Immortal WombatINTERCAL: We have keywords like PLEASE and IGNORE because we think it's funny
Anthropic: We have PLEASE and IGNORE *baked into our source-code* because we have no fucking idea how this thing works
I'm now hearing where it's important to be polite to the idiot robot because then you might get better results CAN YOU FUCKING HEAR YOURSELVES?
samiamsam
Sean Middleditch@pikesley Since we have no idea what the training data actually is, there's a chance it *does* give better results, for some definition of "better" anyway.
Someone should do a study on the quality of results between asking tersely worded questions, asking politely worded questions, and burning the entire industry to the ground.
For science.
Henryk Plötz@pikesley Wait, last year's best magical practices was to *not* say please and thank you because it was costing them millions and making them sad.
@henryk well the cutting-edge Applied Magical Thinking now suggests we have to be nice
I hate it here
EndlessMason@pikesley
"Might"
"Might"
Kevin Granade@pikesley this is very much a problem with gambling based outputs. Every time you change your prompt it's a new pull on the slot machine for a "good" output. As such it gives biases maximum opportunity to kick in, since you also get a new pull on repeat prompts.
It's actually in the vendors interest to keep changing recommendations so that people stay in the, "you're promoting it wrong, just keep making 'corrections' until you like the output" loop. Shifts the blame to the user.
Artemis@kevingranade @pikesley I think this is also where all those "You're a senior <insert technology here> developer, please obey all the best practices" stuff comes from.
Amo Bishop Rodent (@[email protected])
The notion of "best practice" is predicated upon the ludicrous idea that anybody knows what they're doing