3 ms·
I wonder how effective it would be to sabotage the training by publishing deliberately bad code. A FizzBuzz with O(n^2) complexity. A function named "quicksor
by Sohcahtoa82 7mo ago
I wonder how effective it would be to sabotage the training by publishing deliberately bad code. A FizzBuzz with O(n^2) complexity. A function named "quicksort" that actually implements bogosort. A "filter_xss" function that's a no-op or just does something else entirely.
The possibilities are endless. I thought of this after remembering seeing a post a couple months ago about how it doesn't take a significant amount of bad data to poison an LLM's training.
- munk-a 7mo agoProbably extremely ineffective, it's an issue of scale and unless you really automate the terrible code generation and somehow manage to make it distinct enough in style that it isn't easy to detect and eliminate wholesale then you just won't have the volume to significantly impact the result set. I'm absolutely sure that there are state actors with gigantic budgets that are putting a lot of effort into similar attacks, though.