Actually a neural network is just that: data compressed with losses. A transformer makes multiples queries to a large loss-y and stochastically compressed datab

Actually a neural network is just that: data compressed with losses. A transformer makes multiples queries to a large loss-y and stochastically compressed database to determine the next token to generate. The PAQ archiver is famous for being just that: a neural network to predict the next symbol.