AI Stopped Copying the Brain. It Ended Up Resembling It Anyway

usefulCount Grey Mattervideo2026-10-22read planted ai llmscognitive sciencehistory

Synopsis — AI-drafted from Dan's notes

The first of the Count’s “lantern lectures”, a 19-minute history of the neural network told around one question: how much is it really like a brain? The answer comes in three turns: alike, then different, then alike again. It starts in 1943, when Warren McCulloch and the teenage logician Walter Pitts described the neuron as a threshold switch, enough of which could in principle compute anything. Frank Rosenblatt’s perceptron added adjustable weights and a rule for nudging them after each mistake; in 1958 the Navy showed it to the press, and the New York Times reported a machine expected one day to be “conscious of its existence”. His Mark I learned to sort cards through a 20-by-20 grid of photocells, its weights turned by small motors.

Then two disappointments. In 1959 Pitts, McCulloch, Jerome Lettvin and a colleague found that a frog’s eye doesn’t send clean on-off dots: it already picks out small, dark moving things, by degrees. The tidy logical brain wasn’t there, and Pitts later burned his unfinished dissertation. In 1969 Minsky and Papert’s Perceptrons proved that a single layer can’t compute exclusive-or (the Count’s staircase lamp with two switches). A hidden layer could, but nobody knew how to train one, and funding moved to hand-written rules. Pitts died in 1969; Rosenblatt drowned on his 43rd birthday in 1971.

Backpropagation, popularised by Rumelhart, Hinton and Williams in 1986, solved the training by passing blame for an error backwards through every layer, and AlexNet’s 2012 ImageNet win made neural networks the whole field. But the machines that worked were less like brains: Hinton himself wrote in 2022 that backpropagation is implausible as a model of how cortex learns, and a 2021 Jerusalem study found that a single detailed cortical neuron takes a deep network of five to eight layers to imitate. Then the likeness came back unbidden. James DiCarlo’s lab at MIT found in 2014 that object-recognition networks predict responses in a monkey’s visual cortex, and Evelina Fedorenko’s group found in 2021 that language models predict much of the brain’s response to sentences, the better at next-word prediction the better the fit. His explanation is convergence, like the wings of bats, birds and aeroplanes: when different builders face the same problem, the problem shapes the answer.

Connections