ChatGPT’s diagram of how DSD works is very misleading. Even worse than Qobuz’s explanation of high-resolution!
For example, where the ChatGPT waveform is flat, there should be an equal number of alternating 0 and 1 bits, the exact opposite of what the diagram shows.
It is very simple. Keep a running total of where the sound pressure is up to. Take a sample. If it is higher than your running total, add a 1 bit to the stream and your running total. If it is lower, add a 0 bit to the stream and reduce your running total by 1.
By the way, every 64 samples, your running total equals the CD 16-bit number at that point. You can exactly calculate the CD data from the DSD stream, but the reverse does not apply.

