Fluctuating Accuracy in LLM Responses

Dear HN community, I’m brand new here and already feel right at home after just 5 minutes. I have a question for you about my theory:

I’m sure you’ve all experienced the wildly fluctuating quality of LLM responses. My theory: During peak times, the operators gradually reduce the depth of processing to take some of the load off the servers. I’ve noticed this a lot with Claude over the past few months. What do you think?