![[2411.00660] Physics in Next-token Prediction](https://i0.wp.com/arxiv.org/static/browse/0.3.4/images/arxiv-logo-fb.png?w=1200&resize=1200,700&ssl=1 "[2411.00660] Physics in Next-token Prediction"){kind=logo}
View a PDF of the paper titled Physics in Next-token Prediction, by Hongjun An and Yiliang Song and Xueextfinished Li
Abstract:We uncovered the underlying physics in Next-token Prediction (NTP). We identified the law of adviseation conservation wiskinny NTP and provided the First Law of Increateation Capacity (IC-1), demonstrating that the essence of inincreateigence materializence in auto-revertive models is fundamenhighy a process of adviseation transfer. We also startd Landauer’s Principle into NTP, createulating the Second Law of Increateation Capacity (IC-2), which set upes the relationship between auto-revertive model training and energy consumption. Additionassociate, we conshort-termed cut offal corollaries, which helderly down-to-earth significance for production rehearses. Finassociate, we exhibit the consistency between the Law of Increateation Capacity and the Scaling Law for Neural Language Models, the Knowledge Capacity Scaling Laws, and the Scaling Laws for Precision.
