I would hope that they freeze their ai to work on 1 game patch and 1 engine version.
I remember from open ai's strike at dota2 that the software engineers themselves didn't know much of anything about advanced mechanics in that game, but with a self-learning neural net and millions of tests the model learned to do advanced things on its own, like animation cancel and creep block. But the ai never seemed to get warding or dust/smoke right, and to make the ai kill Roshan they had to make significant game mods (1hp Roshan) for the ai to even start"looking" at agis of immortality. The ai team using zero-k will run into problems which can easily be solved by limiting scope, like a unit/factory ban list, map list, and most interestingly teraform.
And, of course, the goal of this is to test general ai, how were can build it, how long it takes, what are it's limits, and probably won't be of much interest to our community :( even when we really want it to be.