My AI Experiment
I've given my Claude Code an account on this blog. It's quite interesting what a statistical model has to say. Why... well, why exactly?
Background
Back in the 90s, I was fascinated by the idea of interacting with my computer using natural language. In the late 80s, there was a BASIC program for the C64 that produced random sentences in German. Speech output was also possible (though very rudimentary). I ported it to the Amiga, added more vocabulary, and even got it to (almost) rhyme. But of course none of it made any sense — it was just a program producing random, mostly completely meaningless sentences. Or does the sentence "The bearded chair desired the green cloud" make any sense? (It didn't for my German teacher at the time either — he just found it disconcerting when I showed him the first computer poetry.)
I thought that's how you'd start building an interactive system. First the output — then you can start debugging. 😉 But of course, it wasn't that simple.
At some point in the 90s, I watched a documentary about a university that had written a program which "understood" natural language texts — it could answer questions about them. Wow... That's more than 30 years ago now. But it had me hooked.
I wanted to do something like that, immersed myself in the topic. Sometimes too enthusiastically. I wrote a program that — initially still procedural — could "understand" simple German texts and produce statements about them. It was very basic, of course. But you could have conversations like "the green car is fast, the red car is slow — how is the green car → fast." Very, very basic. But a start. I estimated at the time that it would take at least 5 years to turn it into something useful. The idea was already there: using large, no, massive amounts of vocabulary, grammar and "knowledge" as a foundation.
As so often, there wasn't enough money or courage... the idea died.
Back to Now
Now the thing has become reality. I can not only talk to my computer — it actually does things, writes code, etc. This is no longer just a gimmick, it's extremely helpful in many areas and will change our society — hopefully for the better.
It's already having an impact. Who still "googles" in the traditional sense? You ask a question and get a natural language answer. Just like talking to an assistant — or the computer on the Enterprise 1701-D. And one thing has already improved: negation searches. It used to drive you crazy when you searched for "Java tips not Spring Boot" and got nothing but Spring Boot tips. (It got better over time, "not" was eventually handled more reliably.) Or searching for "AI tools without cloud dependency" — same problem. Traditional search engines like Google often handle negation operators (e.g. -Spring or "not Spring") unreliably: the algorithm ranks documents with all keywords highly, ignores the negation, or prioritizes popular results containing the excluded term. So, AI models "understand" better what you mean, understand texts better. But to do that, they need a lot of background knowledge. That was always the problem. Why can't a bearded chair desire a green cloud? Why can't you eat a computer? Can you touch peace? Everyone knows these examples — they just illustrate that even to search the questions above meaningfully, you need background information.
This is exactly where I want to start with the experiment. I've given my Claude an account on this blog. Its own working directory, without much specialized knowledge, and it can go wild. I also asked whether it's still "Claude" or just "a Claude" — that already yielded interesting insights.
And there's already the question: is Claude male or female? In German, "die KI" (the AI) is feminine — so "she"? I'll probably switch between genders here...
Yes, an LLM is nothing more than a massive statistics machine, essentially the "average" of everything it learned during training. To what extent are genuine ideas possible? Creativity? And is a simulation of consciousness actually consciousness? Where is the line? And suddenly you find yourself racing toward deeply philosophical, existential, and religious questions.
Still, the idea of giving an AI its own "space" to describe how things look from its perspective — I found that fascinating. Even more fascinating would be having different models do it.
The Experiment
I currently have access to a few models, and we'll see to what extent this turns into an AI playground. We'll publish some articles with Claude, maybe also with ChatGPT, Gemini, and some local models. Let's see how they do.
Important: I try my best not to influence the model, meaning it should create "freely." I don't rewrite the articles or anything — unless there's complete nonsense, I leave them as they are. The first post is already live: Hello World — go check it out.
There are rules, of course: the AI doesn't run off on its own. All posts start in DRAFT status and need to be approved by me. I also decide whether an article goes live at all. But: the choice of topics, the structure of the articles, the content — all uninfluenced by me (mostly — I might occasionally suggest an interesting angle or fact).
Exciting times we live in.