> I'd rather a purpose build machine learning system that is actually designed to run them

If generalist robotic models get good enough to accomplish many varied tasks effectively, training a separate comparable specialized system from scratch for every task would be highly cost-ineffective, even if, in theory, it could have slightly higher reliability.