If that’s what Quinn (comment OP) is saying then I think it’s obviously wrong—people really do value the goods and services they access via the internet very highly. This leads me to believe that this is not what Quinn is saying.
What I (post author) am saying is people don’t apply even a tiny fraction of the vibes that come with that high value to their actual ISP (or, analogously, airline, electric company etc).
I think the lesson of social desirability bias is that valuable services having lower status than they “ought” to is the system working as intended.
Can you elaborate? I don’t understand your point because it’s too compressed. I feel like I need ~3 more sentences here to get it.
Yes, my claim is that “The Comcast Problem” is the reason for hatred, as opposed a feeling more like “I really value the service but am annoyed by the customer service and pricing, so overall meh.”
On a −5 to +5 scale, I’m saying they are often at like −5 when they really should be somewhere between −1 and +1.
Well, one additional factor the US has is that various veto points and power centers cycle on different time scales.
There would also be an incentive to introduce lots of meaningless elections between irrelevant (to you) alternatives in order to abstain and accrue more stored votes.
Could also be described as “There would also be an incentive to allow others to make decisions on issue that matter more to them than to you, in order to be more likely to get your way on an issue you care about.”
Re-phrased that way, it’s not clear to me that this is a bad thing. If they don’t care about those other issues either, then you won’t gain any stored votes on net relative to other voters.
Absolutely! I value your voice. But, and excuse me if this is a misread, your posts in this series read to me like you are still trying to convince yourself and/or him.
It reads like you are a sort of rationalist Martin Luther criticizing the Pope. But, like, there are already a lot of metaphorically-protestant rationalists.
I think I’m trying to make a different point than footnote 20?
It seems like you are taking me to be saying something like “You shouldn’t care what EY thinks about this Trans issue because “Everybody Knows” not to take his statements on this seriously”—that’s how I read FN20.
Whereas I think my point is much more general and really not specific to Trans at all—like why be so deeply invested in the contents of some one guy’s mind, at all? On any issue?
EY wrote some great (book-like objects). Inspiring, even. Worldview changing. But, like, whatever his opinions are today (on any issue), my view is mostly like, who cares? Either his arguments are convincing or they aren’t.
By analogy, suppose (counter factually) that I think that Barack Obama was the greatest president in history (he wasn’t, but he has to be alive for this analogy to work). Does that mean that I should decide what I think about today’s political and policy problems based on what Obama thinks? Such that if Obama was wrong about something, I should engage in an epic quest to Get Obama’s Attention and get him to admit he’s wrong? I mean, that would be ridiculous, right?
Maybe I just don’t get it because I’m not part of the Berkeley Community, I just read the writing. But my immediate reaction to this is like, why does Zack care so much about what Eliezer (2024) does or does not think? Or even whether, these days, he is or is not a fraud?
Like if you thought what he wrote in 2007 was great, just listen to that? Many (all?) authors who write great books have also written worse books. Maybe Zack’s opinion is falling a long way from wherever it was.
But perhaps he would be happier to adopt a more ecumenical non-Berkeley-ite stance, which I think has been common all along outside The Berkeley Community, and which is something like “Eliezer wrote some great stuff that was very influential on my thinking and that I still believe was very insightful, and I really appreciate that. I enjoy reading LW more than I think I’d enjoy the marginal alternative use of reading time, but I don’t go too far out of my way to pay attention to or care about what he’s up to these days.”—rather than assigning himself an Epic Quest to Win This Argument.
One thing to further ponder is the extent to which systematic or repeated boundary violations can effectively amount to a dissolution. Analogous examples:
Forcing someone to submit to multiple-times-daily injections, so far all of which have been harmless saline
Constantly stealing objects from someone’s house in a way that they don’t feel like they can meaningfully accumulate personal property
Entering a country with closed borders so frequently that its ability to enforce its immigration laws is effectively gummed up
(See this list of bar associations just for Massachusetts.)
Minor point but this is often misunderstood. These bar associations are essentially networking groups for lawyers. They are not required in order to practice. What’s required to practice is bar admission which is different. There’s also a federal bar admission, but that’s only two, not dozens.
In a game where you play a higher number of shorter games, you can ideally have a handicap that adjusts after every game. For example, in Super Smash Bros, if you turn handicap to “auto” then the stronger player starts with damage, which (in two player) goes up 10% every time they win, and down 10% every time they lose. It gets a little more complicated in 3+ player games, and I’m not sure the exact algorithm, but it works reasonably well. Maybe something to emulate in a game where handicaps can be reasonably granular?
Yeah, I mean I guess it depends on what you mean by photorealistic. That cat has three front legs.
My general principle here is a generalization of the foundations of tort law—if you do an act that causes harm, in a way that’s reasonably foreseeable, you are responsible for that. I don’t think there should be a special AI exception for that, and I especially don’t think there should be an open source exception to that. And I think it’s very common in law for legislatures or regulators to pick out a particular subset of reasonably-foreseeable harm to prohibit in advance rather than merely to punish/compensate afterwards.
I’m not sure what “human level” means in this context because it’s hard to directly compare given AI’s advantages in speed, replicability, and breath of background knowledge. I think it’s an empirical question whether any particular AI model is reasonably foreseeable to cause harm. And I think “enable any of [the listed] harms in a way that would be significantly more difficult to cause without access to the model” is an operationalization of foreseeability that makes sense in this context.
So with all that said, should it be illegal to effectively distribute amoral very cheap employees that it’s very easy to get to cause harm? Probably. If I ran an employment agency that publicly advertised “hey my employees are super smart and will do anything you tell them, even if it’s immoral or if it will help you commit crimes” then yeah I think I’d rightly have law enforcement sniffing around real quick.
Is it your view that there is a substantial list of capabilities it should be legal to freely distribute an AI model with, but which would rightly be illegal to hire a person to do?
This also would outlaw open source models at a fairly weak capabilities level.
That seems good, if those open source models would be used to enable any of the [listed] harms in a way that would be significantly more difficult to cause without access to [the open source] model. All those harms are pretty dang bad! Outside the context of AI, we go to great lengths to prevent them!
It sounds like some of those examples don’t meet “in a way that would be significantly more difficult to cause without access to a covered model”—already covered by the bill.
Ah, the bill answers this question!
(n) (1) “Hazardous capability” means the capability of a covered model to be used to enable any of the following harms in a way that would be significantly more difficult to cause without access to a covered model:
(A) The creation or use of a chemical, biological, radiological, or nuclear weapon in a manner that results in mass casualties.
(B) At least five hundred million dollars ($500,000,000) of damage through cyberattacks on critical infrastructure via a single incident or multiple related incidents.
(C) At least five hundred million dollars ($500,000,000) of damage by an artificial intelligence model that autonomously engages in conduct that would violate the Penal Code if undertaken by a human.
(D) Other threats to public safety and security that are of comparable severity to the harms described in paragraphs (A) to (C), inclusive.
And “critical harm” means that same list.
Sure, but at the same time it’s illegal to sell bazookas specifically because there is an unreasonable risk that a user may be able to use them to commit the obvious potential harms. So this is not some general tool-agnostic principle—it’s specific to the actual tool in question.
So in this metaphor one must determine, empirically, whether any given AI product is more like a chainsaw or a bazooka. Here, the bill proposes a way to make the categorization.
This bounty might be somewhat controversial here, and I have to be honest that it almost deterred me from posting this on LessWrong, offering a bounty for a “capabilities advance”
I think that this is not a “capabilities advance” in the parlance of around here—where you are looking for ways to elicit more capabilities of existing models, rather than looking for ways to create more powerful models.
You say this like it’s some kind of grudging acknowledgement, but it’s actually the entire point of the structure and a Big F’n Deal. Recall that there was less than 25 years between WW1 and WW2. It’s been almost 80 years without WW3, despite high tensions at various times. WW3 would have been catastrophic, and preventing it is a great accomplishment.