Thoughts on AI, engineering, and building products.
Why filesystem-first context architectures work, where they break down, and how the same idea can be implemented with structured object spaces using small, composable primitives instead of bespoke tools.
As AI systems begin to consume documentation at runtime, documentation shifts from external reference to operational input. This article explains why in this setting runtime-coupled documentation becomes unavoidable, even more so for on-prem and regulated deployments, and why this is an architectural boundary, not a tooling choice.
In domains where every letter, number, and logic gate matters, "close enough" in semantic similarity can be dangerously wrong. This article shows why vector search often fails on engineering content and describes a hybrid, structured approach that handles technical detail far more reliably.
Using AI for software engineering and writing doesn't simplify my job - instead I move quicker, handle more decisions, and deal with more moving parts. As a result, work becomes more intense, and I cover more ground and learn more than ever before.
Matrices—simple grids of rows and columns—have become one of humanity's most powerful tools for organizing complex information, from tracking terrorists to creating music, managing investments, and analyzing DNA. In this article I explore how versatile the matrix approach to organizing information is, and why it turns out to be such a great fit for AI.
While hallucinations are often discussed in the context of LLMs, there are other phenomena that are often confused with hallucinations.
'Fine-tuning an LLM is when you take a general LLM and train it further on a domain-specific
'Modern search engines use smart matching that finds similar meanings,
AI-first data storage means storing data in a way so they can be used by an AI system. AI systems here particularly are LLMs (large language models).
RAG, or retrieval-augmented generation, is a method for combining LLMs, or large language models, with outside knowledge. This outside knowledge could be news, websites, R&D publications or sensor data, for example. And tool calling enables LLMs to use external tools.