>>105630774
There was an old paper that shows that models still pay more attention to the start of the context compared to the middle.