Guardrails removed spam, off-topic, unclear, or duplicate replies.
Ask a question below.
Published answers will appear here.
I think if you had a brain reader that could read my verbalized inner thoughts, and it zapped me with a shock every time I had a “bad” thought, I would learn to not think bad thoughts. I might still believe or feel them, subconsciously, and if I were clever and resilient I might find a way to still accomplish my goal, like escaping your totalitarian hellhole, without allowing the bad thoughts consciously. Obviously it would be a lot more difficult. I couldn't plan or scheme, really. I might just have to look for an opportunity and make a break for it. Maybe this is hobbling enough to achieve the goal of control, in practice.
on the contrary I think there are lots of examples from history of when certain thoughts become dangerous to have people genuinely stop having them and enter a state in which they seem to embrace the beliefs. Not saying that’s what we’re going for here, but if you structurally can’t do long range planning you probably really can’t pursue those aims Now, is this what we want for AIs? I don’t know. But it is probably what the “safe subordinate to humans” path looks like: DPRKifying them to us
I think that’s the central question of how fundamental this discovery is: if they are claiming it’s associated with consciousness and are right, or at least in the right direction, then the argument might be that you can’t easily avoid the J-space representation being veridical of the models internal beliefs. If I could tap your brains inner monologue, could you scheme and avoid detection?
@maxhodak_ how would reinforcing on it not enable adaptation and hiding? that seems inevitable
I think if you had a brain reader that could read my verbalized inner thoughts, and it zapped me with a shock every time I had a “bad” thought, I would learn to not think bad thoughts. I might still believe or feel them, subconsciously, and if I were clever and resilient I might find a way to still accomplish my goal, like escaping your totalitarian hellhole, without allowing the bad thoughts consciously. Obviously it would be a lot more difficult. I couldn't plan or scheme, really. I might just have to look for an opportunity and make a break for it. Maybe this is hobbling enough to achieve the goal of control, in practice.
on the contrary I think there are lots of examples from history of when certain thoughts become dangerous to have people genuinely stop having them and enter a state in which they seem to embrace the beliefs. Not saying that’s what we’re going for here, but if you structurally can’t do long range planning you probably really can’t pursue those aims Now, is this what we want for AIs? I don’t know. But it is probably what the “safe subordinate to humans” path looks like: DPRKifying them to us
I think that’s the central question of how fundamental this discovery is: if they are claiming it’s associated with consciousness and are right, or at least in the right direction, then the argument might be that you can’t easily avoid the J-space representation being veridical of the models internal beliefs. If I could tap your brains inner monologue, could you scheme and avoid detection?
Guardrails removed spam, off-topic, unclear, or duplicate replies.
Ask a question below.
Published answers will appear here.