And AI requires the same thing, we just call them data scientists and ML engineers. Using linear-ish algebra instead of decision trees doesn't change the the fact that you need time and capital to hire experts.

The big difference is that data scientists only work on the model architecture and data sources, whereas expert systems need people who have expertise in the subject matter itself. One of the biggest changes from 'old AI' to modern ML is that we no longer try to use human domain knowledge as much, instead getting the model itself to see the same pattern from data.

Isn't labeling still an important part of modern AI?

Yes, but there is a whole field of artificial intelligence called unsupervised learning that tries to identify labels without pre-defined labels. At the extreme end there are no externally imposed / defined labels and artificial labels are determined by empirical clusters or some orthogonal data pattern or algorithm. Unsupervised learning is much less effective and not as mature as supervised learning. In the case of LLMs the label is "next words" and it's inferred from a corpus of text.

I'd say labels (for supervised ML) are fundamentally different from rules (for expert systems), because

  - labels are easy to decide in many cases
  - rules require humans to analyze patterns in the problem space
  - labels only concern each data point individually
  - rules generalize over a class of data points

Which is very much unlike today’s training set inputs meticulously labeled by domain experts and curated by engineers?