What Runs Locally, Who Sustains Tools, and Which Predictions Hold

Local AI has concrete costs and choices: a desktop client can bundle an entire office suite, while a developer can run a model stack on a Mac Mini. Both make the software and hardware behind the interface worth understanding.

What local AI actually requires

Simon Willison found that the ChatGPT/Codex desktop app includes a full LibreOffice installation, presumably for document conversion and rendering. That is a pragmatic way to handle documents, with a substantial dependency attached. The finding makes installation footprint and local attack surface concrete engineering considerations: an AI client's capabilities may depend on considerably more software than its interface reveals.

For engineers considering running models themselves, one developer's M4 Pro Mac Mini setup provides a concrete account of a local stack. It offers a starting point for evaluating on-device inference where API spending or data privacy matters. The two stories concern different layers of local software: one exposes a desktop application's bundled dependency; the other describes a deliberate hardware and model setup. Neither establishes a universal cost or performance advantage.

Support for alternative development paths

Jujutsu creator Martin has joined ERSC, giving the Git-compatible version control system a dedicated institutional home. For engineers experimenting with jj workflows, the move suggests more sustained investment in the tool. It is a reason to watch its development, while leaving future improvements as an expectation rather than a delivered result.

A separate resource for developers exploring decentralized applications is this curated introduction to AT Protocol. The collection gathers starting material for Bluesky's protocol and offers a shortcut into evaluating federated identity or social data models. Its value is specific: it helps developers find an entry point for a side project or product investigation.

Check predictions against their record

Beyond implementation choices, Dan Luu evaluates the accuracy of Ed Zitron's AI-skeptic predictions. The exercise puts past claims under scrutiny rather than relying on whether their optimistic or skeptical tone feels persuasive. For engineers weighing assertions about AI's durability, that approach offers a useful calibration habit: examine what someone predicted and how accurately it described what followed.