Maighstir: Hmm. That would be interesting, hook an AI/neural network up so that it takes digitised visual and aural information from the game and gives keyboard and mouse input signals to it, try to get it to figure out how to play the games in a manner similar to how we do (ie, taking cues visually and aurally and giving input via the devices we have, rather than directly editing and reading the game's memory locations).

tinyE: Droids play games and Hutts sell them.

Nightblair: Google already do that for real. With 8bit games, so far, but I've seen some video about that here: http://www.theverge.com/2016/6/9/11893002/google-ai-deepmind-atari-montezumas-revenge

Probably more interesting than my own, but who am I to compete with Google?

Wishbone: You don't need to be Google. This guy did it on his own. I'm working on a similar (albeit a lot more complex) concept myself.

Nightblair: I've checked his code. It seems that he tailored the AI for the specific game, he even loads sprites from specific place in rom's memory. That seems more like the bot (like the old Angband Borg), than real adaptable AI's like the Google is developing. That one does not know the internals of the game, nor have any previous information about it.

Wishbone: It's basically a matter of where you put the code that analyzes the game. In Google's case, they're using ALE (Arcade Learning Environment), which does the analysis part and breaks the analysis down into "simple" inputs for the AI. Also, I don't know what you mean by adaptable AI. Do you think a single AI instance is playing many different games? I don't think that is the case.