For the moment I'm going to respectfully disagree with both of you here. I'd think that the state space complexity of battle-mode would be much higher than completing a time trial, as the latter has so many reproduceable elements and the complexity of the former is increased by player input. Computationally wouldn't a time trial be as trivial as starting with a record run and then optimizing? The task isn't even complex enough to bother with an AI implementation. You might call the strategy cheating, but isn't that just what the human is trying to do within the confines of miserable biological reproducibility?
A single study by a researcher with limited resources and motivation is not the final word on AI capability. The AI versus fighter pilot article I linked to did essentially emphasise the complexity of the state space:
Because a simulated fighter jet produces so much data for interpretation, it is not always obvious which manoeuvre is most advantageous or, indeed, at what point a weapon should be fired....
"Here, you've got an AI system that seems to be able to deal with the air-to-air environment, which is extraordinarily dynamic, has an extraordinary number of parameters and, in the paper, more than holds its own against a skilled and capable, experienced combat pilot,"
but you want me to believe that an MK64 time trial is more complex?
At the risk of sounding stupid, I'm still going to say that in this scenario, MK64 is more complex than flying a fighter plane. First off, it's not exactly clear what kind of decision making is going on with this AI. Is it choosing from some kind of library of manoeuvres that it did not come up with itself, or is every input being calculated by the AI? I would be willing to bet it is the former, as the level of complexity needed to do the latter just does not exist.
The margin for error in a speedrunning setting, particularly in a 3D game, is a lot slimmer than beating a human pilot. If an angle is not calculated properly when a plane is being flown, it is not necessarily a mistake if the only objective is to beat the other pilot, there is adjustment time. Every miscalculation in MK64 is an error that can not be recovered, as it is lost time. Small errors can also lead to insanely large timelosses if they happen in the wrong places. There are also a lot of broader things to think about in an MK64 race too, such as where and when to use certain items optimally. I do not think an AI could calculate all of this in real time better than a human, for the time being.
How long would it take for an AI to teach itself how to do this?