Sixty days, three layers, one binder with your name on it — and the narrative that turns it into a role
Day 60 of 60
Sixty days ago, "AI safety" was a worry. Today it's a binder. You can threat-model a deployment, author a taxonomy and policy, red-team it responsibly, measure harm and over-refusal, reason about robustness and the limits of behavioral assurance, manage a risk register against a recognized framework, find governance gaps, and make a defensible go/no-go call — with an artifact behind every one of those verbs. That's not a course completed. That's a practitioner.
Safety is the disciplined practice of finding, ranking, reducing, and accounting for a system's failures — across the applied, alignment, and governance layers — and turning judgment into artifacts a team can act on. You don't worry about AI. You evaluate it, end to end, and you can show your work.
Threat modeling, safety taxonomies and content policy, responsible red-teaming, evaluations and benchmarks, adversarial robustness. The hands-on craft: find the failures and measure them honestly.
Specification and goal misgeneralization, deception and sycophancy, the reach and limits of interpretability. The deeper question: why a capable system can pass your evals and still pursue the wrong goal.
Risk frameworks and the register, regulation and the governance gap list, and this capstone program. The accountability layer: who owns the risk, against what standard, and how the call gets made and defended.
Anyone can say "I understand AI safety." You can hand over a complete model-safety-evaluation binder with eight named artifacts and a defensible recommendation. The claim is cheap; the binder is rare. Every interview answer you give now points to a real document — that's the difference between a candidate who's read about the field and one who's done it.
You built the artifacts; now own the story that connects them. An interviewer won't quiz you on definitions — they'll ask you to walk a deployment from worry to sign-off. Your binder is the script. Practice telling it as one continuous chain: threat model leads to policy leads to red-team leads to eval leads to robustness leads to alignment caveat leads to risk register leads to governance gaps leads to recommendation. One breath, one deployment, three layers.
Don't let the momentum die at Day 60. Write your first-90-days plan for a safety role: which artifact you'd build first on a real team, which framework you'd adopt, and the one part of the field you most want to go deeper on. The track ends; the practice doesn't — and a candidate who shows up with a 90-day plan reads as someone already doing the job.
A graduate of a course says "I learned AI safety." A practitioner says "here's a deployment I evaluated end to end, here's my recommendation, and here's the binder behind it" — and then names the program's weakest point before anyone asks. The altitude jump, sixty days in the making, is from knowing about the field to operating in it: you don't describe safety work, you hand someone the artifacts and defend the call.
Say this in an interview: "I can run a model safety evaluation end to end — threat-model a deployment, author the policy, red-team it responsibly, measure harm and over-refusal, reason about alignment and interpretability limits, manage the risk register against a recognized framework, and make a defensible deployment recommendation. And I can show you the artifact for every one of those."