twenty second July 2026 – Hyperlink Weblog
Are AI labs pelicanmaxxing? (through) Glorious piece of labor by Dylan Castillo, who took a deep-dive into the ceaselessly contemplated query of whether or not the AI labs have been intentionally coaching fashions to attract pelicans using bicycles in response to my deeply unscientific benchmark.
I have been randomly spot-checking this prior to now by testing fashions towards different animals using different forms of car, however by no means with something near the diligence of Dylan’s methodology right here.
Dylan took 8 animals × 6 autos = 48 prompts and ran them thrice every by way of 7 completely different fashions ( GPT-5.6 Terra, Claude Sonnet 5, Gemini 3.5 Flash, Grok 4.5, Qwen3.7-Max, GLM-5.2, and DeepSeek V4 Professional). He then used GPT-5.6 Luna and Gemini 3.1 Flash-Lite to assist consider the outcomes.
There is a neat filter view for exploring the outcomes:
For the fashions he examined he might discover no proof of pelimaxxing:
Pelicans aren’t drawn any higher than different animals. Bicycles aren’t drawn any higher than different autos. And no lab attracts the mix higher than its pelicans and bicycles already predict. GLM-5.2 comes closest: it has the biggest enhance on the precise pelican-bicycle cell, and and its first pelican-on-bicycle pattern caught my eye. However the impact is small and never important, so I wouldn’t put an excessive amount of weight on it.
