“The Transformer architecture, introduced around 2017-2018, enabled models to effectively scale with data and compute, leading to the pre-training era of foundation models.”