I'm no expert at programming or artificial intelligence but is SAO's description of bottom-up and top-down AI a real explanation of how we use AI in games? In reality do we associate these terms in games? How is it possible for us to create a bottom-up AI?
AHAHAHAHAHAHA no, no-ish, that existed since about 1970's-ish
First off, AI is overused as a general term to describe consciousness in machines in such a way that any actual progress in AI is never considered to be "true" AI. Outside an academic context, it has come to mean machine souls. What is a soul/does it exist doesn't matter, it's just not whatever is actually researched in computer science.
Okay, rant over. AI is an offshoot of Machine Learning, which is (in an oversimplified definition) the methods used to detect certain qualities in data via probabalistic methods.
Anyways, we already have both "top down" and "bottom up" AI, although his terms are hilariously simple.
Also, I'm not going to define some terms. They should be easy to look up.
"Top Down" AKA classifiers: These focus on using an existing set of training data and labels (the things about the data you care about) to create a model that can classify future data (test data).
-Examples: Decision trees, perceptron, linear/logarithmic regression
"Bottom Up" AKA deep learning: These focus on creating models that can learn from a training set, and then able to predict labels for test sets.
-Examples: Neural Networks (Recurrent, Convolutional, standard), Markov Chains
The main difference is that while classifiers make predictions from the direct qualities of data, deep learners use data to train a model to recognize patterns in the data probabilistically.
Games use classifiers (or algorithmic approximations) over deep learning for many reasons:
1. Efficiency: Deep learning requires a lot of trial and error, starting from a neutral state. Not the best for things like pathfinding, which already has efficient algorithms to determine a path (Look up A-Star). Not even a classifier, really. Plus, most behaviors have efficient algorithms to produce "smart" AI. Look up how game developers trick players using psychology to make mechanics seem more advanced.
2. Quality: You don't really need deep learning in a game aside from gimmicks and the like, and even then they are usually used for mechanics designed to change over time. And even then, because of the probabilistic nature of deep learning's results, game designers won't use it unless deep learning is the specific gimmick of the game.
Occasional ML gimmicks exist though. Just look at the amiibo.
There are probably more things that can go here, but as you can probably tell, machine learning in games is still an early thing, more used for systems in games (ingame real money shops, ads etc) than the actual gameplay loops.
Hope that answers q1
There is anecdotal reference to those terms, but game programming is heavily centered in computer science, which has other (and more informative) terms for what kawahara means.
Also, I think kawahara got the idea from some psychology theories on consciousness, but overapplied it to computer science.
I think I've answered q3, but in case you are still looking for more, look to the terms that I put in the first part...
Or pick up your phone and say Ok Google/Hey Siri/Alexa/whatever else activates a smart assistant.
But yes, we have made bottom up AI already. It's just that the reality will always be underwhelming from a layman's perspective (but fascinating and a bit scary from an informed perspective/expert perspective).