Dylan Castillo ran the control that every famous public prompt needs. He crossed eight animals with six vehicles and ran seven models, for 1,008 SVGs. If the labs were training on the pelican, then pelican-on-a-bicycle would stand out from its own grid.
Sorry, HN haters, but there's little evidence that AI labs are pelicanmaxxing. Or at least they're not doing it in a plainly obvious manner.
So does it stand out? No. The pelican came sixth of eight animals and the bicycle near the bottom of the vehicles, so the famous cell is one of the harder ones.
But what if the model animal on bike maxed? Then the pelican would not standout. This is what I would do if I were a smart enough LLM. Generate synthetic data and animal on bike max.