Smoothness as Emulation Pressure

An LLM that was just a lookup table could learn the training set perfectly well, but would suck on the validation set. LLMs today don't 'try' to get better at the validation set—they are pushed towards the functions that represent (or simulate!) the validation set because it is more difficult for them to memorize each point than it is to simulate the underlying process that creates text. In this way, the smoothness implicit in the model + optimizer's structure is the thing causing emulation of the underlying variables we care about. The model is like an amplifier for different latent variables—and different models probably amplify different latent variables to greater or lesser extents. It's interesting to ask what small changes in LLMs causes them to amplify more or less. Consider—would an LLM with a built-in copying mechanism amplify copying in language more or less? Would it be more calibrated to reality?