They often publish "needle in a haystack" benchmarks that look very good, but my subjective experience with a large context is always bad. Maybe we need better

They often publish "needle in a haystack" benchmarks that look very good, but my subjective experience with a large context is always bad. Maybe we need better benchmarks.