Originally Posted by
CurtainTwitcher
What exactly are they good at? The leading type of AI, neural nets are exceptional of taking a closed problem and solving it eg
AlphaGo and
AlphaGo Zero
(AGZ). However, these problems were not solved on the fly, they took extensive computational resources to generate it's own training data in the case of AGZ it was a 40 day process simulating playing itself to generate the dataset.
However, neural nets need extensive
clean training data, either from the real world or by simulation. There was no extensive corpus of air returns with double engine damage for A320's unlike the millions of recorded games of Go for Alpha Go training. Even Sully's effort represents a single instance that is effectively useless for future algorithmic training. Billions of simulations would be necessary just to replicate & solve this exact scenario on this day. For self driving cars they actually model intersections and do billions of simulations to generate the training data to enable self driving:
Inside Waymo's Secret World for Training Self-Driving Cars.
There are many things that computers can do exceptionally better than humans, but solving novelty is not one of them using the current leading AI technology.
You're conflating AI with machine learning and with automation.
Furthermore, you're wrong that "AI" requires millions of (recorded games) to learn from. An extensive corpus is not necessarily required. Read about
AlphaGo Zero, which used
zero recorded games (hence the name). In just
three days AlphaGo Zero played 5 million games against itself -- and surpassed the original AlphaGo (the version that beat Lee Sedol), winning against it 100 games to 0. The successor
AlphaZero is even more impressive.
Now, AlphaGo Zero's technique is probably not suited to autonomous vehicles, but we don't even need machine learning to safely return Sully's plane back to La Guardia. For a computer, this is actually a much simpler problem, where a
constraint satisfaction system combining a finite state machine with
path planning would be sufficient.