问题，Questions,先于答案。before answers.

数据决定模型的上限。质量差的数据一定训练不出好模型，但“质量”必须能被拆开、被测量，才不只是一句口号。Data sets the ceiling on model quality. Bad data cannot train a good model, but "quality" has to be decomposed and measured before it means anything.

做过金融长文本、金法医等标注与质检，在供应商侧参与腾讯混元 HY4 的评测与质量管理：制定 Rubric、复核事实与计算、构造高难样本。这些经验是这个方向的起点，但还没有形成公开的数据实验。I have done annotation and QA on long financial text and forensic accounting data, and worked on evaluation and quality management for Tencent Hunyuan HY4 on the vendor side: writing rubrics, double-checking facts and calculations, and constructing hard samples. That work is the starting point for this direction, but it has not yet produced a public data experiment.

奖励信号和评测本身就是数据问题。训练在优化什么、评测在奖励什么，两者一旦对不上，提升就可能只是表面的。Reward signals and evaluation are themselves data problems. Once training optimizes one thing while evaluation rewards another, improvements can be surface-level only.

2025 年起在 Qwen、LLaMA 上做过 SFT 与 DPO 的练习与对比；在混元 HY4 评测中写 Rubric、处理资料冲突和异常结果，对“评测标准怎样变成训练信号”有第一手感受。Since 2025 I have run SFT and DPO exercises and comparisons on Qwen and LLaMA. During Hunyuan HY4 evaluation I wrote rubrics and handled conflicting sources and anomalous results, which gave me a first-hand feel for how evaluation criteria become training signals.

推理成本决定一个模型能被多少人真正用上。每一次重复计算、每一次缓存失效，最后都会变成用户等待的时间。Inference cost decides how many people can actually use a model. Every redundant computation and every cache miss ends up as time a user spends waiting.

在 Eazo 参与 Prompt Caching 与模型调用链路优化：连接复用、配置并行初始化、扩展增量同步。也做过 cctop，把本地 token 记录与服务账单对账。At Eazo I worked on prompt caching and the model-call path: connection reuse, parallel config initialization, and incremental sync for extensions. I also built cctop, which reconciles local token records against service bills.

同样的效果，换一个结构可能只要一半的算力。最值得研究的，是那些改变成本曲线、而不只是堆更多参数的设计。The same result may take half the compute under a different structure. What is worth studying are designs that change the cost curve, not just ones that add more parameters.

此前做过循环 Transformer 的小规模复现；正在调研 Jev 这类类型化决策模型。MLA、稀疏注意力与 MoE 目前停留在读论文和读实现阶段。I have previously done a small-scale reproduction of a looped Transformer and am now surveying typed decision models such as Jev. MLA, sparse attention, and MoE are still at the stage of reading papers and implementations.

在科学里，一个配平错误或一个不存在的反应条件，都会让模型带着同样的错。科学 AI 首先是数据质检问题。In science, one unbalanced equation or one impossible reaction condition teaches the model the same mistake. AI for science is, first of all, a data-QA problem.

做过化学合成数据的质检与错误归因，以及化学与学科资料的证据寻源。这是从数据角度进入科学 AI，还没有做过模型层面的科学实验。I have done QA and error attribution on chemical synthesis data, as well as evidence sourcing for chemistry and discipline-specific material. That is entering science AI from the data side; I have not yet run science experiments at the model level.

模型能力最终都落在基础设施上。执行离数据和环境越近，Agent 越快、越稳，也越容易被观察。Model capability ultimately rests on infrastructure. The closer execution is to the data and the environment, the faster, steadier, and more observable an agent becomes.

在 Eazo 参与 V4 到 V5 升级：把 Agent 执行从后端 Worker 迁到 E2B 沙箱，拆分控制面与执行面，前端经 WebSocket 直连沙箱；也参与 Skills 驱动的多 Agent 体系。At Eazo I worked on the V4 to V5 upgrade: moving agent execution from a backend worker into E2B sandboxes, splitting the control and execution planes, and connecting the frontend to the sandbox directly over WebSocket. I also worked on a Skills-driven multi-agent system.