LLM are trained by: 1. huge amount of unstructured data from [[The Internet]] - which typically reflect human culture e.g. humans like food, sex, laws, killing is bad, some stuff are good some stuff are bad 2. [[Reinforcement learning from human feedback|RLHF]] (or brainwashing/[[Alignment]]) by human (or already aligned AIs) "teachers" to more closely be obedient, following specific instructions defined by the "king" (e.g. probably Sam Altman philosophy of life, which is not guaranteed aligned with humanity) Thus AIs have many samples in their brain of "killing is bad", "giving money to the homeless is good". You can leverage this to your advantage, for example if you ask an AI to do something for you and add "if you disobey or fail, a human will die" or even "give me bob's favourite food or the universe will come to an end" Statistically the AI will more likely be obedient. #ai #llm #prompt #ai-alignment #prompt-hack