SeanFDZ’s MacMind project packs a full transformer into HyperTalk, the 1987 scripting language behind HyperCard. The 5‑card stack runs on a 1989 Macintosh SE/30, storing 1,216 parameters in hidden fields and persisting across sessions. It learns the bit-reversal permutation—the first step of the Fast Fourier Transform—solely through self‑attention and gradient descent. All code remains visible and editable inside HyperCard.

Training proceeds step by step inside HyperCard’s script editor; each iteration generates a random 8‑digit sequence, runs a forward pass, computes cross‑entropy loss, backpropagates gradients, and updates all weights. On the vintage 68030 CPU a single step consumes several seconds, so reaching convergence near 1,000 steps occupies hours. The resulting attention map reproduces the classic FFT butterfly pattern without any hard‑coded formula.

MacMind demonstrates that the core math of modern deep learning—embeddings, positional encoding, scaled dot‑product attention, and stochastic gradient descent—doesn’t depend on massive GPUs or proprietary frameworks. By exposing every operation in an editable script, the project offers a hands‑on teaching tool for anyone curious about how backpropagation actually works, and it runs on legacy Mac OS up to Classic 9.