Ah, but I don't know the rules for Go; maybe with wikipedia I could work out what they were though.

I'd be tempted to assume that an IM would by its very nature be an AI - but I'd probably be wrong.

And yes, the search is hard... but what you can do is make a giant neural net and train it with all the available data (rather than whatever subset you have time for), and you could train as many neural nets as you could invent that might work (obviously you generate these somehow, not just think about them yourself). So I think you'd probably have something much better than what we currently have at the end of it.

An Infinity Machine contains no AI, unless you write it. It isn't qualitatively different from a normal finite computer, except that it has unlimited storage space and you get to ask questions of the form "will this program ever produce an answer, if I run it forever?". It does ordinary mechanistic computation, just very very fast.

Large-scale neural nets are an interesting idea, though, and certainly they neatly dodge my basic methodological objection. I think I wouldn't be confident enough of getting decent results to spend my limited time on that particular approach, but good luck with it if you ever find yourself in this scenario :-)