学习模型蒸馏与专家混合(MoE)技术
研究Embedding空间的几何特性（参考'universal geometry of embeddings'论文）
了解不同训练阶段（预训练/微调/RLHF）对模型能力的影响
实践LoRA等参数高效微调方法
06给 Codex 的 Prompt
Create a technical workflow for building a modular AI system with these steps: 1) Identify core knowledge domains needed 2) Design adapter mechanisms for combining modules 3) Implement knowledge distillation pipeline 4) Test cross-domain reasoning retention 5) Optimize for local deployment. Include Python pseudocode for the module integration mechanism.