Conceptual

Neural Scaling Laws

Power-law relationships between loss, parameters, data, and compute; Chinchilla-optimal token-to-parameter ratios and why frontier labs overtrain past them.