The Brief Was Always the Expensive Part
Production hours collapse when you hand them to a machine. Specification hours don't. Remove the person and ambiguity returns as polished, wrong output.
Read ArticleBlog
Insights on privacy, sovereignty, and building technology that serves humanity.
- Part 1 Is Anything in Software Still Novel? 7 min
- Part 2 The Design→Test Loop Finishes Closable Problems. It Can't Touch Open Ones. 8 min
- Part 3 The Harness Is the Moat: Why That Doesn't Contradict 'Harnesses Aren't IP' 7 min
- Part 4 Human-in-the-Loop Is a Rubber Stamp Unless It's Adversarial 7 min
- Part 5 Allocate Architects by Irreversibility, Not Difficulty 6 min
- Part 6 Ask for Attacks, Not Tests 6 min
- Part 7 The Regression Ratchet Has No 'Done' 7 min
- Part 8 The Last Human Job Is Owning the Predicate 7 min
- Part 1 The Model Is the Easy Part 11 min
- Part 2 Your Hallucination Is a Retrieval Problem 10 min
- Part 3 Users Don't Want Magic. They Want Scaffolding. 10 min
- Part 4 Evals Are the Product Spec 10 min
- Part 5 Stop Upgrading the Model. Upgrade the Pipe. 10 min
- Part 6 The Hardest Part of Agents Is the Plumbing 10 min
- Part 7 You Have a Data Hygiene Problem Masquerading as a Model Problem 10 min
- Part 8 The Best Prompt Engineer Is a Domain Expert 11 min
- Part 9 You're Comparing AI to a Human Standard That Doesn't Exist 13 min
- Part 10 Confidence Is a Liability 10 min
- Part 11 Pilots Lie. Production Tells the Truth. 11 min
All Posts
Nothing Your Business Does Is Novel
Roughly 90% of business-model innovation is recombination. The new part of your company lives in its constraints, and that decides what a machine can own.
Is Anything in Software Still Novel?
Component-level novelty went nearly extinct. The industry's novelty budget moved into constraints, and that decides how to use AI on a real codebase.
The Model Is the Easy Part
Three years of production LLM work moved the leverage downstream of model selection. Teams treating it as the main quality lever optimize the cheapest part.
Your Data Isn't a Moat. Your Loop Is.
A contrarian read on the AI industry's favorite defensibility argument, plus a three-filter test for telling a real moat from data hoarding theater.
Closable vs. Open: Which Business Problems AI Can Actually Finish
AI loops finish problems with countable inputs, a measurable finish line, and no adversary who moves after you commit. Everything else it can only amplify.
The Design→Test Loop Finishes Closable Problems. It Can't Touch Open Ones.
What decides whether an AI loop can own a problem end to end is not difficulty. It's whether the problem closes, and an adversary keeps it open forever.
Your Hallucination Is a Retrieval Problem
Roughly three-quarters of hallucination complaints in production RAG are retrieval failures, not generation failures. A better model is the wrong fix.
The Half-Life of a Prompt Is Shorter Than Your NDA
The industry's prompt-protection playbook rests on a layer the Copyright Office already found doesn't confer authorship. Here's where real protection lives.
The Scorecard Is the Moat
When production collapses to near-zero, leverage moves to whoever can turn a fuzzy goal into a trusted measurement. Your definitions of success are the asset.
The Harness Is the Moat: Why That Doesn't Contradict 'Harnesses Aren't IP'
Trading engineering hours for tokens is not a linear trade. Implementation collapses to zero; specification and validation surface all at once, up front.
Users Don't Want Magic. They Want Scaffolding.
The biggest factor in LLM adoption isn't model quality. It's whether the experience produces calibrated trust, and the research is clear enough to design to.
Agents Don't Have IP. Workflows Do.
The agentic-AI gold rush extended the IP framing onto a layer where it makes even less sense, and is producing very expensive prompt wrappers.
The Brief Was Always the Expensive Part
Production hours collapse when you hand them to a machine. Specification hours don't. Remove the person and ambiguity returns as polished, wrong output.
Human-in-the-Loop Is a Rubber Stamp Unless It's Adversarial
A passive monitor of a reliable process is a rubber stamp with a person's name on it. Oversight that doesn't attack isn't oversight. It's laundering.
Evals Are the Product Spec
Teams that ship reliable AI products have eval harnesses. Teams that ship and pray do not. Here's the maturity ladder, and why most are on rung zero.
What the Lawsuits Are Really About
NYT v. OpenAI, Bartz v. Anthropic, Thomson Reuters v. Ross. Everyone is watching the doctrine. The market has been telling a different story the whole time.
Allocate Architects by Irreversibility, Not Difficulty
Hard problems are now cheap to retry. The cheap-looking calls (a field's type, an event's ordering) are the ones you can't undo. Most orgs route backwards.
Stop Upgrading the Model. Upgrade the Pipe.
Model upgrades produce roughly 3% quality gains. Engineering around the model produces 28-47%. Most teams are optimizing the wrong layer.
Ask for Attacks, Not Tests
Test generation samples the happy path: coverage theater. Vulnerability finding is adversarial search with a real target. The prompt is the difference.
The Hardest Part of Agents Is the Plumbing
Agent failures look like reasoning failures and are almost always integration failures. Tool design, auth and rate limits drive 80% of agent quality.