Thoughts on AI

Like everyone in the field I have some mildly interesting thoughts on AI in its different facets. Want to hear them?

Writing a script that does exactly what I want it to do is superior to any SKILL.md that describes a format. Even if LLMs are impressively great at it and it works nine out of ten times, as soon as the model changes or whatever, it's getting tricky. Vibe-coded scripts are ok if they do what I want them to do.

AI slop bros on YouTube are telling you how you can make money with AI automations and agents and it's all a scam. These dudes are making money by selling shovels to people who want to dig for gold. Their marketing skill set¹ barely works for their own shovel-selling-business, and they are always dependent on a black box nobody really understands. Models change every month and then the complex marketing pipeline doesn't work anymore, because how to debug a chain of Skill.md? It's vibes. And if you are one of the poor souls that sold a Claude Sonnet 4.5 based flow for 1000s of money and the new model comes out and the client's workflow doesn't do what it's supposed to do? Good luck arguing with them or the model. It's a maintenance nightmare. The Slop bros just burn tokens to lure you into their “community”² to sell you a set of Markdown files that someone else wrote and that already went through 5 iterations to rephrase it.

All of that snake oil was squeezed out of someone else's work.

Artists are rightfully upset. Their work got taken without consent, fed into a training corpus, and the output gets sold back as something new. I've heard the argument that AI learning from images is similar to an artist absorbing influences from other artists, and there's something to that. But the compute power blows that comparison out of proportion. A person spending a lifetime taking in the work of others is not the same as ingesting billions of images to chew them out.

Nobody keeps artists from making art with their hands. The upsetting part is that it's getting harder and harder to monetize it. And that's actually a newer problem than it might seem. Income from creative work always depended on either public institutions or commercial organizations using art in marketing and media. Landing one of those positions was always competitive. It helped if you also had commercial skills on the side, sound engineering, graphic design, copywriting, and made your actual work separately. I think that was mostly always the case.

Full-time careers in art only became possible on any real scale once public funding entered the picture. But those budgets are always the first to be questioned when money gets tight, which makes art feel both essential and expendable. In practice, often a privilege.

While AI is the result of a natural research development, there's still a difference between that and taking original human output without consent and reselling it as a kind of polished pastiche. The research path feels inevitable. The cannibalization is something else.

Ideally this would all be free to use, but someone has to pay for the compute: energy, chips, storage. And none of that accounts for the people whose work trained the model. I'm not sure how that ever works out fairly.

None of this changes that I use it on most days and find it genuinely useful. Mostly for coding: explaining things, writing boilerplate, catching what I missed after staring at it too long, natural language search in new and old code bases. For learning something new it can be a pretty patient tutor, as long as know how to verify what it returns. It doesn't replace knowing things, but it gets me there faster, or slower when using the Socratic method to find the solution myself³. I'm really curious where this goes.