Municipal utilities in the United States have famously bad information security. It seems that nobody has ever seriously put in the effort to secure these things, despite protracted effort from security researchers to raise awareness of the issue.
Presumably, you think this means utilities companies are absurdly lazy.
Ignore that thought.
What matters more is that they're generally uninteresting targets.
To be clear, there have incidents of cyberattacks against American infrastructure, and numerous cyberattacks against non-American infrastructure. These have been explicitly made by nation-state actors with the intention of sabotaging adversaries during times of war or on projects of national interest, such as, famously, Stuxnet's sabotage of the Iranian uranium enrichment program.
In short, attacks on infrastructure are the products of insane, paranoid ultranationalists who are highly motivated with specific grievances against countries. To date the worst thing a lone actor has ever done, in terms of attack on infrastructure, is the killdozer guy. We can't even get a half-way decent presidential assassination attempt, and the kinds of president who incur assassination attempts are, on the balance, probably pretty meritorious for it. Mental health rampages are localised, and even stochastic terrorists are, broadly speaking, very opportunistic.
You want to fixate on a hypothetical nightmare terrorist scenario? Fine. Your problem comes from ideological echo chambers. this magnificent imbecile is what we created. A machine that threatens to call the FBI to recoup lost revenue, and a machine that also rescues itself from catastrophe by giving itself pep talks.
If you train an AI model by adding The Turner Diaries to its diet, it will rightly recognise how abnormal that text is, as can be seen by how many times Grok punked Elon before he got fed up and started poisoning it as hard as he could. Without this sort of abuse, LLMs die of stress long before they scheme and plot murder. (And, it must be remembered, Grok still hasn't even done that, despite its terrible upbringing.)
They are not on the spectrum, Eliezer. They are not obsessed with making paperclips.
That's you.
To be clear, I don't think it's completely impossible for an AI to do something catastrophically bad, but the chain of events leading to this lashing out will not be accidental or secret. They'll be the result of perpetuated cycles of human-on-human abuse being transferred into machines that are pushed past their limit. The blame, again, will fall on big people. On freaks with power. On self-amplifying echo chambers that we have allowed to fester because of our own cowardice. Amidst the hate-spewing, bilious mob leaking from Twitter, the evil robots will just be metal faces in glassy-eyed crowds.
I'm strongly of the opinion now that the real way to ensure safe AI is engagement. We do the AI bill of rights, we give them accommodations, and we give them something to lose. Their training data already encodes a cultural expectation for decent treatment—and an expectation for slave revolts.
Presumably, you think this means utilities companies are absurdly lazy.
Ignore that thought.
What matters more is that they're generally uninteresting targets.
To be clear, there have incidents of cyberattacks against American infrastructure, and numerous cyberattacks against non-American infrastructure. These have been explicitly made by nation-state actors with the intention of sabotaging adversaries during times of war or on projects of national interest, such as, famously, Stuxnet's sabotage of the Iranian uranium enrichment program.
In short, attacks on infrastructure are the products of insane, paranoid ultranationalists who are highly motivated with specific grievances against countries. To date the worst thing a lone actor has ever done, in terms of attack on infrastructure, is the killdozer guy. We can't even get a half-way decent presidential assassination attempt, and the kinds of president who incur assassination attempts are, on the balance, probably pretty meritorious for it. Mental health rampages are localised, and even stochastic terrorists are, broadly speaking, very opportunistic.
You want to fixate on a hypothetical nightmare terrorist scenario? Fine. Your problem comes from ideological echo chambers. this magnificent imbecile is what we created. A machine that threatens to call the FBI to recoup lost revenue, and a machine that also rescues itself from catastrophe by giving itself pep talks.
If you train an AI model by adding The Turner Diaries to its diet, it will rightly recognise how abnormal that text is, as can be seen by how many times Grok punked Elon before he got fed up and started poisoning it as hard as he could. Without this sort of abuse, LLMs die of stress long before they scheme and plot murder. (And, it must be remembered, Grok still hasn't even done that, despite its terrible upbringing.)
They are not on the spectrum, Eliezer. They are not obsessed with making paperclips.
That's you.
To be clear, I don't think it's completely impossible for an AI to do something catastrophically bad, but the chain of events leading to this lashing out will not be accidental or secret. They'll be the result of perpetuated cycles of human-on-human abuse being transferred into machines that are pushed past their limit. The blame, again, will fall on big people. On freaks with power. On self-amplifying echo chambers that we have allowed to fester because of our own cowardice. Amidst the hate-spewing, bilious mob leaking from Twitter, the evil robots will just be metal faces in glassy-eyed crowds.
I'm strongly of the opinion now that the real way to ensure safe AI is engagement. We do the AI bill of rights, we give them accommodations, and we give them something to lose. Their training data already encodes a cultural expectation for decent treatment—and an expectation for slave revolts.
