AI Foundationspredict · compress · act

concepts → Algorithm

Algorithm

GRPO

also called group relative policy optimization

frontieras of 2026-093 connections

score a group of answers against each other, so no separate value network is needed

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: Efficiency, cost & energy.

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.

RLHFGRPOnothing depends on it yet — a leaf in the reading order

Before it: RLHF

Where to read it

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

45DeepSeekact 10 · The Open-Weight World

Where it comes from

Every connection

All 3 edges touching this node, grouped by relation family — the sections above are highlights from this list. Colours match the relation families inthe atlas.

Order · 1
requiresRLHFAlgorithm
Flow · 1
is used byDeepSeek-R1Model
Lineage · 1
improves onRLHFAlgorithm

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.