Blog

Same Prompts, Three Macs: Four Models and a Test That Measured Nothing

2026-11-12

Part 3 of Same Prompts, Three Macs. Two DeepSeek releases and two small Qwen models, a choice per machine made on speed and disk, and a domain test on which every model, hosted or local, scored at guessing level.

Same Prompts, Three Macs: Where the Machines Sit

2026-11-11

Part 2 of Same Prompts, Three Macs. Logging thermal pressure beside every run, unstacking two Mac Studios, opening a laptop lid, and then putting everything back to check the first conclusion.

Same Prompts, Three Macs: Higher Precision, Same Answers

2026-11-10

Part 1 of Same Prompts, Three Macs. Two models, each kept in a large file and a small one, answered 40 checkable prompts. The scores were one prompt apart.

Same Prompts, Three Macs

2026-11-09

A short series measuring local models on three Macs: whether bigger model files answer better, how heat and placement change the timings, and what happened when I tried to grade them on a domain task.

Restoring the Models on a Clean Mac

2026-10-23

Part 8 of Oracles, Traces, Triage. Resetting a Mac, restoring eight local model profiles, and separating a copied directory from a verified backup.

Local Models and Tool Permissions

2026-10-22

Part 7 of Oracles, Traces, Triage. The concern that came up while choosing community models, and the tool permissions I need to settle before using them for repository investigations.

Testing Longer Contexts

2026-10-21

Part 6 of Oracles, Traces, Triage. Moving beyond an 8K default with measured inputs, explicit caps, follow-up requests, and rejection tests.