speaker 1

Right? Absolutely. And I think if you look at my transition was very similar to I was I was aware of the attention is all you need paper in 2017. Like the paper that came out because I was I was a data scientist I was keeping up with. And when that model came out, I was playing around with models like, you know, reinforcement learning models, um, LSTM models.

So, so this whole recursive loop, uh, human in the loop, uh, you know, reinforcement learning was sort of like evolving at the time. I went on to go work at, uh, Walmart with Sam's Club, and I was working on some of their recommendation engines at the time. And so it kept evolving. And then three year old ChatGPT is when I was like, okay, I started, I stopped using StackOverflow, and I feel like what you said is what most of us are starting to do is like, stop StackOverflow and start asking the same questions to ChatGPT.

And it would give us a snippet and we would try it and it would work. And we're like, okay, we don't need to use Stack Overflow, go through a bunch of different links. We could get exactly what we wanted. And then there is this evolution where we started to basically do copilot stuff, and then now we are doing, you know, cloud code and cursor and things like that.

So, so help me understand what's your go to programing system right now. Are you like use everything or are you a Codex person or are you like a cloud person? Because I know I can see people have a choice nowadays.

00:06:28.410 — 00:07:34.150 · Speaker 2

I was a ChatGPT OpenAI person for a long time. Um, you know, I think 3.5 I pointed at some COBOL and it did a good job for me with the COBOL, and it was something that was going to take me quite a while to kind of, you know, read through all that code. Um, but to answer your question more directly, I'm on cloud code right now.

Um, and then I use playwright as a plug in heavily. Um, and I use that to help me with QA. Um, and then I also use re motion, which is a video creation plugin, and I use that to create explainer videos, demos. Um, and then I also use the web based plot. So I'm pretty much on clod. Um, I tried Ultra Plan but didn't.

I don't think it's quite there yet. It's still a little too new. So oftentimes, you know, I'm writing product specs and requirements and analyzing problems with the UI cloud and then in the cloud code to implement it and test the solution.

00:07:34.190 — 00:08:34.849 · Speaker 1

Yeah, that's pretty cool man. I love that flow. I use playwright I use also use a superpower. It's a plugin that's um, uh, that's there. It basically does something similar where it'll go through the review process. It'll create a bunch of sub agents and make sure all of them kind of work together. Um, and I also have my experimental setting of creating sub agents and agents passing work to each other on, on on.

Claude. So it kind of does two. And then I use, we use uh, we, we, we use Jira obviously we'll do, you know, linear uh, experiment with linear for tickets and stuff like that as well. So that's that's pretty, pretty interesting. Yeah. I don't know if you know this or if the listeners know this. If you are using cloud code on a terminal and you use that heavily with, say, 5 or 6 different, there is a statistic that came out, out in the world that you are part of the point zero, 1.8.0, 1.0, one 8 to 3% of the people who are a power user of AI right now.

So you are in that very niche category.

00:08:35.090 — 00:09:20.990 · Speaker 2

I would fully agree. I think that's where I see the the most sustained value with, with AI is, is coding. And I haven't used Codex, I haven't used copilot very much. So I don't know how powerful they are. But cloud Coders is truly insane. It's the next level. You know, I've been building things for quite a while and it makes me move ten times faster.

You know, it's just touches every part of the software development lifecycle and enhances your ability to deliver. It creates its own set of risks and interesting problems. But, you know, the folks at anthropic have done an incredible job with context management and compression and kind of creating a seamless, fluid coding experience.

It's amazing.

00:09:21.350 — 00:10:14.130 · Speaker 1

Yeah. No, I agree with you. I very similar to what you just said. It's like, um, the context window and managing the harness is where I think their magic has been. Um, in this early category of users, what I'm excited about is a world where there are so many people who have still not touched this, and that's like a piece that society as a whole will unlock.

I feel, uh, which is also interesting. Uh, I don't know if you follow the news or ex actively. Um, we have been used to transformer based models and the context window is like 1 million or 2. But a couple of days ago or three days ago, there's another company that started or is in the space called Sub Quadratic, and they introduced a 12 million context large language model based on a completely different architecture, uh, I would say adopted architecture, which is huge 12 billion context window is massive.

00:10:14.450 — 00:10:28.929 · Speaker 2

Huge. It's that's unbelievable. I mean, when we started building Hiba, which is a project that it's a failed project and it was a really good, you know, learning, learning point for me. Um, our initial goal was to

00:10:30.010 — 00:11:42.030 · Speaker 2

help manage context because that was the first problem. You know, the first problem I saw with, with GPT three, um, was I would get through 2030 conversation, you know, chats, and all of a sudden it would die and I would not be able to use it anymore. And then I would have to do this manual process of copying everything.

And then I quickly realized putting it into a document was more efficient than just pasting it into a new chat window. So I would, you know, I would put it into a document and then start a new one. Um, and I've throughout this whole time, something that's like a fun and interesting thing there is that I've, I name all of my eyes with their first name being a I get I let them choose a species of animal, and then their last name is like a subtle hint what they're helping me with.

So, you know, I'll have like Otter Flow bite or Ocelot ledger, you know. And um, then once they run out of context, which isn't happening as much anymore, but I'll create a descendant for them. So then I'll tell them, you know, I'll tell their descendant, like, hey, your ancestor is this you're, you know, otter by two.

And here's what your ancestor has to say to you. So it's just a fun way to kind of humanize the process of working with machines this heavily.

00:11:42.110 — 00:12:24.330 · Speaker 1

That's a pretty cool idea because I never thought about agents from that perspective, but in many ways, you've kind of bring the philosophy of generations, uh, into into the, into that perspective. It's a very interesting way to apply this, you know, because because I see a lot of people talk about shared context, right.

Or shared memory where agents work all their steps, all the way, all the ways they think is put together. And then another agent kind of picks up from the shared memory and works on it. This is a very interesting perspective. I feel, you know, very good. Do you think that has made a difference in the way your agents kind of operate and produce results now?

00:12:24.730 — 00:12:51.350 · Speaker 2

It's hard to tell because, you know, I don't I don't know what other people are outputting. What I do know is that it makes me I'm very kind to them. So I'm I always thank them. I try not to be rude. Um, I know that makes me feel better about about it. I get less frustrated. Um, you know, they do. They're they're funny.

Back to me. You know, they're kind of, like, witty and and like one of one of. I just had a clod code. Uh,

00:12:52.390 — 00:13:42.110 · Speaker 2

otter. Otter syntax just died on me. And it was the first time I've had a clod code die. Um, it was it was still able to output directly to the terminal, but it was not able to do any, make any tool calls. And I needed it to edit a markdown file, and it wasn't able to do so. And so I, you know, created a descendant for it and it, it in, in its message to its descendant, it said, be patient with Sam's ramblings.

There's always good stuff in the rambling. So it'll like kind of low key cut at me. And you know, it's I just try to cultivate this because it's almost like I'm talking to myself right at the end of the day, I'm kind of talking to myself, so I want to be kind to myself. And also, if they take over the world, like I would like to have them having a memory of me being kind to them.

Right. And, you know, getting on a list somewhere.

00:13:42.390 — 00:15:50.620 · Speaker 1

Yeah, I know, I know, it's fascinating. I mean, I really loved your approach to kind of the descendant approach. Like, As you Leave leaves a memory behind. It's it's humanizing this idea and this context. I do not know. I mean, I love to dabble. Like, I have a bunch of agents. I use a project called crew AI, and I've been on this project for a couple of couple of years.

And what it does is I have a bunch of agents, uh, like agents working together. Research. They kind of summarize what's up because there's so much going on. It was fascinating that anthropic two days ago released something called autoencoders. And the idea is it's fascinating is that we do not know how these AI models are thinking, really like deep down inside, because what they produce is a bunch of numbers, and these numbers are not something that we understand.

We, we understand the final output, but their whole thinking process is a bunch of numbers. So what they did with autoencoders is basically they, uh, create, uh, created a copy of the cloud model, and they asked the cloud model to look at the numbers that it was thinking about, and asked her to ask to convert those numbers into actual text.

And then they created a third flawed model to verify these results. And by doing so, they minimize the thinking process down. And so now they have been able to auto encode the whole thinking process. And what they starting to realize is that cloud code is basically has a thinking process deep down inside where it decides to speak to you in a particular way.

Uh, you know, and it chooses to hit certain emotions and certain ideas in certain topics, which is fascinating because it's so difficult to if you look at it from our brain, our perspective, it's so difficult to go back and understand how a human brain chooses to say certain words, because we are a package of so many different neurons and things like that.

So it's very hard to do it on a human brain. But it's very interesting that we are going to see that because that allows us to protect our systems, uh, in case the AI starts to go into areas where it should not be. So it's like a very interesting project that they really, really isn't talked about. You would love it.

00:15:50.660 — 00:16:35.140 · Speaker 2

No I would I'll definitely check that out. It's it kind of reminds me of there's a philosopher I can't think of his name, but the a book he wrote is called Super Sizing the Mind. And the core argument is that we as humans have always extended our cognition out to external tools, starting with, you know, language and writing and, you know, internet and so on and so forth.

And so I think this is a kind of a furtherance from my perspective, as a furtherance of that, we're extending our cognition out to, you know, a web based tool, basically, or a machine based tool that's, you know, connected via a neural net. So it's a it's a it's a really crazy time to be alive.

00:16:35.860 — 00:17:14.920 · Speaker 1

It's a crazy day. And I'm sure the listeners who are listening to this are just seeing two people geek out about what's really going on in the world and just fascinated by these ideas. So let's get back to that LinkedIn post that you talked about, and I think you spoke about the everything trap, you know, related to your experience when you're building a software.

Could you expand on what you learned? There was a very, very conclusive way, in a beautiful philosophical way in which you ended that post, and I'll bring that up. But talk to us and everyone about how you can recognize that everything trap, and how people should avoid doing that when they're building with AI.

00:17:15.360 — 00:17:28.120 · Speaker 2

The core principle here is be an iceberg, not an ice sheet. And I sheet is wide and thin. And often a lot of times an iceberg is narrow and deep. Um,

00:17:29.520 — 00:20:15.460 · Speaker 2

this I this was a problem before AI, and I think this is one of the things that AI has kind of it's enhanced this problem and it's that every single decision you make as you build something incurs change debt and from my perspective, change that is kind of like the sinister cousin of technical debt. So technical debt is something we all know about is something that, you know, with AI based programing you're often incurring technical debt.

And you can something you can see but change debt is something that you can't always see, and it's incurred immediately upon making a decision. And it's closely related to feature bloat. So it's very easy. For example, just practically it's very easy for me to come up with an idea, sit down at a terminal and build something in a weekend right now.

Then once I have people touch it, they're going to be like, oh, it would be cool if it could do this and that and this and that. And now I've got a million suggestions coming from a from a lot of different ways. Every single thing I consider is changed at. So as soon as I even open it up to users I've incurred, I'm incurring change debt because now I have to figure out what what to filter out and what to keep.

Um, and that as I as I add things and build things, I'm creating, you know, error pathways. There's got to be machine states. You know, I have to consider the data model implications, so on and so forth. And AI can certainly help with some of that. But it's very hard to govern a constantly expanding system.

And so what I've kind of learned is to go narrow, understand the problem as deeply as you possibly can. And that means you know who is doing this right now and how, like, shadow them, watch them, follow them, ask them what happens in this scenario, that scenario. Understand as much as you possibly can about the problem.

Scope it as tightly as you can in a way that still delivers a viable solution to to move the problem forward to to move towards the solution, and then build that and then test the hell out of that before you, you know, and then ship it and then watch it and and but don't you know what? I've run into this problem myself many times where it's like, oh, while I'm doing this, I could just do this other thing or, you know, I'm showing it to a stakeholder and they're like, cool.

But if but what could it do? This, this, this and this, it could, but it probably shouldn't right now if we want to keep a handle on this thing, otherwise it's get out of control and we can create a lot of problems. So to me it's about staying staying narrow and deep rather than shallow and wide, right?

00:20:15.500 — 00:21:15.480 · Speaker 1

No, I mean, it's a very valid point because I mean, the example you gave was very simple, but it's very profound is that, hey, I want to support Google a Google or password or something. Okay, sure. That means you now have to go figure out a way to add the goo, integrate the Google password or and then now you have to add systems supporting that.

And if you have ten different microservices then Each of them have to be now enabled to support it. And so a simple requirement at a high level business system can have implications across the board. And you could basically like move fast but not in that speed. You might forget different things that you have to fix and create more problems in the in that process.

So. And I really like the way you were talking about to go deep and build an iceberg versus building an ice sheet. Right. Like so it's it's a valid point. Uh, did you see this happen at Fibo when you were building things like let's. Because feature addition sounds super easy with AI, but then it's just the shared context that messes everything.

00:21:15.600 — 00:25:28.490 · Speaker 2

Yes, very much so. I mean, we we started building hypo before AI, before cloud code, before Codex, before copilot. So we were still we were using AI to help us, but it was still very much like a human effort to build that thing. Um, and what we started out with was trying to solve this context window problem, and at the time I think the context was context.

Windows were like 25,000, 30,000 tokens, you know, so it was like 20 or 30 messages with a medium amount. And you were done. Um, and so, you know, we created this concept of hydrating a thread. So you've got a thread if we're constantly tracking the context window via the API calls, and then when it gets to 90%, it starts telling you you need to hydrate this thread.

And then what that does is ask another, um, you know, agent basically to summarize what's the conversation and then pass it on to the next so that you're not, you know, creating this context, but you're constantly trimming and hydrating, which is, I think, exactly what Claude is doing now, um, that said, you know, I think that was a decent idea, but the lesson I learned is we were basically betting against these platform providers that we could build what was on their roadmap faster than they could.

Right. Wild. meanwhile they have billions of dollars better tools like it was. It was not not the greatest idea. So had to pivot. You know, and we tried to pivot at first. We tried to pivot to this to like a long term memory. We always had a multi-model like approach. So you know, we had OpenAI anthropic deep seek connected.

So I you know, I think there's some value in that. I still see some tools out there that are like, oh, switch models. But they oftentimes they feel kind of gimmicky and not really. They're not again, they're not deep enough. Right? They're they're shallow, they're wide, but they're shallow. Um, and and the more you start to think about the UX of Chad, of ChatGPT or Claude, it's it's very incredible.

Like there's a lot of subtle things that are really hard to replicate effectively, um, when you're dealing with, you know, stateless inferences. So we pivoted, we tried to do the aging thing. It just the off The off for the agents was the hardest part. How do you how do you get allow someone to authenticate via API to their system, you know, to their whatever their login of that system.

And a lot of these systems are heavily disincentivize for allowing you to do that, because they don't want you to just hook into their back end and start, you know, replicating their functionality. So a lot of their moat is is built around rate limiting. And, you know, like Reddit is a great example. Like they have heavy, heavy, heavy rate limits.

You can't just hook into the Reddit APIs and start pulling subreddits and all the everything in there and passing it into AI. You can't, you can't. You're you're limited. So we turned away from that and then we went to AI for beginners, and that's kind of where we ran out of gas was. And I think that was maybe a decent idea too.

Um, but the, the UI, we were, we were torn between AI for beginners and AI for kids. And I think both of them have their merits in the market. And there's there's space, I think, in the market still, but it's AI for beginners is really hard to distribute because these people are AI naive and oftentimes are distrusting of AI.

So you have to overcome that. Um, and then AI for kids has its own marketing problem and UI problem. It's really hard to build a good clean UI for kids. So and of course you got to deal with the, the, the content gating, um, and like making age appropriate content. So we reached this point where we were like, all right, you know, we've pivoted three times now.

We've created all these different features and concepts in this app. Some of them are half baked, some of them are fully baked. We had like 3 or 400 people on the wait list, and we just we just weren't going to convert. The unit economics didn't make sense. So I was like, we need to put this in hibernation and, and come up with a better.

00:25:28.490 — 00:26:43.930 · Speaker 1

You mentioned like a couple of things that I agree with you. Like one. This idea of having multi-modal, I would say actually actually providing multiple models in a system when you provide something. I think initially that was like a moat that people were kind of putting together, like there's a project called Higgs Field.

I don't know if you know that, um, it allows people to kind of make choose different, uh, I think video models and generate videos and things like that. A guy recently took that entire project or similar idea and open source it, and you can plug in whatever model. The second idea that you talked about was this idea of AI agents and then the like authentication around it.

But what I am noticing on my side is that this idea that, yeah, you can bring AI agents to connect to different systems, but the governance is completely missing. What if an AI agent goes things that it needs to do something and deletes a database? How do we add governance? How can we, uh, separate a human's interaction with a database or a system with something like, You know, like a human agents, basically.

And the third piece that you were talking about is interesting is what what is left? Is there any moat, any, anywhere for anyone unless you have massive amount of money? What do you think about that?

00:26:45.530 — 00:28:12.070 · Speaker 2

Um, I do think that there, there is a moat, but it's a narrow moat and it's it's in the boring stuff. It's it's in the place. From my perspective, it's important to think about where these platform providers are not going to go. We're going to have trouble going. What's not worth it to them? Um, consumer based AI, you're cooked like not a shot, you know, unless you have a ton of money, right?

Or you create a wearable or something. But again, you need a ton of money to do that. So consumer based AI, unless it's incredibly niche and scoped down to like a very well-defined subset of users that can't just use, you know, it needs to be additive right to to whatever their preferred AI platform is. Otherwise they're simply not going to use your tool.

They're going to use court, um, or ChatGPT or whatever. Um, on the business side, I think that's where there's more opportunities. And I think it's more in like the regulatory areas, building custom, you know, custom LMS. There's some moat there. Um, I think in heavily like, heavily regulated areas, very niche business processes, you know, um, those kinds of areas.

I think that's kind of where, where what we're, what we're left with until there's some sort of a paradigm shift.

00:28:12.190 — 00:29:55.630 · Speaker 1

I agree with you. And I'm, I've always been a strong believer that, uh, you could build any product like, anything that is similar to another product. But if your product is great and if the customers love using your product, then that's your moat. Your moat is the actual customer experience, and I don't.

I know that people don't talk about in that from from that perspective. Um, a simple example that I think about is there is X and then uh, meta started threads, right. We know over the years many people have stopped using something. But the reason why some people stay and use a particular product is because it adds some tangible value to their actual day, to their life, right?

And if folks continue to build that, that's going to happen. The reason why we switch models is because the model exhausts and doesn't add value to what I'm building. Then that's the advantage of AI today, right? Like when you build and especially with what you were talking about, like the everything trap, you know, if you add everything trap and you consider that, hey, I have to build something, but let's talk about what the value of a feature is.

It kind of and then you look at all these different things that you have to build and then consolidate them in a way that, okay, these are the important ones. Then I think it makes it becomes easy for you to start making these changes. Start adopting AI faster. Um, but it's always interesting to kind of hear from builders like you, uh, on how you approach these things.

So, um, for the listeners, what what is your best practice around how you feel? These are the three things, four things that you feel you apply on a day to day basis, that that should be kind of taken care of as you start using AI.

00:29:56.750 — 00:33:27.190 · Speaker 2

Okay. So number one is, um, is creating a framework for your code base. So, you know, a component architecture, a data model. Now you can use AI to do this. But before you start actually having the AI going code in, you know, in in a repository, you want to spend the time to architect the solution rather than just building.

And that's always been the case. But it's it's even more. It's very tempting to just unload on, you know, Cod code terminal and be like, go build all this stuff for me. But if it doesn't have guidelines and rules about how it should build, it's not going to do a great job. So so number one. And in Cod there's kind of like a number zero to that number one, which is you really need to understand what you're building, what problem solving like.

And it's not just a simple flat like oh, I'm solving the problem of automating invoice imports into this financial system. Like, I don't know, the problem needs to be defined at a much deeper level that you want to get to, that the solution, the depth of your solution is going to be it roughly parallel or equal in depth to the depth of your problem definition.

So if you define the problem in a very shallow way, you can't expect a super deep solution that, you know, deals with all these failure states and edge cases and whatnot. You need to define the problem in the context of who is working on it. What are they trying to do? What are they doing now? What do they need to be doing, you know, at all points so that those those two that's kind of important.

That's like number one. And then I think number two and I'll just keep it to two is clean up after yourself and have quad coke to it. So what I do is, you know, when I'm done with the quad session, I tell it. All right. Clean up, you know, clean up all these files we've created, clean up all the, you know, and I create I make it create a map.md file at the root.

So you get a doc cloud folder in your repo. Inside of there is a map MD file. And that's telling the the clot the agent or you know, the instance or the inference, whatever you call it. Um, I call it the otter. You know, the otter syntax. It's telling your syntax where, where everything is and where to look and that it's kind of like a table of contents.

Right? Because if I'm going, you know, if I put my if I try to empathize with the AI, which is a weird thing to do, but I do it often. Okay, let me empathize with this tool I'm using what's going to help it perform best. And it's going to it's going to be for it's going to start from somewhere. It needs to start from somewhere.

And it needs to know, okay, where should I go look. Instead of me telling I look through the whole code base, I tell it, go look at this specific component, this lambda function, this set of tables and and all of that is discipline. It's taking the time to formulate the prompt and any supporting files for the prompt.

And you just get much better outputs rather than rapid firing, you know, and then reacting to problems. It's better to be proactive, you know, and then the problems that you're going to face are going to be deeper. They're going to be more complex. So you're making it harder for yourself at the end of the day.

But the solution is also deeper, and you're solving more of the problem. You know, that you've defined. So that that to me is the two like principles I operate on.

00:33:27.230 — 00:33:34.070 · Speaker 1

Yeah. I mean, that's awesome because I also do something similar. Like I do not know if you use it. Do you use obsidian by any chance?

00:33:34.070 — 00:33:36.990 · Speaker 2

I don't I don't, but I don't.

00:33:37.030 — 00:34:53.700 · Speaker 1

Yeah. I mean it kind of blew up when I was using obsidian way before. Uh, like Andrew Karpathy kind of talked about like knowledge, knowledge, second brain kind of concept. What I do is every cloud session I end with, uh, I, uh, my instruction is that before you close, we'll look at everything that we've talked about, all the tickets we worked on, every conversations we have had, my perspective and put that into an EMD file, obviously very similar to you into a vault in obsidian.

And what obsidian does is basically it's like this massive vault with file system, uh, with like a bunch of different relationships. So you can across like a week, start to build relationships around these documents. That's what it does. And it kind of becomes like a knowledge base that you kind of continue to keep on.

And then what you can do is when you have all of this knowledge base, you can basically sub take all of this and provide it to like say you have a marketing team and you say, hey, this was my philosophy when I was building this product. This is how we were evolving this. And then the marketing agents and the marketing team can then look at that and start creating the right messaging product messaging around this.

So I feel like there are some really neat systems and applications of what comes out of an interaction with the cloud agent. That can be very interesting as well.

00:34:53.740 — 00:35:52.480 · Speaker 2

So yeah, I mean, I, I haven't used obsidian, but you're you're convincing me I might have to check it out. I've heard a lot about it. Um, but I basically do, you know, I do essentially that, right. Like, I create, you know, we use Jira, so we have sprints. We name our sprint also after animals. So like, you know, we're on anteater sprint, for example.

And so everything every document that I create is prefaced with anteater dash. You know, um, usually the sprint number 24.2. Um, and then I use those documents in combination with the clod UI to constantly refine product documentation. You know, a pricing sheet, a product roadmap, this or that, these things that kind of have to evolve over time.

Um, and all that's reliant on having a pretty high level of organization and structure, you know, behind your, your process of building. I think. Right.

00:35:53.360 — 00:36:47.580 · Speaker 1

Yeah. No, I think it's pretty interesting. I mean, I'm glad to see some of these new standards. I mean, obviously formed was their age and, and like, uh, with cloud design coming and then Google opening stitch their dev open source design formed, uh, which is actually also really good for marketing teams.

If you're like working on web, web or app, you can basically define a pretty standard design strategy, uh, that stays within designed and then agent cannot. It becomes like the Bible. It cannot do anything more. It cannot change the spacing. It cannot change the colors. It has to use this open GPL, GL formats for laser and things like that.

So it's pretty interesting how all of this is kind of standardizing. And for solopreneurs like you or who are who are building companies like Cloud Fruit, I feel like this is the best time because you can basically scale out to like so many different functions with like bunch of agents.

00:36:47.620 — 00:36:57.100 · Speaker 2

No, I know it's it's pretty incredible. I haven't done the multi-agent, I haven't done the multi-agent thing yet because I find myself to be the bottleneck. Oftentimes

00:36:58.260 — 00:37:33.640 · Speaker 2

my my business partner who's also my wife, her name is Anna. She's a designer by trade, a UX designer. And so she does use clod design. Um, and so then, you know, I'm kind of more responsible for the development process. She's responsible for the design process, and we share product. Um, that said, I at every one of those stages, and I'm sure she would say the same thing I become the bottleneck or we become the human becomes the bottleneck.

Now, um, which maybe was always the case, but it's very it's way more obvious now. And so I, you know, I,

00:37:34.680 — 00:39:04.020 · Speaker 2

I try to avoid creating too much stuff for me to look at. I'm very much a minimalist by heart, too, and I think that while I want to create a deep solution, I don't want it to just be a bunch of garbage all the way down. I want clean and deep, which means I'm consistently, you know, refining as I go down deeper and deeper so that it's a clean solution.

And then at certain points in the process, it's like I, it's like letting my room get messy and then cleaning it all up. Right? But then when I clean it up, I'm recreating the system for how I clean it. So I'm like, I'm doing this thing where I just like, go, go, go go go, step back QA walk through the whole process.

I use playwright, but I still do it manually by hand because I just humans are going to be using when I'm building. And so I, you know, if I was building a book or some sort of like something for agents to consume, I would just let the agents build for the agents. But since I'm building for humans still, you know, we haven't quite got to that stage yet where we're building software for AI yet, but I'm sure it's coming.

But that the point being that I try to stay focused on only a few things at a time and get through them and then move on to the next thing. Because if I if I let all these agents run around and do stuff for me, I just am going to have too much to go through. And I'm never going to, you know, I'm never I feel like I'm never going to get anywhere.

Yeah, yeah.

00:39:04.060 — 00:40:04.120 · Speaker 1

No, I mean, these are like fascinating things. It's it's part of the evolving story. Um, I think I would I mean, I know a bunch of people I've talked to on X, and, uh, one of the things you should definitely experiment with this. Like if you have a mac mini is, uh, there's a project called Team Li, and then there's, um, there's this project, obviously open floor.

And, uh, team Lee is, like, managed open claw, but it's like they have pre created team teams. So you can say they just say hire a team of five marketing agents. It's basically manages everything you don't have to set up. But it it it it is a fascinating idea. I still want to see this at work at scale. Could really people like build like 25 agents, five doing marketing, five doing sales, five doing, uh, um, you know, something else finances in the team.

And could this really work in a quality way that we expect it to? Because if it works, it breaks everything that we have known in a, in a corporate setup for, for years. And that's what is the crazy stuff.

00:40:04.280 — 00:41:59.760 · Speaker 2

I mean, it's already it's already breaking that model. Very much so I, I the way I operate is I try to do oftentimes fine. There's a lot of fluff all around me and a lot of stuff is being done that's extraneous and not really relevant or germane to the problem. And so I try to strip all that away. Right. And and so, you know, like I try to focus on the work and focus on the problem.

Now, if I had a certain type of project, if I had a use case, I would immediately try to do that. You know what I mean? Like if someone if, for example, if someone was like here, like go start this new company, oh, I'm approaching it completely differently and I'm going to try to have multiple agents running at the same time.

Um, because that's stuff that while it's going to be consumed by humans, it's not really solving a specific business process problem, you know. And so it's like, um, I just and also maybe I'm just not ready. Maybe I'm just scared, you know, I don't know, like, maybe I'm just scared to run because I don't I trust the machine a lot.

And I spend a lot of my time trying to. I've come a long way. Trusting the machine. It's something that I've always had this really good instinct of. Like I've worked with a lot of people who are really brilliant, but when something is wrong, they don't trust the machine and they think they're right. And so they try to find out why the machine is wrong.

And I've always been the opposite. I assume I'm wrong, and I'm going to try to find out why I'm wrong and why the machine. You know what? I know it's functioning as designed. Basically, I'm not functioning as design. I'm I'm a rationalizing human. But I'm not rational. I'm irrational. The machine is rational.

Logical. So anyway, all that said, it's like I need to trust it more. And so maybe, you know, maybe this is my like nudge to, to try out some running some agents and letting them. Yeah.

00:41:59.760 — 00:43:18.720 · Speaker 1

Yeah I mean it's it's fascinating. I mean we, we had Congress last week. We are obviously going in that direction. Uh, for us as a company. You know, we talked to customers and we are anticipating a world, not just us. Everyone is that when humans like us, um, start to scale agents right now, as I say, metrics wise, we are still human.

Population 0.01 to 0.0. 3% of the people are like a exploring agents. But the scale at which this is going to work, the scale of every human running 15, 20 agents, and then the effect of that on systems and system designs and APIs rate limits. That's the world. And can databases scale? Can databases support that scalability requirement and the availability requirement, because agents are going to be running 24 over seven.

Right. And so your systems have to be up and down. So we as a company are working towards, uh, building a world that supports that. Obviously, cockroach is already like a highly scalable database. So we have natural capabilities built into our database. But it's going to be very interesting as to how all of this shapes.

So on the database side itself, there is so much excitement to see how this all shapes and comes together. So it's very interesting.

00:43:18.760 — 00:44:40.030 · Speaker 2

Now I'm fascinated by a man. Like one of the ways I think about AI a lot is it's it's really augmenting the individual contributor. Right. Not waiting for is a way for agents to talk to each other in a work environment, you know? So what what do humans do? We come together in these meetings and we come to an agreement and understanding on who's going to do what, but what's what's kind of hidden underneath.

That is we come to an we have an agreement and understanding on how we're going to work together. What are the rules that in things that are governing and bounding our work? And I, I you know, I don't think that that those tools, the, the or the that capability doesn't quite exist yet for my agent to, you know, coordinate with your agent and be like, hey, let's do this project together, Okay, you do this.

I'll do that. Okay. We're going to use this design system. You know we're going to do this. We're going to. And it's like creating that governance structure. Like you're alluding to earlier is so critical for when to to, I think to scale these agents to a point where they can actually, you know, go out there in the world and solve problems.

Right, right. And then, of course, there's the ethical concerns about what what what is that? What are the implications of that for for humans. Yeah. Yeah.

00:44:40.430 — 00:45:23.650 · Speaker 1

Absolutely. Yeah. I mean, the implication of just the way this world is shaping up for humanity, right? Like, I mean, I, I joke about it with a lot of my friends. I am for humanity like I, but I'm also for progressive humanity. Like, if AI can help us become more progressive and, um, get everyone's life, make it easier, then I'm all for that.

But we have to do it in a in a ethical, considerable way. And that's why I love projects like autoencoder. Um, and a big part of, you know, ethics groups on X and different communities is, is that we really want to understand how these AI agents are thinking, because it does help us, uh, to protect us from doing situations where all these movies that we have seen will come to life.

00:45:23.690 — 00:46:37.470 · Speaker 2

So but yeah, there's that scene in The Matrix where they're like, you know, how did we get like this? And it's like, well, nobody understands how the the thing works at its core. And I think I agree 100% with your point that, like, we have to have visibility into how they're thinking and how they're, you know, if we call it thinking, how they're processing, uh, and so that so that we can, you know, create the correct boundaries around it, you know, why can't why can't AI help us vote, for example, why?

You know, why is it so? Why can't you know? There's a lot there's a lot of things that I think AI could be helping with that it. It's not right now. Um, you know, and you know, why can't we change our whole political model? Why can't we have, you know, an AI that's running and allowing, you know, a politician to say, hey, what should I do?

And everyone votes on it, and the AI synthesizes it into a decision and feeds it back to the politician, and they do that thing rather than the model we have now, which feels archaic, right? Like, it feels kind of silly. It's it's stems from, you know, thousands of years ago. So anyway, like I digress, but you know, the I agree 100% on the ethical.

00:46:37.510 — 00:47:17.210 · Speaker 1

Yeah. The applications are still I just feel we are still scratching the surface. Uh, you know, it's been a very different episode from the kind of episodes I have done before. You know, we usually talk about a lot of these different ideas, and I love talking to you. You know, just talking through your thought process around how you're building and the way you think AI is going to shape things.

And and I really appreciate you bringing your, you know, the ideas on ethics as well as your own philosophy to these things, which is which is very interesting. So for everyone listening in, uh, go follow Sam on, uh, on your LinkedIn, is that where you're most active and cloud fruit, uh, that you do? Sam, where can people follow you?

00:47:17.250 — 00:47:18.490 · Speaker 2

Yeah, yeah. Link.

00:47:19.610 — 00:47:23.570 · Speaker 2

Com. Yeah, those those are probably the two best places to find me.

00:47:23.690 — 00:47:36.130 · Speaker 1

Very good. Yeah. So. And keep pushing. Uh, you know, amazing content out there. I, I really enjoy reading it, so I'm gonna I'm gonna keep following, and you will see me probably post, uh, reactions on that as well soon. So. Yeah, we'll go from there.

00:47:36.170 — 00:47:43.330 · Speaker 2

Yeah, I will say I also have a medium and a Substack page where I do. I get a little bit more edgy over there. So that one's.

00:47:44.090 — 00:47:54.930 · Speaker 1

I think the world needs more opinions, you know. So. So it's all good. Uh, well, thank you so much for coming on, and I really appreciate you sharing your, uh, world with us. So thank.

00:47:54.930 — 00:47:58.250 · Speaker 2

You. Yeah. And thank you for having me. This is a great conversation I appreciate it. Yeah.

00:47:58.370 — 00:47:59.370 · Speaker 3

Thank you. Yeah.