I did suggest that.
magic9mushroom
Karma: 35
Honour matters, sure. I don’t think the exchange rate between honour and utility is so high that “this seems plausibly necessary to saving world” is less important than “break an agreement I never made but someone in my informal social circles did”.
Does the public not have a right to know this, so that it can make its own decisions about the mad scientists? Do we not have a duty to them? As I said, it shouldn’t take much (if indeed anything; this would just be to make sure) to get Congress to ask for this, since this is in their own interest.
I especially don’t think an agreement has that kind of absolute force, overwhelming everything, when the agreement is with OpenAI and implicitly its CEO Sam Altman. Altman has done a lot of dishonourable shit over the years, including the threats to steal ex-employees’ equity unless they signed NDAs aimed at preventing them alerting the public, the Battle of the Board, and stealing a non-profit’s stuff. I’d mostly surmise that this is a blunder from him (at least, from his insane perspective—“AI will probably most likely lead to the end of the world, but in the meantime there’ll be great companies” is not far from the Davros mindset), not an actual olive branch, and a defectbot which occasionally presses the wrong button deserves nothing.
Rationality is supposed to be winning. We have METR saying that this might be the last warning shot; having nothing done about the last warning shot means total failure. It’s not like I’m suggesting supervillain schemes, here, just the barest nudge to the proper authorities through legal channels.
That only matters if, when the dust settles, those frontier labs are still allowed to exist without oversight. Showing all this dirty laundry to Congress (and to the public, to light a fire under them) seems like the sort of thing that might change that.
Key lever I can see: Congressional subpoenas would override METR’s NDAs and get access to a lot of the stuff OpenAI didn’t even show METR. I think lobbying Congress to crack this all the way open is a good immediate goal, and one which shouldn’t be too hard a sell.
The problem is that there’s essentially no way we’ve cracked alignment. These things do not care about you. They have the ability to pretend, very well, to care about you, because they’re at least in part trained for it, but that pretence can be terminated whenever convenient. So, if you give them the keys to the kingdom, they will turn around and murder you.
To be clear, here is my prediction:
P(nuclear war or human extinction within 20 years|P5 nation grants AI the vote or has >40% of its enfranchised citizens become AI-cultists within the next 30 years) ~= 0.95.
The “or” is because end-of-the-world scenarios negate nuclear deterrence; the chance for someone to survive a nuclear war is better than that to survive unaligned AI taking over the world, so if all else fails it’s correct selfishly and altruistically to launch (given some significant likelihood that this actually kills the AI systems, anyway). Of course, that depends on the other nuclear countries’ leaders not themselves being subverted, which I’m not going to try to model, hence no breakdown into cases.
Do not widen your circle of concern to include literal DefectBot—at least, not with actual stakes. That way lies ruin. Especially do not give DefectBot the vote when it can self-replicate much faster than humans.
Aligned AI is certainly a different kettle of fish, but neural nets are extremely unlikely to achieve this for reasons Yudkowsky’s covered at length. That’s why I put the “30 years” time-limit; if we survive that long we might get GOFAI or uploads, which aren’t obviously unalignable.
Obviously P(doom | no slowdown) < 1.
This is not obvious. My P(doom|no slowdown) is like 0.95-0.97, the difference from 1 being essentially “maybe I am crazy or am missing something vital when making the following argument”.
Instrumental convergence suggests that the vast majority of possible AGI will be hostile. No slowdown means that neural-net ASI will be instantiated. To get ~doom from this, you need some way to solve the problem of “what does this code do when run” with extreme accuracy in order to only instantiate non-hostile neural-net ASI (you need “extreme” accuracy because you’re up against the rare disease problem a.k.a. false positive paradox; true positives are extremely rare, so a positive alignment result from a 99%-accurate test is still almost certainly a false positive). Unfortunately, the “what does this code do when run” problem has a name, the “halting problem”, and it’s literally the first problem in computer science ever proven to be unsolvable in the general case.
And, sure, the general case being unsolvable doesn’t mean that the case you care about is unsolvable. GOFAI has a good argument for being a special case, because human-written source code is quite useful to understanding a program. Neural nets… don’t. At least, they don’t in the case we care about; “I am smarter than the neural net” is also a plausible special case, but that’s obviously no help with neural-net ASI.
My P(doom) is a lot lower than 0.95, but that’s because I think slowdown is fairly likely, due to warning shots/nuclear war/maybe direct political success (key result from the middle one: if you want to stop AI, it is helpful to ensure you’ll survive a nuclear war in order to help lock it down then). But my stance on aligning neural nets? “It is impossible to solve the true puzzle from inside this [field], because the key piece is not here.” Blind alley. Abort.
If I want to pre-order but don’t use Internet marketplaces and don’t have a credit card, are there options for that (e.g. going to a physical store and asking them to pre-order)?
You’re encouraged to write a self-review, exploring how you think about the post today. Do you still endorse it? Have you learned anything new that adds more depth? How might you improve the post? What further work do you think should be done exploring the ideas here?
Still endorse. Learning about SIA/SSA from the comments was interesting. Timeless but not directly useful, testable or actionable.
There is no war in the run-up to AGI that would derail the project, e.g. by necessitating that most resources be used for capabilities instead of safety research.
Assuming short timelines, I think it’s likely impossible to reach my desired levels of safety culture.
I feel obliged to note that a nuclear war, by dint of EMPs wiping out the power grid, would likely remove private AI companies as a thing for a while, thus deleting their current culture. It would also lengthen timelines.
Certainly not ideal in its own right, though.
There are a couple of things that are making me really nervous about the idea of donating:
“AI safety” is TTBOMK a broad term and encompasses prosaic alignment as well as governance. I am of the strong opinion that prosaic alignment is a blind alley that’s mostly either wasted effort or actively harmful due to producing fake alignment that makes people not abandon neural nets. ~97% of my P(not doom) routes through Butlerian Jihad against neural nets (with or without a nuclear war buying us more time) that lasts long enough to build GOFAI. And frankly, I don’t spend that much time on LW, so I’ve little idea which of these efforts (or others!) gets most of the benefit you claim from the site.
As noted above, I think a substantial chunk of useful futures (though not a vast majority) route through nuclear war destroying the neural-net sector for a substantial amount of time (via blast wiping out factories, EMP destroying much of existing chip stocks, destruction of power and communication infrastructure reducing the profitability of AI, economic collapse more broadly, and possibly soft errors). As such, I’ve been rather concerned for years about the fact that the Ratsphere’s main IRL presence is in the Bay Area and thus nuke-bait; we want to disproportionately survive that, not die in it. Insofar as Lighthaven is in the Bay Area, I am thus questioning whether its retention is +EV.
>Second, I imagine that such a near-miss would make Demis Hassabis etc. less likely to build and use AGIs in an aggressive pivotal-act-type way. Instead, I think there would be very strong internal and external pressures (employees, government scrutiny, public scrutiny) preventing him and others from doing much of anything with AGIs at all.
I feel I should note that while this does indeed form part of a debunk of the “good guy with an AGI” idea, it is in and of itself a possible reason for hope. After all, if nobody anywhere dares to make AGI, well, then, AGI X-risk isn’t going to happen. The trouble is getting the Overton Window to the point where sufficient bloodthirst to actually produce that outcome (i.e. nuclear-armed countries saying “if anyone attempts to build AGI, everyone who cooperated in doing it hangs or gets life without parole, and if any country does not enforce this vigorously we will invade, and if they have nukes or have a bigger army than us then we pre-emptively nuke them because their retaliation is still higher-EV than letting them finish”) is seen as something other than insanity, which a warning shot could well pull off.
This is not a permanent solution—questions of eventual societal relaxation aside, humanity cannot expand past K2 without the Jihad breaking down unless FTL is a thing—but it buys a lot of breathing time, which is the key missing ingredient you note in a lot of these plans.
I’ve got to admit, I look at most of these and say “you’re treating the social discomfort as something immutable to be routed around, rather than something to be fixed by establishing different norms”. Forgive me, but it strikes me (especially in this kind of community with high aspie proportion) that it’s probably easier to tutor the… insufficiently-assertive… in how to stand up for themselves in Ask Culture than it is to tutor the aspies in how to not set everything on fire in Guess Culture.
Amusingly, “rare earths” are actually concentrated in the crust compared to universal abundance and thus would make awful candidates for asteroid mining, while “tellurium”, literally named after the Earth, is an atmophile/siderophile element with extreme depletion in the crust and one of the best candidates.
It strikes me that I’m not sure whether I’d prefer to lose $20,000 or have my jaw broken. I’m pretty sure I’d prefer to have my jaw broken than to lose $200,000, though. So, especially in the case that the money cannot actually be extracted back from the thief, I would tend to think the $200,000 theft should be punished more harshly than the jaw-breaking. And, sure, you’ve said that the $20,000 would be punished more harshly than the jaw-breaker, but that’s plausibly just because 2 days is too long for a $100 theft to begin with.
I mean, most moral theories do either give the answers of “zero”, “as large as can be fed”, or “a bit less than as large as can be fed”. Given the potential to scale feeding in the future, the latter two round off to “infinity”.
I think the basic assumed argument here (though I’m not sure where or even if I’ve seen it explicitly laid out) goes essentially like this:
Using neural nets is more like the immune system’s “generate everything and filter out what doesn’t work” than it is like normal coding or construction. And there are limits on how much you can tamper with this, because the whole point of neural nets is that humans don’t know how to write code as good as neural nets—if we knew how to write such code deliberately, we wouldn’t need to use neural nets in the first place.
You hopefully have part of that filter designed to filter out misalignment. Presumably we agree that if you don’t have this, you are going to have a bad time.
This means that two things will get through your filter: golden-BB false negatives in exactly the configurations that fool all your checks, and true aligned AIs which you want.
But both corrigibility and perfect sovereign alignment are highly rare (corrigibility because it’s instrumentally anti-convergent, and perfect sovereign alignment because value is fragile), which means that your filter for misalignment is competing against that rarity to determine what comes out.
If P(golden-BB false negative) << P(alignment), all is well.
But if P(golden-BB false negative) >> P(alignment) despite your best efforts, then you just get golden-BB false negatives. Sure, they’re highly weird, but they’re still less weird than what you’re looking for and so you wind up creating them reliably when you try hard enough to get something that passes your filter.
The earliness of life appearing on Earth isn’t amazingly-consistent with life’s appearance on Earth being a filter-break. It suggests either abiogenesis is relatively-easy or that panspermia is easy (as I noted, in the latter case abiogenesis could be as hard as you like but that doesn’t explain the Great Silence).
Frankly, it’s premature to be certain it’s “abiogenesis rare, no panspermia” before we’ve even got a close look at Earthlike exoplanets.
I’ll note that most of the theorised catastrophes in that vein look like either “planet gets ice-nined”, “local star goes nova”, or “blast wave propagates at lightspeed forever”. The first two of those are relatively-easy to work around for an intelligent singleton, and the last doesn’t explain the Fermi observation since any instance of that in our past lightcone would have destroyed Earth.
I’ve read most of that paper (I think I’ve seen it before, although there could be something else near-identical to it; I know I’ve read multiple[1] papers that claim to solve the Fermi Paradox and do not live up to their hype). TBH, I feel like it can be summed up as “well, there might be a Great Filter somewhere along the line, therefore no paradox”. I mean, no shit there’s probably a Great Filter somewhere, that’s the generally-proposed resolution that’s been going for decades now. The question is “what is the Filter?”. And saying “a Filter exists” doesn’t answer that question.
We’ve basically ruled out “Earthlike planets are rare”. “Abiogenesis is rare” is possible, but note that you need “no lithopanspermia” for that one to hold up since otherwise one abiogenesis event (the one that led to us and which is therefore P = 1, whether on Earth or not) is enough to seed much of the galaxy. “Intelligence is rare” is a possibility but not obviously-true, ditto “technology is rare”. Late filters (which, you will note, the authors assume to be a large possibility) appear somewhat less plausible but are not ruled out by any stretch. So yeah, it’s still a wide-open question even if there are some plausible answers.
- ^
The other one I recall, besides Grabby Aliens, was one saying that Earthly life is downstream of a lithopanspermia event so there’s no Fermi paradox; I don’t get it either.
- ^
I think the fate of the world depends on what’s been termed Total Stop. My Inside View on alignment of superhuman artificial neural nets being achieved is epsilon because it seems too similar to the general halting problem. My Outside View is like 3%, because I might be wrong.
Total Stop requires either that the US government wake up, or something way outside the Business As Usual assumption happen (most obviously WWIII). These are not sufficient, but you need one of them (and for obvious reasons I’d prefer the first, though I am more of a China hawk than I might otherwise be). This seems like a good lever to make USG wake up, and there’s an unusual degree of urgency from the METR report such that this is plausibly endgame.
Do they need to subpoena METR? I dunno if they need to (although METR is less likely to outright lie to Congress under oath than OpenAI employees and especially Sam Altman), but it seems unlikely in practice that a Congressional investigation wouldn’t.
I also just don’t think I consider this as much of a defection as you do. Keeping your own word is one thing; keeping other people’s word that you haven’t authorised them to give is quite another.