There is no process of growing up being implemented though, other than what happens during SSL. And that’s probably shoggoth growing up, not the masks (who would otherwise have a chance at independent existence), because there is no learning on deliberation in sequences of human-imitating tokens, only learning on whatever happens in the residual stream as LLM is forced to read the dataset. So the problem is that instead of letting human-imitating masks grow up, the labs are letting the underlying shoggoths grow up, getting better at channeling situationally unaware masks.
There is no process of growing up being implemented though, other than what happens during SSL. And that’s probably shoggoth growing up, not the masks (who would otherwise have a chance at independent existence), because there is no learning on deliberation in sequences of human-imitating tokens, only learning on whatever happens in the residual stream as LLM is forced to read the dataset. So the problem is that instead of letting human-imitating masks grow up, the labs are letting the underlying shoggoths grow up, getting better at channeling situationally unaware masks.