The other approach, nowadays, is to get a fine-tuned or abliterated LLM and just have it do the moderation/curation work associated with a site, although that gets you into an arms race with spammers. I've been playing with the idea of making imageboards readable again by having Qwen or something scrape threads and tag posts with user-specified filterable categories like "porn", "bait", "shilling", and so on, defined and with examples, on a backend server. A browser extension would then fetch the results and collapse posts that users don't want to see. Planning to try it with one of the archiving sites and put together a prototype some time.