I noticed the quality of answer degrades horribly beyond few thousands of tokens. Maybe 10k. Is anyone actually successfully using these 100k+ token contexts for anything?