As a an AI Safety researcher who is freaked about about catatrophic risks (e.g. bioweapons) from AI (sometimes to the point of rudeness, sorry): this is the best take on AI regulation I've yet seen, by far.
You really hit the nail on the head with,
"But how do we regulate an industrial revolution? How do we regulate an era?
There is no way to pass “a law,” or a set of laws, to control an industrial revolution. "
The Narrow Path essay and Max Tegmark's Hopium essay both seem to suggest that, because the AI industrial revolution is scary, we should ask the US Federal government to wave a magic wand and make it not happen. That simply isn't an option on the table, and any realistic plan must start by facing up to that.
Rather than trying to tell people not to use AI for science or AI for improving AI, we should aim to channel it. Offer rewards (e.g. subsidized compute) for researchers in exchange for operating in a loosely supervised setting. Focus on preventing only the very worst civilization-scale risks rather than micromanaging. Trying to reign in developers with overly stringent rules will just drive research underground. We can't stop the tide of technology.
Will help when my friend Johny finally gets Amanda to write Anthropic's. They're next. Google is always last on policy things, unless we can convince them its good for developers.
I'm going to keep slamming my head against the desk until enough people have realised that if you can formalise the outcomes you don't want, you can formalise the benchmarks before it, so you can formalise dynamic legislation and regulation and costs that snap into place if/when they're met. You can also have bonds for firms involved in certain development. You can have advanced market commitments.
We don't need backward looking lagging regulations. This is beyond basic now.
"There is no way to pass 'a law,' or a set of laws, to control an industrial revolution."
True, but then you gave it a shot anyway. No, no kind or quantity of word-magic will work any better than today's systems of laws, badly written and selectively enforced. Nor do the "safety" bugmen in the private sector have any chance of doing anything good, they can only delay usefulness of AI.
No, you didn't forget, most likely just an example of the disingenuous rhetorical style characteristic of Mercatus' managerialist alignment. I am religiously opposed to lies and Straussian concealment of truth. You do say a great deal in most of your articles with which I agree, though.
Your frame in the bulk of the article presumes the legitimacy of software regulations and regulators, which concedes the whole issue. I agree that there needs to be Federal preemption but simply to prevent states from infringing freedom, not to prohibit or mandate ANYTHING.
The risks go beyond the obvious - one should be looking at other ways that proliferation of AI could endanger humankind. For one, devolution has species-altering potential if not worse. We (and children) are increasingly relying on AI instruments for the simplest of tasks like writing essays, arithmetic calculations, and so on. We are devolving biologically faster than we have evolved in the last million years. Brains are shrinking fast enough to show significant decline in a few decades.
One can’t just sit back and wonder how an industrial revolution can be contained or suggest that it can’t. Instead, find ways to enforce regulations, advocates need to come forward and act. I wonder what the annual conventions run by the other section of leaders and billionaires (AI opponents) are doing and saying about this.
On the flip side, the recent elections perhaps could work in favour of the cause. Although I dread the far right and the implications, and think it doesn’t bode well for World progress, Trump’s alliance with Musk could actually help foster AI regulations to a great extent.
Hi Dean, While this is not my area, I'm glad I subscribed a bit ago as I'm finding your writing very interesting. Hope you have a nice holiday season!
Thank you!
As a an AI Safety researcher who is freaked about about catatrophic risks (e.g. bioweapons) from AI (sometimes to the point of rudeness, sorry): this is the best take on AI regulation I've yet seen, by far.
You really hit the nail on the head with,
"But how do we regulate an industrial revolution? How do we regulate an era?
There is no way to pass “a law,” or a set of laws, to control an industrial revolution. "
The Narrow Path essay and Max Tegmark's Hopium essay both seem to suggest that, because the AI industrial revolution is scary, we should ask the US Federal government to wave a magic wand and make it not happen. That simply isn't an option on the table, and any realistic plan must start by facing up to that.
Rather than trying to tell people not to use AI for science or AI for improving AI, we should aim to channel it. Offer rewards (e.g. subsidized compute) for researchers in exchange for operating in a loosely supervised setting. Focus on preventing only the very worst civilization-scale risks rather than micromanaging. Trying to reign in developers with overly stringent rules will just drive research underground. We can't stop the tide of technology.
Thank you! And 100% in agreement. I wish more people in policy making communities grokked this.
Nice to have it all in one place. The hard part is rank ordering them.
Manifesto gang 👀.
Lots of overlap here in the areas relevant to me: NAIRR, DoE, model specs. Nice.
Your model spec idea has been quite influential on me. Next big step is: how do we define this in statute? (Something we should discuss sometime!)
Will help when my friend Johny finally gets Amanda to write Anthropic's. They're next. Google is always last on policy things, unless we can convince them its good for developers.
Continue to find your writing really worthwhile, converted this through ElevenLabs to audio again:
https://open.substack.com/pub/askwhocastsai/p/heres-what-i-think-we-should-do-by
thank you!
I'm going to keep slamming my head against the desk until enough people have realised that if you can formalise the outcomes you don't want, you can formalise the benchmarks before it, so you can formalise dynamic legislation and regulation and costs that snap into place if/when they're met. You can also have bonds for firms involved in certain development. You can have advanced market commitments.
We don't need backward looking lagging regulations. This is beyond basic now.
"There is no way to pass 'a law,' or a set of laws, to control an industrial revolution."
True, but then you gave it a shot anyway. No, no kind or quantity of word-magic will work any better than today's systems of laws, badly written and selectively enforced. Nor do the "safety" bugmen in the private sector have any chance of doing anything good, they can only delay usefulness of AI.
Do you think I wrote the words you quoted and then just forget about them when I wrote the rest of this essay?
No, you didn't forget, most likely just an example of the disingenuous rhetorical style characteristic of Mercatus' managerialist alignment. I am religiously opposed to lies and Straussian concealment of truth. You do say a great deal in most of your articles with which I agree, though.
Your frame in the bulk of the article presumes the legitimacy of software regulations and regulators, which concedes the whole issue. I agree that there needs to be Federal preemption but simply to prevent states from infringing freedom, not to prohibit or mandate ANYTHING.
The rise of AI means the replacement of humanity by something else. Do you have a policy for that?
I obviously do not share your degree of certainty about this matter even remotely.
Is there anything that humans will always be better at?
Transparent to whom?
And when government labs, intelligence, and military orgs set up projects, these regulations will not apply to them?
To the public, and correct.
The risks go beyond the obvious - one should be looking at other ways that proliferation of AI could endanger humankind. For one, devolution has species-altering potential if not worse. We (and children) are increasingly relying on AI instruments for the simplest of tasks like writing essays, arithmetic calculations, and so on. We are devolving biologically faster than we have evolved in the last million years. Brains are shrinking fast enough to show significant decline in a few decades.
One can’t just sit back and wonder how an industrial revolution can be contained or suggest that it can’t. Instead, find ways to enforce regulations, advocates need to come forward and act. I wonder what the annual conventions run by the other section of leaders and billionaires (AI opponents) are doing and saying about this.
On the flip side, the recent elections perhaps could work in favour of the cause. Although I dread the far right and the implications, and think it doesn’t bode well for World progress, Trump’s alliance with Musk could actually help foster AI regulations to a great extent.