We've had those since the days of FORTH though.
   Self-replication, self-healing, error-checking is easy enough to
   build in. We already have programs that run on automatic without
   human intervention.

   Writing a program that can look at itself and change its own
   code is old-hat in some programming languages; it 'seems' brand
   new in newer languages like C++ and Python, but programmers have
   been doing this stuff for years.

   Erlang which keeps the telephone switchboards operational and
   underlies a good portion of the current Internet is AMAZING at
   parallel, non-interfering message passing, with billions of
   simultaneous threads that are ... suicidal as it were, without
   effecting the rest of the systems already are purring along,
   doing their thing.

   It depends on the programming language used and the
   specifications of the designers.

   Neural networking always had promise and it's a rather mature
   field finally, after years of stagnation during the AI Winter of
   the 1990s, which was rather dreadful because I was a child of
   the 80s that saw the promise of 5th Generation computers fail
   miserably and back to 4th generation we went, and we're STILL
   stuck on 4th generation computers...

   so, I suppose I've seen many disappointments in the promises of
   AI through the years, and I cheer when we make some progress,
   but then we end up stepping backwards 10 paces after making one.