Writing

2026-08-01 Running a 26B MoE on an 8 GB Jetson by streaming experts from SSD My Jetson Orin Nano has 8 GB of memory and Gemma 4 26B-A4B needs 13.3 GiB at Q4. I patched llama.cpp to stream the routed experts off the SSD instead, with logits bit-for-bit identical to the stock path, and recorded the model decoding on device.
2026-07-31 My context selector beat grep. An agent with grep beat it. I built cognitive-cache to pick which files an LLM should see, then spent two months measuring it. It beat naive grep by a statistically real margin, lost to a coding agent holding nothing but grep and read, and one of its six signals turned out to be worth exactly zero. Here is the whole arc, including the part where I killed it.
2026-07-10 J-space in practice: using Anthropic's Jacobian lens to decide what an LLM can forget Anthropic's global workspace paper introduced J-space and the Jacobian lens. I turned the workspace readout into a KV cache eviction signal that beats SnapKV and H2O across three Qwen models, and shipped it in EVOKE.