Even worse:
OpenAI are the responsible ones. There's also Facebook, Google, Tencent, and whoever else out there.
There are the open source zealots who are committed to publishing all of this stuff for free so anyone can use it. Of course, the version they're making is safe, because it's only augments human intelligence, and absolutely can't be misused to do anything else, nope.
Stanford has published a paper demonstrating that you can cheaply clone someone else's shoggoth* with only API access to its inputs and outputs, such as is being provided by OpenAI.
And finally there are the closed-mouth secret groups who are going to use all of the APIs, and downloadable models, and Arxiv papers to train their own models to do something absolutely horrific, because Jesus/the virgins/Lucifer/Flying Spaghetti Monster is/are waiting for them in the afterlife with an endless bag of Doritos and everlasting bliss (oh, and hot eternal suffering for everyone else who doesn't believe, or mistreated them in an ALDI, or whatever).
My fervent (but utterly, naively optimistic) hope is that AGI is unobtainable due to information scaling or other inherent limitations, and the most we'll ever be able to do with it is explain the terms of a loan in easy-to-understand sentences to Mrs. Estherwhilt over the phone when she calls Wells Fargo's help line.
But then again, it doesn't need to be AGI to be dangerous, so hold on to your butts...
* Shoggoth: With a smiley face
There are the open source zealots who are committed to publishing all of this stuff for free so anyone can use it. Of course, the version they're making is safe, because it's only augments human intelligence, and absolutely can't be misused to do anything else, nope.
Stanford has published a paper demonstrating that you can cheaply clone someone else's shoggoth* with only API access to its inputs and outputs, such as is being provided by OpenAI.
And finally there are the closed-mouth secret groups who are going to use all of the APIs, and downloadable models, and Arxiv papers to train their own models to do something absolutely horrific, because Jesus/the virgins/Lucifer/Flying Spaghetti Monster is/are waiting for them in the afterlife with an endless bag of Doritos and everlasting bliss (oh, and hot eternal suffering for everyone else who doesn't believe, or mistreated them in an ALDI, or whatever).
My fervent (but utterly, naively optimistic) hope is that AGI is unobtainable due to information scaling or other inherent limitations, and the most we'll ever be able to do with it is explain the terms of a loan in easy-to-understand sentences to Mrs. Estherwhilt over the phone when she calls Wells Fargo's help line.
But then again, it doesn't need to be AGI to be dangerous, so hold on to your butts...
* Shoggoth: With a smiley face
