AI

2026.08.19 Do Instructional Fingerprints Produce Stable Expert-Routing Signatures in a Mixture-of-Experts Model?A controlled negative: fingerprints implanted in Qwen1.5-MoE-A2.7B shift expert routing but no stable trigger signature emerges above matched nulls.

2026.08.08 Measured Pruning Damage Depends on the Evaluation Corpus: A Renaming Control for Mixture-of-Experts Expert PruningExpert-pruning damage on code depends on the eval corpus: a renaming control on CPython stdlib moves reported damage 1.6x at constant program structure.

2026.07.15 qMLX: Dogfood Tales #1 - Multiplexing & Cache StabilityRipping the hot cache out of qMLX, splitting DeltaNet checkpoints so they restore from SSD, and using multiplexing to run concurrent Qwen 122b sessions on a single 96GB Mac Studio.

2026.07.10 qMLX: Maximising my AI psychosis by minmaxing my Mac StudioDS4 Flash looked promising on paper. In practice it was too slow for long contexts and unreliable for agentic coding. Here is why I switched to Qwen 3.5 122B and the three bugs I …

2026.06.19 Charted: a zero-dependency charting library (and an accidental agentic sidequest)Why I built charted, a zero-dependency Python library for beautiful SVG charts, and how a constraint I picked for fun turned out to be the perfect fit for agentic workflows.