Learning AI By Writing Three Programs At The Same Time

Once we stopped calling large language models AI, made a harness and put the LLM in it…

We only needed the human on the wire, and we nailed that in 1770 with the Mechanical Turk.

And more recently with automatic checkout machines, that have call a human to fix it button.

The AI version is you having a chat with AI, on any subject, and then the AI summarizing it an nothing what it learned.

Thus enhancing it self on the subject, related subjects, and by subtle analogy other things.

If you look at AGI as a network of unique harnesses, and having a humans, as tools, I think you can call it AGI, though I ma sure it is not the correct use of the term.

And this is where I was today with my three projects, coding them all at the same time.

My blog generator has AI to fix things, but I also made room for an AI that watches things.

The program is named odor, and the blog builder us under odor build, agent is under odor agent, odor it self becomes the entity in charge.

Like Hendrix in Altered Carbon (Poe if you saw the series), it is a general intelligence to watch over the blog, and eventually take ever.

The second program, is actually vibe-coded, I never looked at the source, but I do remind the AI that it can refactor, and we are on rewrite #4.

We did four from scratch creations, we called the second a failure.

AI took over the third one, and made it work its own way.

This is a very good project, that comes from the observation that, Visual Programming language interface is not 2D.

I just told the AI upstairs to freeze rotation of the graph, but keep the mouse wheel zoom, that’s still 3D.

And I don’t use multiple ports in my latest visual programming, they just have have input and output ports.

Packets that flow into them though input can set multiple properties, this way I can represent the visual programs as ffmpeg filterfraphs.

Made in 3D printed as text, where named edges connect the graph.

The third program is a poem narration application, it is very nice, it parses my writing style.

I write in couplets, a singe compute sentence, with an alfa and an omega.

AI split my poems by two consecutive new lines, called each a segment, and presented it with cue cards.

I asked for the scroll-wheel to perform the movement, and R key to record, the UI feels wonderful.

But this is one of those projects where I was humbled, because the first time I tried to do this, I ruined it.

I didn’t know how a microphone actually works, a mic can be hot, warm, or trouble.

In my first attempt I foolishly begun recording, as I clicked the segments que card.

And a cold mic, has a warm-up period that in Firefox is almost a second, and even in Audacity a popular audio recorder, it is way too long.

Couple of weeks ago, I asked AI how professionals use microphones, to eliminate the warm-up period where the mic is calibrating it self.

I did not expect what the AI had in store for me, it turned out in a studio the microphones warm up when the light turn on.

Long before talent arrives to make chicken noises, studio microphones are away on, but not recording storing that.

They are warm and warmed up, and when you press record, you just start storing the data that the microphone was already processing.

I never had trouble with code quality, because I either ask the AI to rebuild from scratch, or carefully use Signals.

Or just program in the style of the web browser, MDN style.

And with odor the catpea blog builder, I asked AI to rearrange the code to industry standards.

As the result of that and using muriel, a filtergraph language in JavaScript, odor has 120 unit tests, which is a lot, and this in fac t, has no stink at all.