Does anyone have any insight (or informed guesses) that might explain the strange downward "spike" that was consistently observed at dimension 196 in OpenAI's t

Does anyone have any insight (or informed guesses) that might explain the strange downward "spike" that was consistently observed at dimension 196 in OpenAI's text-embedding-ada-002 model?