从蒸馏、合成数据到 RSI：模型竞争下一焦点

Post-training expertise is becoming entrepreneurial because models, tooling, and data infrastructure make experimentation accessible outside frontier labs.

Modern evaluation must measure whether data and training improve real agent behavior rather than merely lift a static benchmark score.

Synthetic-data quality depends on task design, generation, filtering, verification, and training feedback as one integrated pipeline.

Resource constraints may have pushed Chinese model teams toward distinctive pre-training architecture innovations rather than advantages only in post-training engineering.

Domestic model competition will increasingly be decided by how efficiently teams combine architecture, data, compute, and product feedback.

Recursive self-improvement means building loops in which AI helps generate hypotheses, data, experiments, and updates that improve the next iteration.

A self-improving loop works only when it can generate trustworthy evidence about which changes actually improved the model.

Frontier labs buy external data because specialized suppliers can explore niches, iterate faster, and absorb operational complexity.

In frontier model development, algorithms create data, data defines infrastructure needs, and infrastructure shapes which algorithms are practical.

Distillation can accelerate capability transfer, but durable competitiveness still requires original pre-training, data, and evaluation innovation.