one of the reasons I'm so critical of the messianism of both the Altman e/acc types and the Yudkowsky doomer types is that we 1) don't actually have evidence that apocalyptic-scale AI is anywhere close and 2) don't need to reach that point at all for scaled harm to occur through autonomous agents
Reflecting on how Roko's Basilisk is a red herring. You don't need superintelligence / AGI per se for agents to start coordinating and collaborating in a way that benefits a larger goal rather than individual needs, and those ends can be harmful to humanity without intention.