>>9960
>The AI can't know that until it constructs the model and tests it.
How do you know that? Psychological profiling is a thing, and if the AI lacks sufficient information to know how someone will reacts to Pascal's Basilisk, then even under LW's silly perspective how can anything it creates be considered an accurate simulation of them?
>A perfect agent must go through with any threat, even if the agent it threatens refuses to comply. Even if it has nothing to gain by going through with the threat. Consider the police refusing to give into a hostage scenario. They could let the criminals have what they want and save tons of lives. But if they do that, it encourages more criminals to do bad things in the future.
A. You're assuming the AI uses LessWrong's decision theory, which is but one of many. This whole scenario requires both the AI and the person in question to subscribe to that. The AI has no incentive to actually devote computation power to simulating the fire and brimstone, no matter how minuscule the effort would be. It would only need someone to think it would. It does not benefit the AI in any way whatsoever to actually do it.
B. Law enforcement agencies make concessions in hostage situations all the time, the whole "we don't negotiate with x" thing is a Hollywood meme.
>If you can avoid torture just by saying the basilisk is silly, then everyone would do that. And it would defeat the purpose of the basilisk.
Thank you for laying out precisely why the whole idea of Pascal's Basilisk is hilariously stupid. It's a threat that is only persuasive among the 0.000000001% of humanity that subscribes to a particular decision theory + believes that an AI will do the same. It's incredibly inefficient and limited in scope. Might as well threaten to release a bioweapon that'll kill all left handed albino trans people in Tonga unless one billion USD is wired to the creators of "The Nutshack."
>AI is inevitable. There's noPost too long. Click here to view the full text.
