If anyone has great examples of using llms.txt, CLAUDE.md (etc) to poison LLMs, please share. I’m aware Anubis tried this at one point. I’m thinking Nuclear semiotics, “nothing of value here”, that kind of thing!
- replies
- 1
- announces
- 4
- likes
- 3
@jmtd I should probably blog the details, but it boils down to: building a markov chain from a select set of public domain books, generate a bunch (~50) HTML "articles" from that, gzip ’em, put them on the website with "Content-Encoding: gzip" (so I save bandwidth, but the receiving end does not).
The tricky part is to direct scrapers to that, but a dispatch based on User Agent sent around 1700 requests there, so I think it works.