Thoughts, tutorials, and insights about software development, modern technologies, and problem-solving.
Turn the eval pyramid into a working TypeScript harness with deterministic graders, isolated repeated trials, calibrated model judges, human review, and release gates.
LLM and agent systems need more than unit tests, but that does not mean every check should become an LLM judge. The eval pyramid starts with deterministic checks and adds broader judgment only where the product needs it.
A system that succeeds once is capable. A system that succeeds repeatedly is reliable. Here is how to measure both, choose the right repeated-run metric, and use LLM judges without trusting them blindly.
A chat model only knows what it saw in training. RAG is the plumbing that lets it answer from your private documents, today's news, or anything it was never trained on: a second kind of model that turns text into coordinates, a database of those coordinates, and a search step that runs before the LLM ever sees your prompt.
A stack of transformer layers is not yet ChatGPT or Claude. This is the rest of the path: how text becomes tokens, how a raw next-word predictor turns into an assistant across three training phases, how LoRA customizes a model on a budget, and why everyone is racing for data and compute.
We use models built on transformers every day, yet the architecture itself usually stays a black box. It doesn't have to, and following it doesn't take heavy math. A visual, ground-up walkthrough: how words become numbers, how attention lets them shape each other, and how it all turns into a next-word prediction.
Reading notes from Addy Osmani on the difference between using AI as a tool and letting it quietly replace your thinking. Key research findings and heuristics worth keeping.
I've been shipping with AI tools for a year. At some point the gap between using them and actually understanding them stopped feeling fine. This is about what I decided to do about it.
The AI job market is noisy and competitive. Here are the ideas worth keeping from a talk by Andrew Ng and Laurence Moroney on careers, technical debt, and what it takes to ship real AI work.