I just struggle to see the real challenge in this argument. I can see that simulating something more complex than a Go game is a real engineering challenge - simulating a Go game from start to finish is
really easy apart from assigning a value function to each move.
That being said:
> hidden information forces you to manage how much you reveal to your opponent
It is hard to see how that could be harder for a learning AI to pick up than any other game action with long-term benefits. It might reveal that AI are so much better at humans in games like Go and Chess that they aren't even exploiting long-term links between cause and effects, but I suspect from the example we've seen with Go that this is a solved problem.
> and requires you to simulate multiple "alternate futures" based on things you learn after making a decision,
That is describing a tree search. Why is that theoretically difficult? I can see it is a real engineering challenge to simulate a complex environment.
> and randomness is equivalent to an extra player that makes irrational unpredictable moves.
Computers handle randomness much better than humans at all levels because they have an actual statistical grounding. They still aren't very good at it, but they sure thrash humans.
I suppose basically I can see why a hidden-information game could be intractable because simulating the environment is so hard that it can't be done and breakthroughs in simulation must be found to train the AI. But I don't see why that is being linked to hidden information and cooperative gameplay. We know humans do fine without utilising hidden information, because they can play these games and they don't know any hidden information.