Large Language Models
Quantization
Fewer bits per weight, almost the same model.
You can jump straight in, but this star assumes The Transformer. Not recommended as a first stop.
Four layers of depth
Each layer ends with a quiz. Finish layer 4 and you own this concept.
- L1IntuitionHow shrinking every number in a model from 16 bits to 4 makes it faster and cheaper to run — and why it usually doesn't make it noticeably dumber.6m
40 XP - L2MechanicsScales and zero-points, symmetric vs asymmetric quantization, why outlier channels break naive schemes, and what GPTQ/AWQ/GGUF actually do differently.10m
70 XP - L3CodeImplement group-wise int4 quantization and dequantization from scratch, and measure the actual reconstruction error on realistic weight data.11m
110 XP - L4FoundationsThe scale/zero-point math in full, the variance of rounding error, and why decoding is memory-bandwidth bound — the real reason quantization speeds things up.10m
180 XP
33 stars in the atlas.