Are you setting the bar too high? That's an honest question.
For example:
OpenCV - Great it can recognize a face, however training models were largely done on white people, so they have white-bias for detecting faces.
Humans are notoriously bad at recognizing people from other races. "They all look the same" has been a punchline for a long time. Failing the same way humans do, and for the same reason, seems like a vote in favor of the deep learning solutions.
They are all universally designed for commercial applications (eg phone IVR's) and thus there is no standardization and you end up retraining your data, wasting months of processing time when a better NN vocoder or synth comes out.
Should we be looking for standardization at this point? I could see arguments on either side. We want to try lots of things vs. we need to be able to compare the different things we're doing.
Also they use very low quality inputs, which results in some really low quality voice synths that "sound a little better than telephone conversations."
So we need better inputs. That means the pretty-impressive results we're already getting will only get better.
The AI can eventually figure out how to solve these games better than a human because it's FASTER at making decisions, not because it's better.
Chess masters study previous games and situations so that when they see an arrangement on the board it looks like a solution they've already studied. How is that different from the AI doing it in real time?
Chatbots - Can not solve customer's issues, they are primarily designed to play queue-bounce. Chatbots can be designed to help customers pick the right solution, but they are largely (and websites of the same companies) are designed to bury human contact by trying to get the customer to help themselves, but really the result is more frustration.
Many CSRs work from scripts designed to do the exact same thing. Is there a functional difference between a chatbot that isn't able to improvise and a human who isn't allowed to?
Deep Learning however has no plasticity once it's put into production. Quite literately, when it's not in training mode, it can't learn.
This one I completely agree with you. As long as the hardware required for training is significantly greater than the hardware required to run the agent, it's going to run up against edge cases that it can never handle.