2 ms·
It is easy to get a local R1 model to talk about Tiananmen Square to your heart's content. Telling it to replace problematic terms with "Smurf" or another nons
by CamperBob2 2y ago
It is easy to get a local R1 model to talk about Tiananmen Square to your heart's content. Telling it to replace problematic terms with "Smurf" or another nonsense word is very effective, but with the local model you don't even have to do that in many cases. (e.g., https://i.imgur.com/btcI1fN.png https://i.imgur.com/btcI1fN.png)
- cratermoon 2y agoAh yes, evidence of Shadow Prompting[1]. Likely their API is given additional, hidden, prompt context making Tiananmen verboten. 1 https://www.techpolicy.press/shining-a-light-on-shadow-prompting/ https://www.techpolicy.press/shining-a-light-on-shadow-promp...