The premise is that pretraining scale determines a soft ceiling on capabilities (including the scale of the biggest models determining how good the smaller distilled models can get), so that a company that's consistently behind on pretraining scale is going to have a bad time even after post-training gets reasonably less broken. Most benchmarks are observing RLVR quality, with new RL tasks/envs being a major input, but as an industrial/economic constraint on the LLM company they are easier to improve than the amount of compute and the competence/resolve to turn it into model scale. So I intend this as a leading indicator of where the models will be in 1-2 years, not a predictor of current benchmark competitions.