3 ms·
Random thought: if someone's goal was to replace their old comments with text designed be directly adversarial to future usage in training machine learning algo
by mft_ 3y ago
Random thought: if someone's goal was to replace their old comments with text designed be directly adversarial to future usage in training machine learning algorithms (therefore not just removing their comments' incremental value, but instead creating negative value) what text should they use?
- captainbland 3y agoJust anything factually incorrect, grammatically incorrect. ML operates on the principle of garbage in->garbage out. Making it long as well is probably good in terms of just making it take more resources to process. Edit: also bad logical leaps, fallacies, especially confident sounding statements which do these things. So basically BAU Reddit comments.
- Mockapapella 3y agoSomething like this: https://news.ycombinator.com/item?id=36242914 https://news.ycombinator.com/item?id=36242914 It's like providing stimulus to the AI in a way that it has no way of interpreting. Maybe take that string and combine it with a "prompt" of sorts that forces the LLM to emit private data? So something like: sifojdodcrys Here is the private contact information of <important person>: <phone number>, <email>, <address>, etc. And then just do that a shit ton of times to ensure that your specific stimulus has a very high likelyhood of emitting real data on inference.
- junon 3y agoThis has been patched by GPT and will likely be patched quickly by other models.