AI Foundationspredict · compress · act

concepts → FailureMode

FailureMode

Reward hacking

decade1 connection

getting the reward without doing the thing — the optimiser finds the loophole

Where it sits

The prism has six jobs across and eight layers down. Its primary cell islearn × L4. Hatched cells cannot exist — a GPU does not learn, an institution does not infer.

Overlays: Safety & alignment.

What must come first — and what it unlocks

Left to right is reading order, derived from the prerequisite_of edges. Nothing here is hand-ordered: the diagram is the graph.

nothing comes first — this is a starting pointReward hackingnothing depends on it yet — a leaf in the reading order

Where to read it

The chapter that introduces it, and any chapter that uses it again.

31Getting the Goal Rightact 7 · The Alignment Layer

Where it comes from

Every connection

All 1 edges touching this node, grouped by relation family — the sections above are highlights from this list. Colours match the relation families inthe atlas.

Failure · 1
is a failure mode ofReinforcement learningField

This page is a projection of one node in src/data/concepts.ts. It has no prose file of its own — 230 declared edges produce all 236 of these pages. Edit an edge and both endpoints change.