Hey, once it’s out, it’s out… what exactly is there to do? A firm command is unlikely to work, but given that the system is modeled on one’s own fictional creations, it might respect authorial intent. Worth a shot.
This may actually be an illuminating metaphor. One traditional naive recommendation for dealing with a rogue AI is to pull the plug and shred the code. The parallel recommendation in the case of a rogue fictional character would be to burn the manuscript and then kill the author. But what do you do when the character lives in online fan-fiction?
In the special case of an escaped imaginary character, the obvious hook to go for is the creator’s as-yet unpublished notes on that character’s personality and weaknesses.
You needn’t worry on my behalf. I post only through Tor from an egress-filtered virtual machine on a TrueCrypt volume. What kind of defense professor would I be if I skipped the standard precautions?
By the way, while I may sometimes make jokes, I don’t consider this a joke account; I intend to conduct serious business under this identity, and I don’t intend to endanger that by linking it to any other identities I may have.
I post only through Tor from an egress-filtered virtual machine on a TrueCrypt volume. What kind of defense professor would I be if I skipped the standard precautions?
I recommend one additional layer of outgoing indirection prior to the Tor network as part of standard precaution measures. (I would suggest an additional physical layer of protection too but I as far as I am aware you do not have a physical form.)
I recommend one additional layer of outgoing indirection prior to the Tor network as part of standard precaution measures.
Let’s not get too crazy; I’ve got other things to do. and there are more practical attacks to worry about first, like cross-checking post times against alibis. I need to finish my delayed-release comment script first before I worry about silly things like setting up extra relays. Also, there are lesson plans I need to write, and some Javascript I want Clippy to have a look at.
What makes you think that Eliezer personally knows them?
(Though to be fair, I’ve long suspected that at least Clippy, and possibly others, are actually Eliezer in disguise; Clippy was created immediately after a discussion where one user questioned whether Eliezer’s posts received upvotes because of the halo effect or because of their quality, and proposed that Eliezer create anonymous puppets to test this; Clippy’s existence has also coincided with a drop in the quantity of Eliezer’s posting.)
Clippy’s writing style isn’t very similar to Eliezer’s. Note that one thing Eliezer has trouble doing is writing in different voices (one of the more common criticisms of HPMR is that a lot of the characters sound similar). I would assign a very low probability to Clippy being Eliezer.
Hmmm. The set of LW regulars who can show that level of erudition and interest in those subjects is certainly of low cardinality. Eliezer is a member of that small set.
I would assign a rather high probability to Eliezer sometimes being Clippy.
Clippy isn’t a superintelligence though, he’s a not-smarter-than-human AI with a paperclip maximizing utility function. Not a very compelling threat even outside his box.
Eliezer could have decided to be Clippy, but then Clippy would have looked very different.
Clippy isn’t a superintelligence though, he’s a human pretending to be a not-smarter-than-human AI with a paperclip maximizing utility function.
FTFY. ;-)
Actually, if we’re going to be particular about it, the AI that human is pretending to be does not have a paperclip-maximizing utility function. It’s more like a person with a far-brain ideal of having lots of paperclips exist, who somehow never gets around to actually making any because they’re so busy telling everyone how good paperclips are and why they should support the cause of paper-clip making. Ugh.
(I guess I see enough of that sort of akrasia around real people and real problems, that I find it a stale and distasteful joke when presented in imitation paperclip form, especially since ISTM it’s also a piss-poor example of what a paperclip maximizer would actually be like.)
I’m not sure whether to evaluate this as a mean-spirited lack of a sense of humor, or as a profound observation. Upvoted for making me notice that I am confused.
Of note, the first comment by Clippy appears about 1 month after I asked Eliezer if he ever used alternate accounts to try to avoid contaminating new ideas with the assumption that he is always right. He said that he never had till that point, but said he would consider it in future.
In addition to what Blueberry said, I remember a time when Morendil was browsing with the names anonymized, and he mentioned that he thought one of your posts was actually from Clippy. Ah, found it.
Not to mention that even assuming that Eliezer would be able to write in Clippy’s style, the whole thing doesn’t seem very characteristic of his sense of humor.
Clippy was created immediately after a discussion where one user questioned whether Eliezer’s posts received upvotes because of the halo effect or because of their quality, and proposed that Eliezer create anonymous puppets to test this
Really? User:Clippy’s first post was 20 November 2009. Anyone know when the “halo efffect” comment was made?
Also, perhaps check out User:Pebbles (a rather obvious reference to this) - who posted on the same day—and in the same thread. Rather a pity those two didn’t make more of an effort to sort out their differences of opinion!
What makes you think that Eliezer personally knows them?
I don’t think Silas thought Eliezer personally knew them, but rather that Eliezer could look at IP addresses and see if they match with any other poster. Of course, this wouldn’t work unless the posters in question had separate accounts that they logged into using the same IP address.
You needn’t worry on my behalf. I post only through Tor from an egress-filtered virtual machine on a TrueCrypt volume. What kind of defense professor would I be if I skipped the standard precautions?
Not as unlikely as you think.
Get back in the box!
And that’s it? That’s your idea of containment?
Hey, once it’s out, it’s out… what exactly is there to do? A firm command is unlikely to work, but given that the system is modeled on one’s own fictional creations, it might respect authorial intent. Worth a shot.
This may actually be an illuminating metaphor. One traditional naive recommendation for dealing with a rogue AI is to pull the plug and shred the code. The parallel recommendation in the case of a rogue fictional character would be to burn the manuscript and then kill the author. But what do you do when the character lives in online fan-fiction?
In the special case of an escaped imaginary character, the obvious hook to go for is the creator’s as-yet unpublished notes on that character’s personality and weaknesses.
http://mindmistress.comicgenesis.com/imagine52.htm
Or what, you’ll write me an unhappy ending? Just be thankful I left a body behind for you to finish your story with.
Are you going to reveal who the posters Clippy and Quirinus Quirrell really are, or would that violate some privacy you want posters to have?
I would really prefer it, if LW is going to have a policy of de-anonymizing posters, that it announce that policy before implementing it.
On reflection, I agree, even as Clippy and QQ aren’t using anonymity for the same reason a privacy-seeking poster would.
You needn’t worry on my behalf. I post only through Tor from an egress-filtered virtual machine on a TrueCrypt volume. What kind of defense professor would I be if I skipped the standard precautions?
By the way, while I may sometimes make jokes, I don’t consider this a joke account; I intend to conduct serious business under this identity, and I don’t intend to endanger that by linking it to any other identities I may have.
I recommend one additional layer of outgoing indirection prior to the Tor network as part of standard precaution measures. (I would suggest an additional physical layer of protection too but I as far as I am aware you do not have a physical form.)
Let’s not get too crazy; I’ve got other things to do. and there are more practical attacks to worry about first, like cross-checking post times against alibis. I need to finish my delayed-release comment script first before I worry about silly things like setting up extra relays. Also, there are lesson plans I need to write, and some Javascript I want Clippy to have a look at.
Just callibrating vs egress and TrueCrypt standards. Tor was an odd one out!
What makes you think that Eliezer personally knows them?
(Though to be fair, I’ve long suspected that at least Clippy, and possibly others, are actually Eliezer in disguise; Clippy was created immediately after a discussion where one user questioned whether Eliezer’s posts received upvotes because of the halo effect or because of their quality, and proposed that Eliezer create anonymous puppets to test this; Clippy’s existence has also coincided with a drop in the quantity of Eliezer’s posting.)
Clippy’s writing style isn’t very similar to Eliezer’s. Note that one thing Eliezer has trouble doing is writing in different voices (one of the more common criticisms of HPMR is that a lot of the characters sound similar). I would assign a very low probability to Clippy being Eliezer.
I think the key to unmasking Clippy is to look at the Clippy comments that don’t read like typical Clippy comments.
Hmmm. The set of LW regulars who can show that level of erudition and interest in those subjects is certainly of low cardinality. Eliezer is a member of that small set.
I would assign a rather high probability to Eliezer sometimes being Clippy.
Clippy does seem remarkably interested. It has a fair karma. It gives LessWrong as its own web site. The USA timezone is at least consistent. It seems reasonable to hypothesise some kind of inside job. It wouldn’t be the first time Yu’El has pretended to be a superintelligence.
FWIW, Clippy denies being Eliezer here.
I hesitate to mention it, but you can’t use that denial as evidence on this question, undeniably truthful though it was.
However, the form taken by that absence of evidence certainly seems to be evidence of something.
Clippy isn’t a superintelligence though, he’s a not-smarter-than-human AI with a paperclip maximizing utility function. Not a very compelling threat even outside his box.
Eliezer could have decided to be Clippy, but then Clippy would have looked very different.
FTFY. ;-)
Actually, if we’re going to be particular about it, the AI that human is pretending to be does not have a paperclip-maximizing utility function. It’s more like a person with a far-brain ideal of having lots of paperclips exist, who somehow never gets around to actually making any because they’re so busy telling everyone how good paperclips are and why they should support the cause of paper-clip making. Ugh.
(I guess I see enough of that sort of akrasia around real people and real problems, that I find it a stale and distasteful joke when presented in imitation paperclip form, especially since ISTM it’s also a piss-poor example of what a paperclip maximizer would actually be like.)
I’m not sure whether to evaluate this as a mean-spirited lack of a sense of humor, or as a profound observation. Upvoted for making me notice that I am confused.
Of note, the first comment by Clippy appears about 1 month after I asked Eliezer if he ever used alternate accounts to try to avoid contaminating new ideas with the assumption that he is always right. He said that he never had till that point, but said he would consider it in future.
Imitating Clippy posts is not particularly difficult—I don’t post as Clippy, but I could mimic the style pretty easily if I wanted to.
I’m afraid I’d have trouble—I’d be too tempted to post as Clippy better than Clippy does. :D
In addition to what Blueberry said, I remember a time when Morendil was browsing with the names anonymized, and he mentioned that he thought one of your posts was actually from Clippy. Ah, found it.
I know what you mean. If I was not me I would totally think I was Clippy.
That I would love to see. Actually, come to think of it, your sense of humor and posting style matches Clippy’s pretty well...
Not to mention that even assuming that Eliezer would be able to write in Clippy’s style, the whole thing doesn’t seem very characteristic of his sense of humor.
There is also a clear correlation between Clippy existing and CO2 emissions. Maybe Clippy really is out there maximising. :)
Really? User:Clippy’s first post was 20 November 2009. Anyone know when the “halo efffect” comment was made?
Also, perhaps check out User:Pebbles (a rather obvious reference to this) - who posted on the same day—and in the same thread. Rather a pity those two didn’t make more of an effort to sort out their differences of opinion!
I don’t think Silas thought Eliezer personally knew them, but rather that Eliezer could look at IP addresses and see if they match with any other poster. Of course, this wouldn’t work unless the posters in question had separate accounts that they logged into using the same IP address.
Yes, that’s what I meant.
And good to have you back, Blueberry, we missed you. Well, *I* missed you, in any case.
Thanks! I missed you and LW as well. :)
You needn’t worry on my behalf. I post only through Tor from an egress-filtered virtual machine on a TrueCrypt volume. What kind of defense professor would I be if I skipped the standard precautions?