Writing

ML, AI systems, research notes and product observations — written while the work was still warm. 22 notes.

2026

Rules should be learnable, not hard-codedNeuro-Symbolic27 Apr The compliance stack is the engineering stackEU AI Act9 Feb

2025

MCP is a context budget protocol in disguiseMCP15 Dec GraphRAG solves a different problem than RAGGraphRAG28 Oct Observability is the primitive, not a featureLLMOps3 Sept Evals are compressed production failuresEvaluation22 Jul RAG systems decay silentlyRAG6 Jun Chunking is choosing the unit of meaningRAG30 Apr Spend cheaply on hypotheses, expensively on verificationInference19 Mar Agents fail at orchestration, not reasoningAgents25 Feb Context engineering is memory managementPrompting14 Jan

2024

Chain-of-thought rents computation with tokensReasoning3 Dec Fine-tuning is a last resort, not a first moveFine-tuning21 Oct Small-model behavior does not predict large-model behaviorScaling11 Sept RAG quality is a data problem firstRAG19 Aug Hallucination is architectural, not a bugReliability30 Jul LLM knowledge has three shapesRAG17 Jun Temperature is not a creativity dialSampling22 Apr Attention is weighted averaging, and that's enoughTransformers4 Mar Gemini and the shape of multimodal evaluationMultimodal15 Feb Embeddings are coordinates, not encodingsHugging Face9 Jan

2023

Your model never reads text, it reads tokensTokenization14 Nov