@feld i think people have really just not used these systems at all. most frontier llms can read the room much better than the average programmer

@feld that it answered the question correctly does not mean this is the approach it would take in an actual someone-is-ordering-pizza situation.

@jeremiah an AI model deployed for this purpose when done properly would be trained well for this specific task and would be set with a temperature=0 and some other parameters that make it fully deterministic

This is not a fantasy, it's totally possible but not with the stock inference engines people are using day to day. You'd need a purpose built one like @SlicerDicer made. Then you get the exact correct response every time, not some random answer. And you would be able to have a vast test suite simulating calls to grade its performance to find any deviations when you change the model in any way

@feld @SlicerDicer i read the point as one about interpreting improvization. you just described massive tests and retesting on a case. what if i didn't know the pizza code but thought of chinese food, or avoiding getting the electric shut off -- scenarios where i would be providing my address?

@jeremiah @SlicerDicer it's not like you have to hardcode every possible codeword into the model. You can tell it what to do when the caller is requesting something that is not emergency services and is not directly addressing your questions.

@feld @SlicerDicer @jeremiah modern models can absolutely read between the lines. these are not 1950s style beep boop robots who say "DOES NOT COMPUTE".