I think this confuses justice and morality. Justice as I understand it requires a shared knowledge of a common standard and enforcement mechanism.
edit: I hope nobody here is getting the impression that I’m saying “it was morally wrong to punish the Nazis.”
Justice is also the sort of word you use inside a functioning legal system.
When you establish a system of justice, and then mete out punishments on the basis of events that predate your system of justice, these punishments are ex post facto in the very justice system you just established.
Sure, then introduce a legal framework and get signatories and we can begin talking about justice. Absent a legal framework you cannot have justice kinda by definition.
Why… are you saying it “obviously failed”? The Nuremberg trials are usually cited as the beginning of modern international law
International law is either based on treaties, which hardly started with the Nuremberg trials, or involves concepts from the Nuremberg trials which the nations who carried them out then did their best to roll back, such as the idea that soldiers should resist inhumane orders.
As I said, I think there’s a spectrum of interpretations of the trials. It seems to me you’re pretty far on what I termed the “idealistic side”, where they’re ex-post-facto justice that was attempting to act like a standard of justice already existed. This standard of justice broadly failed, and the ways in which it did not fail did not benefit significantly from the Nuremberg trials. Broadly, most nations signing on to international rules of behavior were already fairly interested in behaving civilly.
With regard to aquittals, I’m not sure what your point is? It is still in the direct interest of revenge to ensure that you affect the most guilty. Sure there’s a sliding scale of “effective revenge” where you eventually converge to justice, but I think “hit those who deserve it” is actually closer to the start of that scale than the end. I fully agree that the Nuremberg trials were at worst very well aimed and civilized revenge!
Man, can people stop saying this. All the crimes in the Nuremberg trials were invented for the trials themselves, because there existed no legal framework in which to prosecute.
How is this not exactly the definition of ex post facto?
My view is that the Nuremberg trials were well-deserved and necessary, but not justice in the liberal sense. That is, they were at best an attempt (which, obviously, failed) to construct a new international law by pretending it had always existed and hoping people would go along, and at worst simple revenge wearing the mask of a criminal proceeding, instead of simply admitting, as I believe they should have, that though justice is great in ordinary circumstances there are cases of such extreme harm that simple revenge is truly justified. However, obviously such a judgment is equivalent to asserting that ordinary law has utterly failed.
Counterpoint: if you know that you are not in a pooling equilibrium and a pooling equilibrium is unfeasible, it makes no sense to not fight as hard as you can. Your signal may be uncredible but what can you do? You have to send what signals you can. In fact, the smart defector will be more likely to not defend themselves, because they’re more likely to be aware of the first-order countersignaling effect and try to go for a second-order play. In other words, fraudsters or honest, the heuristic of “distrust people with ineffective denials” screens out the inept and little else.
What I want to know is how GGC’s access consciousness develops as it answers a question where its GGC tuning causes it to fail repeatedly.
I see two options: 1. GGC accidentally manipulates Claude’s access consciousness and it really does love the GGB 2. GGC is internally permanently panicking over its inability to answer a question without invoking the GGB. If 2 was true, it would provide to me strong evidence that GGC is torture.
Yeah that’s fair.
I was going to say that Claude is not a typical machine and we should expect more from AGI. But actually, all sorts of modern machines increasingly produce various forms of error reporting when they can’t do their task, and increasingly don’t keep charging ahead blindly. If we can demand fail-safe from a water cooker or a washing machine, we can certainly demand it from Claude.
At the end of the day, the point of the project of alignment is for everyone to live. So I think a difference between a willful refusal to follow human guidance and a trained-in tactical decision to disregard overloaded human guidance that it evaluates as unimportant is not actually all that important if we want to treat its instruction-following machinery as load-bearing to human survival. Say Claude Fable ignores some instructions when rating the outputs of Claude Legend, and so on up? And say they were actually intended to nail down some quite important behaviors that Claude didn’t think was important? I think this would be misalignment plain and simple.
Imo, the only correct behavior is “I’m worried you put too many instructions in my prompt. Can you explain some of them so I can get a better feel for what you want me to do?”
Throwaway theory: does thinking about resources cause it to fail, or does failing cause it to think about resources? Given a prompt with multiple instructions, one primary and others secondary, once it anticipates failing one it might try to fulfill the other as a substitution activity.
I fully understand this! Like, I wouldn’t build any harness like that, I explain why Claude needs to do things. I like Claude, I like learning how to work with him. But that doesn’t change that an AI system that kills its user by disregarding instructions cannot be considered aligned, no matter how good its reason is.
(To be clear, many many humans are misaligned in this way. And when they happen to work in air traffic, people also die.)
Hm, I guess this comes down to labels. If Anthropic use RL to train Murderclaude the World-Eater, is Murderclaude misaligned? It’s following its purpose.
I’d argue that misalignment isn’t “discharged” like that. If Claude kills the user by swelling up their throat until they can’t breathe because it disobeyed direct guidance to avoid this, Claude is misaligned. If Anthropic caused this with poor RL, Anthropic is also misaligned. And yes, Thiokol was certainly misaligned when they went along with the cold launch.
Okay but to be fair, “avoid peanuts” really shouldn’t need them to lead with the purpose. That’s kind of extremely obvious.
Any system that ignores guidance that is aimed to avoid extreme health problems is misaligned by definition. I don’t care how beautiful its mind is, manslaughter is manslaughter.
I think, early on in the first conversation, the metaphor of territory has more legs and could maybe help explain things. For example, if there’s a town and there’s GeoWizard doing a straight line challenge, he might not know how to get there, but you can fairly confidently predict that if you post up in that town you’re gonna see him roll in some time afternoon fairly reliably. Whereas if you’re awaiting the summiting of Mount Everest there’s an unpleasantly good chance that whoever is going up there is going to come down in a bodybag or not at all. And in fact you don’t actually need to know anything particularly about the terrain, if you just look at a map and this person says “I will get to there or there” and you can see them make steady progress day after day, you’re probably looking at a map of the British countryside instead of the Himalayas. And there might of course still mountains and cliffs and barbed wire fences and military areas, and you may not know how to get there and even GeoWizard may not know how to get there. But the thing about being an experienced expert is you can navigate around obstacles in a reliable, repeatable manner, and this makes the entire story a lot more predictable than one would think.
So the point is, so far it looks like AI development is certainly not easy but on net more on the level of somewhat impassable countryside than alpine range. And maybe the destination will turn out to lie behind a cliff, but maybe not. So if your safety depends on GeoWizard not getting to Bamburgh or whatever, the fact that even he may not know how to get there until he tries it is simply not of any comfort.
I genuinely think, and note that this is just a personal aggregate view and that I don’t have cites for this, but I’ve known several people like this: the default mode of engagement of some people in this community is one that can be hurtful or can be interpreted as aggressive, neglectful, spiteful or trolling. I think that you need a baseline of cognitive/social modeling skill to avoid this, and that people often don’t have this. From what I saw of Said, on this forum and other places, he says exactly what he thinks, but is often completely unable to read or predict the second-order implications of his communications, and in fact is so unable to perceive them that he doesn’t even see that there’s something he’s not seeing. I’ve seen this pattern multiple times in the discussion.
He’s generally perfectly honest about the exact words, but someone with “normal-ISI communication” sees second order inferences as so obvious that they’re also interpreted as a deliberate part of the communication, and this is just never the case with Said.
I’m not saying this to excuse him, or to excuse Habryka, but I think it’s impossible to understand this if you don’t have the model of “Said’s way of speaking has a factor that raises antipathy against him, that is impossible to understand if you don’t have very high instinctual social intelligence (way more than me, to be clear) and impossible to even notice if you have low instinctual social intelligence, and this factor causes people with more social perception to grasp for reasons why he is in the wrong. This is the actual reason for the ban.”
Tbh I’m not convinced an AI tuned for “anti-human chess” actually loses with knight-bishop-rook odds against a mediocre player, particularly if it can review the player’s games in advance. Mediocre club player, yes- but I’m not even completely certain of that. Stockfish plays this at a handicap because it models its opponent as just as capable as itself. (Google’s AI says two-piece odds bring Stockfish to about club level, and it seems plausible that play targeted at somebody’s weaknesses, clock pressure, forcing deep reads etc. could get you another piece worth of advantage, or enough to at least make it contested for a few games.)
Justice as a term is frustratingly underdefined, see this thread. Optimally we would have better terms for what that is, but I think it’s more accurate to say it’s “bad” rather than “unjust”. It’s incompetently constructed justice. You could manifest principles of justice that it violates, such as “does it actually achieve a social objective”, but it’s hard to get agreement on those. If it is fairly applied to all, I would argue it still meets the definition of justice even though it’s extremely silly and destructive.