

It’s wildly difficult to control the output of the black box and that’s hardly llms showing signs of self-preservation. These cries are from people in the industry trying to pretend the models are something that they are not, and cannot ever be. I do agree with the sentiment that we should be prepared to pull the plug on them though, for other reasons.







Just double checked, and no they are very much talking about LLM’s. Specifically they were testing gpt-4o, gemini-1.5, llama-3.1, sonnet-3.5, and opus-3 o1. https://arxiv.org/pdf/2412.04984 And the concerns raised in that paper are legit, but not indicative of consciousness or intent.