AI Safety
Where I'd Put the Next Hour of Work: A Gap Map of AI Safety
Where would an hour of independent work do the most good in AI safety? A gap map, a test for my main hypothesis, and how I would choose a first project.
Read Article→
READ
BRAND DRIFT
Field notes on brand identity in the age of AI.
AI Safety
Human Oversight Is a Design Problem, Not a Job Title
A person in the loop is not a safeguard by definition. The arithmetic of how much review it takes to catch a rare failure, and what a checkable oversight claim looks like.
Read Article→
AI Safety
Safety Evaluations Don't Travel: The Language and Context Gap
Safety tests built in English may not travel to Darija, Arabic or French. The evidence, what is missing, and a pre-registered pilot I would run.
Read Article→
AI Safety
AI Has No Near-Miss System, and Aviation Does
Aviation learns from near-misses. AI mostly learns from screenshots. What a near-miss report for deployers would need, sketched against three public chatbot cases.
Read Article→
AI Safety
Compute Thresholds: What They Catch and What They Miss
A compute threshold is a tripwire for paperwork, not a verdict on risk. The arithmetic of why fixed lines age, and a layered design that expects to be revised.
Read Article→
AI Safety
What a Passing Eval Score Does and Doesn't Tell You
A model passed its dangerous-capability test. What can you conclude? Four ways a pass misleads, and a seven-question rubric for reading one.
Read Article→
AI Safety
Deciding When You Don't Know: How Policymakers Can Act While AI Risks Stay Uncertain
The Collingridge dilemma, two California bills, and what a decision record for rules made under uncertainty could look like.
Read Article→
AI Safety
Open Weights Are Not Open Source, and the Difference Matters for Safety
Open weights, open source and the EU's exemptions are three different things. What the difference means for safety, and how I would structure a release decision.
Read Article→
Never miss a post