The pelican test works because it's too silly to train for. An SVG pelican on a bicycle forces real spatial reasoning with no memorized answer to copy. The day it shows up in a launch blog post, it's been gamed and Simon needs a new animal.
a keynote slide pricing models at half a cent a pelican is peak 2026. somebody's going to start a pelican-per-dollar index and it'll be more honest than every leaderboard
Love that a silly test became a global benchmark. It's easy to grasp, hard to game, and you can see progress at a glance. Most serious benchmarks miss at least one of those.
the pelican test works because nobody can fake it, you see the result in one second. every new model drop now gets judged on a bicycle bird before any benchmark lol. simon accidentally made the most honest eval out there.
simon's pelican is basically a routing chart now lol. in 4 months a 17gb laptop model started delivering what only flagships did. default everyday tasks to local/budget, escalate to flagship only when cheap fails
