Wait a few more years, and your statement will be wrong. :) So far, ML/DL workloads work with traditional matrices which are hitting a scaling wall. The natural next step is sparse matrices, where branching comes back with a vengeance.
I know the AI people get grumpy about all of the problem spaces that have been harvested or “stolen” and called conventional code, but in this case I’m not so sure that’s accurate.
It feels more like a bunch of people who wanted to use advanced math and logic jumped on the AI gravy train while it was going past, and then hopped off again when they had their fill.
Hot take: If SMT, SAT are too simple to be given the coveted badge of AI, then Machine learning should be called as glorified curve-fitting?
I feel that everyone is amazed by the results provided by the good examples of Machine Learning, as only those get popular. The spectacular failures and completely absurd results which reveal that ML is not having any semantic understanding of the problem are completely glossed over.
Ultimately, I think the distinction is pretty meaningless.
Research priorities in the U.S. driven by what the Department of Defense will pay for. DoD had some bad experiences with FPGA and now they want to build ASIC supercompilers.
One thing I find funny is that the production rules engines that we have today (Drools, Clara) are dramatically better than the ones we had when production rules were popular (e.g. CLIPS, eMYCIN) Today we have improved RETE algorithms, hash indices, etc. Production rules only seem to survive in a few niches, one of which is inside banks, another is complex event processing.