Show HN: MacMind – A transformer neural network in HyperCard on a 1989 Macintosh

Background

MacMind is a single-layer transformer neural network implemented entirely in HyperTalk, the scripting language for the classic Macintosh application HyperCard. Designed to run on 1980s-era hardware like the Macintosh SE/30, the project demonstrates that the fundamental mathematics of modern artificial intelligence—including self-attention and backpropagation—can be executed on vintage processors without external libraries or compiled code.

Community Perspectives

The discussion surrounding MacMind centers on the realization that modern artificial intelligence is less a product of new mathematical breakthroughs and more a result of massive scaling. Commenters noted the historical irony that while the "Attention" mechanism is a relatively recent innovation, the underlying backpropagation algorithms were published in 1986, just a year before HyperCard itself debuted. This led to a philosophical comparison where one user likened the project to introducing germ theory to the ancient Greeks; the hardware was capable of the logic long before the specific conceptual framework was applied to it.

A significant portion of the conversation focused on the untapped potential of legacy hardware. Participants pointed to the "demoscene" and retro-console homebrew communities as proof that hardware often possesses capabilities that remain undiscovered for decades. By implementing a transformer on a 68000 processor, the creator has "unlocked" a capability that was technically present in 1989 but remained dormant because the industry had not yet pivoted toward the transformer architecture. This sparked a broader reflection on what current hardware might be capable of in twenty years once developers move past today’s software abstractions.

Technical curiosity also drove the dialogue, particularly regarding the performance of interpreted HyperTalk. While the original hardware requires significant time for training, the community quickly adapted the project to modern environments. One developer imported the stack into a web-based HyperCard simulator, confirming that the pure HyperTalk implementation—free of external XCMDs—runs efficiently in a browser. This accessibility allowed users to verify the model’s ability to learn the bit-reversal permutation, a task essential to the Fast Fourier Transform, which the model discovers independently through gradient descent.

Finally, the project was praised for its educational transparency. By storing weights and activations in visible HyperCard fields, the author removed the "black box" nature of AI. Commenters appreciated that the math is entirely inspectable, allowing anyone to "open the hood" and see that intelligence, at this scale, is simply a collection of 1,216 numbers being updated through trial and error.

Further Reading

• HyperCard Simulator: A web-based tool where the MacMind stack can be run and inspected in a browser.
• 8088 MPH: A landmark IBM PC demo mentioned by commenters as an example of pushing obsolete hardware to its absolute limits.
• N64 Homebrew: A resource for modern developments in retro-console programming.