Stephane Hamel
Digital marketing & analytics shaped by data governance, privacy and ethics | Educator · Speaker · Consultant
Follow me
December 13, 2024
Another day, another AI challenge popping up in my LinkedIn feed: “AI can’t generate an image of a completely filled glass of water.”
Well, here’s a twist for anyone up for the ultimate AI test:
Try generating an image of a partially completed puzzle where:
a) The puzzle piece shapes actually make sense.
b) The number of missing pieces in the puzzle matches the pieces laid out on the side.
Spoiler alert: it’s impossible.
• AI models struggle to ensure the edges of each piece align correctly because these models aren’t “aware” of geometric rules. They generate pixels based on learned patterns, not detailed logical constructs like puzzle mechanics.
• AI is not inherently good at exact object counting or maintaining logical relationships between elements in an image. For example, matching the number of missing pieces in the puzzle to those laid out on the side requires a logical rule, not just artistic interpretation.
• Most AI models are trained on general images of puzzles but not on detailed examples that explicitly demonstrate how partially completed puzzles with logical piece arrangements should look.
December 13, 2024