AI and Home-Cooked Software

“Every line of AI-generated code is a plausible-looking liability. It may pass basic tests, only to fail spectacularly in production with an edge case you never considered.”

Every time I read something along the lines I have to wonder whose code these people review during code reviews. It’s not like the alternative is bulletproof code.

I've made this point before, and in the short to medium term, I really do think it's one of the biggest and most underrated uses of AI. If I am making a tool for myself, and only myself, and if I deeply understand both the inputs and the expected outputs, and if it's a one off script or tool designed to solve a specific problem I have, then a huge swath of the issues with AI coding go away.

It's still not perfect, but it is dramatically easier to be fast and productive, and it is a huge leap in capabilities for people who previously couldn't code anything at all, but had deep enough domain knowledge to know what tools they wanted, and approximately how they should work, what kind of information they should ingress, and what kind of information they should egress.

Also a lot of the AI code reviewer tools catch bugs that you wouldn't catch otherwise

I feel personally targeted :D

Programming classes didn't work out for me in college, so I went into sysadmin with a dash of Devops.

Now I can make small tools for things like syncing my living room PC to a big LED panel above the TV (was app-only but there's a Python reverse engineering which I vibe-coded a frontend for) or an orchestration script which generates a MAC address, assigns a DHCP reservation in OPNsense, and created the VM or CT using that MAC and a password which gets stored in my password manager.

I could have done either of these projects myself with a few weekends and tutorials. Now it's barely an evening for proof of concept and another evening to patch it up to an acceptable level. Knowing most of the stuff around coding itself definitely helped a lot.

I've been thinking in a similar way over the past year or so—we're seeing the emergence of more widespread access to custom software for use cases that previously never would have justified the investment.

There are so many situations where a little program can save one or a few people hours of work on tedious tasks, but wouldn't make sense to build or support given traditional software development overhead, but that becomes realistic with AI-assisted development. This is the idea of a sysadmin who has automated 80% of his job with a variety of shell scripts, borne out into many other roles.

We've seen the early phases of it with Replit and Lovable et al, but I think there's a big playing field for secure/constrained runtime environments for non-developers to be able to build things with AI. Low/no-code tooling increasingly seems anachronistic to me: I prefer code, just let an AI write and maintain it.

There's also a whole world of opportunity around the fact that many people who could benefit greatly from AI-built programs are simply not particularly suited to building them themselves, barring a dramatically more capable AI, where I think enterprising software engineers can likely spin out tons of useful stuff and address a client base they might never have previously been able to address.