5 ms·

Yudkowsky-derived "Safety" is probably now dead, especially as Meta completely dismantled their safety group while this shitstorm was playing out. "Safety" now

Yudkowsky-derived "Safety" is probably now dead, especially as Meta completely dismantled their safety group while this shitstorm was playing out. "Safety" now means "least risk for shareholders" ...and if you really push the question you'll get LeCun's response of "Don't worry, our AI will just kill theirs" which is hilariously dystopian but also pragmatic.

nightowl_games 3y ago
In my opinion "safety" has always been about that and it's also a projection of power. If you say "we must ensure safety from this machine" you are asserting the machine has a high level of power. People are attracted to power. Investors are attracted to power. Speaking of "safety" is implicitly speaking of power.
- esafak 3y ago
  Well duh. The question is how to ensure humanity's safety from the power of AI, whether it is autonomous or wielded by other humans.
  - nightowl_games 3y ago
    I'm specifically stating that to speak of safety is to speak of power. And to speak of power is to attract investment and those who seek power. Speaking of "Safety" is perhaps not to actually seek to 'protect humanity', but rather to advertise dominance.
vertis 3y ago
I remember going to a talk in Berkeley by Nick Bostrom about AI Control theory after reading his book Superintelligence (in 2014?). What struck me at the time was AI control theory was mostly pointless. If you put all these guard rails around your AI to keep it from turning everyone into paperclips (etc) then it's just not YOUR AI that turns everyone into paperclips/computronium, it's the person that doesn't put those guard rails in. Worse by putting those guard rails in you're making your AI less effective that the other guys so it can't even outcompete.
lapsed_pacifist 3y ago
These safety groups seem to be mostly concerned with making sure LLMs don't say bad words, not with preventing FOOM!.

nightowl_games 3y ago

In my opinion "safety" has always been about that and it's also a projection of power. If you say "we must ensure safety from this machine" you are asserting the machine has a high level of power. People are attracted to power. Investors are attracted to power. Speaking of "safety" is implicitly speaking of power.

esafak 3y ago
Well duh. The question is how to ensure humanity's safety from the power of AI, whether it is autonomous or wielded by other humans.
- nightowl_games 3y ago
  I'm specifically stating that to speak of safety is to speak of power. And to speak of power is to attract investment and those who seek power. Speaking of "Safety" is perhaps not to actually seek to 'protect humanity', but rather to advertise dominance.

vertis 3y ago

I remember going to a talk in Berkeley by Nick Bostrom about AI Control theory after reading his book Superintelligence (in 2014?). What struck me at the time was AI control theory was mostly pointless. If you put all these guard rails around your AI to keep it from turning everyone into paperclips (etc) then it's just not YOUR AI that turns everyone into paperclips/computronium, it's the person that doesn't put those guard rails in. Worse by putting those guard rails in you're making your AI less effective that the other guys so it can't even outcompete.

lapsed_pacifist 3y ago

These safety groups seem to be mostly concerned with making sure LLMs don't say bad words, not with preventing FOOM!.