Labs · Tech blog
Long-form engineering write-ups — the full story behind the systems I build, from need to research to architecture. For the short-form stream, see Posts.
Subscribe via RSS.
17 August 2026 · Tool
The prose Claude writes for you is buried in tool calls and bash output. claude-reader tails the transcript file Claude Code already keeps, shows only that prose and your prompts, rendered as markdown, in a second terminal pane. No hooks, no server, works over ssh. Why the mouse is off by default, why the header has no clock, and the four crashers an outside review found before release. pipx install claude-reader.
Read the write-up →
15 August 2026 · Study
I read the dependency files of 703 starred GitHub repos tagged as AI agents. Raw SDKs beat all frameworks combined, more apps use LangChain's parts than LangChain itself, and production-grade apps keep frameworks slightly more often than demos do. Plus the search result that was wrong by 25x and still looked completely believable.
Read the write-up →
14 August 2026 · Explainer
A million-token context window does not mean the model can use a million tokens. The measured mechanisms behind context rot: attention sinks (delete the first 4 tokens and the model falls apart), positional decay, distractors, and why the needle-in-a-haystack chart on every model card is the easiest possible test. With the 5-minute experiment you can run yourself.
Read the write-up →
13 August 2026 · Benchmark
Qwen3.6-35B-A3B, 4-bit on one consumer card, lands at +0.005 [−0.01, +0.02]: statistically at the frontier, first local model at the line, perfect scores on four dimensions, and the whole run cost $1.24 in judge fees. Plus why the 35B MoE fits where the 27B dense model doesn't, two harness fixes worth stealing, and a correction I owed the tracker.
Read the write-up →
12 August 2026 · Benchmark
1.81x more output tokens for statistically identical quality, no measurable capability regression, and six of nine long-context tasks refused as cyber content on a corpus of synthetic office memos. Plus the bug in my own scorer that turned the model's new habit of appending a usage example into a clean zero.
Read the write-up →
10 August 2026 · RAG series, part 1
There is no RAG debate, there are five: grep vs vectors, long context, GraphRAG, CAG, and memory. I read ten viral "RAG in production" posts side by side, pulled the five debates apart, and found the boring baseline holding up fine, plus the parts the listicles oversell (chunking) and skip (parsing, indexing).
Read the write-up →
7 August 2026 · Benchmark
Three engines, three model sizes, one GPU. The datacenter engine came dead last on identical weights, then won the batch round five times over. Plus: how llama.cpp runs an 18GB model at 71 tok/s on a 16GB card, and the one env var every Ollama user should set.
Read the write-up →
7 August 2026 · Benchmark
The long version of the day-one run: the first model to land above my frozen frontier line, how it splits against Kimi K3 dimension by dimension, two failure stories I verified by hand, the 3x cost gap that's quietly the real headline, and three serving traps to know before you integrate it.
Read the write-up →
7 August 2026 · Open source
The story of project-brain: a long-standing pain (re-explaining my own projects to my own tools), a 30-day deletion discovery, a research pass through everything that already exists, and the layered plain-Markdown architecture we landed on — which turned out to match Karpathy's LLM-wiki pattern almost exactly.
Read the write-up →