Editorial read
In depth · editorial + model
Model-generated corpora, self-play traces and verifiable problem sets used to augment or replace scarce human data in training and post-training. It is becoming the primary way to keep scaling training and to manufacture the verifiable data that reasoning RL requires. Credible projections put exhaustion of high-quality human text in the late 2020s, so frontier labs now lean heavily on synthetic and self-generated data. Low-quality or unfiltered synthetic data degrades models (model collapse), quietly capping capability gains.