"Akin to being lost in translation, the model chose slightly wrong numbers, which produced word sequences that made no sense. More technically, inference kernels produced incorrect results when used in certain GPU configurations."
Several of the examples I saw involved ChatGPT going into what looked like repetitive (but not completely so) output loops. I'm not sure that "explanation" matches the things.
I can imagine it'd get screwy when the incorrect output token selection gets fed back into the model with the correct incorrect token selection with nonsense tokens. It's plausible.
I remember Bing Chat doing that sometimes in the first days when it was rolled out. Could it be the "temperature" set too high (or interpreted incorrectly) in some instances?
Several of the examples I saw involved ChatGPT going into what looked like repetitive (but not completely so) output loops. I'm not sure that "explanation" matches the things.