I've shipped a lot of AI. Now I'm learning how it actually works.
The gap I finally admitted to
I've shipped AI into products for months. Trading agents, coaching apps, automation pipelines, CLI tools. Real things, in front of real users.
And for most of that, I couldn't have told you what was actually happening inside the model. I called an API, got a response, wrapped it in good product. That works — until you want to do something the wrapper can't.
So I stopped pretending the gap wasn't there.
What I'm actually doing
Not a course I watch at 2x and forget. I'm building the small stuff by hand.
- A linear model trained with plain gradient descent, no framework — just to feel what a gradient is.
- A tiny neural net forward and backward pass written out, so backprop stops being a word and starts being arithmetic.
- The math underneath: derivatives, matrix multiplies, why a loss curve does what it does.
The rule is simple: if I can't rebuild it small, I don't understand it. Watching someone else's notebook run is not understanding. It's spectating.
What's clicked so far
The abstractions were hiding more than I thought. .fit() and .predict() are three lines that quietly contain everything hard about the field. Once you write the loop yourself, you see where models actually break — bad data, wrong scale, a learning rate that's off by an order of magnitude. Things no prompt fixes.
Debugging changes. When you know what's under the call, a weird output stops being magic and starts being a bug you can trace. That alone is worth the time.
Product intuition gets sharper. Knowing what a model can and can't do — really, not marketing-really — changes what you decide to build in the first place.
Why now, honestly
I'd rather be the person who builds on top of models and understands them than the one who only knows the API surface. When the abstraction leaks — and it always leaks — the second person is stuck. The first one keeps going.
I'll write more as I go. Real progress, real dead-ends. Not a highlight reel.