How I work today

I’m not a terribly nostalgic person, and while I do appreciate how programming used to be and the personal benefits I received from doing it manually, I also realize that my passion for technology and the will to build more and better is stronger than the pure activity of writing code. Writing code itself has changed a lot over the years, from the daunting task of writing terse assembly code trying to shave every cycle to high-level languages and package systems that allow for rapid development of fairly complex applications.

LLMs are the latest step in that direction, and they are now a normal part of how I work. What I get out of them is roughly proportional to what I already know.

Two things changed in practice. One is reach: AskMei.ai, Little Control Room, and Fractal Strike are written in languages I never formally learned, which used to be a good reason not to start at all. The other is that experiments and tools became cheap. A utility that was never worth a week of work now takes an afternoon, and an idea I would have talked myself out of can be tried and thrown away in a day.

OVERSWEEP is the opposite case. It is C++ and Vulkan, in a domain I have worked in for decades, so there is no language barrier to cross. What I get there is speed on work I could do anyway, which makes it a fair way to judge how much these tools are actually worth.

What does not carry over is the product itself. Cheap experiments and good tooling get you to something that works, and that is where most of it stops. Getting to something people actually want to use still takes expertise, taste, and a lot of iteration by hand: how it feels, what to cut, when something is nearly right but not right yet. That part has not become any faster, and it is most of what separates a product that stands out from the many that merely work.