Foundations & Models
Neural Network
Many simple computing nodes are connected in layers; each connection has a weight that is adjusted during training. The model shares little with the biological brain beyond the metaphor. What matters is something else: networks learn relationships that nobody has explicitly written down — and for exactly that reason nobody can say completely what they have learned.
An example
A single building block in such a network does something astonishingly simple: it takes in numbers, multiplies each by its own factor, adds everything up and then decides whether the result is large enough to pass on. That is the whole calculation. What a modern language model can do arises from chaining this calculation together several hundred billion times with different factors. There is no place where anyone wrote meaning into it.
A closer look
A network consists of nodes that form weighted sums of their inputs and pass the result through a nonlinear function. During training, the weights are adjusted backward from the error at the output. What this has in common with the brain is the vocabulary, and little else.
Why it matters
The analogy to the brain is the source of most misunderstandings in discussions about AI. It dates from the 1940s, was a crude simplification even then, and today carries almost nothing. Anyone who takes it seriously expects understanding, intent or consciousness and will be disappointed or alarmed in the wrong place. A more useful description is the sober one: a very large, differentiable function whose factors have been tuned to fit examples. That can be planned around sensibly.
Commonly misunderstood
The borrowing from biology is a metaphor from the 1950s. Anyone who draws conclusions from it about consciousness or intentions is overstretching it.