I don't think this argument is a good one though, as it would be quite natural for a lab rhat want to macimize the performance of their model on the pelican bench to train it for “text-to-svg simple image generation” rather than just “pelicans on bicycle”.
https://dylancastillo.co/posts/pelicanmaxxing.html
https://simonwillison.net/2026/Jul/22/