This is a most excellent place for technology news and articles.
cross-posted from: https://programming.dev/post/35852706
Is there nightshade but for text and code? Maybe my source headers should include a bunch of special characters that then give a prompt injection. And sprinkle some nonsensical code comments before the real code comment.
Anubis isn't supposed to be hard to avoid, but expensive to avoid. Not really surprised that a big company might be willing to throw a bunch of cash at it.
I feel like at some point it needs to be active response. Phase 1 is a teergrube type of slowness to muck up the crawlers, with warnings in the headers and response body, and then phase 2 is a DDOS in response or maybe just a drone strike and cut out the middleman. Once you've actively evading Anubis, fuckin' game on.
I think the best thing to do is to not block them when they're detected but poison them instead. Feed them tons of text generated by tiny old language models, it's harder to detect and also messes up their training and makes the models less reliable. Of course you would want to do that on a separate server so it doesn't slow down real users, but you probably don't need much power since the scrapers probably don't really care about the speed
I love catching bots in tarpits, it's actually quite fun
Some guy also used zip bombs against AI crawlers, don't know if it still works. Link to the lemmy post
Can there be a challenge that actually does some maliciously useful compute? Like make their crawlers mine bitcoin or something.
Did you just say use the words "useful" and "bitcoin" in the same sentence? o_O
Bro couldn't even bring himself to mention protein folding because that's too socialist I guess.
LLMs can't do protein folding. A specifically-trained Machine Learning model called AlphaFold did. Here's the paper.
Developing, training and fine tuning that model was a research effort led by two guys who got a Nobel for it. Alphafold can't do conversation or give you hummus recipes, it knows shit about the structure of human language but can identify patterns in the domain where it has been specifically and painstakingly trained.
It wasn't "hey chatGPT, show me how to fold a protein" is all I'm saying and the "superhuman reasoning capabilities" of current LLMs are still falling ridiculously short of much simpler problems.
The crawlers for LLM are not themselves LLMs.
Crawlers aren't LLMs; they can do arbitrary computations (whatever the target demands to access resources).
The Monero community spent a long time trying to find a "useful PoW" function. The problem is that most computations that are useful are not also easy to verify as correct. javascript optimization was one direction that got pursued pretty far.
But at the end of the day, a crypto that actually intends to withstand attacks from major governments requires a system that is decentralized, trustless, and verifiable, and the only solutions that have been found to date involve algorithms for which a GPU or even custom ASIC confers no significant advantage over a consumer-grade CPU.
Is there a migration tool? If not would be awesome to migrate everything including issues and stuff. Bet even more people would move.
Codeberg has very good migration tools built in. You need to do one repo at a time, but it can move issues, releases, and everything.
I mean, we really have to ask ourselves - as a civilization - whether human collaboration is more important than AI data harvesting.
