Do newer models perform well as context fills? | Not Hacker News!