Rovers

Blog

What we are measuring.

  1. 10 min

    The one from yesterday.

    A colleague needs eight words. A prompt box needs fifty. Conversation gets from one to the other in six passes, and the average conversation with a model is two turns long. The cost is not the typing. It is the requests you stop making.

  2. 11 min

    WER is the wrong number.

    A real transcription error costs almost nothing. A clean-up that improves the transcript recovers nothing. A fluent rewrite that scores beautifully costs twenty-four points. The metric cannot tell them apart.