In a hands-on diary published by The Verge, reviewer Antonio G. Di Benedetto describes his first serious attempt to use AI locally, installing the open-source Hermes Agent on an M5 Ultra Mac Studio and running the 125-billion-parameter Qwen 3.8 Flash Next model. The experience, he writes, is exciting, overwhelming, and frustrating — in part because the sheer number of available models makes it hard to know where to start, and in part because an empty AI text box remains an intimidating prompt to fill.
The main appeal for the writer is privacy: a local AI keeps sensitive data off cloud services and answers only to the user. That trust barrier led him to test tasks he would not send to a cloud provider, including analyzing financial records and building a spec-comparison spreadsheet for an unreleased laptop. His early projects had mixed results — a morning briefing initially failed because the Mac was asleep when the job ran, while reorganizing a 400-game Steam library worked well once a revocable API key was provided.
The piece lands as Apple and PC makers push local AI as a selling point for new hardware, including Windows machines with up to 128GB of RAM. But Di Benedetto notes you do not need a $12,000 computer to make a simple briefing, and his experience suggests the bigger obstacle is not raw spec sheets but usability. If local AI is to move beyond enthusiasts, the article implies, vendors may need to simplify model selection and setup rather than just advertise bigger memory pools.