Let me be honest: before we built our first own model, I thought AI was mostly a question of enough data and enough compute.
Then came a project for Apple. Object recognition, on the device, no cloud. And one number that decided everything: 80 MB. The model was not allowed to be any bigger.
That was the moment AI went from theory to craft for me.
The task
A model that recognises objects is nothing special in itself. There are ready-made architectures, pre-trained weights, tutorials. If you have unlimited memory and a server, it's a weekend.
On the device it's a different story. The model has to be small, answer fast, not drain the battery, and still recognise reliably. Each of those goals pulls in a different direction.
Bigger model: better recognition, but too slow and too large. Smaller model: fits, but recognises worse. And somewhere in between is the point where it works.
We went through that loop a lot of times. Not because we did something wrong. Because there is no other way.
What you learn along the way
The first thing you learn: a model is not a program. You can't look inside and fix a bug. You can only change the data, change the parameters, train again and see if it got better.
The second: the data matters more than the model. We spent more time collecting, checking and cleaning examples than on the architecture. One bad image in the training set costs more than a clever layer gains.
The third: making it smaller isn't just "leaving things out". When you shrink a model, it loses accuracy first where it has seen the fewest examples. So you need to know exactly which cases matter to you, and add more of those.
The price of a percentage point
At the beginning it went fast. From "recognises nothing" to "recognises most things" is a big step, and it doesn't take long.
After that it gets expensive. Every further percentage point of accuracy needs more data, more training runs, more evaluation. And at some point a single percentage point costs more than the first 80 combined.
At 80 MB there's another layer: every megabyte you give the model has to be saved somewhere else. You trade accuracy for size, size for speed, speed for battery.
Once you've been through that, you read benchmarks differently. "Two percent better" is not a small thing. It's often the difference between a project that pays off and one that doesn't.
Why this matters more today than back then
Today everyone talks about AI. Most people use it through an API and pay per token. That's fine. But it hides what happens behind it.
Fact is: whoever has never trained a model sees the price per request and takes that for the cost. Whoever has done it once knows: behind it are data that someone had to collect. Training runs that someone paid for. And decisions about what the model should be able to do and what not.
Exactly that helps me on every project today. When a client says "the AI should just recognise that", I hear: what data do we have for it? How many examples? How do we measure whether it's good enough?
Back then the 80 MB were a constraint. In hindsight they were the best education in AI I ever got.
A model that has to fit into memory forces you to understand every decision. That is exactly what many people who only use AI today are missing.
Written by me. The thoughts, the values, the learnings, the mistakes: all mine. Grammar and spelling are corrected by our own twin model, trained on my texts. Sometimes a stumble stays in. That is mine too.