Tootfinder

Some very insightful comments in this thread i.m.o.
"It's classic Skinnerian operant conditioning with intermittent (variable rate) rewards. You want whatever it's outputting to be good (code, text, image, etc). Sometimes it isn't but sometimes it is, and you can't usually understand why. When it is good, you experience the reward. The fact that the reward is intermittent and inscrutable makes the desire to repeat the behavior extremely strong."
"Normally an intermittent reward is a sign of a skill that one can master. ... It makes sense that as a species for which tool use is so fundamental that we'd be especially prone to this. ... But we really aren't prepared for when the thing can't be mastered, where it's fundamentally unreliable."
"the slot machine cycle... if I can just figure out exactly how to word this prompt..."
"It feels so overwhelmingly good to some % of people they don't even bother to measure if their AI stuff is actually doing anything useful, because of course it must be, because the feeling is so strong."
#LLMs #addiction #Skinner #behaviourism #OperantConditioning #VariableReinforcement