blankcheck
AI safety, Duolingo style

Don't hand AI a blank check.

Four levels of hands-on lessons on reward hacking, deceptive alignment, interpretability, and verifiable rewards - taught like a game, built for people who ship models.

Overall progress0 / 16 lessons

Level 1

Outer Alignment & Specification Gaming

Why a reward is never quite what you meant

Locked

Inner Alignment & Deception

What the model actually wants inside

Locked

Mechanistic Interpretability & SAEs

Reading the model's mind, neuron by neuron

Locked

Verifiable Rewards & RLVR

Rewarding what you can actually check

Progress is saved in your browser.