Trust,
but Trace
A journal on what it takes to make ML systems trustworthy in production: traces, fixtures, contracts, replays. Written from the engineering side.
“Most of what an LLM agent does in production has never appeared in any evaluation set.”
Latest
ML fundamentals
Loss and gradients without the formulas: backprop is a graph walk
- 02From GPT-2 to Llama: four changes and the reason for each ML fundamentals · 11 min
- 03GPT-2 is the transformer plus five decisions, and inference is a different program ML fundamentals · 11 min
- 04A transformer block is six tensors and a bus: read it like a pipeline ML fundamentals · 12 min
- 05Attention is a learned weighted average, and the cost is in the square ML fundamentals · 12 min
- 06Shapes are the type system of ML, and nobody checks them for you ML fundamentals · 12 min
- 07An embedding is a lookup table, and everything else is how you fill it ML fundamentals · 11 min
Topics
Recurring threads
4 threads that keep coming back. Click one to see every piece tagged with it.
On the desk
Reading list
What’s on the desk this season.
- 01BookAI Engineering: Building Applications with Foundation Models
- 02PodcastLatent Space — Artificial Analysis on independent LLM evals
- 03BookMultimodal, Real-Time AI Agent Systems
- 04ReferenceEU AI Act, Article 15: Accuracy, robustness and cybersecurity
- 05BookHands-On Large Language Models
- 06PaperConstitutional Classifiers
- 07PodcastPragmatic Engineer — Inside engineering at frontier labs