If RL is enough, then there is no physically realizable way to actually train a RL based GAI in the near future. RL based learning requires evaluating the outco

If RL is enough, then there is no physically realizable way to actually train a RL based GAI in the near future. RL based learning requires evaluating the outcome of millions or billions of scenarios over time in order to optimize the network. Given that requirement you'd have to either find a way to accurately model the world and all of those interactions in silicon, or you'd have to build millions of robots that can report back the results of billions of interactions each day. It's not impossible to do that, and maybe it would even be likely that we would eventually accomplish that but the cost would make it prohibitive for anyone but a nation to even attempt today. It's almost certainly outside the realm of what is possible in the near future. Maybe when robotics has progressed enough that robots are capable of interacting with the world with basic AI will we see the rise of something like a GAI.