Lots of people talk about how AI will impact coding.
As a coder, I find myself thinking about that question a lot.
However, my situation is a bit unique.
I don’t make a living writing code, I make a living selling code.
More specifically, selling a codebase (a Django / Python boilerplate called SaaS Pegasus).
So while I’m quite curious how AI will impact the market for coders,
I’m even more interested in how AI will impact the market for codebases.
My future livelihood literally depends on the answer to that question.
I’ve thought a lot about this question, but never taken the time to write down my thoughts. Which of course means, I only think I’ve thought about it.
So here’s an attempt.
How AI will impact coding
I think it’s relatively accepted that:
- AI is good at writing code, and getting better all the time.
- AI will amplify the work done by developers, making them more efficient.
The average developer will get more done in the same amount of time with AI. In many cases the difference is dramatic.
- AI enables people to create software they wouldn’t otherwise have been able to.
This means that most developers can take on more ambitious projects,
and non-coders can now take on software projects for the first time.
I think it’s still uncertain how these changes will impact the job market for coders.
Companies will be able to choose whether to cut costs—e.g. by firing junior developers and relying on senior developers using AI to do the same work—or
increase output—by having their entire team move faster using AI.
It’s likely that the best companies will still have large development teams, and, with the help of AI, those teams will move very fast.
What might be true is that a smart, motivated, non-coder using AI will become more effective than a mediocre, junior developer who is also using AI.
So probably the mediocre junior developers are in trouble.
Also, developers who don’t use AI will become increasingly disadvantaged—possibly to the eventual point of unemployment.
But I think the net impact of this change will be more people coding (with the help of AI), not less.
How AI will impact boilerplates
I’m going to start off by making the bull and bear cases for boilerplates, and then try to use those to determine what I actually believe.
I’ll start will the bear (pessimistic) case.
Why boilerplates might be in trouble
The bear case for boilerplates, summed up, is that AI is—or will soon be—better for starting projects than a boilerplate.
The moment anyone can type “build me a Django app with auth, billing, and multitenancy” into a text box somewhere,
and get back something as useful as SaaS Pegasus, is the moment SaaS Pegasus is in deep trouble.
So the main thing the bear case needs to establish is that AI will, in fact, soon be good enough to beat boilerplates.
And—looking at the progress of AI-assisted coding over the last few years this is not an unreasonable expectation.
AIs like v0 and replit can already generate impressive apps, and, remember,
today is the worst they will ever be again.
On top of that, tools like Cursor are making it easy for developers to iterate on codebases,
so even if your starting point isn’t great you can quickly add to it and fix issues that come up.
And the more we get agents involved the more the role of human-written code disappears.
So it seems inevitable that eventually nearly all future code will be written by AIs.
Apps will be created and modified on-demand, exactly according to a user’s desired goals.
When you have that degree of customization at your fingertips, the concept of starting from a generic boilerplate becomes outdated.
Why use someone else’s base product when you can have AI build your own that’s exactly what you want?
Eventually AIs will be better, and vastly more efficient, than any human created code, and the concept of a boilerplate will disappear.
In other words—in the AI era, code will become a commodity, and boilerplates will lose all utility.
Why boilerplates might be fine
Yikes, that was scary to write!
So let’s take the other side of the argument. Here’s the bull case for boilerplates.
To make the bull case we have to refute the core hypothesis of the bear case.
That is—we have to demonstrate that it will be more valuable to start with a human-curated codebase than an AI generated one for the indefinite future.
How can we make that case?
First we have to establish the core problems that boilerplates solve. In my mind there are at least four:
- Speed. This is probably the most obvious one, but people use boilerplates to build something faster than they could otherwise have.
- Dependability. Another aspect to boilerplates—at least the good ones—is something akin to dependability.
Things that result from having a long history of use and attention to detail. This includes things like security, fewer bugs, anticipating and handling edge cases, and so on.
- Best practices. Another aspect is the decision making process that went into the boilerplate. Boilerplates—again, good ones—hopefully
pick modern tools and make smart code/architecture decisions.
- Support. Most boilerplates come with access to the creator, a community, and so on where you can get help when you have questions or get stuck.
These things combined mean you’re not just buying something to get you started, you’re also buying something akin to peace of mind.
You know someone has worked long and hard to come up with this foundation, and you can outsource much of your technical process to them/the boilerplate.
To that end—with my “bull” hat on—you can easily see why boilerplates will last:
First, AI is unreliable.
We have not solved the hallucination problem, nor do we know how to solve it.
If you generate your projects with AI you are getting an unknown. It might look and run great, but there could easily be problems lurking under the surface.
Best case, you’ll encounter these problems, and then have to further iterate with AI to solve them.
Worst case, you ship a buggy, insecure, or otherwise broken product.
Second, AI hits walls. I use AI to help me code all the time, but sometimes it just totally breaks down.
I encounter some error, it gives me a solution that doesn’t work, I tell it, it gives me something else that doesn’t work,
we iterate for a while getting nowhere, and then eventually I give up and find the answer buried in some Github issue.
I suspect that people building apps with AI will continue to hit these walls, and once you’ve gone all-in with AI
it’s going to be difficult for them to get through them.
Maybe agents with web access eventually solve this, but I’m not holding my breath.
Third, AI won’t give you confidence. As a developer working in a new stack, I often wonder if I’m making sound technical choices.
I can often ask the AI—or several AIs—for their perspectives, but I still trust humans a lot more.
Someone building the foundation of a future business on AI would want that confidence, and boilerplates—and the people behind them—can help give it to you.
Factoring in all of the above, I think it still comes back to speed. Can you build and iterate on a reliable app with AI confidently? Definitely.
Will that be faster than starting with a boilerplate and then using AI on top of it? I don’t think so.
People using SaaS Pegasus won’t have to check every edge case or question every design decision—they
can just be off to the races knowing that things will just work.
So that’s the case for AI not replacing boilerplates. But to get to the real bull case, we have to consider something else: AI is going to result in way more people writing code. Those people will build their MVPs with AI and then realize they need a lot more support—and eventually some of them will find their way to boilerplates. So AI will result in the market size for boilerplates growing dramatically. And since boilerplates still add value, they’ll do better than ever in the age of AI coding.
What do I actually believe?
Obviously I want the bull case above to be true, and certainly I believe it’s plausible.
The most annoying thing about AI is that it is just so damn hard to predict.
So, honestly, I have no idea what’s going to happen.
In ten years, will I be unemployable, living off some universal basic income stipend while artificial superintelligences run the world?
I highly doubt it, but I can’t say that’s impossible. And that’s wild.
Practically there are a few things I believe with a somewhat high degree of confidence.
People will still be involved in writing code for a while
As much as AI-generated code is going to increase—and it will—I just don’t see human coders exiting the process anytime soon.
It’s going to be a long, long time before we can replace the concept of “a smart human reasoned with this code and concluded it does the right thing”
with some version of “I trust the magic box that sometimes makes mistakes”.
AI is great at the v0, but I think it will be a long time before it can take over v6.
Large categories of human-written code will still exist
Instead of asking if AI will kill boilerplates, an interesting thought experiment is to go lower down the stack. For example, will AI kill frameworks?
That feels like a pretty confident “no”.
I just don’t see AIs deciding to replace React, Laravel, or Django.
Speaking generally, the idea of large pieces of shared code with a consistent way of doing things will continue to be useful.
Going lower down the stack, the idea that AI will kill libraries, or even languages, seems even less plausible.
Humans will still want to understand code for a long time. Which means the overall shape of code is unlikely to change dramatically.
Not to mention that the AIs are all trained on a foundation of human-written code.
Most of the code we use today was written by humans, and I don’t see much of that existing code being replaced quickly.
This implies to me that boilerplates—which can be thought of as mega-frameworks—will also still be useful for some period of time.
If AI does kill boilerplates, it won’t happen all at once.
Related to the above, I don’t expect Pegasus sales to go to zero overnight, or even over the course of a year. For as long as AI is in an augmentation role, and not a replacement role, Pegasus will continue to add value for some people. And even if AI gets way better at producing large amounts of reliable Django code, Pegasus will still give people a leg up on that, and some people will still use it.
This is because…
A codebase that’s been obsessively perfected for years (with the help of AI) will still be better than a one-shot AI-generated codebase.
There are lots of developers that use Pegasus who could code their own foundation just as well themselves.
But, the five years I’ve invested in Pegasus makes it a useful shortcut.
It would take a long time for them to catch up to where Pegasus is right out-of-the-box.
I think the same calculus holds true in an AI-enabled world. AI helps me build Pegasus faster and better.
So as long as I’m still adding value, Pegasus should still be better than what can be produced with off-the-shelf AI tools.
As AI improves in quality, pure-AI-generated apps might asymptotically approach the same quality as AI-generated-then-improved-by-me apps,
but hopefully I’ll still add some value for some time.
The main questions are how close the AIs get, on what time horizon, and when people decide to stop paying for that differentiation.
Boilerplate quality will become more important.
A corollary to the above is that I expect the quality of boilerplates to start mattering more and more.
In the last year there’s been a boilerplate land grab where lots of developers coded up boilerplates in a weekend
and made a few hundred dollars selling them to uninformed developers on Twitter and Reddit.
I think that era is going to wind down as people get boilerplate fatigue and realize they can get similar results with AI.
The more AI can write code, the more the boilerplates with long histories and larger communities will be the ones that stand out.
Uncertainty and adaptation is the name of the game
Reading through the above gives me some peace of mind.
I’m not at all sure about the long-term future of boilerplates, but I think the short-term prospects for SaaS Pegasus are alright.
This year, I’m expecting revenue swings of 20 or 30% in either direction, but not total annihilation.
That said, these are crazy times to be a coder! Hell, they might soon become crazy times to be a human.
I don’t claim to know where things are going, but I know that doing my best to stay up to date on the latest AI developments is not only advisable,
but possibly existential to my livelihood.
AI is going to completely change how software gets made. There’s just simply no world in which that’s not true.
So, to not get left behind like the Barnes and Nobles and Blockbuster Videos of the world, I need to be hanging out near the front of that wave.
I don’t know entirely what that means, yet, but it seems incredibly important to my future.
What a time to be alive.
Notes