Neural Networks, Gently

Underneath almost all modern AI sits a neural network. You don't need the mathematics to use AI well, but the picture is worth having, because it explains why AI feels fuzzy rather than exact.

Start with one unit. It takes some numbers in, multiplies each by a weight, adds them up, and if the total is high enough it passes a signal on. That's the whole thing — a weighted vote. On its own it can do almost nothing interesting.

Now stack them. Put units in a layer, put layers behind each other, and connect them. Feed a handwritten digit into the first layer as raw pixel brightnesses. The early layers learn to respond to simple things — an edge here, a curve there. Middle layers combine those into shapes: a loop, a vertical stroke. The last layer combines shapes into a verdict: this is an 8. Nobody defined "loop" or "stroke". Those intermediate concepts were invented by the network because they turned out to be useful for getting the answer right.

That's the actual magic, and it isn't in any single unit — it's in the connections. Training adjusts the strength of millions or billions of those weights until the whole arrangement produces useful output. "Deep" learning just means many layers, which allows more of this build-up from simple features to complex ones.

You'll see models described by their parameter count — billions of these weights. Bigger models generally know more and handle harder reasoning, but the relationship isn't clean, and it's the single most over-read number in AI. A well-trained smaller model routinely beats a poorly-trained larger one, and for most everyday tasks the smaller, faster, cheaper model is genuinely the right choice. Judge by results on your own work, not by the size of the number.

The picture also explains the strangeness. There is no line of code that says "if the user asks X, answer Y." There's a landscape of weights, and your prompt is a path through it. Nudge the wording and you take a slightly different path and can land somewhere noticeably different. This is why prompting works at all, and why two nearly identical questions sometimes get differently useful answers.

It explains the failures too. The network has no separate store of facts to consult and no mechanism for noticing that it's off the edge of what it learned. It will produce an answer for a question far outside its training just as fluently as for one squarely inside it — same confident tone, same clean formatting, dramatically different reliability. Nothing in the architecture flags the difference.

Which is the honest summary of this whole level so far: a neural network is not a mind and not a database. It's an enormous, tuned pattern-matcher. That's less magical than the headlines and considerably more useful, because pattern-matchers have predictable strengths — and predictable is something you can work with.