In the high-stakes arms race of Artificial Intelligence, we usually talk about 'scaling laws' and 'emergent properties.' But a new, weirder term has entered the lexicon: pelicanmaxxing. It sounds like a niche internet meme, but it actually points to a growing anxiety in the tech community: Are AI labs actually getting smarter, or are they just getting better at passing the test?
The Bird, the Bike, and the Benchmark
The term stems from a specific, quirky benchmark created by Simon Willison, which asks AI models to generate an SVG of a pelican riding a bicycle. Because this specific prompt became a popular way to test a model's spatial reasoning and coding abilities, a suspicion grew that labs might be 'cheating' by including this exact scenario in their training data to inflate their scores.
Data Contamination or Genuine Growth?
To find out if labs were simply memorizing the 'pelican-bicycle' answer, researcher Dylan Castillo conducted a deep dive. He generated over 1,000 SVGs across seven frontier models using a grid of different animal and vehicle combinations. By testing whether models could handle other weird pairings as well as the pelican, he aimed to see if the AI possessed general intelligence or just a specific 'pelican' shortcut.
This isn't the first time 'maxxing' has hit the AI discourse. Palantir CEO Alex Karp recently accused labs of 'token-maxxing'—prioritizing the consumption of tokens to signal growth over actually solving customer problems. Whether it's tokens or pelicans, the fear is the same: a pivot from utility to optics.
The Verdict
While the term 'pelicanmaxxing' is a bit of a joke, the underlying question is serious. If models are optimized for benchmarks rather than real-world reasoning, we aren't building AGI—we're building the world's most expensive game of Simon Says.
Sources
Media
