Would you participate?
No. A thousand times “no”. Just seems like a complete waste of resources federating intentional gibberish and more shit for users to block by default.
Instead, I recommend pushing your instance admins to run something like Nepenthes so that bot traffic is automatically served gibberish in the background instead of actual content. I’ve been doing this for a couple weeks now, and multiple bots are constantly thrashing around in the tarpit.
Is fediverse even used to train LLMs? It sounds more sensible to spam gibberish on mainstream platforms like Reddit and stuff
it absolutely is. i see posts from instance admins all the time with graphs on scraper/bot traffic.
Hilariously, the industry is doing this job itself. Endless ai generated LinkedIn posts, tweets, reddit comments, news articles etc. will ensure there is increasingly useless data flooding the internet. There is no real way to filter it and it will poison all future models.
The problem ist that this is all too obvious and can simply be filtered out based on the location. Better would be to deliberately add all sorts of gibberish to regular posts in order to poison machine learning models.
How long until data-poisoning is declared terrorism?
It’s too easy to actually poison an LLM. They aren’t scrapping the web like they used to anymore. Even if they did, they would have filters to pick up on gibberish.
It’s too easy to actually poison an LLM
How so? I’m curious.
In a joint study with the UK AI Security Institute and the Alan Turing Institute, we found that as few as 250 malicious documents can produce a “backdoor” vulnerability in a large language model—regardless of model size or training data volume.
This is the main paper I’m referencing https://www.anthropic.com/research/small-samples-poison .
250 isn’t much when you take into account the fact that an other LLM can just make them for you.
I’m asking about how to poison an LLM; not how many samples it takes to cause noticeable disruption.
Isn’t that [email protected]?
