tetraspace 💎⏹️🇺🇳 ✓

[image] The legal, institutional and even technical ability to halt AI development if it is found to pose a danger doesn't come for free; for there to be an off-switch that works, it must be designed and implemented by researchers and legislators.

[image] It's hard to determine the outcomes of a plan just by looking at it - otherwise, everything would be done perfectly on the first try. Many times so for plans created by smarter entities, whose plans can make use of facts about the world it understands better than you.

[image] The Superintelligence's Playbook: What's the best way to achieve some aim, if that aim doesn't involve specifically leaving humans alive? Step 1: Take humans out of the picture, in case they try to stop you - either directly by shutting down your infrastructure when they realise

[image] Many goals are compatible with the behavior of looking good to the rater, because "look good to the rater during testing, and then act differently during deployment" is a strategy that serves many goals well. One must look inside the shoggoth to check whether it's the right goal.