👋 Heya!

🧐 The Challenge: "Draw an SVG of a Pelican Riding a Bicycle”

If you really wanna test an LLM’s true intelligence, here’s a spicy prompt:

📝 "Generate an SVG of a pelican riding a bicycle."

Simple, right? Not really. When Simon Willison threw this at GPT-3, Claude, and Grok, their outputs looked like something a toddler would sketch—except toddlers have better spatial awareness.

Here’s what went wrong and why AI still struggles with actual creativity.

🤖 Why Is This So Hard for AI?

1. Spatial Reasoning? Never Heard of It.

LLMs are text-based, meaning they predict words, not visuals. So when asked to "draw" something, they’re basically guessing in the dark.

A pelican on a bike? That requires understanding how wings, feet, and handlebars interact. Instead, LLMs end up generating a pelican awkwardly levitating near a bike or straight-up fusing into it.

2. Coordinates? AI Is Just Throwing Darts.

SVGs need precise positioning—where the pelican’s feet go, where the pedals start, how the handlebars connect. AI doesn’t intuitively "see" these relationships, so it ends up plotting random coordinates that make zero sense.

🔢 Example: If the pelican’s foot is at (216, 48), the pedal should be at (216, 47). LLMs? They put it at (50, 300). Bruh.

3. Anatomy? Who Needs That?

Ask AI to describe a pelican, and it’ll give you an essay. Ask it to draw one? It’ll make an amorphous bird-blob on wheels. Merging objects (like a bird and a bike) while preserving their anatomy? Absolute nightmare.