A model is a machine that makes a guess

In the last item, a machine learned the shape of junk mail from examples. That sounds like something a mind does. It isn’t. Underneath, a model is a plain piece of arithmetic with a few numbers you’re allowed to change. To see that clearly, we’ll build the smallest possible model by hand — one that guesses a flat’s price from its size — and watch that the same skeleton scales all the way up to the giants.

A model is a rule with adjustable numbers

Suppose you want to guess a flat’s price from its floor area. A first attempt: price goes up by some amount for every extra square metre, on top of a base price. Write it as a rule:

price = size × A + B

A is how many pounds each square metre adds. B is the base price a flat starts at before you count any space. Pick A = 3,000 and B = 20,000. Now the rule is a working model. Feed it a 60 m² flat:

60 × 3,000 + 20,000 = £200,000.

That’s it. That is a model. A and B are its parameters — the adjustable numbers, the knobs. Everything the model “knows” is stored in those two numbers.

The knobs are the whole model

Change the knobs and you change the model’s mind. Set A = 4,000 and the same 60 m² flat is now guessed at £260,000. Set A = 2,000 and it drops to £140,000. The formula never moved. Only the two numbers did. The skeleton — size, times a knob, plus a knob — is fixed; the setting of the knobs is where all the behaviour lives.

This is the idea that scales. The models behind face recognition or finishing your sentences are the same thing wearing a much bigger coat: instead of two knobs there are millions or billions, and instead of one multiply-and-add the numbers flow through a longer, more tangled calculation. But the shape is identical. Input goes in, numbers get multiplied and added, a guess comes out. When people say a model “has 70 billion parameters,” they mean it has 70 billion knobs. That is the whole secret behind the scary number: it counts the knobs.

It computes a guess — it doesn’t look one up

Here’s a distinction worth holding onto, because it comes back again and again. The model does not store a table of flats. It never memorised “60 m² costs £200,000.” It computes the answer fresh, every time, from the knobs. Hand it a size it has never seen — 63 m², 108 m², 41.5 m² — and it produces a number just the same, by running the arithmetic. A model is a guesser, not a filing cabinet. It gives you an answer for inputs it was never shown, because it isn’t recalling; it’s calculating.

That’s a strength: it generalises, filling in the gaps between the examples. It’s also the seed of every strange thing AI does later — because a machine that always computes an answer will always give you an answer, whether or not it should.

A worked example: same skeleton, two minds

Take five flats that actually sold:

Size	Sold for
50 m²	£170,000
60 m²	£210,000
70 m²	£230,000
80 m²	£250,000
90 m²	£300,000

Now try two knob settings on price = size × A + B.

Setting one: A = 3,000, B = 20,000. For the 50 m² flat it guesses 50 × 3,000 + 20,000 = £170,000 — bang on. For the 90 m² it guesses £290,000, close to the real £300,000. Not perfect, but the line runs right through the middle of the real prices.

Setting two: A = 1,000, B = 20,000. Now the 50 m² flat is guessed at £70,000 and the 90 m² at £110,000 — far too low across the board. Same formula. Same flats. A completely different, and much worse, set of guesses. The only thing that changed was two numbers.

Which brings the real question into focus

So a model is not a mind and not a memory. It’s a shape with knobs, turning an input into a guess. You can feel, from the worked example, that some knob settings are good and some are bad — setting one hugs the real prices, setting two misses badly. That comparison is doing quiet, important work: it means “good” and “bad” knobs can be measured, not just eyeballed.

And that is the hinge the whole field turns on. Nobody sets a billion knobs by hand. The machine has to find good settings on its own — which first requires a way to score how wrong any setting is. That score is the next item, and once you have it, the entire mystery of “how does it learn” becomes a single, almost mechanical idea. You are two steps from the centre of the machine.