I think they meant fields of research. If you do anything in NLP, CV, inverse-problem solving or simulations, things have changed drastically.
Some directly, because LLMs and highly capable general purpose classifiers that might be enough for your use case are just out there, and some because of downstream effects, like GPU-compute being far more common, hardware optimized for tasks like matrix multiplication and mature well-maintained libraries with automatic differentiation capabilities. Plus the emergence of things that mix both classical ML and transformers, like training networks to approximate intermolecular potentials faster than the ab-initio calculation, allowing for accelerating molecular dynamics simulations.
Some directly, because LLMs and highly capable general purpose classifiers that might be enough for your use case are just out there, and some because of downstream effects, like GPU-compute being far more common, hardware optimized for tasks like matrix multiplication and mature well-maintained libraries with automatic differentiation capabilities. Plus the emergence of things that mix both classical ML and transformers, like training networks to approximate intermolecular potentials faster than the ab-initio calculation, allowing for accelerating molecular dynamics simulations.