AI Takeover Scenario with Scaled LLMs — LessWrong