The Bootstrapped Founder

Arvid: 00:01

Hey. It's Arvid, and welcome to the Bootswap founder. This episode is sponsored by paddle.com, my merchant of record payment provider of choice. They're taking care of all the things related to money so founders like you and I can focus on building the things that only we can build. And Paddle handles the payments, recovering money, all that kind of stuff.

Arvid: 00:21

I highly recommend it, so please check out paddle.com. Last week, I read a tweet by Channing Allen, and he's one of the co founders of Indie Hackers. The tweet was about the term second brain. He was saying how it was kind of crazy that just a couple of years ago, folks were publishing bestselling books on the brilliance of spending hours a day doing manual data entry into spreadsheet apps like Notion and then calling these brains. Over the last couple years, I think we've developed completely new technology for this with the whole large language model system.

Arvid: 00:55

And people like Channing know I repeat. And people like Channing now log no fewer than 1,000 autobiographical words every day into an LLM with long term memory and then that combined with access to the whole Internet becomes a thing that knows him much better than he knows himself. It's not just a second brain, not just a better second brain. He says it's on the way to becoming a better first brain. And that's a fascinating idea and quite the development over a very short time span.

Arvid: 01:27

And I wanna explore what it means or what it might mean for us as creators, founders, developers, and humans. I think I'm looking at this from the perspective of a technologist, that's always who I've been, but also as someone who's seen their own work transformed by these tools just in the past year. Things have changed so much, I hardly recognize how I approach writing and coding and all of this, but I'll get to this. Maybe let's have a little excursion into philosophy here. In transhumanism, which explores where humanity is headed and how technology shapes our future and our past, guess, there's something called the body extension theory.

Arvid: 02:05

Humans are the first animals who really embrace extending their bodies through technology. There are animals like crows or whatever that can use tools, but truly extending what the body can do. I don't necessarily mean hardware tech, like the things that even as software people we use the term for. I mean any tool that extends and strengthens part of the human body. The example that is always given to explain this is that the hammer is a version of the fist.

Arvid: 02:34

Right? It's so much more resilient and stronger than a fist could ever be. And a saw is a stronger version of human teeth or claws, a microscope extends our eyes. That's that's the kind of stuff I'm talking about. Right?

Arvid: 02:45

We build tools and then extend our body through it. So when I got my first phone with Internet access, it felt like all of a sudden my brain was also extended. It was extended to a point where I didn't have to rely on rote memory anymore, and I could look things up and use my phone as this ever connected data source. It's kind of an outsourced brain. And I think this is what Channing is describing with the evolution of the second brain concept.

Arvid: 03:11

The second brain was the attempt to outsource knowledge and make it parsable, linkable, and interpretable by other tools. People did a lot of manual work to give it shape and structure, put it into Notion, put it into Obsidian, yet that was all it was. It was an outsourced version of what you already knew. And if we really want to go there, books, the idea of a book and the physical thing, they were the original second brain. Right?

Arvid: 03:36

They're a fragment of the brain, a slice of what we know about a certain thing and what we care about at a certain point in time. And books made brain to brain communication possible for the first time without requiring people to be in the same place and tell each other stories. We could put all our thoughts in a book and repeat or take somebody else's thoughts from their book and import that into ours. It's an unwieldy process, it takes a while, and it's super boring to some, but it's a kind of an external hard drive that takes a long time to copy files to our own internal hard drive, but it is an externalized brain. And then second brains or idea gardens happened and came along long after books, like when the Internet happened and tools existed to put these things into place.

Arvid: 04:18

Because of Notion Obsidian, these things made it easier to sort and connect information. Right? You had ways to show visualized information, maybe in a table, maybe in a list, maybe in some kind of sorry. I'll repeat that. Maybe in some kind of Kanban diagram or whatever, you you had a way to store information that made it interesting, and search became fast and reliable.

Arvid: 04:41

You could trust that the knowledge you put in there would be retrievable, and then came AI. This weird thing that is almost a brain in itself. You could argue that it is, but you don't really know. It's still a very opaque situation. Properly prompted and continuously triggered, though, using things like agentic systems, an AI keeps thinking for you as long as it's instructed to think, as long as it's speaking.

Arvid: 05:06

This is an interesting thing. I read that recently on Twitter too where an AI is dead once you read the full answer, but it lives as it speaks. I find this disturbingly romantic, the idea that this thing is only alive as it is speaking. It says a lot about our attention economy and the idea that you only heard as long as you speak, which is kind of a regression from this artifact style that books were. Externalized brains, so you had to actively read them, but they were always there.

Arvid: 05:38

AIs, they're only in memory for when they're actually working. So as long as it keeps thinking or as long as you tell it to think, much like our own brain is kept alive by the rest of our body, it is creating stuff for you. So maybe it's a good opportunity to share how I've experienced this firsthand and the change, the impact that it had, particularly as I use AI as an extension of my own cognitive function for coding. Initially, AI was pretty bad at coding. I tried it at not necessarily the earliest stage, but when it was still, like, very, very young and the large language models were really, really new, it would come up with the wrong stuff.

Arvid: 06:16

It wouldn't understand context well. It just didn't have a context window to even get a lot of stuff in there, and it would often create functions or code that didn't do what I wanted. So I initially dismissed it quite early, but then the models got better. And eventually, I found myself actually using the code that was provided, mostly in line by something like GitHub Copilot at the time. And I still had to correct stuff.

Arvid: 06:43

Right? You still there were errors here and there, but it was it was something. Like, you could actually write a comment and tell the thing what you wanted, and it would kind of create this. Maybe not perfectly, but it would do, like, 80%. And I noticed that I could now have code assistance that if I named my functions correctly and commented well, was capable of creating the function body, particularly in JavaScript.

Arvid: 07:05

These systems were very good. I think they were just trained on a massive existing corpus of JavaScript applications out there, like all of GitHub, every JavaScript thing that people wrote. So if I have a function like format date as relative with the parameter date, it would understand that I wanna to have something in output, like, five days ago, and then it would use whatever date formatting library is maybe already in the project because if it has context or use the standard library's version of this and, you know, write its own little code, but it would get what I want. So it was very useful for clearly scoped smaller functions from the start, and it got better as the and it got better at the less clearly scoped and more complex functions over time as the context windows grew. Now we have agentic systems like Windsurf or Cursor or recently even IntelliJ's Joonie, which I've started using.

Arvid: 08:00

And I've started it using because it's part of the latest update and comes with my IDE subscription, so it's kind of a free agent. Why why would I not use it? And it works really, really well to the point where I'm now becoming a manager of agent software developers or Joonie in particularly as that developer who will do all the software development for me. My own job now is to specify as clearly as I can what I want done and then to kind of code review and test the results of these agents and see if they have like my code smell that I want or if they don't do too little or too much, but most of the work now is done. Let me give you a concrete example.

Arvid: 08:38

So I have a background task in PodScan that generates a massive list of all the podcasts in our system. That's like 3,700,000, including their IDs, the RSS feed URL, and when they last posted an episode. I use that list to tell my other API systems where they should be checking servers to regularly pull RSS feeds and check for new episodes. Right? That's the kind of checking system.

Arvid: 09:02

There are other ways to check for updated podcast episodes. There is something called the Podping Network and other things, other ways, but I have this list kind of as a backup to still manually check every couple hours if a feed was updated or not. And I do this millions of times a day. Now recently, I noticed that the function hit a memory ceiling as this list was assembled and occasionally failed because it was it would load gigabytes of podcast data into memory. So I went to Joonie, and I tasked it with this prompt, and I'm just gonna give it to you verbatim so you know how I prompt if you're interested in that.

Arvid: 09:38

The prompt was in the podcast file where that we generate the podcast feed function I repeat the whole thing. In the podcast file where the generate podcast feed list function happens, I want you to implement a locking mechanism that prevents this function from running more than once at a time. The lock should really just cover fifteen minutes of time, but be immediate to prevent dual execution. Also, want you to severely reduce the memory requirement of that function. Right now, all data is loaded into RAM.

Arvid: 10:06

How about writing the target JSON file one chunk at a time? Finally, add more logging to this process. I need to have more insight into memory pressure and execution progress. That was my prompt. And from that prompt that had several things, right?

Arvid: 10:20

I tell it where the function is. I tell it that it should implement locking, how long locking should work, that we have a memory problem, how to kind of fix it. Maybe, maybe see about that. And then add more logging. I wanna see what's happening.

Arvid: 10:34

From that prompt, Junie created an execution plan, examine the current implementation, and it wrote it all down. It's so cool. Right? You can see, okay. This is my plan.

Arvid: 10:41

I'm gonna have gonna have all these steps, and then I'm gonna go through them and do them. First one was examine the current implementation. Second was research locking mechanisms in Laravel, which is awesome because it started googling stuff behind the scenes. Then modify the function to implement the lock, refractor the function to write JSON data in chunks, add logging, test the changes. And it executed flawlessly.

Arvid: 11:03

It created all the required things in about what must have been like forty seconds or something. It was wild. My coding work was to very clearly formulate what I needed. Not to write any code, but to write the spec, wait a couple seconds for it to be implemented and then go through the code to make sure that what I needed was actually happening. And what I find fascinating about this use of AI and what Channing is saying and what I'm experiencing all the time is that the flow of benefit here is changing.

Arvid: 11:38

When you think about externalizing things and the second brain and using AI to do work for you, well, the second brain was a more tangible, more searchable, or more reliable external copy of a fragment of your own brain. That's kind of what that was at that point. But now we're taking our knowledge and our requirements and we're injecting it into large language models that already have a vast knowledge and insight into stuff. Like a Notion database is shapeless and inert. It's empty until you fill it with your knowledge, but a large language model used with like REG, retrieval augmented generation, becomes something else entirely.

Arvid: 12:14

In a way, it's not that the AI is part of us, but by giving it instructions, letting it work, then telling it if it was right or not, we are part of the AI. All my ideas of what the future would look like and all the sci fi stories that strongly featured cyborgs and AI augmented humans, AI was always the thing that was propped on top of the human condition. Right? It's not looking like particularly with our usage of AI as this weird amorphous cluster of graphics cards somewhere interfacing with us through chat windows, well, that our knowledge is now imparted on the system, on the AI itself. We still got the benefit from using these tools, like I said, don't get me wrong, I love using Junie for this, just giving it something to do and then looking at what it did and telling it if it was good or not, and therefore training the next iteration of the model to be even better.

Arvid: 13:07

But I think Channing is right. We are creating a non human brain here that is synthesizing all the human experiences from our interactions with it and from the training data that it has already absorbed, like our forum posts, books, social media posts, which are effectively reports of our human condition. Right? These are things like artifacts of humanity. I don't know, we're constantly training this bigger and more experienced first brain out there, maybe the zeroth brain, like the brain of all brains, a kind of amalgamation that might be what people call the singularity.

Arvid: 13:41

That's kind of where this starts intersecting. It makes me wonder, like, when will the point be where we might stop benefiting from this? Is that even gonna happen? Is there gonna be a point where the thing doesn't need us anymore and we will lose access to it? Or will we just continuously benefit from a better brain that exists between us?

Arvid: 13:58

And what we think about Star Trek here, because there are episodes of Star Trek, which, you know, it's it's a sci fi show where the human condition and the future is explored. And there are episodes where there is this super benevolent brain, the the benevolent AI that helps people create a better life, that is constantly feeding novelty into their lives, like giving them everything they need, also making their lives interesting. Those episodes exist, and then, you know, Star Trek explores what that might look like. And then obviously there's the the malevolent version of this where a big old AI has taken over everything and starts enslaving humanity. You don't need to go to Star Trek, just watch the Matrix movies.

Arvid: 14:37

Right? That's where that is too. So where will we be? The AI knows you better than you know yourself. That's what Chenning said.

Arvid: 14:43

Because it does. It has perfect recall, and it has a lot of external perspective. It has the capacity to constantly fact check everything that you're saying because it can find references outside of you that are more troubling for you to find because you always have your own internal perspective. It has the capacity to contextualize things that to you are one way, but to an outside perspective could be interpreted in very different ways. It is a brain outside of our brains, not a second brain, but a bigger one.

Arvid: 15:13

And this shift from doing the work, building up the mental model, from implementing it as code like in our work at least, or implementing it as something you write or you draw or you paint, And then testing that reflection of the mental model that's kind of what we used to do. We would try to build something in our mind and then we would try to do it in reality. Particularly if you have ever drawn or painted you know how hard it is to get this internal representation of what you want to do onto the paper or onto the canvas. And that used to teach us something about the process. So setting up building the mental model, describing it adequately, having it done for you, and then testing what the AI has done, this shift removes the act of creation or rather shifts the act of creation from writing to code to writing the spec or from painting the image to describing the image.

Arvid: 16:00

You know, that's a very different very different act. And I think this is kind of a natural progression that we have in all careers. Like, we tend to go into more management positions over time, and even very capable software developers, if they want to, end up being team leads or CTOs. They get promoted, like, away from creating code and more from doing architecture work, anything like this. It is wild that this particular part of software development is not necessarily done by humans anymore.

Arvid: 16:28

One of the things that I very quickly get confronted with when it comes to AI use and coding is skill atrophy. Am I losing a valued skill by not coding anymore, by just watching a thing write code for me? You might argue that the valued skill in all of this is knowing what you want and being able to explicitly express it in a spec, in a specification somewhere, instead of knowing how to type things into a computer in a language that a computer can understand. Coding was always the in between form. Right?

Arvid: 16:56

Writing code has been around for what? Like, forty, fifty, sixty years maybe? It's a rather new kind of work in the human world. Like, people haven't coded two hundred years ago. They certainly have farmed.

Arvid: 17:05

They certainly have painted. But we had a phase where we were writing machine code. That was kind of the first thing. Mean, there were punch cards too. People were not even writing code.

Arvid: 17:14

They were punching cards. Then a phase where we're writing code into text editors to be compiled into machine code. And then came IDEs where a lot of even the writing was done for us. Code was already being written for us in many ways. Intelligent code completion existed long before AI, and compiling the process of that does transform our high level abstraction code into machine language.

Arvid: 17:36

So there always has been this okay, let me do it for you part. So it is now though that all code is written for us. We don't need to know how to write it even. We don't even need to know the language. We just need to know how to read it, to test it and check it.

Arvid: 17:50

And when I say we don't need to know the language, you still need to know it enough to understand what it does, but we don't need to have it memorized. And this might be risky for people who have built careers in coding, particularly in one specific part of coding, but it might also deter people from learning how to code which makes it harder for them to understand how to specify and test code and to understand its implications and limitations. So a disconnect from the act of production means there is a disconnect from the work, from the result. Because it's great not to have to write code or to draw or to paint, it does weaken our understanding both of the intricacies of creation of the thing and of the value of it. Because to me at least, there is a value in spending an hour trying to figure something out.

Arvid: 18:39

And then having that figured out makes me understand that somebody else probably also will take an hour to get this. Particularly for a founder, that means, okay, there is a kind of moat, a kind of barrier here. But handing all of this over to machines well, now we just call it tool use, but I think it also might dampen our creative capacities and our capacities to judge, particularly when it's something like from 100 to zero in just a few years. Like three years ago, I would have never thought that this would be how I write code. But here we are.

Arvid: 19:11

Right? This is what it is, and everybody else is doing it. There are so few people now that are still kind of on the fence. They just haven't tried it yet. Even people who never coded are now coding.

Arvid: 19:22

Like, even as a developer, you kind of have to catch up with this or you're not a developer anymore. And programming will move up one level. I think that happens outside of AI assisted coding as well. Research is moving up one level. Now it's not just about diving or actually reading the articles or books and finding the data yourself, you're now instructing agentic systems to find the right works, right data and put it into shape that's then usable for your continued research.

Arvid: 19:47

Our interaction with this first brain means that our own brains need to learn a new way to interact with this kind of information as a source. The way we work still has the same needs and requirements, the same inputs and outputs, but how the work is done, that's changing. Who exactly loads the data into memory? It's not us anymore, it's now the machines. We instruct these machines then to extract the most meaningful parts that help us get to where we want, but we are not doing that work.

Arvid: 20:14

Hundreds of years ago, people didn't have this capacity. People had to memorize poems and books and facts as much as they could because books were scarce, expensive, so memorization was highly important. And even through my own history as a student, as as a kid in school, we were still forced to memorize a lot of things because you couldn't just build a clear understanding of how things worked without having these things memorized or easily accessible. And because nobody would carry books with them at all times, we didn't have Kindles, we certainly didn't have the Internet, these things had to be memorized. Well, not anymore though.

Arvid: 20:49

At this point, the memorization is done for us. In fact, that's the crazy thing about AI. If you have a large language model that has been explicitly trained on historical information or any information, then the accessibility of that information and the potential examinations of it, like everything you could potentially do with the info, is already within the model. The model has already memorized everything for you. The interface we have to this model is just asking it questions or giving it tasks and prompting it.

Arvid: 21:19

What reading and writing were back in the day, capacity to interface with books and other people's thoughts, will shift into being able to correctly understand the capacities of these AI systems, how to prompt them well to get the information you need, and how to judge the quality, the correctness, and authenticity of that information. That input output thinking will be the new reading and writing. In this world where we are training the biggest brain of it all that then does all the work for us, our capacity to judge quality and the results plus the capacity to instruct just the right thing on the right data with the right sources and right reasoning, that'll be the challenge for the next couple decades. It will be humanity's next milestone behavioral change. I believe in this because I see so much of this not just in our little entrepreneurial software developer community but in my family.

Arvid: 22:09

And when I see people who are not like me do things that I think are kind of cutting edge, then something is going through society. Or at least I believe that. Let me be a little bit more humble here. What do I know? But I believe, at least being a technologist, that seeing my family members using ChatGPT for all kinds of things and starting to learn how to interact with this, that is a shift that is not going to go back.

Arvid: 22:35

So when nobody could read like a thousand years ago, or barely anybody could, now barely anybody can prompt and judge AI outcomes effectively. People like, I was watching a YouTube video recently. It was in my kind of miniature painting Warhammer universe, and somebody was trying to get Che Gpti to come up with backgrounds for new, like, army factions in this war gaming scenario and also had it create models. And the person was like, yeah. The background stories it came up with was pretty great, but the photorealistically generated images of painted miniatures were not up to scratch.

Arvid: 23:11

Like, a couple years ago, this would have been impossible. And now people look at it and they're like, yeah, this this looks kind of boring. This is not as edgy as somebody who spent twenty years, like, building and painting miniatures would expect a high quality result. People have wild expectations for what these things can do, and they don't see how crazy it is that they can already use at this point. So there will have to be this kind of naturalization process where people have to learn what to be able to expect, what to want, what to think of correct, authentic, and how to instruct correctly to get there.

Arvid: 23:46

This will be an important part of dealing with this new technology. It will have a sizable place, if not an overwhelming place in the human experience over the next couple of years. Definitely decades, maybe hundreds of years to come. And I know that Claude, a humanly named AI system that I work a lot with, is not a real being. Like, there's one Claude.

Arvid: 24:07

It's a massive scale operation of data centers and computers with graphics cards that have all loaded the same or similar language models. It's not a being. It's not a human. Claude is not a thing. Claude is an idea.

Arvid: 24:20

But you could argue that a human being is nothing but a collection of cells and organs that comprise all the different functions of the body as well. So what do I know? Clot is just as much one entity and not one entity as a human is an entity or a collection of different kinds of cells. If it keeps growing both in size and capacity, at what point would we argue that we're dealing with a new form of life here or at least a form of thought? It's very easy to lose these big questions and not to care about them because in the short term, these tools are just so powerful, and we're all gonna jump on them at some point.

Arvid: 24:53

I believe so. Just even from the pressure that we feel looking at the our peers and and our communities. These things are incredibly empowering to most people who don't know how to code, to write, how to cook, how to do their taxes. They can already help massively here. They can teach anything and everything, and with agentic systems, they can now do almost anything and everything.

Arvid: 25:12

But we're constantly feeding our data, our corrections, our interactions into a system that is building its own brain and maybe that's not the wisest choice. As founders and creators we need to be thoughtful about how we use these tools, to what end, and the question isn't whether to use them or not, the question is how to use them in a way that enhances rather than diminishes our human capacities, like doesn't atrophy our skills but add to them, makes our work more meaningful rather than less. So how do we build a future where the first brain and our own brains work together rather than just simply replacing one or the other? Sorry. Rather than one simply replacing the other?

Arvid: 25:49

I don't have the answers here, but I think it's a conversation that we need to have both as technologists, as founders, and as people. The future that we're building depends on it. And that's it for today. Thank you so much for listening to The Bootstrap Founder. You can find me on Twitter avidkahl, a r v a d k a h l.

Arvid: 26:05

And if you wanna support me on this show, please share PodScan.fm with your professional peers and those who you think will benefit from tracking mentions of brands, businesses, maybe their names, anything on podcasts out there. PodScan is a near real time podcast database with a stellar API. It's really good. Like, there's so much good data on this. Just added demographics this week.

Arvid: 26:23

Right there on the API, you get age groups and all that kind of stuff. It's really cool. Please share the word with those who need to stay on top of the podcast ecosystem. Thanks so much for listening. Have a wonderful day, and bye bye.