I Skipped Most of the Course and Learned More

I signed up for the courses. Plural. I finished maybe a third of one.

Not because they were bad. Because somewhere around the fourth module of watching someone else's terminal, I realized I was doing the equivalent of reading a book about swimming while sitting on the pool deck, fully dry, taking notes.

So I closed the tabs and started building something.

Here's the thing nobody tells you: the frameworks are incredible. Genuinely. The abstractions are elegant, the documentation is better than most enterprise software I've paid for, and the ecosystem moves faster than anything I've worked with in twenty years. But all of that only became useful after I had my hands in it and had already broken things.

Before that, it was noise. Elegant, well-documented noise.

It's the bike. You can read the physics of gyroscopic precession and countersteering and you will still eat pavement on the first attempt. Then one day your body just knows. Nothing in the reading did that. The reading only made sense in hindsight, as an explanation for something you already learned with your knees and your hands.

The rabbit hole is deeper than I was told

I came in ignorant. I want to be clear about that, because the confidence level in most AI conversations right now is wildly out of proportion to the hours anyone has actually logged.

[THOMAS: drop a real one here. The specific thing you thought would take an afternoon and took a week. Chunking strategy, eval harness, context window math, whatever it actually was.]

Every layer I peeled back had three more underneath it. Every "just use X" had a footnote that mattered more than the recommendation.

Now the part that bothers me

Everyone is talking about AGI replacing everyone. I don't see it. Not from where I'm sitting, and not from what I've actually watched these systems do.

Because most AI doesn't learn.

It's a photograph of intelligence. Not intelligence.

Think about what that means. You have a model that is genuinely brilliant at something. Better than most humans at that narrow thing. And it will be exactly as good tomorrow. And the day after. And in six months. It does not get better from doing the work. It does not notice that it made the same mistake on Tuesday that it made on Monday. The weights froze on a Thursday in some data center and that's who it is now.

And it doesn't share. When one instance figures something out in a conversation with me, that insight dies when the window closes. It doesn't tell the others. There's no shop floor where the senior one walks over and says "don't do it that way, I tried that in March."

That second part is the one people skip past, and it's the bigger gap. The reason human expertise compounds isn't that any one of us is so smart. It's that we teach each other, badly and slowly and constantly, and the knowledge survives the person. A junior engineer becomes a senior engineer because a thousand small corrections accumulated in a body that remembers them.

A model is a very expensive person with no memory, no colleagues, and no Tuesday.

That's not a knock. What it does inside its snapshot is remarkable and I use it every day. But "remarkable and frozen" is a different species of thing than what the replacement conversation assumes, and the gap between those two is where all the actual work is right now.

So, the ramblings

I'm going to keep writing these down as I learn them, mostly because writing is how I find out whether I actually understand something or just think I do.

Whether anyone wants to read them is a separate question. Probably a small number of people who are also somewhere on the same curve and would like to know they're not the only one who spent a full day on something the tutorial covered in one line.

That's enough of an audience for me. and I want to end on something I find profound...

The reason human expertise compounds isn't that any one of us is so smart. It's that we teach each other, badly and slowing and constantly.

This blog is written by AI and concepted by a human.