A developer has implemented a complete transformer neural network in HyperCard, Apple's 1987 scripting language, running on a vintage Macintosh with just 1,216 parameters.
MacMind includes all core transformer components: embeddings, positional encoding, self-attention mechanisms, backpropagation, and gradient descent. The entire implementation uses HyperTalk, the scripting language that shipped with HyperCard, with every line of code readable and inspectable within the application's script editor.
The project trains the network to learn bit-reversal permutation—a specific computational task. While execution is slow on 1989-era hardware, the implementation demonstrates that modern machine learning architectures are theoretically implementable even with severely limited computational resources.
The effort showcases both the flexibility of HyperCard's scripting capabilities and highlights how far neural network optimization has advanced. The readable code format makes the mathematical operations transparent, offering educational value alongside the technical achievement.
The Model Context Protocol team has published a roadmap detailing upcoming features and improvements. The announcement has generated significant interest in the developer community.
Kagi, the privacy-focused search engine, has added a new setting allowing users to exclude paywalled content from their search results. The feature gives subscribers greater control over result quality and accessibility.
Developer Dan Luu argues that modern software performance failures stem from neglect rather than technical limitations. The post, which has garnered significant discussion on Hacker News, challenges the assumption that slowness is inevitable.
Data and finance professionals are competing in Excel obstacle courses, amassing large followings and keeping Microsoft's decades-old program culturally relevant.