Writing

ML, AI systems, research notes, and product observations. Articles can embed tweets, charts, dashboards, tables, code, and linked source tags.

Agents fail at orchestration, not reasoningFebruary 25, 2025RAG systems decay silentlyJune 6, 2025Evals are compressed production failuresJuly 22, 2025GraphRAG solves a different problem than RAGOctober 28, 2025MCP is a context budget protocol in disguiseDecember 15, 2025Observability is the primitive, not a featureSeptember 3, 2025Spend cheaply on hypotheses, expensively on verificationMarch 19, 2025Context engineering is memory managementJanuary 14, 2025The compliance stack is the engineering stackFebruary 9, 2026Rules should be learnable, not hard-codedApril 27, 2026Your model never reads text, it reads tokensNovember 14, 2023Embeddings are coordinates, not encodingsJanuary 9, 2024Attention is weighted averaging, and that's enoughMarch 4, 2024Temperature is not a creativity dialApril 22, 2024LLM knowledge has three shapesJune 17, 2024Hallucination is architectural, not a bugJuly 30, 2024Small-model behavior does not predict large-model behaviorSeptember 11, 2024Fine-tuning is a last resort, not a first moveOctober 21, 2024Chunking is choosing the unit of meaningApril 30, 2025Chain-of-thought rents computation with tokensDecember 3, 2024Gemini and the shape of multimodal evaluationFebruary 15, 2024RAG quality is a data problem firstAugust 19, 2024