Learning · Efficiency

Stop bouncing on jargon.

One click adds plain-English hover definitions of AI safety terms — “RSP”, “scalable oversight”, “mesa-optimiser” — to any article you read.

See it in action

This is what an article looks like with the glossary on. Each tinted term below shows a plain-English definition when you hover it — you never leave the page.

Researchers working on AI safety treat AI alignment as the core challenge: making sure a frontier model pursues the goals we actually intend. Techniques like RLHF help shape behaviour, but they don't solve scalable oversight — how we supervise systems smarter than ourselves. Open problems include reward hacking, a mesa-optimizer pursuing goals of its own, and deceptive alignment, where a model hides its intentions. Getting this wrong is part of why some worry about existential risk from advanced AI, and why mechanistic interpretability tries to open the black box.

What it is

A tiny bookmark that highlights known AI-safety terms on whatever page you're reading and shows a short definition (and its source) when you hover. You stay in the article — no new tabs, no looking things up.

Nothing installed, no permissions, no tracking; it runs only when you click it.

Install it (once)

Show your bookmarks bar if it's hidden — Ctrl+Shift+B (Windows) or ⌘+Shift+B (Mac).
Drag the button below up onto your bookmarks bar.
That's it — open any article and click the bookmark to read it with definitions.

AI Safety Glossary

Like this — drag it up to the bar.

Can't drag it? Add it by hand instead.

Copy the code below, then right-click your bookmarks bar → Add page…, name it “AI Safety Glossary”, and paste the code as the URL.

Try it now

You've installed it — give it a spin. Click your new AI Safety Glossary bookmark and the terms in the sample article below will light up. Hover one to read its definition.

Why labs test their models before release

Before releasing a new model, a lab puts it through careful evaluation. Red-teaming probes for ways the system could be turned to misuse, and interpretability research tries to explain why the model behaves as it does. Training with RLHF nudges it toward helpful answers, yet a model can still learn reward hacking — scoring well on a test without actually doing the intended task.

Work like this matters more as systems edge closer to AGI. Testing for robustness and alignment is one of the few ways we have to catch problems early, long before a model reaches millions of people.