Hi synx, Recently I've also been thinking about different kinds of trust, and I really like your idea of associating trust with a category. However, I have a few concerns.
First, are we considering the common case? It's certainly possible that someone might write wise words in category A and nonsense in category B, but it seems to me that the urgent problem in newsgroup-like systems is filtering out automated spam. Anything that makes that problem harder while making it easier to make fine distinctions about humans is probably a step in the wrong direction - I might prefer a blunt instrument that says "never show me messages from this identity in any category" to a sharp one that forces me to mark down the same spammer in 100 categories, just in case they said something useful in one of them. Second, can we solve the problem of carrying over trust from one category to another without an a priori hierarchy of categories? Maybe we could use something like collaborative filtering to detect similar categories - to the extent that my explicit trust ratings in category A agree with my ratings in category B, use my ratings from A to assign initial trust values to identities in category B, and vice versa? But to the extent that such an approach reduces the number of explicit trust ratings I need to make, it also reduces the information available for detecting similarity between categories. Can we solve that problem without feeding the estimated trust ratings back into the similarity detector, which would seem likely to pull certain categories ever-closer together? Third, you mention the meta-category "recommending trustworthy people". Is this actually a single category, or does the assumption that trust is category-specific mean that I need to know who's good at recommending trustworthy people in category A, category B, etc? Cheers, Michael synx wrote: > Trust networks are difficult to get working right, especially in a > decentralized fashion. Trying to figure out what exactly I mean by > "trust network" is hard in of itself, even for me! But here's my best > shot at what I think we would want in a trust network, and what would > make it most likely to work. > > What I want to do is make a trust network that gets separated into > semantic categories or "actions". For instance a man who brought me > flowers a thousand times might not be a man I trust with my life. But a > man who saved my life I would be inclined to trust more to do so a > second time. Contrarily, a man who saved my life, I would still have no > reason to trust him to be a reliable source of flowers. So trust itself > is a categorical sort of thing. It depends not just on how much they did > for you in the past, but on how much of what they did for you in the > past. If a man murdered your friend with an axe, you would have very > good reason to trust that this man will now murder the rest of you, and > hopefully take measures to take his axe away. You certainly would want > to preserve that trust relationship (to avoid being axe murdered) and > yet you don't want it to spill over into your other trust relationships, > as an axe murderer may be a very untrustworthy person with loaning money to. > > I think there would have to be therefore some kind of semantic language > behind any trust network. Each person would have a list, not just of who > they trust, but who they trust to perform what action. If the action is > beneficial, such as someone repaying your loan, or if the action is > malignant, such as someone skipping out on payments, it's still a matter > of trust. You're trying to predict how they will act in the future, and > the more reliable your predictions can be, the less you get fooled by > scams and con artists. > > Newsgroups have one such language. Each group is named by hierarchical > topic. A poster in one group might act trollish and brusque, but while > posting in another group would act proper and modest, and they might use > the same PGP key for both groups. That happens a lot actually, that the > environment of the group determines what attitude a poster will bring to it. > > But the PGP Web of Trust has no such categories. In fact it doesn't > refer to trust at all, but merely a way to extend already centralized > identity tracking systems. What I'm talking about is a different kind of > trust network, where in the example of newsgroups you would have three > values for each trust entry: the signing key you're trusting, which > newsgroup you're talking about, and how much you trust them to post that > which you prefer to read. The action in effect is always "posting what > you want to read" but it adds the clause "in group _____" since many > people post different kinds of messages in different groups. A sage at > theoretical physics could be a blind wisecracking idiot at fly fishing. > > Another possible way to conceptualize these actions are those online > text dungeons accessible through telnet. Not the Nethack style ones, but > the Interactive Fiction style ones. Generally they work by having a > number of actions you can perform in any given situation. "sit" "stand" > "attack" "go" "jump" and so forth. Each of these "action objects" > regardless of the actual words or syntax used for them, is a separate > semantic concept, which you could use as the basis for trusting people. > If someone benefited you in the past with the "give" action, you would > not be then obliged to trust them not to abuse the "steal" action. > > So that's pretty much what I would like to do in trust networks, is > implement a trust network that determines real value, or at least value > as closely as humanly possible to predict. In doing so there would be > different "kinds" of trust, perhaps including meta-categories that > affect the trust network itself (such as the often sought after action > of "recommending trustworthy people to me"). Each category of trust is a > separate semantic concept, though not necessarily independant. For those > concepts which depend on each other, the trust can bleed over to some > degree, but some concepts may be so diametrically opposed (axe murdering > and money lending) that trusting in one will actually reduce the trust > in the other. > > In the end I think we can fashion a sort of trust network, where we have > a good idea what we can provide from each other in the future, and a > good idea of who we can depend on to do what. Their government identity > isn't as important as whether or not they will behave benevolently to > you, and even someone with no identity at all would be able to create an > identity with zero reputation, and spend time committing selfless acts > to build up enough of a reputation that others will trust them in return. > > There are a few problems with such a network. People building up trust, > then betraying all at once for instance. I think those problems are > present in any of our current systems however, and though this system > might not be able to catch all of the problems that society faces, it > would still be a great improvement over the haphazard helplessness we > all now suffer from, never knowing exactly who to trust. > _______________________________________________ > p2p-hackers mailing list > p2p-hackers@lists.zooko.com > http://lists.zooko.com/mailman/listinfo/p2p-hackers _______________________________________________ p2p-hackers mailing list p2p-hackers@lists.zooko.com http://lists.zooko.com/mailman/listinfo/p2p-hackers
