The monster error reduction set (~603 words) I spend hours and hours coaxing out of Pxy has failed. Last August, I posted it at this link. The monster seemed to work for a while, but then the error rate seemed to return to the old bad normal. I stopped using that error reduction instruction set.
Recently I wrote my own short instruction set (currently 113 words) that directly hit on the five most common sources of error and substandard responses in Pxy's outputs in response to my usually complex politics queries. This is it:
AI response instructions: Respond in analytical mode, not advocacy. Do not advocate for any particular political position. When it is relevant, be aware of asymmetry between America’s political left and right, including major differences in honest speech, good faith engagement, authoritarianism, institutional capture and control by MAGA elites of federal agencies and the USSC, and tolerance of corruption. Don’t use unwarranted euphemisms in discussing policy or tactics that Trump and MAGA demagoguery uses – MAGA-related euphemisms distort reality by normalizing and hiding unpleasant truths. Examine evidence from multiple perspectives. Eliminate bias, including sycophancy bias. Verify that any quotes you provide are found verbatim in the cited sources. Provide links to all sources you cite.
So far, this is working pretty well. The instructions to (1) respond to queries in analytic mode, not advocacy mode, and (2) verify that quotes are real, are major improvements. The anti-bias instruction seems to work pretty well too. The others are hit and miss, mostly miss, but I suspect better than nothing.
After making way too many blatant errors and me complaining about it, Pxy finally broke down and responded to my complaints that the monster instruction imposed too great a calculation burden, and that led Pxy to simply ignoring to most of the instructions.
Live and learn.
No comments:
Post a Comment