A bit belated, but since I know a few DLers saw the AI Dungeon 2 posts (e.g. stuff in this thread:
https://twitter.com/JanelleCShane/status/1202968242286784512 ), I should recommend Shane's book
"You Look Like A Thing And I Love You" if you want to catch up on the latest in AI. It's a very funny, readable book that isn't impenetrable that helps separate what deep learning AIs are incredibly good at, and what they aren't.
Easy example:
https://aiweirdness.com/post/175110257767/the-visual-chatbothttp://demo.visualdialog.org/if you want to try it yourself, noting that the 2019 version is more up-to-date than the one in that post.
Anyway, I partially bring it up because famously, AIs have been thrown at old Atari games and the like and done crazy stuff with mastering them, but also gotten stuck in weird areas (like AIs who get flat walled by stage 2-2 or so of Super Mario Bros., because they get stuck on a ledge and had never ever learned to go backward.). It can even be tricky to set up a reward system properly in more complicated games than the Atari ones, where you can just say "maximize score." One of the cool ways to properly carrot & stick the AI through such games: make one AI a predictor whose job is simply to guess what will happen next in-game. If it has no idea, it guesses wildly and is probably wrong. The other AI actually controls the character, and it gets rewarded whenever the first AI is wrong. So... death is fine! But boring. Die enough and the first AI knows that Mario is going to go up then down then the level will restart and it's seen it all. The second AI is thus encouraged to explore more and more of the game to seek out unknown levels the first AI can't predict. However! Take an example of "find your way out of a maze." The above plan works pretty well to ensure the AI explores the maze properly. If you stick a TV in your maze however with YouTube cat videos or random static or whatever, and the AI will just stand there entranced by the TV that it can't predict, deciding that this is the best game ever, it never knows what is coming next.
Anyway, if you think that snippet is cool, then go read the book! eBook edition available on Google Play or the like if you want to read it on a mobile phone, as usual.
--
On the note of different kinds of AIs - reading
Saturn's Children right now and pondering that the general idea would make for a good RPG space opera. This was written before Nier: Automata, too, but general conceit is that the humans are all dead so it's just robots having a classic Heinlein-esque planetary intrigue romp. But in all seriousness, it would make a lot of sense for an RPG too in that you could explain much better why Our Heroes are fearlessly heading into extraordinarily hostile environments, why they might be able to revive, and why there might be hordes of enemies. Maybe the robots were sent as a first wave to terraform a new solar system and are waiting for the Creators to come, but until then, go build a human-like place but also live in it themselves.
Anyway, the other vaguely relevant thing about the book - a lot of Japanese stuff in particular, although it shows up elsewhere too, involves "magic girlfriends" and the like that are bonded to Our Hero for reasons. The Western equivalent might be more like magical servants, a la Harry Potter house elves. Anyway... the romance on this is always a bit problematic since one side kinda doesn't have a choice. The Heroine of Saturn's Children is basically a sexbot made to be a "escort" for humans, except the humans are all dead, so she will never find true love (a la XG Ramsus). But hey, screw that, go make your own fate! I approve, it's a rare way to make the concept work in interesting ways.