“The transformer architecture, introduced around 2017-2018, enabled the pre-training era of foundation models by scaling effectively with data and compute.”