(LW) LLMs as currently trained run ~0 risk of catastrophic instrumental convergence even if scaled 1000x more compute