The Quantization Model of Neural Scaling — LessWrong