> develop a robust enough model of reality if it can't learn by affecting reality This seems like comparing AlphaGo with AlphaZero, with the former learning fr

> develop a robust enough model of reality if it can't learn by affecting reality This seems like comparing AlphaGo with AlphaZero, with the former learning from professional games and the latter from self-play. AlphaZero is better but both understand the game well enough to bead pros. In particular most believed for quite a while that something like AlphaGo would not be possible along similar lines of reasoning about AGI. Or consider the thought experiment if the machine was given the ability to perform experiments and measurements in the real world of its own choosing. Would this make AGI more possible or still only similarly possible?