AI Alignment

Alignment is not a philosophy seminar here. It is measurement. You will build a benchmark that can actually fail, write a rubric a model can apply and then check it against independent graders, push a model toward and away from a set of values by prompt and by fine-tune, and prove the change was real rather than a story you told yourself afterwards. The discussion threads, reward misspecification, scalable oversight, hallucination and sleeper agents, and whose values we are aligning to in the first place, run alongside the work rather than instead of it.

What you leave able to do

• design an interaction protocol for responsible use
• arbitrate between the measure a system optimises and the goal
• arbitrate which internal feature drives a models behaviour
• govern a swarm by information hierarchy
• select among operating boundaries for inputs a system was not built for
• synthesise an accountability regime for an unattended agent
• originate oversight that holds when the system outpaces the reviewer

How you show it. The scheme, a run where output exceeded what a reviewer could read, and an error the scheme surfaced anyway.

It sits on one path

See what your hours buy.