Writer tests private AI agents on costly high-memory computers
The writer is testing whether local AI is useful enough to justify buying expensive, high-memory computers, while keeping personal data away from cloud services. They have begun experimenting with AI agents on a Mac Studio, drawn by the prospect of a private assistant that runs on their own hardware.
They installed the free, open-source Hermes Agent and chose Qwen 3.8 Flash Next, a 125-billion-parameter model that takes about 105GB of storage, on a Mac Studio with 256GB of unified memory. A Telegram bot lets them control it by phone, and their first project is a daily briefing that checks email and calendar and reports the weather; the writer says it is not especially useful so far. They plan to try smaller models on other Mac and Windows machines as the experiments continue.
- Local AI offers more privacy from cloud services.
- A 125-billion-parameter model is running on a 256GB Mac Studio.
- The first agent project is a daily email, calendar and weather briefing.
New here? Start with this
Most people use artificial intelligence services like ChatGPT that process their information on company servers. Some prefer to keep information private by running AI directly on their own computers instead. This story follows a writer testing whether that approach can work well enough to justify the expense.
The writer is testing free software and AI models that run entirely on their own Mac Studio, which has 256GB of memory. Their first project is a daily assistant that checks email and calendar, and reports the weather, controlled from their phone via a messaging app.
The main challenge is that powerful AI requires enormous amounts of computer memory, making it far more expensive than using internet-based services. The experiments will help show whether home-based AI could become practical for those who want privacy and are willing to invest in expensive hardware.
Both sides, in good faith
The strongest fair case each way — we don't pick a winner.
The case for
Privacy and data sovereignty are fundamental values that justify the hardware investment for those who can afford it. Sending personal emails, calendar entries, and location data to cloud services means trusting external companies with intimate information about your life, exposing yourself to data breaches, policy changes, and surveillance. Running AI locally offers genuine control and protection by keeping sensitive information on your own hardware, rather than vulnerable to corporate or governmental access.
The case against
Whilst privacy concerns are understandable, cloud services offer proven efficiency, reliability, and cost-effectiveness that make expensive local hardware difficult to justify. The writer's own experience demonstrates this—investing over £3,000 in a Mac Studio has not yet produced meaningful utility compared to readily available cloud alternatives that cost far less and require no hardware burden. For most users and tasks, cloud services provide adequate privacy protections whilst offering greater convenience, lower total cost of ownership, and regular updates without the complexity of maintaining local infrastructure.
AI Geopolitics Politics Technology
Read the full article at the source →
Originally published by The Verge as “Learning to use local AI is exciting, overwhelming, and frustrating”.