6809
-
Titles, a game, a fact, and a tune
Post five of five: four exhibit demonstrations beyond computer names. Sixteen episode titles that were never filmed, a game opponent that updates counts during play, a fact you hand it right now without training, and a tune it finishes from the bar you give it. Then what the whole thing is for.
-
The prompt, the size, the training data
Post four of five: how prompts change context, larger models change capacity and cost, and training examples and their order change the output probabilities.
-
Twenty epochs, then the assembly
Post three of five: what 1,160 nudges do to the model, checkpoint by checkpoint, and then the 6809 instructions that do one of them: two unsigned multiplies, one subtraction, and four shift pairs.
-
Where the numbers live
Post two of five: the two tables where the model's numbers live, why each word gets three, what a parameter is, what one step of training changes, and how a score becomes something you can roll a die on.
-
A 45-year-old computer learns to invent computer names
My Tandy Color Computer from 1981 starts from random numbers and, a minute later, invents computer names that never existed. Post one of five on how it does that.