Gemini 3 and the Art of the Pelican: When AI Gets Creative
Here’s something unexpected: Google’s latest Gemini 3 model can draw a pelican riding a bicycle—and it’s actually good. This seemingly trivial benchmark reveals something profound about where AI image generation has arrived.
The Core Insight
Simon Willison, a respected AI researcher and creator of the Datasette tool, put Gemini 3 through its paces with an unusual test: generating SVGs of pelicans on bicycles. The results were impressive enough that he called it “the best one I’ve seen so far.”
But this isn’t just about birds on bikes. It’s about what happens when frontier AI models encounter specific, unusual requests that push them beyond their training distribution.
Why This Matters
The pelican-on-bicycle test is deceptively challenging. It requires:
– Understanding animal anatomy (the distinctive pelican pouch)
– Understanding mechanical objects (bicycle frame, spokes)
– Composing these into a coherent scene
– Following precise instructions about pedaling position and breeding plumage
This is exactly the kind of edge-case prompt that reveals whether a model truly “understands” what it’s generating or just statistically predicts likely image tokens.
Gemini 3’s success here suggests Google’s model has genuinely learned to decompose complex visual concepts into SVG instructions—not just memorize common compositions.
Key Takeaways
- Gemini 3 produces significantly better SVG output than previous models on unusual prompts
- The “pelican on bicycle” benchmark has become a useful test case for AI visual reasoning
- These tests reveal genuine understanding vs. statistical pattern matching
- Google positions Gemini 3 as built for “science, research, and engineering” but it shows creative capabilities too
Looking Ahead
As AI models continue improving on these unusual benchmarks, the gap between “intelligent” and “just pattern matching” continues to narrow. The next time someone asks what AI “really understands,” we might point them to a pelican on a bicycle—and ask them to do better.
Based on analysis of “Gemini 3 Deep Think” by Simon Willison