deleted by creator
I think this is the right path towards making AI that has actual understanding of the world and that we could interact with meaningfully. Training AIs in virtual environments like this allows the agents to build a world model of their environment, and then it would be possible to teach the agents language by ascribing symbols to different aspects of the environment. Having a shared world model and language to talk about it would mean that we could actually talk to the AI and teach it about the world via language the way we teach each other.
Obviously these kinds of experiments are still far off from that, but fidelity of these simulations will improve drastically in the coming years. Once an AI is trained it can be copied many times and further trained to specialize in different fields. This will also raise many potential ethical concerns as well regarding things like free will and self determination of AIs.
A YouTube link was detected in your post. Here are links to the same video on Invidious, which is a YouTube frontend that protects your privacy:
