AI Foundationspredict · compress · act

concepts → Property

Property

Generalization

timeless6 connections

doing well on data you have never seen — the only thing that counts

Where it sits

The prism has six jobs across and eight layers down. Its primary cell islearn × L0. Hatched cells cannot exist — a GPU does not learn, an institution does not infer.

Overlays: Evaluation & methodology.

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.

Inductive biasGeneralizationnothing depends on it yet — a leaf in the reading order

Before it: Inductive bias

What kind of thing it is — and what it is made of

It is made of: Training set.

How it goes wrong — and what answers that

known failure modes
Overfittingmemorising the training set instead of learning the pattern
Underfittingthe model is too simple to capture the pattern, so it is wrong everywhere

Where to read it

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

11Why Memorizing Failsact 2 · The Instruments

Where it comes from

paperScaling Laws for Neural Language ModelsJared Kaplan, et al. · 2020

Every connection

All 6 edges touching this node, grouped by relation family — the sections above are highlights from this list. Colours match the relation families inthe atlas.

Structure · 1
has as a partTraining setDataset
Order · 1
requiresInductive biasProperty
Flow · 1
is measured byBenchmarkBenchmark
Contrast · 1
trades off withCapacityQuantity
Failure · 2
fails byOverfittingFailureMode
fails byUnderfittingFailureMode

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.