Context beats parameters: why a 9B model holds up on routine workRoutine work has clear boundaries, material in hand and checkable answers. Those three properties shrink the gap between a 9B model and the frontier.16 min read·karmabox.hashnode.dev2d ago
Two homes for agent memory: in the weights, or in a file you keepFine-tuning bakes what your agent learned into the model. A memory file keeps it outside. A comparison for the week everyone is eyeing a switch.16 min read·karmabox.hashnode.dev3d ago
The MIT license survived. The repo didn't.ONNX Runtime vanished from GitHub this week. Every existing copy is still legally yours to use. Getting a copy is now your problem.15 min read·karmabox.hashnode.dev4d ago
A residency promise you rent vs. a data boundary you runThree situations pull the two setups apart: a disclosure order, telemetry nobody commissioned, and a renewal that rewrites the terms.10 min read·karmabox.hashnode.dev5d ago
Anatomy of a Self-Hosted AI Agent: Four Layers You Keep, One You RentCredentials, memory, policy and logs stay on your machine; inference can live anywhere. Why the approval layer only counts outside the model11 min read·karmabox.hashnode.dev5d ago