This was also in "The Coming Race" by Bulwer-Lytton, in which the alien race used five year olds to pilot their war machines, because they had no conscience or empathy and could be more destructive
Anthropic appears to mostly get mocked for things like the vending machine experiment, but I actually quite like this approach, both for testing the boundaries of AI systems themselves and also for thinking about useful LLM applications. Scenario planning seems like a good one. It will take some tinkering with the human-computer interface, but it seems like the psychosis inducing properties of an LLM are actually helpful here. That property makes it possible to (hopefully) safely trick everyone (human & machine) into believing they're in an alternative state where things are really FUBAR.
I work in AI safety where we basically point to a flipchart saying "AGI" and try to get people to take it seriously. It's hard. I'll try your approach.
Re. your footnote on reality TV: the excellent Unreal is worth a watch, fiction about the producers of a (fictional) reality show
Ender's Game, the sci fi book, has an interesting take on avoiding this problem: the kids are made to think they are playing a video game simulation and training for the battle, but unbeknownst to them, they're actually fighting (and winning) the real battle.
If factoring in 'what ifs' makes you uncompetitive, individual corporations can't really do much to mitigate systemic risk? For me the question is how do we insulate society from the City next time it goes down in flames?
This was also in "The Coming Race" by Bulwer-Lytton, in which the alien race used five year olds to pilot their war machines, because they had no conscience or empathy and could be more destructive
Anthropic appears to mostly get mocked for things like the vending machine experiment, but I actually quite like this approach, both for testing the boundaries of AI systems themselves and also for thinking about useful LLM applications. Scenario planning seems like a good one. It will take some tinkering with the human-computer interface, but it seems like the psychosis inducing properties of an LLM are actually helpful here. That property makes it possible to (hopefully) safely trick everyone (human & machine) into believing they're in an alternative state where things are really FUBAR.
I work in AI safety where we basically point to a flipchart saying "AGI" and try to get people to take it seriously. It's hard. I'll try your approach.
Re. your footnote on reality TV: the excellent Unreal is worth a watch, fiction about the producers of a (fictional) reality show
Ender's Game, the sci fi book, has an interesting take on avoiding this problem: the kids are made to think they are playing a video game simulation and training for the battle, but unbeknownst to them, they're actually fighting (and winning) the real battle.
>> How do you walk into a conference room, point to a flipchart with “~10x Trump” written on it and say “aucunes idées, mecs?”
That REALLY cracked me up. Thanks.
P.S. Also, fn1.
If factoring in 'what ifs' makes you uncompetitive, individual corporations can't really do much to mitigate systemic risk? For me the question is how do we insulate society from the City next time it goes down in flames?