Blog
- Qwen3.8 on One Card ·Part 1
Fitting Qwen3.8-27B on One RTX 3090: The Memory Math Is Wrong in Both Directions
A 27B model with 262K context on a single 24 GB card. The obvious memory arithmetic overestimates the cost by 4x, then underestimates the ceiling. What actually fits, how fast it runs, and where the efficiency knee sits.
- Evaluation Methodology ·Part 1
Behavioral and Mechanistic: AI Evaluation's Missing Loop
Behavioral evals and mechanistic interpretability are usually framed as two camps. They are not. They are two halves of a method neuropsychology has been refining for decades.
- Home AI Lab ·Part 2
Building a Home AI Inference Node, Part 2: Adding llama.cpp and Going From Chat Server to Research Workstation
Part 1 made the node stable. Part 2 makes it useful for research: adding llama.cpp as a second runtime, GGUF model control, and the OpenAI-compatible API that ties it together.
- Home AI Lab ·Part 1
Building a Home AI Inference Node, Part 1: WSL, Ollama, and the Windows Problems Nobody Mentions
What I actually ran into turning a Windows desktop with an RTX 3060 Ti into a persistent local AI server accessible from my MacBook over LAN, and why the hardest problems had nothing to do with AI.