Sandcastles Made of Bits
The future will be full of temporary software
Back in 2018, a friend showed me something wild. In Indio, California, there were houses where renting them out for just the two weekends of Coachella would pay the mortgage for the entire year. That meant you could either pocket the rental cash for the rest of the year or keep the place as a “free” vacation home.
I figured this little arbitrage would already be priced into the market. But a quick look on Zillow said otherwise. Many HOAs in Indio banned short-term rentals, but a few didn’t—and the market hadn’t figured that out yet. Some of these homes were undervalued, and the opportunities were sitting there, hiding in plain sight.
This got me thinking: if this was happening in Indio, it was probably happening in other markets. So, I wrote a small program to scrape real estate listings and compare them against Airbnb rental prices. The goal was simple: find homes where their rental potential made them more valuable as a business than the price they were being sold for on the market.
The program was rough and hacky but good enough to solve the problem —it was like a little sandcastle made of bits.
What Are Sandcastles?
If you know how to code, you’ve probably built something like this—a throwaway program to solve a specific problem. It’s not built to last, it’s not polished, and it definitely isn’t scalable. But it works. And once the problem is solved, you move on. That’s a sandcastle: software that’s quick and disposable.
I’ve built hundreds of these sandcastles over the last 25 years of writing code. Here are a few random examples:
A custom Mac app that helps me with my email
A tool to manage a podcast I ran for a while
A coding agent that edits software based on Github issues
Most people don’t realize how common this is. Coders build sandcastles all the time. The reasons for building our own tools are numerous. Sometimes you just need a quick and dirty solution, sometimes you don’t want to pay for someone else’s software, or sometimes you just build something for fun on the weekend.
The complexity of these projects range from full blown applications to 5 line scripts. But building these types of things historically has been a freedom unique to those who know how to code.
But does it scale?
Today, most software companies aren’t building sandcastles—they’re building systems to serve thousands, millions, or even billions of users. Scale is expensive. You need performance, reliability, security, error handling, etc. Google employs tens of thousands of engineers to maintain a search engine that, on the surface, is just a white page with a text box. Why? Because supporting billions of users is hard.
But sandcastles can do less.
A sandcastle isn’t built to handle millions of users—it’s built to solve one problem for a small group of users. So scale is much less of an issue.
Where’s my agent?
Right now, LLMs are making existing coders 10x more effective. LLMs fine tuned for various tasks and chained together in workflows have created agents. These agents can write boilerplate code, debug issues, and even learn APIs on the fly. A program that used to take hours or days can now be spun up in minutes.
As an example of how sandcastles can come together within this new paradigm, I threw together a quick little route planner that lets me plan out my cycling workouts based on the type of training that I want to do. This sandcastle was created for the soul purpose of this substack.
Prompt (which was a stream of consciousness): i need an app that helps me decide on a cycling route for exercise. I should enter my current location, it should give me an out and back or circular route with map that is optimized for exercising. I should be able to put in my target heart rate zone and the app should estimate the speed / cadence I need to ride on the route to maintain that zone.
After a stream of consciousness dump into Vercel's v0 and a little bit of tweaking I converged on something usable. This took me 15 minutes end to end. Here is a screenshot.
But I know how to code, the real impact comes from when people that don't know how to code can use agents to build their own sandcastles like I can.
In the near future LLMs will help you build a sandcastle to book your European vacation. Or manage the hiring process at your company. Or manage your companies expense approval process. Or submit your expenses, Or teach your kids how to read, or whatever.
Based on the trajectory and speed that I see in the world of coding agents we probably aren't far off from being able to allow normal people to build sandcastles. The fact that my Mac email helper was written in Swift (a language I don't know) pretty much entirely by GPT-3.5 means in many ways we are already there and just lack the tools to piece everything together.
Whats the TAM?
In a world of sandcastles, the TAM (Total Addressable Market) for many classes of software will start to collapse. Traditionally, software companies succeed by scaling: solving problems for millions of people at once. But sandcastles operate at the opposite extreme—solving problems for just a few people at most.
When software is tailored so specifically to a few individual’s needs, there’s no need for mass-market solutions. Software stops being something you consume—it becomes something you create.
But then what happens?
What is likely is that as agents get stronger sandcastles will evolve into something more: living systems.
With stronger AI, sandcastles won’t just solve the problem —they’ll adapt and grow with you over time. Agents will continuously reform, repair, and optimize your sandcastles based on your needs.
For instance, the agent managing my cycling training could notice that I’ve been spending more time mountain biking rather than road biking. Unprompted, it might automatically adjust my workouts to focus on improving my strength for short-duration climbs.
Similarly, the agent managing my email helper could recognize when I’m describing something that would benefit from a visual. It could automatically adapt by creating a simple whiteboard app, allowing me to sketch a quick diagram to include with the email.
One of the dominant narratives in tech right now is that “coding agents will enable the one person billion dollar software business”. This is probably wrong.
We’re heading into a world where anyone can build software. This doesn’t just change how software is created and sold—it changes what software is. It will no longer be a product you need to buy. It will be something you create, adapt, and evolve with.
What a time to be alive.
🤖 AI
With OpenAI's new ChatGPT Pro you get access to OpenAI's state of the art reasoning model which is pretty amazing. The model is currently priced at $200/mo which is out of reach for most people.
With o1-mini and o1-preview rate limited this gives a glimpse of what kind of future we could end up in. One where access to strong systems could be the realm of the wealthy. This would be very bad and makes me want to spend more time working on open source models / infrastructure.
💡Steal this idea
I've been using Anki to help my kids study for spelling tests. I wish there was an LLM powered Anki style app which teach concepts using interactive prompts & voice by spacing the concepts out over time. In the spirit of "sandcastles" maybe I'll have the Cursor agent cook something up.
📊 Data
🔗 Cool $hit
Apollo research was able to setup a situation where Claude 3 Opus attempted to copy its weights to another server to avoid being turned off.
NVIDIA's Self Paced Training looks pretty awesome for those of us who choose to remain technical in a world of sandcastles.
Tavus rolled an app that lets you video chat with Santa.
🔈What I'm listing to
Acknowledgments
Thanks to the friend who came up with the term "sandcastles" to describe what I was pitching him over dinner last week. You know who you are. Thank you.