Anthropic: "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. "
Discussion
Loading...
Post
Stop uploading gibberish to the cloud
You're making the LLMs die for crying out loud
Putting gibberish into your docs is driving them mad
And then your search results get very very bad
They'll think that gravy is just fine
To pour and mix in with your wine
And they'll tell you to drink castor oil
And wrap your head up with aluminum foil
They'll think that birds can emit radio noise
And that live sharks make very good toys
And they might even think sea horse emojis are real
And that the best kind of food to eat is arsenic meal
#poem (human)
bonfire.cafe
A space for Bonfire maintainers and contributors to communicate
Automatic federation enabled