Last week I covered Anthropic’s relatively strong submission, and OpenAI’s toxic submission. This week I cover several other submissions, and do some follow-up on OpenAI’s entry.
Could you elaborate on what you _mean_ by "vice signalling"? It is fine to have new terms for shorthands, but I find this one unclear. I get the impression that it means some action to signal not cooperating with some group or some point of view, but I'm quite uncertain which group or point of view it is signalling dissent from. AI existential safety groups? AI prosaic safety concern groups? Woke groups? Anyone connected to the Democratic Party? Someone/something else?
Some researchers managed to accidentally create an AI model which was basically a mustache-twirling villain. For any given question, it would recommend the most evil action it could think of: murder, enslavement, elimination of the human race, etc.
"Vice signalling" means publicly advertising that the speaker is a villain, someone who does bad things on principle or for the lulz. And you want everyone to know that you are a villain.
This is largely independent of most political philosophies, and it can only take root where a large enough chunk of society admires villains because they break every moral rule.
"and it can only take root where a large enough chunk of society admires villains because they break every moral rule"
Umm... The moral rules aren't physical constants, they are a kind of political position, so if a large chunk of society disagrees with them, then they aren't some sort of generally agreed on position either. This sounds much more like dissent or pushback against control freak moralists (currently, that is usually the Woke, but if some _other_ group of moralists is being pushed back, it would help to be clearer on just who or just what position is being pushed against). Either we need that, or we need a photo of the mustache being twirled.
I encourage you to read about the model described in "On Emergent Misalignment." It's a pretty startling case of a model having what amounts to a global "good/evil vector", where training it to transgress against a single rule made it transgress against every moral rule. Deliberately writing insecure code apparently leads to advocating murder, at least for AI models. It's all correllated.
I am not a moral relativist. I believe that a large majority of people from the ancient Egyptians to the native Americans to modern liberals agree on a huge number of moral issues. Yes, there are some significant disagreements. For example, most ancient civilizations practiced slavery, but very few modern ones will openly support it. But murdering your neighbor is nearly universally frowned upon, and even the ancient Egyptians knew that being an exploitive boss made you a bad person.
"Vice signalling", in this framework, means deliberately advertising that you are bad person doing bad things.
Many Thanks! I did read Zvi's post about the "On Emergent Misalignment" paper. It was interesting, but my suspicion is that the "good/evil" vector might mostly be a Woke/caricature_of_antiWoke vector.
In general, I am extremely skeptical of ethical systems, seeing them frequently promoting batshit insane positions like Kant's or Singer's. I think that there is far less agreement on ethics than ethicists try to claim. Slavery is not the only big disagreement. Remember that classic civilizations also e.g. exposed unwanted infants. I'd love to hear what classic Spartans would have said re Singer's "drowning child" scenario. Singer wrongfully asserts universal agreement that does not exist.
The quote:
John Pressman: It's more like
"well the entire Trump administration seems to be based on vice signaling so".
is just flatly completely partisan. No, the half of the nation that voted for Trump is NOT vice signalling. The left needs to actually LISTEN to the people who disagree with them, who have their own concerns and grievances, and not to dismiss half the nation as mustache-twirling villains or, in Hillary Clinton's word, "deplorables".
As opposed to the right wing, particularly Trump? If you think it's just the left that has that problem, you are either not paying attention or just as blinded by partisanship as those you're sneering at.
Just because something/someone is popular doesn't make it ethical. An awful lot of people voted for Hitler after all. And then there are the "elections" in Russia and other authoritarian countries.
"Just because something/someone is popular doesn't make it ethical."
In a nutshell, I could not disagree more. "Ethics" are not constants of nature. To the extent that they mean anything at all, they mean the mores and practices of one's community. A MAGA administration following MAGA mores is _not_ vice signalling.
The term vice signaling is basically the opposite of virtue signaling: A way of showing how you are willing to go along with the ingroup, even when (especially when) it goes against what is right. More at Wikipedia, including a link to the 2019 article that originated the term: https://en.wikipedia.org/wiki/Virtue_signalling#Vice_signalling
Many Thanks! There is no such thing as "what is right", sitting in a vacuum. Going along with the mores of one's community is _virtue_ signalling within that community. That some other community is pissed at it does _not_ mean that the person within their community, be it MAGA or Woke, is a mustache-twirling villain. It just means that the two communities disagree about mores, which is a _very_ common situation.
It is indeed frustrating trying to deal with the If You Really X Then You'd Absurd Thing Y argument, even outside the realm of AI. Not just because it shows a profound failure of typical mind/ITT, but because pretty much every rhetorical tactic in the category "define the frame that your opponent can operate in" feels fundamentally illegitimate, Dark Arts-y. Yet refusing to accept that frame tends to get labeled by audiences as Dodging The Question. (And the purpose of such discourse is, of course, less to convince the actual interlocutor than to sway the opinions of everyone else reading. You know the rules, and so do I...) That's what makes these Miami Vice Signalling attempts scan so poorly, to me, that they've already ceded the field to the administration's preferred framing. You can absolutely write in an America, Fuck Yeah register without making the implicit subtext into explicit text and beating audiences over the head with it. Hell, I bet o3 could do it better...
Tangential, but wanted to note that I'm glad you took up the Username (explain the context): suggestion when quoting tweets. It really does make the format much easier to read.
One could say that this encapsulates the problem with AI safety in a nutshell. People, even smart technology-focused reporters, don’t “feel the AGI” because honestly, how could they? Normalcy bias is deeply ingrained, and the supremacy of human brains is baked into our understanding of the world in a thousand different ways.
Podcast episode for this post:
https://open.substack.com/pub/dwatvpodcast/p/more-on-various-ai-action-plans
What are the practical timelines on when we see concrete plans/legislation/EOs from the gov once these are submitted?
Could you elaborate on what you _mean_ by "vice signalling"? It is fine to have new terms for shorthands, but I find this one unclear. I get the impression that it means some action to signal not cooperating with some group or some point of view, but I'm quite uncertain which group or point of view it is signalling dissent from. AI existential safety groups? AI prosaic safety concern groups? Woke groups? Anyone connected to the Democratic Party? Someone/something else?
Zvi used the term earlier, here: https://thezvi.substack.com/p/on-emergent-misalignment
Some researchers managed to accidentally create an AI model which was basically a mustache-twirling villain. For any given question, it would recommend the most evil action it could think of: murder, enslavement, elimination of the human race, etc.
"Vice signalling" means publicly advertising that the speaker is a villain, someone who does bad things on principle or for the lulz. And you want everyone to know that you are a villain.
This is largely independent of most political philosophies, and it can only take root where a large enough chunk of society admires villains because they break every moral rule.
Many Thanks!
"and it can only take root where a large enough chunk of society admires villains because they break every moral rule"
Umm... The moral rules aren't physical constants, they are a kind of political position, so if a large chunk of society disagrees with them, then they aren't some sort of generally agreed on position either. This sounds much more like dissent or pushback against control freak moralists (currently, that is usually the Woke, but if some _other_ group of moralists is being pushed back, it would help to be clearer on just who or just what position is being pushed against). Either we need that, or we need a photo of the mustache being twirled.
I encourage you to read about the model described in "On Emergent Misalignment." It's a pretty startling case of a model having what amounts to a global "good/evil vector", where training it to transgress against a single rule made it transgress against every moral rule. Deliberately writing insecure code apparently leads to advocating murder, at least for AI models. It's all correllated.
I am not a moral relativist. I believe that a large majority of people from the ancient Egyptians to the native Americans to modern liberals agree on a huge number of moral issues. Yes, there are some significant disagreements. For example, most ancient civilizations practiced slavery, but very few modern ones will openly support it. But murdering your neighbor is nearly universally frowned upon, and even the ancient Egyptians knew that being an exploitive boss made you a bad person.
"Vice signalling", in this framework, means deliberately advertising that you are bad person doing bad things.
Many Thanks! I did read Zvi's post about the "On Emergent Misalignment" paper. It was interesting, but my suspicion is that the "good/evil" vector might mostly be a Woke/caricature_of_antiWoke vector.
In general, I am extremely skeptical of ethical systems, seeing them frequently promoting batshit insane positions like Kant's or Singer's. I think that there is far less agreement on ethics than ethicists try to claim. Slavery is not the only big disagreement. Remember that classic civilizations also e.g. exposed unwanted infants. I'd love to hear what classic Spartans would have said re Singer's "drowning child" scenario. Singer wrongfully asserts universal agreement that does not exist.
The quote:
John Pressman: It's more like
"well the entire Trump administration seems to be based on vice signaling so".
is just flatly completely partisan. No, the half of the nation that voted for Trump is NOT vice signalling. The left needs to actually LISTEN to the people who disagree with them, who have their own concerns and grievances, and not to dismiss half the nation as mustache-twirling villains or, in Hillary Clinton's word, "deplorables".
As opposed to the right wing, particularly Trump? If you think it's just the left that has that problem, you are either not paying attention or just as blinded by partisanship as those you're sneering at.
Just because something/someone is popular doesn't make it ethical. An awful lot of people voted for Hitler after all. And then there are the "elections" in Russia and other authoritarian countries.
Many Thanks!
"Just because something/someone is popular doesn't make it ethical."
In a nutshell, I could not disagree more. "Ethics" are not constants of nature. To the extent that they mean anything at all, they mean the mores and practices of one's community. A MAGA administration following MAGA mores is _not_ vice signalling.
The term vice signaling is basically the opposite of virtue signaling: A way of showing how you are willing to go along with the ingroup, even when (especially when) it goes against what is right. More at Wikipedia, including a link to the 2019 article that originated the term: https://en.wikipedia.org/wiki/Virtue_signalling#Vice_signalling
Many Thanks! There is no such thing as "what is right", sitting in a vacuum. Going along with the mores of one's community is _virtue_ signalling within that community. That some other community is pissed at it does _not_ mean that the person within their community, be it MAGA or Woke, is a mustache-twirling villain. It just means that the two communities disagree about mores, which is a _very_ common situation.
If a junior member of a gang murders an innocent person to prove their dedication to the gang, is that nonetheless still virtue signaling?
Many Thanks! From the perspective of the gang, yup. BTW, this scales to snipers in armed forces. https://en.wikipedia.org/wiki/Vasily_Zaitsev_(sniper) 265 kills
"Hero" and "Serial killer" can describe the same person
No ScarJo for this round?
It is indeed frustrating trying to deal with the If You Really X Then You'd Absurd Thing Y argument, even outside the realm of AI. Not just because it shows a profound failure of typical mind/ITT, but because pretty much every rhetorical tactic in the category "define the frame that your opponent can operate in" feels fundamentally illegitimate, Dark Arts-y. Yet refusing to accept that frame tends to get labeled by audiences as Dodging The Question. (And the purpose of such discourse is, of course, less to convince the actual interlocutor than to sway the opinions of everyone else reading. You know the rules, and so do I...) That's what makes these Miami Vice Signalling attempts scan so poorly, to me, that they've already ceded the field to the administration's preferred framing. You can absolutely write in an America, Fuck Yeah register without making the implicit subtext into explicit text and beating audiences over the head with it. Hell, I bet o3 could do it better...
Tangential, but wanted to note that I'm glad you took up the Username (explain the context): suggestion when quoting tweets. It really does make the format much easier to read.
Shameless plug of what we wrote at Ai2 on supporting open models, let us know if you have thoughts: https://allenai.org/blog/OSTP
Just the two concerns that Zvi has raised regarding open models in the past:
1. Is it a concern that open models are basically giving away your secrets to competitors and rivals?
2. What do you do if an open model turns out to be dangerous after you've released it?
One could say that this encapsulates the problem with AI safety in a nutshell. People, even smart technology-focused reporters, don’t “feel the AGI” because honestly, how could they? Normalcy bias is deeply ingrained, and the supremacy of human brains is baked into our understanding of the world in a thousand different ways.
SP: "Cate Blanchett" repeated...
"Cate Blanchett, Phoebe Waller-Bridge, Bette Midler, Cate Blanchett"