MacMind trains transformer in HyperCard

Retro demo, real math. The platform is the hook, but the important part is that it makes transformer mechanics legible in an environment never meant for neural nets.

• –Everything is inspectable inside HyperCard, which turns a normally opaque model into something you can read and modify line by line
• –The bit-reversal task is a smart choice because it forces the model to learn positional structure instead of memorizing a shortcut
• –Saving the trained stack makes the model portable and persistent, so the weights behave like a real artifact rather than a throwaway demo
• –The project is a strong reminder that attention and backprop are hardware-agnostic math; only the scale changes