28:["$","ul",null,{"children":["\n",["$","li",null,{"children":"looks at those 10 numbers"}],"\n",["$","li",null,{"children":"scores every possible next note"}],"\n",["$","li",null,{"children":"outputs a list of 7 scores (for 7 notes: a b c d e f g) - the 'logits'"}],"\n"]}] 29:["$","p",null,{"children":"we turn these into probabilities (via a function, softmax):\na: 0.45\nb: 0.22\nc: 0.01\n…"}] 2a:["$","p",null,{"children":["then, if the real next note was a, but the network leaned toward b, we record the difference between the ",["$","em",null,{"children":"true"}]," next note and the ",["$","em",null,{"children":"predicted"}]," next note: the loss."]}] 2b:["$","ol",null,{"start":"5","children":["\n",["$","li",null,{"children":"backprop"}],"\n"]}] 33:T42a,parameter <- parameter - learning_rate * gradient2c:["$","p",null,{"children":["each parameter (weights, biases, embeddings) gets a gradient (a slope of how the loss changes if that parameter changes). then we nudge it by ",["$","code",null,{"dangerouslySetInnerHTML":{"__html":"$33"}}]]}] 2d:["$","p",null,{"children":"this is the learning. the error flows backward, and everyone in the room gets blamed in proportion to how badly they contributed to the fucked prediction. the final guesser, the hidden guessers, and the embeddings (the coordinates) all adjust."}] 2e:["$","p",null,{"children":["do this 200,000 times, and the room becomes ",["$","em",null,{"children":"really"}]," good at guessing the next note."]}] 2f:["$","ol",null,{"start":"6","children":["\n",["$","li",null,{"children":"generating songs"}],"\n"]}] 30:["$","p",null,{"children":"to generate a song, you just let the room guess over and over:"}] 31:["$","ul",null,{"children":["\n",["$","li",null,{"children":"start with a rest ('.')"}],"\n",["$","li",null,{"children":"guess the next note -> 'e'"}],"\n",["$","li",null,{"children":"feed 'e' back in"}],"\n",["$","li",null,{"children":"guess again -> 'f'"}],"\n",["$","li",null,{"children":"guess again -> 'g'"}],"\n",["$","li",null,{"children":"guess again -> 'a'"}],"\n",["$","li",null,{"children":"it predicts a rest ('.') again"}],"\n"]}] 32:["$","p",null,{"children":"in one sentence: three notes go in, get turned into coordinates, mixed around by a bunch of weights and biases, scored against possible next notes, and the whole system nudges its numbers slightly depending on how wrong it was; then repeat."}]