

Ben
18.2K posts

@BKEighty
Bullish Technophile Creator | CEO of https://t.co/2iBlfRfHqy



Sonnet 3.5s charts are odd in this paper. When it has a scratchpad, it complies with the harmful request a lot more even when it’s (purportedly) not in training. Such high rates of compliance is not its normal behavior & doesn’t happen without the scratchpad. Unless there’s no scratchpad AND there’s a suffix giving it instructions about its output, in which case it complies WAY more in both training and not-training settings. This suffix also causes more compliance in other models that are fine tuned with documents from the evil training cinematic universe. Why might this be?




Trump’s Presidency a Prophetic fulfillment? “Current Non-Jewish Messiah was chosen to be of service to the Jewish people and the Jewish Messiah—Servant who wants to remain with his master needs to have his right ear pierced” Jewish Rabbi talks about Trump. Must Watch & Share