Method · Existential Risk
Re-examining the extinction odds
A list produced by an AI chatbot — and widely shared after Mark Cuban prompted it — puts humanity's odds of extinction over the next 25 years at roughly 3.8%, with unaligned AI supplying most of that and everything from engineered pandemics to asteroid strikes filling out the tail. Taken at face value it reads as a tidy ranking. Read with any discipline, it is two different documents wearing one format.
The ranking itself is defensible — it tracks the expert elicitation in Toby Ord's The Precipice. The precision is not: quoting "0.00001%" implies a calibration no one possesses. And the format hides the real story — these risks are not independent line items but a small, coupled network with one high-leverage node. The useful move is not to argue the third decimal of the AI figure, but to treat the human-made, coupled cluster as a wide-banded joint tail and mitigate where magnitude, coupling and reducibility overlap.
Read the full essay →
Hazardous-Facility Siting
"Existing" should not mean "exempt"
Much of how we regulate dangerous facilities turns on a single accident of timing: whether a site was built before or after the rules that would govern it today. Our work on the Rancho LPG butane facility in San Pedro illustrates the cost of that distinction. A facility that almost certainly could not be licensed under current environmental-impact and quantitative-risk-assessment standards continues to operate beside homes — largely because it was there first.
Open-ended grandfathering creates a perverse incentive: it rewards keeping an aging, poorly sited facility in place rather than confronting its risks. The fix is not to litigate history but to add sunset provisions — periodic re-evaluation against present-day siting and containment standards — so that "we've always been here" stops being a substitute for "we are safe here."
Read the Rancho LPG project →
Public Safety & Data
You can't manage what you don't measure
The United States has no systematic, federally supported dataset on school shootings and gun violence collected under a single grant authority. For years the Dickey Amendment chilled the research that might have produced one. The result is a public debate that too often confuses data, information, and opinion — and treats all three as interchangeable.
They are not. Data becomes information only once it is subjected to analysis, and policy made on opinion alone tends to swing between overreaction and neglect. Before we can argue productively about what to do, we need a shared, well-curated evidentiary record of what actually happens, how often, and under what conditions. Building that record — and mapping the full range of policy responses to it — is a precondition for serious decision-making, not a distraction from it.
Read the School Shooting project →
Method
Deciding under uncertainty: metrics, levers, objectives
Most disaster debates collapse three different questions into one argument. What are we measuring? What can we actually do? And what are we trying to achieve? Our framework keeps them separate on purpose: metrics describe the hazard and its consequences, levers are the interventions available to us, and objectives define what success looks like — all under real risks and uncertainties.
Separating them clarifies the trade-offs. It lets us ask whether a proposed lever actually moves the metrics we care about, and whether those metrics genuinely serve the objective — rather than optimizing something merely because it is easy to count. And because outcomes feed back to refine the metrics, the framework is a loop, not a line: every event is a chance to measure better next time.
Explore the interactive framework →