My Why
The AI failure I actually worry about, and what I do about it.
When people talk about AI risk, they usually picture something sudden and dramatic. The part I actually worry about is quieter and slower. AI is already moving into companies, codebases, legal work, and public institutions, faster than we’re making it reliable. Nothing blows up. It just starts making more of the decisions, a little at a time.
That’s when it gets real. When an AI speaks for a company, writes the code, or advises someone, its mistakes stop being theoretical. People trust answers that are wrong. Automation runs confidently past its limits. Guardrails get quietly bypassed. And the part that worries me most is how our own judgment fades: we hand off one decision, then another, until we’ve given up more control than we meant to, and it’s hard to get back.
So the work I care about isn’t the model itself, it’s the part that gets skipped: making AI trustworthy as it takes on real responsibility. Some of that is research, figuring out where our safeguards actually break. Some of it is building the systems that keep an agent inside clear limits, easy to inspect, and possible to stop. And a lot of it is measuring whether the human oversight we lean on holds up under load, because usually nobody checks. That’s the thread through everything I do. Headroom measures where oversight breaks. OperatorLock enforces limits you can’t override in the moment. And at AI Objectives, I worked on keeping people in the loop at scale.