claude explains why describing the technology as moving fast is ridiculous: the core is just the associative memory / content-addressable lookup that hopfield described in 1982, and before that the basic dot-product similarity measure goes back to the 60s information retrieval literature. the softmax normalization is boltzmann weighting from statistical mechanics, 1868. backpropagation is from the 70s. the feedforward network layers are perceptrons from 1958. the 2017 contribution was removing the recurrence (no more RNNs/LSTMs processing sequentially) and showing that pure attention over the whole sequence, parallelized on GPUs, scaled better. that's an engineering insight, not a mathematical one. "what if we just computed all the pairwise similarities at once instead of sequentially" - which is only an insight if you have the hardware to brute-force it.