Humans can't reliably precommit.

Humans don't follow any decision theory consistently. They sometimes give in to blackmail, and at other times resist blackmail. If you convinced a bunch of people to take acausal blackmail seriously, presumably some subset would give in and some subset would resist, since that's what we see in ordinary blackmail situations. What would be interesting is if (a) there were some applicable reasoning norm that forced us to give in to acausal blackmail on pain of irrationality, or (b) there were some known human irrationality that made us inevitably susceptible to acausal blackmail. But I don't think Roko gave a good argument for either of those claims.

From my last comment: "there are probably some decision theories that let agents acausally blackmail each other". But if humans frequently make use of heuristics like 'punish blackmailers' and 'never give in to blackmailers', and if normative decision theory says they're right to do so, there's less practical import to 'blackmailable agents are possible'.

This implies that the future AI uses a decision theory that two-boxes in Newcomb's problem, contradicting the premise that it one-boxes.

No it doesn't. If you model Newcomb's problem as a Prisoner's Dilemma, then one-boxing maps on to cooperating and two-boxing maps on to defecting. For Omega, cooperating means 'I put money in both boxes' and defecting means 'I put money in just one box'. TDT recognizes that the only two options are mutual cooperation or mutual defection, so TDT cooperates.

Blackmail works analogously. Perhaps the blackmailer has five demands. For the blackmailee, full cooperation means 'giving in to all five demands'; full defection means 'rejecting all five demands'; and there are also intermediary levels (e.g., giving in to two demands while rejecting the other three), with the blackmailee prefer to do as little as possible.

For the blackmailer, full cooperation means 'expending resources to punish the blackmailee in proportion to how many of my demands were met'. Full defection means 'expending no resources to punish the blackmailee even if some demands aren't met'. In other words, since harming past agents is costly, a blackmailer's favorite scenario is always 'the blackmailee, fearing punishment, gives in to most or all of my demands; but I don't bother punishing them regardless of how many of my demands they ignored'. We could say that full defection doesn't even bother to check how many of the demands were met, except insofar as this is useful for other goals.

The blackmailer wants to look as scary as possible (to get the blackmailee to cooperate) and then defect at the last moment anyway (by not following through on the threat), if at all possible. In terms of Newcomb's problem, this is the same as preferring to trick Omega into thinking you'll one-box, and then two-boxing anyway. We usually construct Newcomb's problem in such a way that this is impossible; therefore TDT cooperates. But in the real world mutual cooperation of this sort is difficult to engineer, which makes fully credible acausal blackmail at least as difficult.

This implies that the future AI will have a deontological rule that says "Don't blackmail" somehow hard-coded in it, contradicting the premise that it will be an utilitarian.

I think you misunderstood point 3. 3 is a follow-up to 2: humans and AI systems alike have incentives to discourage blackmail, which increases the likelihood that blackmail is a self-defeating strategy.

Shut up and multiply.

Eliezer has endorsed the claim "two independent occurrences of a harm (not to the same person, not interacting with each other) are exactly twice as bad as one". This doesn't tell us how bad the act of blackmail itself is, it doesn't tell us how faithfully we should implement that idea in autonomous AI systems, and it doesn't tell us how likely it is that a superintelligent AI would find itself forced into this particular moral dilemma.

Since Eliezer asserts a CEV-based agent wouldn't blackmail humans, the next step in shoring up Roko's argument would be to do more to connect the dots from "two independent occurrences of a harm (not to the same person, not interacting with each other) are exactly twice as bad as one" to a real-world worry about AI systems actually blackmailing people conditional on claims (a) and (c). 'I find it scary to think a superintelligent AI might follow the kind of reasoning that can ever privilege torture over dust specks' is not the same thing as 'I'm scared a superintelligent AI will actually torture people because this will in fact be the best way to prevent a superastronomically large number of dust specks from ending up in people's eyes', so Roko's particular argument has a high evidential burden.

A few misconceptions surrounding Roko's basilisk

by Rob Bensinger 5 min read5th Oct 2015130 comments

58


There's a new LWW page on the Roko's basilisk thought experiment, discussing both Roko's original post and the fallout that came out of Eliezer Yudkowsky banning the topic on Less Wrong discussion threads. The wiki page, I hope, will reduce how much people have to rely on speculation or reconstruction to make sense of the arguments.

While I'm on this topic, I want to highlight points that I see omitted or misunderstood in some online discussions of Roko's basilisk. The first point that people writing about Roko's post often neglect is:

  • Roko's arguments were originally posted to Less Wrong, but they weren't generally accepted by other Less Wrong users.

Less Wrong is a community blog, and anyone who has a few karma points can post their own content here. Having your post show up on Less Wrong doesn't require that anyone else endorse it. Roko's basic points were promptly rejected by other commenters on Less Wrong, and as ideas not much seems to have come of them. People who bring up the basilisk on other sites don't seem to be super interested in the specific claims Roko made either; discussions tend to gravitate toward various older ideas that Roko cited (e.g., timeless decision theory (TDT) and coherent extrapolated volition (CEV)) or toward Eliezer's controversial moderation action.

In July 2014, David Auerbach wrote a Slate piece criticizing Less Wrong users and describing them as "freaked out by Roko's Basilisk." Auerbach wrote, "Believing in Roko’s Basilisk may simply be a 'referendum on autism'" — which I take to mean he thinks a significant number of Less Wrong users accept Roko’s reasoning, and they do so because they’re autistic (!). But the Auerbach piece glosses over the question of how many Less Wrong users (if any) in fact believe in Roko’s basilisk. Which seems somewhat relevant to his argument...?

The idea that Roko's thought experiment holds sway over some community or subculture seems to be part of a mythology that’s grown out of attempts to reconstruct the original chain of events; and a big part of the blame for that mythology's existence lies on Less Wrong's moderation policies. Because the discussion topic was banned for several years, Less Wrong users themselves had little opportunity to explain their views or address misconceptions. A stew of rumors and partly-understood forum logs then congealed into the attempts by people on RationalWiki, Slate, etc. to make sense of what had happened.

I gather that the main reason people thought Less Wrong users were "freaked out" about Roko's argument was that Eliezer deleted Roko's post and banned further discussion of the topic. Eliezer has since sketched out his thought process on Reddit:

When Roko posted about the Basilisk, I very foolishly yelled at him, called him an idiot, and then deleted the post. [...] Why I yelled at Roko: Because I was caught flatfooted in surprise, because I was indignant to the point of genuine emotional shock, at the concept that somebody who thought they'd invented a brilliant idea that would cause future AIs to torture people who had the thought, had promptly posted it to the public Internet. In the course of yelling at Roko to explain why this was a bad thing, I made the further error---keeping in mind that I had absolutely no idea that any of this would ever blow up the way it did, if I had I would obviously have kept my fingers quiescent---of not making it absolutely clear using lengthy disclaimers that my yelling did not mean that I believed Roko was right about CEV-based agents [= Eliezer’s early model of indirectly normative agents that reason with ideal aggregated preferences] torturing people who had heard about Roko's idea. [...] What I considered to be obvious common sense was that you did not spread potential information hazards because it would be a crappy thing to do to someone. The problem wasn't Roko's post itself, about CEV, being correct.

This, obviously, was a bad strategy on Eliezer's part. Looking at the options in hindsight: To the extent it seemed plausible that Roko's argument could be modified and repaired, Eliezer shouldn't have used Roko's post as a teaching moment and loudly chastised him on a public discussion thread. To the extent this didn't seem plausible (or ceased to seem plausible after a bit more analysis), continuing to ban the topic was a (demonstrably) ineffective way to communicate the general importance of handling real information hazards with care.


On that note, point number two:

  • Roko's argument wasn’t an attempt to get people to donate to Friendly AI (FAI) research. In fact, the opposite is true.

Roko's original argument was not 'the AI agent will torture you if you don't donate, therefore you should help build such an agent'; his argument was 'the AI agent will torture you if you don't donate, therefore we should avoid ever building such an agent.' As Gerard noted in the ensuing discussion thread, threats of torture "would motivate people to form a bloodthirsty pitchfork-wielding mob storming the gates of SIAI [= MIRI] rather than contribute more money." To which Roko replied: "Right, and I am on the side of the mob with pitchforks. I think it would be a good idea to change the current proposed FAI content from CEV to something that can't use negative incentives on x-risk reducers."

Roko saw his own argument as a strike against building the kind of software agent Eliezer had in mind. Other Less Wrong users, meanwhile, rejected Roko's argument both as a reason to oppose AI safety efforts and as a reason to support AI safety efforts.

Roko's argument was fairly dense, and it continued into the discussion thread. I’m guessing that this (in combination with the temptation to round off weird ideas to the nearest religious trope, plus misunderstanding #1 above) is why RationalWiki's version of Roko’s basilisk gets introduced as

a futurist version of Pascal’s wager; an argument used to try and suggest people should subscribe to particular singularitarian ideas, or even donate money to them, by weighing up the prospect of punishment versus reward.

If I'm correctly reconstructing the sequence of events: Sites like RationalWiki report in the passive voice that the basilisk is "an argument used" for this purpose, yet no examples ever get cited of someone actually using Roko’s argument in this way. Via citogenesis, the claim then gets incorporated into other sites' reporting.

(E.g., in Outer Places: "Roko is claiming that we should all be working to appease an omnipotent AI, even though we have no idea if it will ever exist, simply because the consequences of defying it would be so great." Or in Business Insider: "So, the moral of this story: You better help the robots make the world a better place, because if the robots find out you didn’t help make the world a better place, then they’re going to kill you for preventing them from making the world a better place.")

In terms of argument structure, the confusion is equating the conditional statement 'P implies Q' with the argument 'P; therefore Q.' Someone asserting the conditional isn’t necessarily arguing for Q; they may be arguing against P (based on the premise that Q is false), or they may be agnostic between those two possibilities. And misreporting about which argument was made (or who made it) is kind of a big deal in this case: 'Bob used a bad philosophy argument to try to extort money from people' is a much more serious charge than 'Bob owns a blog where someone once posted a bad philosophy argument.'


Lastly:

  • "Formally speaking, what is correct decision-making?" is an important open question in philosophy and computer science, and formalizing precommitment is an important part of that question.

Moving past Roko's argument itself, a number of discussions of this topic risk misrepresenting the debate's genre. Articles on Slate and RationalWiki strike an informal tone, and that tone can be useful for getting people thinking about interesting science/philosophy debates. On the other hand, if you're going to dismiss a question as unimportant or weird, it's important not to give the impression that working decision theorists are similarly dismissive.

What if your devastating take-down of string theory is intended for consumption by people who have never heard of 'string theory' before? Even if you're sure string theory is hogwash, then, you should be wary of giving the impression that the only people discussing string theory are the commenters on a recreational physics forum. Good reporting by non-professionals, whether or not they take an editorial stance on the topic, should make it obvious that there's academic disagreement about which approach to Newcomblike problems is the right one. The same holds for disagreement about topics like long-term AI risk or machine ethics.

If Roko's original post is of any pedagogical use, it's as an unsuccessful but imaginative stab at drawing out the diverging consequences of our current theories of rationality and goal-directed behavior. Good resources for these issues (both for discussion on Less Wrong and elsewhere) include:

The Roko's basilisk ban isn't in effect anymore, so you're welcome to direct people here (or to the Roko's basilisk wiki page, which also briefly introduces the relevant issues in decision theory) if they ask about it. Particularly low-quality discussions can still get deleted (or politely discouraged), though, at moderators' discretion. If anything here was unclear, you can ask more questions in the comments below.

58