Making an AI bigger keeps improving its reasoning long after its variety of responses has already stopped growing.

This explores why capability and diversity stop improving at different model scales, and the corpus suggests the two aren't even measuring the same thing — they saturate separately because they're driven by separate mechanisms. The clearest evidence comes from decomposing 'capability' into distinct skills: when you split model performance into 12 separate skills, they scale on wildly different curves — metacognition saturates around 7B parameters while logical efficiency keeps climbing to 30B, and reasoning and knowledge improve continuously past both Do all AI skills improve equally as models scale?. So 'saturation' is never one event. Surface-level skills (style, form) flatten early because they're cheap to imitate; substance skills (reasoning) keep rewarding scale. Diversity behaves like the surface-skill camp — it's about the shape of the output distribution, not the correctness of any single output, so it hits its ceiling on a different schedule than the hard reasoning skills.

The deeper reason they decouple is that diversity is often destroyed by the very training that boosts capability, independent of scale. Outcome-based RL sharpens the policy toward correct answers, and that sharpening bleeds diversity loss from solved problems onto unsolved ones — capability rises while the distribution collapses Does outcome-based RL diversity loss spread across unsolved problems?. Relatedly, RL post-training converges on a single dominant output format within the first epoch, and — crucially — which format wins depends on model scale, not on which format performs best Does RL training collapse format diversity in pretrained models?. That's a direct answer to the 'different scales' puzzle: the mechanism that suppresses diversity is scale-sensitive in a way that's orthogonal to the mechanism that lifts capability.

The learner's real surprise here is that quality and diversity aren't two dials on the same machine — they produce distinct downstream effects. One study separates synthetic-data properties and finds quality drives in-distribution generalization while diversity drives out-of-distribution generalization, and that collapsing them into a single metric is exactly what makes self-improvement loops silently degrade How do quality, diversity, and complexity affect synthetic data differently?. If diversity buys generalization to new territory and capability buys accuracy on familiar territory, there's no reason their saturation points should line up — they're paying for different things.

This also explains why 'just make the model bigger' doesn't rescue diversity. Pure self-improvement stalls specifically on diversity collapse and the generation-verification gap, not on raw capability — reliable gains require smuggling in an external anchor (a judge, a past model, tool feedback) rather than more scale Can models reliably improve themselves without external feedback?. And the tradeoff isn't even fixed: whether preference tuning shrinks or grows diversity flips by domain (RLHF compresses diversity in code but expands it in creative writing), because each domain incentivizes convergence or distinctiveness differently Does preference tuning always reduce diversity the same way?.

The thing worth carrying away: some of what looks like a saturation ceiling is a measurement artifact of how we aggregate. The exploration-exploitation 'tradeoff' in RLVR nearly vanishes once you stop measuring at the token level and look at hidden-state rank instead — the two can be enhanced simultaneously Is the exploration-exploitation trade-off actually fundamental?. So before concluding capability and diversity have hit fixed and different ceilings, it's worth asking whether the ceilings are properties of the model or of the metric.