My guess is they thought they were 99% done but that the 1% (“just gotta deal with these edge case hallucinations”) ended up requiring a lot more work (maybe even an entirely new sub-system or a wholly different approach) than anticipated.
I know I suggested the issue might be hallucinations above, but what I’m genuinely curious about is how they plan to have acceptable performance without losing half or more of your usable RAM to the model.
My guess is they thought they were 99% done but that the 1% (“just gotta deal with these edge case hallucinations”) ended up requiring a lot more work (maybe even an entirely new sub-system or a wholly different approach) than anticipated.
I know I suggested the issue might be hallucinations above, but what I’m genuinely curious about is how they plan to have acceptable performance without losing half or more of your usable RAM to the model.