Day 0 · 2026-09-30
14 million parameters, a CPU, and answers at chance.
Morning
Prime started as 14.4 million parameters running on a Mac: 256-wide, six layers, one layer of separate streams for each kind of input. Each step took 0.8 seconds. Its first curriculum had 34,184 samples: ten kinds of things from CIFAR-10, names read by the Mac’s system voices, characters rendered in twelve fonts.
Afternoon
Report card · 40 questions each · chance 10%
Essentially chance. Even on pictures it had seen, it named one in eight. The first lesson of the diary was not about the model at all: a 14-million-parameter model on a CPU was too small and too slow to tell us anything.
Note to self
Move to the eight-card server. Make the model bigger. And look harder at the picture book: twenty images per thing, read again and again, might be exactly how a toddler is taught, or exactly how a model learns to recite.