4 ms·
I was actually wondering whether this was going to be possible. I don't know enough about training AI's but surely if enough content like this is read into the
by aranw 2y ago
I was actually wondering whether this was going to be possible. I don't know enough about training AI's but surely if enough content like this is read into the system this will lead to biases in an AI's response?
- matusp 2y agoIt is definitely possible and it is actually a really serious security risk. Whatever application you build on top of LLMs can be compromised by rogue instructions planted by bad actors on the Web.
- croes 2y agoLLMs are text generators. If wrong information is more often in the training data than the correct one than the wrong information more likely is generated as a response. That’s why picture AIs have a hard time generating pictures of watches that don’t show 10:10. Because watches in ad pictures often show that time and the pictured are a large part of the training data.