Recent Themes
A summary of what I've been writing about lately, generated from my 10 most recent essays using Claude on AWS Bedrock. Updated every time the site is published.
The tame/wicked frontier of AI. AI advances fastest on problems with checkable answers — code, engineering, analysis — and stalls at the boundary where standing, accountability, and competing interests must be settled. That boundary doesn't disappear as capability grows; it sharpens. (The Wicked Frontier)
Measuring AI by outputs, not inputs. Tokens and licenses measure consumption, not value. The right unit of ROI is the full cost of producing an accepted, verified result — and that framing lets organizations compare providers, surface real tradeoffs, and separate the machinery from what it actually delivers. (The Unit of Return, The Field and The Frontier)
AI compounds through evaluation, not generation. Generation is now nearly free; the constraint is the cost of knowing whether the work got better. Organizations that build persistent evaluators — rubrics, test cases, structured feedback loops — accumulate judgment as a durable asset. Those that only automate output are running a checking account, not a savings account. (The Cost of the Answer)
Assumptions expire faster than organizations update them. AI is shortening the half-life of decisions about what technology can do, which workflows to automate, and which models to use. The fix isn't better forecasting — it's building a notice system: recording why decisions were made, setting explicit triggers for revisiting them, and treating correction as routine rather than exceptional. (The Half-Life of an Assumption)
Adoption requires a fair exchange, not just a working tool. Technologies become infrastructure when they return enough value, visibility, and control to the people asked to rely on them — not just to the organization deploying them. Efficiency gains on one side that remove legibility on the other stall adoption until the bargain is rebalanced. (The Barcode Bargain, How This Was Made)
As AI handles more execution, human judgment concentrates rather than shrinks. When agents can run processes autonomously and surface unlimited plausible options, the scarce resource becomes the willingness to choose — to name what the work is for, stand behind that direction, and own what follows. That question is the one no system can answer. (The Silent Governor, What The Garden Is For)