Muse Spark Open, Meta + Apple, Thin and Thick
Muse Spark Open, Meta + Apple, Thin and Thick
Ben Thompson starts with Meta’s downloadable Muse Spark Open model and its conditions of use. At an Apple demonstration, two Mac Studios served a model that altered an app on request and worked on large video files. He argues local inference could matter when uploads are difficult or users want control over model use, even if cloud models remain stronger.
The coding demonstration made one prompted change. Thompson’s separate proposal is for a harness to build alternative feature paths overnight, with follow-on code ready for the path a person chooses in the morning. Multiple independent candidate patches are already possible in Cursor’s worktrees; precomputing dependent paths is the untested extension. Apple measures Mac inference speed, not feature acceptance.
Thompson concedes dedicated Macs may be less economical than cloud services. A local branch still costs electricity, machine capacity, testing, comparison and human integration, even with no per-token API bill. Neither the demonstration nor the launch measures review time, discarded work, merges or later requester satisfaction. Compare single- and multi-path builds for one request, counting the human decision and maintenance alongside model spend.