The rational methods used to establish a large P(doom).
To definitively falsify the claim, humanity would need to live alongside superintelligence forever without it wiping us out. That doesn't mean we can't be pretty darn sure though, in the same way that we can be pretty darn sure that nuclear war could end civilization.
I wish you'd used "very large asteroid strike on Earth" rather than "nuclear war" because although nuclear war will probably make thing "uncivilized" for a while, I'm pretty sure human civilization bounces right back after even the largest realistic nuclear war.
I wouldn't be so sure about that. 95% of the population will starve to death, global supply chains will collapse, and the world will regress to a pre-industrial-revolution lifestyle - with the critical difference that all shallow oil and coal deposits have been mined out, not to mention many other resources. You're going to have a very tough time bootstrapping it all again. All vital knowledge for running a modern civilization will be lost - most of it isn't even written down now, but is instead implicit in distributed social structures which will all be destroyed. They won't re-form quickly either - travel will be even more risky than in the Dark Ages thanks to fallout. Even a healthy society can coast in a mostly steady state for thousands of years without technological takeoff, and we've had several civilizations in the past who seemed to hold most of the pieces - empiricism, healthy economy transcending food production, precision metalworking - and yet never seemed to quite get there. This won't even be that - it's a restart in a world depleted of natural resources and blighted by death, radiation, mass extinction, and climate change.
"95% of the population will starve to death" is very unlikely although I can't completely rule it out as a possibility. The research on nuclear winter is of a poor quality, and we should be funding better research.
"The world will regress to a pre-industrial-revolution lifestyle" seems to be a retreat from "civilization will end" (not that I agree with it).
>travel will be even more risky than in the Dark Ages thanks to fallout
The radioactive contamination caused by Chernobyl has persisted for decades. In contrast, the radionuclides created by nuclear explosions are of a very different nature and persist for weeks. 3 weeks after the end of the explosions, there will be no place except ground-zero craters (i.e., very small areas) where the levels of radioactivity remain at deadly levels. (But I definitely agree with you that during those critical 3 weeks, almost all people who decide to travel will end up dying unless they know a lot about nuclear war and have good radiation detectors.)
To repeat my overall point: the worst-case scenario for the outcome of AI research should be compared to an asteroid strike by an extremely large asteroid, not to nuclear war—because nuclear war is too tame in its effects to make a good comparison.
The reason I got involved in this discussion is that some people use the possibility of nuclear war as one reason for us to go forward with AI research as fast as possible. Such people argue that a sufficiently mature AI technology would eliminate the possibility of nuclear war. I think that is a fallacious argument become AI "progress" is the more potent of the two dangers.
Those "rational methods" are essentially just pascal's wager. You stick some ~infinitely bad outcome (nasal daemons torturing simulations of you and everyone you love for all time) and/or ~infinitely good outcome (perfect paradise for as many sentient minds as could fit in the physical universe), then you can get whatever conclusion you want by multiplying any finite probability by these things, and even give yourself bonus asspats for being conservative by adding several order of magnitude fudge factors against your claims.
The rationalist doom and sex cult uses convoluted mathematics inspired language to justify their faith-- but at the end of the day it's just faith no matter how they dress it up.
No, it's not Pascal's wager because Pascal's wager works even when the probabilities are tiny whereas Eliezer's p(doom) is about 99% unless AI "progress" is shut down and kept shut long enough that the current crop of AI researchers is going to have to find something else to do for a career.
In other words, it's not Pascal's wager if the very bad outcome is what you naturally expect to happen.
And what I just wrote applies to not just Eliezer and his allies: I don't know a single person or a single published argument that has ever used Pascal's-wager-like argumentation to argue that AI "progress" needs to be stopped. Eliezer is the author of a blog post titled "Pascal's mugging", but that blog post and any follow-up discussion are completely silent on the dangers of AI.
How cynical. What if the fears of extinction are _rational_? Given the stakes, we're obligated to treat the threat as a real possibility until proven otherwise.
People will go to great lengths to avoid believing truths that dramatically challenge their worldviews. Consider the case of Rudolf Vrba and Alfréd Wetzler, who escaped from Auschwitz and were initially met with disbelief when they described the atrocities they'd witnessed.
Nobody ever gives a real reason why extinction is a plausible threat from current model development, and Chinese researchers don't seem to agree that it's plausible either.
> If you actually thought your product will kill all humans, then you could, you know, like, just stop building the torment nexus. All by yourself, no government regulations required.
Wrong. This is a classical prisoner's dilemma. _Everyone_ has to stop, otherwise the undesirable outcome happens faster. It requires coordination, which can be achieved through regulation.
Isn't it telling that Musk, Altman, Hassabis, Amodei are all signaling fear of what unchecked AI development could do? These guys can't normally agree on the color of the sky. Let's give them credit for trying to sound the alarm.
If you choose to slow down (cooperate) and the other side doesn't, you lose more than if you didn't try to slow down. In order to succeed (gain mutual advantage) both parties have to choose to cooperate without knowing beforehand what the other party will do.
Chemist here, what bioweapon do you know of that can wipe out 8 billion people before anyone can do something about it? Never mind one release by a single person.
It's really not 'trivial' to slaughter millions of people.
Such a bioweapon (probably) doesn't exist yet, but that doesn't mean it's impossible in principle. It's "trivial" in the sense that it's easy to _imagine_ such a thing... but a superintelligence is by definition much much smarter than humans, so whatever it comes up with will likely be very surprising to us.
The rational methods used to establish a large P(doom).
To definitively falsify the claim, humanity would need to live alongside superintelligence forever without it wiping us out. That doesn't mean we can't be pretty darn sure though, in the same way that we can be pretty darn sure that nuclear war could end civilization.
reply