Why Nostr Moderation Needs To Happen In The Client
For every user, there are four kinds of content you will have encountered on Nostr:
- Content that nobody actually wants to see, like random DDoS and spam
- Content you want to see, but other people may not.
- Content you DON'T want to see, but other people do.
- Content you simply don't care about either way.
Veganism. Profanity-laced discussions between buddies. Nudity. Furry. Jesus. Taylor Swift. Donald Trump.
Filtering at the relay level can remove the obvious content nobody wants to see, and content that could get the operator's door kicked in by police. That's a necessary part of the ecosystem.
But a relay that tries to do more filtering than that has to be all things to all people. This is clearly impossible, even in an environment as conformist as a HR Department or a Marvel fan forum.
So choices get made, some people are silenced, others are not, and if governments or multinational companies built the filter in secretive and opaque ways ( Reportinator (npub14h2…e2sh) yes I'm talking about your LLM) then ideological and commercial censorship sneaks in and takes over.
So we can't filter at the relay level without significant harm to the things that make Nostr great.
We can get around this, though, by filtering and moderating on the individual user's client app. Users each create plenty of data on what they want to see (interactions) and what they don't (reports).
No system is perfect, but by training the filter on the individual's preferences we can make sure the things they like get through the filter, the things they hate are kept out, and that over-filtering, when it happens, hits the content the user didn't care about anyway.
Obviously anything running in the app needs to be very frugal with resources, but small models can be very accurate on specific tasks with enough data to train with.
I've done some work on Minitru to build these. It has a website of sorts now, thanks to DanConwayDev (npub15qy…yejr) and ngit. Its not ready for production use, but power users can "git clone" it from https://gitworkshop.dev/repo/minitru
You don't need to be any kind of programmer, but you'll need to be able to install Git, Tor and Python, and use a command-line.
Right now Minitru can:
- Download your reports and interactions into a dataset
- Train a auto-moderator model based on this
- Allow you to test the moderator against your posts, your follows' posts, and Global.
Low Information Voter
npub149…tzmp0
2024-04-20 11:07:09
Author Public Key
npub1494rtg3ygq4cqawymgs0q3mcj6hucvu4kmadv03s5ey2sg32df5shtzmp0Published at
2024-04-20 11:07:09Event JSON
{
"id": "974363acd7ae2061e0759f27695325c344c5d6e5dfdba6689240eea161a79220",
"pubkey": "a96a35a224402b8075c4da20f0477896afcc3395b6fad63e30a648a8222a6a69",
"created_at": 1713611229,
"kind": 1,
"tags": [
[
"p",
"add5190be4673768546c18b565da3a699241f0e06a75e2dbc03f18663d1b7b27",
"",
"mention"
],
[
"p",
"a008def15796fba9a0d6fab04e8fd57089285d9fd505da5a83fe8aad57a3564d",
"",
"mention"
],
[
"r",
"https://gitworkshop.dev/repo/minitru"
]
],
"content": "Why Nostr Moderation Needs To Happen In The Client\n\n\nFor every user, there are four kinds of content you will have encountered on Nostr:\n\n- Content that nobody actually wants to see, like random DDoS and spam\n- Content you want to see, but other people may not.\n- Content you DON'T want to see, but other people do.\n- Content you simply don't care about either way.\n\nVeganism. Profanity-laced discussions between buddies. Nudity. Furry. Jesus. Taylor Swift. Donald Trump.\n\nFiltering at the relay level can remove the obvious content nobody wants to see, and content that could get the operator's door kicked in by police. That's a necessary part of the ecosystem.\n\nBut a relay that tries to do more filtering than that has to be all things to all people. This is clearly impossible, even in an environment as conformist as a HR Department or a Marvel fan forum.\n\nSo choices get made, some people are silenced, others are not, and if governments or multinational companies built the filter in secretive and opaque ways ( nostr:npub14h23jzlyvumks4rvrz6ktk36dxfyru8qdf679k7q8uvxv0gm0vnsyqe2sh yes I'm talking about your LLM) then ideological and commercial censorship sneaks in and takes over.\n\nSo we can't filter at the relay level without significant harm to the things that make Nostr great.\n\nWe can get around this, though, by filtering and moderating on the individual user's client app. Users each create plenty of data on what they want to see (interactions) and what they don't (reports).\n\nNo system is perfect, but by training the filter on the individual's preferences we can make sure the things they like get through the filter, the things they hate are kept out, and that over-filtering, when it happens, hits the content the user didn't care about anyway.\n\nObviously anything running in the app needs to be very frugal with resources, but small models can be very accurate on specific tasks with enough data to train with.\n\nI've done some work on Minitru to build these. It has a website of sorts now, thanks to nostr:npub15qydau2hjma6ngxkl2cyar74wzyjshvl65za5k5rl69264ar2exs5cyejr and ngit. Its not ready for production use, but power users can \"git clone\" it from https://gitworkshop.dev/repo/minitru\n\nYou don't need to be any kind of programmer, but you'll need to be able to install Git, Tor and Python, and use a command-line.\n\nRight now Minitru can:\n\n- Download your reports and interactions into a dataset\n- Train a auto-moderator model based on this\n- Allow you to test the moderator against your posts, your follows' posts, and Global.",
"sig": "d1cda14dc58e984629ed81b2611667c43949e0ae53190686fe6ce06aac05c38a26d347a310dd5d27bbf63f73863dccb5fea34a894b1d60b6b76d8a593733f669"
}