Blog
Notes on building and running AI.
Essays and guides on AI you can trust, how to measure what it changes and where to start in your business.
The Barbell Is a Mirror
A fashionable argument says most workers will never get good at AI. The economics say it mistook one failure mode for a law of nature.
The Body Around the Brain: The Harness
A model that can call tools still is not something you would leave alone with real work. This is the story of the harness, the program around the model that turns a sequence of guesses into a job you can trust.
The Model With No Hands: MCP
A language model can reason brilliantly about your business and cannot touch a thing inside it. This is the story of the standard socket that gives it hands, and when it is worth wiring your own.
How Do You Know It Got Better?: Evals
You change the prompt, the demo still works, and you ship it. This is the story of how to replace that feeling with a number, and why anyone building with AI eventually has to.
The Counterfactual Problem
Every claim about AI’s return on investment hides the same question: compared to what? The organisations best equipped to answer it keep forgetting to ask.
This Upgrade Is Not for You
Every few months a frontier lab ships a new model and the timeline declares that everything has changed. For most of the people watching, Monday morning will be exactly the same.
The AI glossary: 110 terms in plain English
Plain-English definitions of 110 AI terms, from agents to world models, in 11 groups you can scan in minutes.
Why Discovery beats a sales pitch
Most AI consulting starts with a pitch and a proposal full of assumptions. We start with Discovery, a paid audit of how one workflow runs today, so what we build and run for you rests on your own numbers.
Before you automate a workflow: a practical checklist
Five areas to check before anyone builds AI into one of your workflows, and what we do with your answers in Discovery.