10-19-2026, 04:38 AM
I tested this tonight. I couldn't reproduce the slowdown until I pushed context much higher. The main thing that moved was latency, not quality.
|
comparing models without fooling yourself
|
|
« Next Oldest | Next Newest »
|