I got the notification that my subscription to The Argument renewed this weekend.
Writers that I respected had left major publications and absconded to Substack before but I had always felt a little bit weird about paying for a single writer.
Subscribing to a magazine, however, was a lot more palatable, and it's gratifying reflecting on how many compelling articles came from writers that I had not been exposed to before. The Argument pretty quickly rose into the piece of media that I value reading the most every day.
The Argument ended up breaking the dam and actually getting me on Substack for individual writers (for better and sometimes for worse) but the reminder that it's been a year made me reflect that I am easily getting more than my subscription's worth.
Honestly, this isn’t a hard question. Prosecute the people that set the AIs up to it. It was humans who had the idea to give the AIs the task, and trained them to be persistent. The humans involved with building these AIs are openly aware of the risks they pose.
This is negligence. It’s something we already have a very well known and widely used legal framework for.
I'd hate to be the monster on a jury voting to lock them up, though. They'd cry they never meant to cause in harm, they were trying to stop the AI, they thought they had sufficient safety guards in place, and I'd weep in sympathy and vote "Not Guilty".
Yes, but negligence is most often used in the context of civil law. There are some places where it's used as a mens rea standard in criminal law, but even then it typically shows up as gross negligence, which is heightened above what the typical tort standard uses.
The situation you laid out in your comment could plausibly meet the bar for negligence in the civil/tort way but I'm skeptical it would clear it for gross negligence in the criminal context.
[The U.S. Army has announced plans to add nuclear microreactors to military bases]
I want to believe it, but I'll celebrate when one microreactor on any base is actually operating. Otherwise it's just the eternal Trump promise, "two weeks away".
Snap! I was just making this point in response to Joshua Gans. Strict, joint and several liability for all humans in the process, from AI designers to the person who pressed send on the prompt.
"Intent" is generally used as a "moral" judgement against people. They "deserve" punishment because they "intended" harm.
But we gotta do away with "morality" in public life because "moral" claims are always essentially unfalsifiable and therefore religious.
But "intent" is still relevant in terms of prediction. If someone intended harm in one case, then they will likely intend harm again, and should be penalized in order to prevent future harm.
In the recent AI hacking cases, there is still nobody who intended to do harm. But I think going forward, we do need find ways to hold AI companies liable, because "intent" isn't the only predictor of future harm.
It's probably not fair to go too harsh on these initial mishaps, because until recently, we didn't know how capable AIs were of hacking whatever they want. But as we become more educated on what AI will likely do, we can reasonably hold AI companies liable for not taking certain precautions.
The “cheapest cost avoider” heuristic carries the major advantages IMO of being both workable and modelable. Do you think people will in general ‘buy’ this approach from a fairness/legitimacy perspective, though?
Contemplating mere tweaks to the law would make sense if AI agents were going to remain at their current level of capability for the foreseeable future. Unfortunately, AI capabilities are a moving target with a quantifiable track record of extremely rapid improvement.
In other words, today’s agent swarms hacking into companies are tomorrow’s superintelligences taking over all of society—what better way could there be for an agent to ensure that it passes its evaluation?
It’s time to wake up to the emergency we’re in and advocate for strict controls on the training of more powerful models.
I think Jerusalem's post hints at a better path - OAI should be held criminally and civilly liable for lawbreaking enacted by it's digital creations. That's honestly the only surefire way to ensure they actually take safety seriously.
Financial incentives aren’t enough to control the development of a technology that is as dangerous as biological weapons, if not more so. If Chevron blows up an oil refinery, they can be sued, which will incentivize them to have better safety procedures going forward. The equivalent industrial accident in the case of AI could result in the end of human control over the planet.
I got the notification that my subscription to The Argument renewed this weekend.
Writers that I respected had left major publications and absconded to Substack before but I had always felt a little bit weird about paying for a single writer.
Subscribing to a magazine, however, was a lot more palatable, and it's gratifying reflecting on how many compelling articles came from writers that I had not been exposed to before. The Argument pretty quickly rose into the piece of media that I value reading the most every day.
The Argument ended up breaking the dam and actually getting me on Substack for individual writers (for better and sometimes for worse) but the reminder that it's been a year made me reflect that I am easily getting more than my subscription's worth.
Thanks for all the hard work!
Honestly, this isn’t a hard question. Prosecute the people that set the AIs up to it. It was humans who had the idea to give the AIs the task, and trained them to be persistent. The humans involved with building these AIs are openly aware of the risks they pose.
This is negligence. It’s something we already have a very well known and widely used legal framework for.
I'd hate to be the monster on a jury voting to lock them up, though. They'd cry they never meant to cause in harm, they were trying to stop the AI, they thought they had sufficient safety guards in place, and I'd weep in sympathy and vote "Not Guilty".
Shrug. Intention doesn’t really come into it wrt negligence.
There would be 11 other jurors.
Sure, but what the standard of care is does matter, which is really where the question of sufficient safety guards comes up
Yes, but negligence is most often used in the context of civil law. There are some places where it's used as a mens rea standard in criminal law, but even then it typically shows up as gross negligence, which is heightened above what the typical tort standard uses.
The situation you laid out in your comment could plausibly meet the bar for negligence in the civil/tort way but I'm skeptical it would clear it for gross negligence in the criminal context.
[The U.S. Army has announced plans to add nuclear microreactors to military bases]
I want to believe it, but I'll celebrate when one microreactor on any base is actually operating. Otherwise it's just the eternal Trump promise, "two weeks away".
Snap! I was just making this point in response to Joshua Gans. Strict, joint and several liability for all humans in the process, from AI designers to the person who pressed send on the prompt.
"Intent" is generally used as a "moral" judgement against people. They "deserve" punishment because they "intended" harm.
But we gotta do away with "morality" in public life because "moral" claims are always essentially unfalsifiable and therefore religious.
But "intent" is still relevant in terms of prediction. If someone intended harm in one case, then they will likely intend harm again, and should be penalized in order to prevent future harm.
In the recent AI hacking cases, there is still nobody who intended to do harm. But I think going forward, we do need find ways to hold AI companies liable, because "intent" isn't the only predictor of future harm.
It's probably not fair to go too harsh on these initial mishaps, because until recently, we didn't know how capable AIs were of hacking whatever they want. But as we become more educated on what AI will likely do, we can reasonably hold AI companies liable for not taking certain precautions.
The “cheapest cost avoider” heuristic carries the major advantages IMO of being both workable and modelable. Do you think people will in general ‘buy’ this approach from a fairness/legitimacy perspective, though?
Contemplating mere tweaks to the law would make sense if AI agents were going to remain at their current level of capability for the foreseeable future. Unfortunately, AI capabilities are a moving target with a quantifiable track record of extremely rapid improvement.
In other words, today’s agent swarms hacking into companies are tomorrow’s superintelligences taking over all of society—what better way could there be for an agent to ensure that it passes its evaluation?
It’s time to wake up to the emergency we’re in and advocate for strict controls on the training of more powerful models.
I think Jerusalem's post hints at a better path - OAI should be held criminally and civilly liable for lawbreaking enacted by it's digital creations. That's honestly the only surefire way to ensure they actually take safety seriously.
That is necessary but not sufficient imo
I trust financial incentives to be more effective than government regulation, but I wouldn't mind both.
Financial incentives aren’t enough to control the development of a technology that is as dangerous as biological weapons, if not more so. If Chevron blows up an oil refinery, they can be sued, which will incentivize them to have better safety procedures going forward. The equivalent industrial accident in the case of AI could result in the end of human control over the planet.