GPT-4-TURBO misses instructions sometime

Did anyone notice that. I am curious to know if is it because GPT-4-Turbo is trained using approximate self attentions like MAMBA and hence the low quality output.