Not to also mention that much of this would still be in the stone age if it were not for MUCH faster and accessible hardware (GPU/TPU/QPU/XPU), what I think makes up for the high salaries is that these people are quite aware of how to allocate for hardware on the ML/RL/AI training side and how to allocate time for the people responsible for tagging and optimizing models. They know how to reach the objective(s) in training/testing/optimizing already.