Behold, an Axolotl!
I’ve been teaching my ten-year-old daughter spec-driven development using GitHub Copilot. Picture this in your mind: The main character is an axolotl. It is represented as a pink circle, 4% of the width of the play area, with blue dots for eyes and two half-circles side by side with no gap for a smiling mouth. The eyes and mouth are blue.
Did you picture a wobbly smile, like a 3 turned on its side? continue/Sonnet got pretty close. Take a good look: two semicircles, side by side. Nobody said which way they're oriented! Technically, we matched the spec. The AI made a mistake that only an AI or Diogenes could make.
We're using the Kaplay Javascript library to make a Pac-Man clone with an axolotl for a main character. The graphics are drawn using polygons; gives me fond Apple //C nostalgia. This also means that we can use the spec-driven approach to graphic design – not something I would necessarily recommend in production, but a fun experiment.
This stands as a reminder of the limitations of AI development tools and their ability to validate their work visually. There's progress happening on that front, but it's far from being baked into development tools like Copilot or continue.More importantly, and frequently overlooked with or without AI: developers need to understand the context for even a fairly detailed specification to yield a good result. Yes, the shape of a cartoon character's smile is something that a human is fairly likely to grasp intuitively, and carbon-based developers can smooth over these semantic gaps more effectively than silicon. But the gaps in your project are probably a lot more meaningful than the direction an axolotl's semicircles are pointing. When wetware meets silicon, that's where your tribal knowledge is put to the test: Have you really specified what you want to build, and why?
