ah its from this
"Nice logo, eh? It was created by Liz Manicatide, a very nice artist friend as a commissioned work- she's a hired-gun artist. You can hire her for artistic, web, and user-interface work as well:
lizm at emphasiscreative dot com ."
The tool, itself, gets good results with a fairly small training set, but it's a lot of work
Support Vector Machine should outperform similarity hash once the spam base diversifies. The trick will be to make the labor for training accessible for instance admins. There's 2 ways to fork a classifier if the training file is transferable. That could help. The ultimate benefit would be sharing the training harness with tagging tasks other than spam. It doesn't take much more effort to select all relevant tags than it does to sort spam
The OCaps structure being developed by @emacsen and @cwebber has an unspecified proof of work requirement (stamps). The proof of work could be a ReCaptcha style training of classifier. You don't want spammers training your spam filter, but having untrusted senders suggest tags and presort a training set (that would be verified before training) could be useful
ReCapcha is using these challenges as a way of finding out if someone is a human, so it hates smart AI. But if I only care about the correct answer, I don't care if a bot did it or a human did it, therefore if Stamps used a ReCapcha-type challenge (ala Distributed Proofreader or Maproulette), then we'd get the benefits of ReCapcha and the benefits of Stamps.
But that's complex for a first go around. Hashcash is simpler at first.
@cwebber @emacsen @yaaps @Gargron ive liked some of the email ideas that involve having to put a "stamp" on the email which is a cpu-time-intensive operation that isnt much for a single sender, and largely not necessary if the participants involved have whitelisted each other, but incredibly prohibitive for spam
i feel like it hasnt been implemented because it kneecaps a lot of stupid fucking ~THE CLOUD~ shit and concern trolling from whomever implemented the most profitable variations on bayseian filters at the moment
I don't think that these ideas haven't been implemented because of those reasons.
I wasn't involved in AP's development early enough to have been able to bring my experience regarding the issues that I've later addressed in my Unwanted Messages paper, and (frankly) I probably wouldn't have had the basis to build a good idea off of w/o @cwebber and OcapPub. Stamps is imperfect and needs to be flushed out, but the current fediverse is too subject to abuse.
Machine learning is that blue puzzle piece that people want to fit everywhere. I think we finally have consensus that it's not an edge piece, but it still draws the eye. The community would benefit from positive direction on where it goes
A witchy space for most any face! Whether a witch or a witch-respecter, join the coven that is free of fash, TERFs, feds, and bigots