The a priori argument that using money is important doesn't stand up under closer >examination. If you are incapable of generalizing from in-game currency to dollars, you >won't be capable of generalizing from poker to other activities. And player behavior >does not seem to be grossly different - take for example the fact that prediction markets >work the same with real money or fake money.

It may not be much different for prediction markets, but it is VERY different for online poker. Even if you play exactly the same with or without money, your opponents will not, and therefore you will be "training" on different data than you think. This applies especially to NL games; risking 1,000 points on one bet is a lot different than risking 1,000 dollars.