@Malcolm_Ocean i'd be more worried if we got a sudden datacenter / hardware / energy overhang while simultaneously seeing agents tending towards sociopathic / antisocial totalizing behavior as scaled up what i think we will see instead is more "laziness" & satisficing & dynamic smol hive models that can run local+remote hybrid & more complex architectures leading to less internal alignment (like among experts in MoE, etc) & more stochastic behavior & finding that interpretability has some fundamental limitations