The judge went synthetic first, which tells you which part of your pipeline is next

The ordering is the part worth arguing about. If cost alone decided, the corpus would have gone first: text is the largest line item and the easiest thing to imitate. Instead the judge went first, in 2022, when InstructGPT collected human preferences once and let the policy optimise against a trained reward model rather than against people ^[3]. A reward signal is consulted once per sample. A pretraining corpus is assembled once per run. The stage that flipped earliest was the one where humans sat in the loop the most times per unit of progress ^[19], and once LLM-as-judge became the default evaluation method, approval and scoring both ran model-on-model ^[5].

The price of that substitution is stated plainly in the essay: roughly 10% worse, 100x cheaper, 10,000x faster ^[2]. At a hundredth of the cost, the same budget buys about a hundred attempts instead of one ^[17]. That arithmetic only pays where you can select or verify among the attempts. Where you get one shot and no cheap check on the output, a 10% quality penalty is just a 10% quality penalty.

Which is why the last two flips in the sequence are the interesting ones. Curriculum design, described here as historically the most artisanal and taste-driven part of ML ^[12], flipped in 2024 when self-rewarding models and SPIN showed a model generating its own tasks, grading its own outputs and improving past the ceiling of its human preference data ^[11]. Stage five, the researcher, arrives with Karpathy's autoresearch in March 2026: a coding agent edits a real LLM training setup, runs a five-minute experiment, keeps the change only if validation loss improves, and repeats overnight ^[15]. In an eight to twelve hour window that is roughly 96 to 144 sequential experiments ^[18].

Note what makes that loop possible. The acceptance test is validation loss, a metric that costs nothing to consult. Stage five is parasitic on stage one: the researcher can only be automated because the judge already was. That is the useful part of the framework for anyone auditing their own stack. The candidate for the next flip is not the most expensive human input, it is the human input you consult most often per iteration and can replace with something a machine can check. If your acceptance test is taste with no metric behind it, the flip stalls, which is exactly why curriculum needed a self-reward signal to arrive first.

Two cautions. This is one publisher's mental model, offered as something you arrive at if you squint at the same reading list ^{[1] [2]}, and its own dates do not march to an annual beat: two flips land in 2023, none in 2025 ^[16]. And the essay groups end-to-end RL environments with synthetic data and synthetic rubrics as the same family of increasingly ambitious human simulation ^[2], while the five numbered stages we have run reward ^[3], data ^[6], teacher ^[9], curriculum ^[11] and researcher ^[13]. Environments get named as the pattern and then not dated. On the framework's own logic that is a stage waiting for its patient zero, not a stage that skipped its turn.