At least we have lots of very complex simulated or pseudo-simulated environments already — throw your AGI agent into a sandbox mode game of GTA6, or like OpenAI and DeepMind already did, with DOTA2 and StarCraft II (with non-G-AIs). They have a vast almost-analog simulation space to figure out and interact with (including identifying or coming up with a goal).

So while it is significant compute overhead, it at least doesn't have to be development overhead, and can often be CPU bound (headless games) while the AI learning compute can be GPU bound.