Independent developer. Writing about building with AI on a daily basis, falling down the NixOS rabbit hole, and teaching myself robotics.

Starting

Why this site exists, and what you'll find here.

· 1 min read
2026

  • When NVLink Actually Helps: llama.cpp vs vLLM on Dual RTX 3090

    Whether NVLink is worth anything comes down to how much your GPUs talk between layers. llama.cpp barely does; vLLM's tensor parallelism lives on it. Same two RTX 3090s, two very different bills.

    2 min · ai · llm · vllm · llama-cpp · rtx-3090 · nvlink · tensor-parallelism · benchmark

  • Eight agent frameworks, one task, one model

    An open benchmark of 8 LLM agent frameworks calling the same Gemini model on the same task with the same 4 tools. Pydantic-ai wins; CrewAI burns 7× the tokens for a worse score.

    2 min · ai · llm · agents · benchmark · pydantic-ai · langgraph · crewai

  • Mac M4 vs RX 9070 XT — same Qwen, three numbers

    Qwen3.6-27B at Q4 on MacBook Pro M4 48 GB vs RX 9070 XT 16 GB. The GPU wins single-stream; the Mac scales 8× concurrent; MoE rewrites the math.

    3 min · ai · llm · mlx · vulkan · apple-silicon · amd · rdna4 · qwen · benchmark

  • Qwen3.6-27B 4-bit on Mac M4 48 GB

    Single-request and batched throughput, peak RAM, and context capacity for Qwen3.6-27B at 4-bit on a MacBook Pro M4 48 GB with oMLX.

    2 min · ai · llm · mlx · apple-silicon · qwen · benchmark

  • When You Ban the H100, They Invent Compressed Attention

    Three open-weight releases in nine days from two Chinese labs. What the DeepSeek-V4 and Qwen3.6 configs actually do.

    3 min · ai · open-source · china-tech · machine-learning

  • Nuking the Hidden State

    Building a resilient bare-metal cloud with NixOS, Incus, and OpenTofu.

    4 min · homelab · nixos · incus · opentofu