Jason,
Yes, chatbots are actually away from true AI. They are basically consisted of "dirty tricks" to make you feel like you are actually talking to someone smart, while they are just being a kind of parrots. But I'm surprised how successful they are in this deception. Anyway, it could be a way to go, if you pair a chatbot engine with a logical reasoning engine. I suspect this is what Sophia does, but don't take me by the word.
But there is something else going on in another corner of my mind...
There is an idea I have for a while here, to exploit neural networks in a certain way, and these days, GPT2 just concreted my thoughts on it. Trained NN is actually a way of packaging which function results correspond to which function parameters. I consider these functions as black boxes (Turing complete, I hope) which magically do the right stuff we expect from them when posing some parameters. Ok, now, consider an imaginary human-machine conversation flow:
user says: ...
computer says: ...
user says: ...
computer says: ...
...
What could be interesting here is how the user behaves. What she/he says in cycles is actually always a result of the same function parameterized with previous relative inputs/outputs. The problem we are trying to solve here is what should the machine say (or do in some further development). And the answer is fairly simple: the machine could use the same function that the human use to respond to the machine. And that function could be learned by artificial neural network by observing the user's input relative to previous computer's output. Simple, isn't it?
I'd have to investigate neural networks more thoroughly to actually test this concept, but GPT2 keeps convincing me that the whole thing could work very well. The thing would represent a mirror of all responds it collected from its environment. It could even be placed online to gather communications with random users, learning and reflecting their responses in future dialogs. Initially, just to boost up a start of learning, it could be trained on Reddit, but later, it could be switched to alive user interaction. And it would reflect a collective hive of humanity thoughts.
All of this is just a conceptual thought you could judge after learning some of advanced properties of neural networks. But if you decide to test it before me, let me know how it went, I'd like to know if it works.
All well,
Ivan V.