Steal the fire from complexity. Distill it. Give it away.
Fire,
distilled.
I go into complex systems, ideas and technologies, find what actually matters, and bring it back in a form humans can understand.
Current thesis
Code is becoming easier to produce. Understanding what should be built, and why, is not.
Complexity, on purpose → structure, on demand
Retrieval finding meaning
- embeddings
- vectors
- chunking
- semantic search
- BM25
- HNSW
- reranking
- recall@k
Inference running models
- tokens
- attention
- context windows
- latency
- caching
Agents acting on purpose
- tool calling
- planning
- memory
- structured outputs
Reliability holding under load
- queues
- retries
- backpressure
- indexes
- transactions
Evaluation knowing it works
- groundedness
- nDCG
- MRR
- evals
Recently distilled
Attention
Attention lets every token look at every previous token, so the work grows with the square of the context, while the weights in each row must still sum to one.
Tokenization
A language model does not read characters; it reads tokens, so counting letters is a reconstruction task, not a lookup.
From the notebook
What I mean by “distillation”
A distillation is not a simplification. Nothing is lost. It is a separation: the idea is pulled apart until the part that carries the meaning sits alone, and the rest is understood as packaging.
grex
One workspace to run Claude, Codex, Cursor and Gemini agents in parallel — and keep their work from colliding.
Not a portfolio. A laboratory where I learn things, then explain them.
Start with a distillation