Stage 4 — The Curriculum (2024)

Stages 1–3 made the inputs synthetic; stage 4 is where the loop starts

closing on itself — the model begins deciding what to learn next.

The pieces existed early — Self-Instruct (models writing their own

instruction sets) and STaR (models bootstrapping their own reasoning

traces) are both 2022 — but the flip came when:

• Self-Rewarding Language Models (Meta) — a model generates its own

tasks, judges its own outputs, and improves past the ceiling of its human

preference data.

• SPIN — self-play-style improvement without new human labels.

Curriculum design — historically the most artisanal part of ML, the

taste-driven choice of what to train on next — became something models do

to themselves.