Machine Learning Basics
Your Brain Didn't Write the Rules Either
The most powerful pattern-recognition system on the planet — including the one you're using right now to read this — learned everything it knows without anyone explaining the rules.
The Idea
Here's what makes machine learning genuinely strange: the programmer never writes down what the machine should do. Instead, they show it thousands — sometimes billions — of examples, and the system figures out the structure on its own. This is a radical departure from how software used to work, where every decision was explicitly coded. Rule-based systems are brittle; the world is too messy for exhaustive instructions. Machine learning sidesteps that problem entirely. The core mechanism is almost insultingly simple in principle. A model makes a prediction, compares it to the correct answer, measures the gap (the 'error'), and then adjusts its internal parameters — millions of numerical dials — very slightly in the direction that reduces that gap. Repeat this billions of times across enormous datasets, and something remarkable emerges: the model begins to capture genuine structure in the world. It learns that certain arrangements of pixels tend to mean 'cat.' That certain sequences of words predict what comes next. That certain patterns in medical scans correlate with disease. What the model learns is never a tidy set of rules you could read and check. It's distributed across a vast web of weighted connections — more like an intuition than an argument. This is why these systems can be so capable and so opaque at the same time. They don't know why they got the answer. They just got it.
In the World
In 2012, a team at the University of Toronto entered a machine learning model called AlexNet into a global image-recognition competition called ImageNet. The competition had been running for years, with researchers painstakingly handcrafting features — telling their programs things like 'look for edges, then curves, then combinations of curves.' AlexNet did none of that. It was a deep neural network trained on raw pixels and labels, left to discover whatever patterns it found useful. It won by a margin that stunned the field — nearly 11 percentage points ahead of the next competitor. That gap was so large it looked like a measurement error. It wasn't. AlexNet hadn't been told what a dog's ear looks like, or how fur differs from fabric. It had simply seen enough labelled photographs that its internal parameters had rearranged themselves into something that could reliably tell the difference. What made this moment a genuine rupture in computing history was the implication: tasks that had resisted decades of expert hand-engineering — recognising faces, transcribing speech, translating language — might all yield to the same basic approach. Show the system enough examples. Let it adjust. Repeat. The 2012 ImageNet result didn't just win a competition; it triggered a wholesale rethink of what software could be, and set in motion the AI landscape we're navigating now.
Why It Matters
Understanding this one shift — from rules to examples — changes how you think about almost every AI system you encounter. When a recommendation algorithm surfaces something you didn't know you wanted, it's not following someone's explicit theory of your taste. It found a pattern in the behaviour of millions of people that happens to predict yours. When a language model writes something that sounds authoritative but is completely wrong, it's not lying — it's doing exactly what it was trained to do, which is produce text that statistically resembles correct text, not text that is verified as true. This distinction matters practically. It means these systems inherit the biases and gaps in whatever data they were trained on. It means they can be spectacular in one domain and baffling in an adjacent one. And it means that asking 'how does it know that?' is often unanswerable — not because anyone is hiding something, but because the knowledge is genuinely distributed in a form no one can directly read. Once you see this, you stop being surprised when these systems are uncannily good and startlingly brittle in the same breath. That's not a bug being worked out. It's the nature of the thing.
A Question to Ponder
If a system learned everything it knows from examples rather than rules, is there any meaningful sense in which it 'understands' what it's doing — and does your answer change anything about how much you should trust it?
Get a new one of these every morning.
Start learning with Thinkable