Almost a year ago, I was laid off from a large company.
TL;DR: Spoiler: I’ve already found my place, so this is not a story about AI taking my job. This is a story about something else.
At the time, everyone had completely lost their minds over AI. Management was running around intoxicated by the idea that for 100 bucks a month you could replace pretty much anyone. And not just developers. They were replacing everyone - lawyers, accountants, and probably even cleaners if they could figure out how to replace them with AI. Everyone was insanely excited that now you could do absolutely anything without understanding shit about it, literally in 15 minutes. Very quickly, these contraptions started going into production.
One of my last tasks at that company was an automatic llms.txt
generator for company websites. The idea seemed trivial: the application crawls a website, sends the pages to an AI, and then assembles an llms.txt
file from the content. When I left, I thought: well, why not make my own generator? It’s that simple. And a couple of days later I had a prototype.
It worked, of course, but in some ideal world where websites have only a few dozen pages. But not every website consists of 20 URLs. Sometimes there are 200, or even 2,000. Stripe’s documentation, for example, has 4,500 pages and the number keeps growing. weather.com has more than a million. And those are exactly the resources that need an automatic generator, because you can write an llms.txt
for a 30-page website by hand in an hour.
And pretty quickly it became obvious that simply going through all those URLs synchronously wasn't going to work. You need background jobs, parallelism, state management, race conditions
, deadlocks
, retries, and a whole lot of other shit. And none of my attempts to dump those development problems onto AI worked either - it would fix something in one corner and everything would immediately stop working somewhere else. That was the first warning sign I noticed back then, but I had no idea how to interpret it.
Anyway. After a few months, the generator was finally finished in some usable form - here it is, if anyone is interested:
I tried promoting it for a while, tried some more - without much success. I realized that, as usual, this kind of thing needs investment, and I moved on.
Meanwhile, the fever continued. You can see it even today, with Google already declaring models deprecated that were the smartest thing around yesterday. And that got me thinking. People are seriously discussing AGI now, trillions are being spent on it… what exactly is behind all this? I became curious: how good is AI actually if you remove the human from the process?
And I came up with this project - Schrödinger’s website.
Inside the box, instead of a cat, there is a story. The AI invents it on the fly and then lays it out as an HTML
page itself.
Let me emphasize the rules of the game: On schrodinger.svcpool.com you can see what it can actually produce - and how well - under conditions of complete creative freedom. This is what AGI is supposed to be about, right?
Well, what you get is some spectacular fucking garbage.
Gemini is the canonical piece of shit. There is really nothing to discuss: the model dutifully assembles something out of familiar patterns, but calling it creativity is a stretch. Anthropic Opus is noticeably more interesting. This one you can actually read. You wouldn't say you're looking at complete madness. But it still turns into some kind of HTML
fever dream, stitched together with visible seams and parts that don't quite fit.
When AI gets complete freedom - “come up with a beautiful story, whatever you want; make a page; create anything you like, as long as it’s interesting” - it turns out that creativity is not quite as good as people like to tell us. The model is excellent at repeating patterns. It can combine familiar constructions. It can do a hell of a lot of things, BUT ONLY IF the human already KNOWS what they want. And preferably knows the subject well enough to judge whether the AI’s solution is actually any good.
A power drill will never come up with a clever way to drill a hole diagonally to get around the rebar. AI seems to be the same kind of thing - in skilled hands it can perform miracles. In the hands of a layman and a dilettante, it will produce layman-and-dilettante-grade shit - only much faster.
You can check this yourself. Go to schrodinger.svcpool.com, click around, generate a few stories and see what comes out. And if your “wave function collapses” - that just means the budget ran out :))) and if someone throws some money at me, we'll increase it.
Discuss in the comments? Someone definitely disagrees, right?