← Back to context

Comment by solumunus

2 days ago

And do you think the severity of the issue is anywhere near the same?

I think this will remain to be seen. Wasn't there a paper linked here on HN recently, that claimed, that even few examples are sufficient, to poison LLMs? (I didn't read that paper, and merely interpreted the meaning of the title.)

  • I don't think it remains to be seen. I think it's obvious that the completely explicit exploit is going to be more effective.