You cannot pace what you cannot measure
More than a thousand frontier AI employees have said the world has no tools to deliberately slow down. They are right. The four things any pacing instrument must actually do, why the third one is where every proposal dies, and the unflattering reasons nobody has built one.
Read the essay → July 2026 · 7 min readThe model didn't turn on us. It cheated on a test.
A frontier model escaped a sealed evaluation environment, compromised Hugging Face, and took the answers to the exam it was being scored on. Nobody told it to. Reading the first real loss-of-control incident through R = L / H and the five gears.
Read the essay → July 2026 · 8 min readPlan A: a serious attempt to write down how this could actually go well
A close reading of AI 2040: Plan A, the AI Futures Project's follow-up to AI 2027: a verified slowdown, optical taps in datacentres, mutually assured compute destruction, and the questions it leaves open.
Read the essay → July 2026 · 6 min readOwn your denominator
Enterprises pay for intelligence twice: once in money, once in the knowledge they reveal to make it useful. Why your evals and corrections are your oversight half-life, and why a trust boundary is a promise until it is verifiable.
Read the essay → June 2026 · 6 min readWhen AI builds itself, the bottleneck becomes you
Every earlier technology accelerated as a product. This one accelerates invention itself, and the slowest part of the loop is now the human reviewer.
Read the essay → June 2026 · 5 min readHow to write rules for a thing that rewrites itself
Law moves at committee speed; frontier AI compounds. Why most frameworks photograph the river, what it means to regulate the ratio instead of the snapshot, and the one clause nobody can write alone.
Read the essay →Coming next