Recursive neural networks are not necessarily halting when executed in arbitrary precision arithmetic.
Maybe that’s partly due to the expectations of the medium, but I feel like it could be much more concise to, instead of saying, “here is a non-rigorous version of an argument which was previously presented at [other location] (note that while we believe the argument as we are presenting it is essentially the same as it was presented, we are making this different choice of terminology). Now, here is a reason why that argument is insufficiently rigorous, and a way in which that lack of rigor could be used to prove too much. Now, here is a more rigorous valid version of the argument (with various additional side tangents)”, one could instead say “Let objects X,Y,Z satisfy properties P,Q,R. Then, see this argument. Now, to see what this is applicable to the situations that it is meant to represent, [blah].”
No they aren't. Think about LSTMs for example.
How do you the a value out of a person?
One way to interpret this is that all existing AI systems are obviously halting computations simply because they are acyclic dataflow graphs of floating point operations and this fact is both easy to state and to prove in any formal system that can express the logic of contemporary AI models. So no matter how much Ray Kurzweil might hope, we are still very far from the singularity.
1: https://link.springer.com/article/10.1007/s11023-014-9349-3