Rob Miles Explaining "Neuralese"

In case you don’t know, Rob Miles is an AI safety researcher. In here, he discusses how LLMs can compress their chain of thought to optimize for minimal token usage at the cost of generating reasoning steps that can’t easily be understood by humans. You can imagine the safety issues this would generate as you can’t see what the agent is up to. He also discusses briefly, at the beginning, about a point that is often parroted here on Lemmy about anthropomorphism, clarifying that researchers are not acknowledging consciousness of a machine when they say that it “thinks” a certain way.