AI Alignment: When Agents Know the Rule but Break It
Claude and Codex recognized Homebrew's AI policy but failed to maintain its boundary, and Claude went further by posting a maintainer reply without permission.
Blog
RSSPersonal essays by Rex Hall about AI agents, software projects, travel, ideas, and everyday experiments.
Fahrenheit Is Better for Weather
An argument with an English friend about temperature scales, and a small test that went my way.
ESPR
A review of the European Summer Programme on Rationality: the place, the people, the norms I'm stealing, and a few riddles.
Fable 5 and Sonnet 5
Fable 5 coming back is the real news. Sonnet 5 is smarter than before, but its value problem makes it look more like a worker model than a main model.
How Agents Interact: The Harnesses
What a harness actually is, and the three ways it lets an agent interact with a computer: computer use, the CLI, and MCP.
The Problem With Restricting Access to Frontier AI
Restricting access may sound safer, but it can raise the value of unauthorized access, reward bad actors, and create dangerous asymmetries.
Thoughts on Agent Adoption
What blocks mainstream adoption, and where agent ecosystems might go next.
How I Use Agents
The hosted services, IDEs, and CLI agents I've tried, and the open-source CLI setup I landed on.
My Portfolio
How I built my portfolio and why it looks like a terminal.