Perfect summary.

It's like we invented a world where you can finally, _finally_ speedrun an enormous legacy codebase and all patted ourselves on the back like that was a good thing.

AI is very useful, but it so far doesn't write the type of code I can trust. Thus I use it but I carefully review everything it does.

He told it to calm down and just use php, it gave him 100 lines with no dependencies that worked the first time.

The Pieter Levels stack :)

Of course, this is ideal for solo entrepreneur. If you are employed, then you cannot finish it in 100 lines. How will you get paid to maintain it for the next ten years, and hire all your friends to help you?

I think this difference in incentives explains most of what we've been complaining about for the last twenty years.

Very few understand deeply what's happening within the computer between the cou and the bridges and the rest.

The fdiv bug in 1994 took us all by surprise because we were in a situation where bug couldn't exist in hardware, because it either works or it doesn't.

When I'm using firebase or aws, I don't know the underlying system, I don't know why some resources can be created with an underscore or other can't start with a number.

Yet it works.

We're working in layers where usually we only touch the last one. Yes, understanding the others is great to debug.

I'm even wondering whether we need tests when they are written by the same llm that wrote the code.

But I don't see it. Where is this glut of software?

Not saying that's how it should be, that's just the world I predict in the not too distant future.

I love thinking, but most people I know seem to experience it as a form of physical pain.

by managing the risk of failure and technical debt with a financial instrument, i have a lot more freedom to move fast and break things, and scale aggressively

Others are pointing out that you cannot understand everything - and that's true enough.

But you only need to understand what's important. The experience of a good expert helps you to find that out.

As a systems administrator the recent AWS outage in the Middle East is the best recent example. There will be roughly three types of companies, separated by their understanding:

- Don't Understand - these companies thought that the cloud would handle this kind of thing for them, and are probably going to be doing a lot of finger-pointing in the near future.

- Do Understand, Don't Care - these companies did understand that high availability meant going multi-region, but decided against it for whatever reason. Probably cost vs perceived likelihood. These companies know that they've made a mistake. Short term they're wondering how to survive it, long term they'll be re-assessing their risk acceptance. Many may decide to stay single-region, but at least understand why.

- Do Understand, Do Care - these companies will simply be checking that their procedures worked for any manual parts of their failover, plus possibly looking at any improvements they can make given the real-life experience they've gained.

An LLM is just going to tell you how to implement it. It's not going to be thinking "what sort of availability do we require?", it may never start that conversation unless explicitly prompted. And even then it's going to return consensus opinions, which may not be what you want when evaluating risk.

I'd love to think a lot of companies will be looking at this event and updating their own risk register or justifying their existing risk decisions for hosting. But let's be honest - most won't even have thought about it, and won't until it goes wrong.

No one understanding what's going on inside of complex systems in financially constrained environments built and maintained by average, at best, engineers is the norm and is what keeps the world running.

None of that is a symptom of AI. The only change AI brings is that even first person developers don't know anymore what the fuck they just deployed.

ASM was machine specific. C was portable but required expert programmers. C++ was even more user friendly, but still very hard for normal people. Today, most anyone can write a program in Go.

The more we abstract, the less knowledge/expertise is needed. So yes, programs are being built by people who don't really understand what they are doing. That is intended.

There are reports of industries trying to use these tools to generate as much as possible, sure. There are also people generating bad art and unpleasant prose and using llms to generate nonsense they don’t pay attention to.

I don’t see why that implies that you or I lost interest in tinkering with toys we build. If I want to spend 4 weeks understanding oauth a little better by implementing a client, I still can and I still do.

Automating our builds absolutely didn’t create a cathedral of complexity while nobody noticed. It did mean I can open an Free Software project, read the build file and understand how to build the thing. That’s the opposite of generating complexity.

I worry about our future generations as much as the next person, but this low effort pabulum doesn’t represent the thoughtful industry and hobby that I love.

No. Software is being centralized. If the snake oil AI companies are selling about the coming agentic age were true, then the end result is not "anyone can produce software" it is "anyone stupid enough to rent the ability to run their business from an AI vendor can produce software".