Just me, a keyboard, and a computer

Or, why I care about open-weight AI models

As a ten year old in 1998, the Internet held great possibility. I was lucky that my family had more than one computer at home, and even an extra phone line I could tie up for hours connected to our dial-up service. Personal websites were all the rage, and, at the annual Scholastic Book Fair at my school, I was thrilled to find a book titled Make Your Own Web Page! A Guide for Kids. I took it home, fired up Netscape and SimpleText (we had a Mac), and I was off to the races.

Sadly that first web page I built is lost to time, but it featured a tiled background of blue sky and clouds, a picture of my dog Nadja, and a MIDI tune that would play on load. I learned, somehow, that our ISP provided free hosting for personal websites. I connected to it with an FTP client, uploaded my files, and I was live on the web. But what was that cgi-bin folder? And how did that cool DHTML cursor effect I’d installed on my page via copy-paste actually work? The explanations came in the form of programming languages: Perl, and JavaScript, which further opened my eyes to the possibilities of what I could build. With the help of Matt’s Script Archive I was able to add interactive features to my site, and, through brute-force trial and error, customize them a bit myself.

This introduction to programming opened an important door for me—I’ve been building things on the web ever since, and the delightful combination of creative problem solving and near-instant feedback has never gotten old to me. So much has changed in that time, but one thing has been constant: with only my brain, a keyboard, and a computer, I can create useful, fun, interactive things for other people to use, and even pay for.

The last two years or so have been a period of incredible change in the world of software development, driven by the rise of AI coding assistants. Putting aside for at least this moment as much of the debate about utility, impact, and ethics as possible, I will say that I’ve had fun “increasing my ambition”: putting together successful software projects I never would have previously thought I could tackle on my own, or with such a small team.

There are myriad pitfalls and flaws, but ultimately I think these tools offer the opportunity to democratize access to the kind of things I’ve been building for myself for ages. In the same way I’ve written small scripts to automate tedious tasks, we can now build entire apps that will only ever be used by one person to automate a tedious part of their job (and we’re doing this at Melior).

This paradigm has, however, introduced something more troubling to me: the slowly dawning realization that operating at the top of one’s game in the professional world of developing software today increasingly requires ongoing payments to an AI model provider.

We can debate the value of software artisanship, the impact of an ever-expanding garbage patch of vibe-coded slopware, or whether we’re staring down a decade of skill atrophy that will leave us without anyone who understands the fundamentals, but increasingly I do find it hard to argue with the idea that AI is moving the baseline of what can be done by an individual programmer in a given period of time. Programming professionally without using AI is seemingly headed towards being a niche role, at least in the grand scheme of software jobs. Many people seem to be confronting that feeling, these days, and I’d be lying if I said I didn’t have periods where I deeply resent what AI has precipitated here, even as I embrace the tools day to day.

What does excite me, though, is the rapid advancement of open-weight models. I’ve been experimenting with them for a few years now, downloading the newest ones that could reasonably be run on my desktop Linux machine with a consumer GPU¹. This was just for fun at first, but increasingly out of a sense that participating in the world of self-hostable AI and supporting the organizations that are working in it can be a way to shape a future that I think we’d all benefit from: one where the power of these tools is not simply available pay-per-use from a small handful of corporations, but something that you can own and shape to your personal taste.

For a long time these models have been mostly a bit of a novelty (at least for my purposes)—slow and constrained by my hardware in either their raw capability or the amount of context they could process. The smart models I could run would have so little VRAM left over for context that they were only useful for working with short pieces of text or generating small scripts. The smaller models left room for a bigger context, but weren’t smart enough to really do much. Mixture-of-experts models took some pressure off the GPU itself, but the open ones were still way behind the frontier.

The last few months, however, have marked a remarkable shift: the Gemma 4 series and Qwen 3.x models crossed a threshold of functionality to where they’re genuinely useful, and can run locally with enough context to do real work. Qwen 3.8 27b, new as of a month ago, is the first one that has had me looking seriously at options to upgrade my computer to run it faster.

Spending $1,500 on a GPU might not make any sort of financial sense as long as the major AI labs continue to offer $20/month plans, or perhaps even versus just paying per token from an inference provider, but to me it presents an exciting return to a sense of personal digital sovereignty over my tools: the idea that I can still sit down and build, just me, a keyboard, and a computer.

All the better if I can run it on solar power.