Discussion
@dngrs @mathew @fasterandworse @davidgerard @maracuya
> mastodon-crawler: creates a trust graph based on who users are following
> web-graph-sim: Dioxus visualizer for friendship/trust graphs
you have instantly failed the assignment
trust is an _implicit_ property (decided upon unto themselves by the parties involved, perhaps codified in some types of agreement). you cannot "automatically" derive this
please stop what you're doing and learn from prior art
@froztbyte @mathew @fasterandworse @davidgerard @maracuya
current code is merely a proof of concept, I (mostly) agree with your position! The final version is planned to offer high flexibility in what to "subscribe" to, so YES re: nothing implicit. I do remain convinced that transitive trust (by trusting you I also trust the people you trust) is actually how human society works/contrary to the rugged individualism of "MY block/allowlist". pls do link prior art, always interested!
@dngrs @froztbyte @fasterandworse @davidgerard @maracuya I don't want to name names, but I have friends who I trust, who trust a person I absolutely do not trust. Surely this isn't unusual and you have examples in your own social circles?
@mathew @dngrs @froztbyte @davidgerard @maracuya it’s such a flimsy concept, it’s impressive how much effort went into codifying it rather than considering it.
@fasterandworse @mathew
good news, this is -of course- considered. (Advogato calculates trust using two metrics, one network-global capacity constrained flow, and one user-specific based on Eigenvectors)
You can find more details in the links I posted in this reply: https://chaos.social/@dngrs/117190845786240693
if you want to get down to the nitty gritty, there's also the thesis paper:
@dngrs @fasterandworse @mathew I'm looking at the Wikipedia article and the trust metrics section notes the failure mechanisms I remember Advogato having at the time
https://en.wikipedia.org/wiki/Advogato
tl;dr you can't actually mechanise trust and if you do it becomes something to game, and this should be obvious
@davidgerard @fasterandworse @mathew Levien discusses this in the postmortem I linked - he argues that it's less a failure mode and more a matter of perspective (will add a timestamp later); in general I find he's pretty honest about shortcomings, but of course he's still subjective. I 100% agree that "solving" trust is impossible (IMO he would too)! What's important to me is that the network flow paradigm prevents someone with 100k bots from having more impact than a single individual.
@davidgerard @fasterandworse @mathew
"advogato has failed" (doesn't keep out cranks) - he agrees that under this critic's assumptions it's a failure, but that wasn't his goal: https://youtu.be/9KgqtU25dBI?t=2270
botnet resilience even in the face of gullible users who trust a spammer: https://youtu.be/9KgqtU25dBI?t=1192
@davidgerard @fasterandworse @mathew I want to emphasize that criticism/threat models are probably the thing I'm interested in *most*, so please everyone keep bringing it, I want to build something that's *actually useful* (and it's gonna be an experiment with various pathways, fully expecting some or even the majority of subprojects to fail, this is a research project)
@dngrs @davidgerard @mathew “actually useful” is not a concrete purpose. What’s the purpose definition?
@fasterandworse @davidgerard @mathew
1. flag targeted harassment/spam/abuse material/scams so it doesn't flood your notifications (network-global score via Advogato-1 capacity constrained flow)
2. personalized curated feed ("posts especially interesting to myself") (customized score with Advogato-2 Eigenvectors)
@dngrs @fasterandworse @mathew
> 1. flag targeted harassment/spam/abuse material/scams so it doesn't flood your notifications (network-global score via Advogato-1 capacity constrained flow)
this sort of tech solutionism is playground equipment for online Nazis
in 2026, your threat model needs to be Nazis, not just spam
you cannot solve the most malignant bastards online with an algorithm
@davidgerard @fasterandworse @mathew
I would be *extremely* interested in a link collection of Nazi Fedi instances/users/posts so we can model against them, I do think they'd score too low because not a lot of non Nazi Fedi follows them - if that hunch is wrong, we absolutely do need a different approach
@dngrs @mathew @fasterandworse @davidgerard @maracuya I look forward to reading your Our Incredible Journey sunset post the first time a scammer fleeces you blind
(put otherwise: you may remain convinced for all you want, but it's simply not true. save yourself the effort)
@froztbyte we're essentially reimplementing Advogato which ran successfully for many years.
so yeah there's actually prior art we're working with here, proven in the wild, with a mathematical model built *specifically* for attack resistance (capacity constrained flow).
https://en.wikipedia.org/wiki/Advogato
https://www.youtube.com/watch?v=9KgqtU25dBI
edit:
I've updated the README to clarify current POC status and added a mention+link to Advogato, so thanks for pointing out what was missing, even if a bit rude ;)
Install bonfire.cafe
Get the full app experience
