I really don't think there is a power creep. In 2016, knowing just script-kiddy level ML knowledge like scikit-learn would not qualify you for full-fledged ML roles at FAANGs anyways. In general, FAANGs always looked for 2 qualities in ML roles - rigorous and principled mathematical background, and solid SWE skills.
Architectures are becoming simpler - Transformers are being used instead of LSTMs for sequence modeling. Code is becoming easier to run and streamlined. CUDA is much easier to install these days. Jax is basically streamlined tensorflow 1.
I really don't think there is a power creep. In 2016, knowing just script-kiddy level ML knowledge like scikit-learn would not qualify you for full-fledged ML roles at FAANGs anyways. In general, FAANGs always looked for 2 qualities in ML roles - rigorous and principled mathematical background, and solid SWE skills.
Architectures are becoming simpler - Transformers are being used instead of LSTMs for sequence modeling. Code is becoming easier to run and streamlined. CUDA is much easier to install these days. Jax is basically streamlined tensorflow 1.