一个人、两周、数百美元训出榜首小模型：数据工程、聪明蒸馏与主权 AI

A single researcher can now post-train a competitive small model in weeks rather than needing a traditional AI lab.

A model can improve sharply on a target benchmark while losing performance on unrelated evaluations.

The hardest part of personal model training is data judgment and execution, not merely running a training script.

Synthetic data is useful only when it targets real model errors instead of multiplying the model's existing biases.

A robust fine-tuning mix may require a majority of real data, with synthetic examples used to repair specific weaknesses.

The capacity gap between teacher, student, and task can make data effectiveness difficult to predict in advance.

Smart distillation selects teacher outputs and training signals around the student's actual failure modes.

Training your own model makes sense only when inference savings, control, and proprietary value exceed data and maintenance costs.

Application companies with proprietary workflows and feedback data may evolve into small, domain-focused AI labs.

Local AI could become a PC-scale opportunity if privacy, latency, cost, and sufficient intelligence align.