我喜欢做真正走进生产环境的东西。在小红书做大模型对齐和 Agent 产品、在荷耳智能做 LLM 策略产品、在阶跃星辰做 AIGC 产品、在 TikTok 做内容生态——每一段都让我更理解 AI 产品是怎么长出来的。I like building things that actually ship. LLM alignment and Agent products at Xiaohongshu (RED), LLM strategy at an AI-hardware startup, AIGC product at StepFun, content ecosystem at TikTok — every stop taught me a little more about how AI products actually grow.
本科到硕士一路读人类学,中间在 UC Berkeley 交换。业余把观察人的方法拿来观察 AI:写了首篇 Agent Harness 系统综述(GitHub 200+ stars),也在社会学年会上讲 AI 如何反向规训创作者。Trained as an anthropologist from undergrad through my master's, with an exchange at UC Berkeley. Off hours I turn ethnographic methods onto AI itself: first systematic survey on Agent Harness (200+ GitHub stars), plus a sociology-conference talk on how AI disciplines creators.
19 岁做出过百万粉丝的抖音矩阵。工作之外喜欢葡萄酒、植物,和文化有层次的城市。At 19 I grew a Douyin matrix past a million followers. Off work: wine, plants, and cities with layers.
重度 AI native——这个网站和里面的每个 demo,都是我指挥 coding agents 一起做的。Heavily AI-native — this site and every demo on it were built by me and my coding agents, together.
四段实习,一个百万粉丝的起点。从做内容的人,到训练模型的人。Four internships, one million-follower origin story. From making content to training the models that read it.
负责高流量搜索场景的答案理想态、标注标准与评估体系,从 rubric、训练数据到上线验证全链路;顺手用 vibe coding 给团队造内部工具和产品 demo。Own answer-quality design, annotation standards and eval systems for high-traffic search scenarios — the full loop from rubrics and training data to online verification; build internal tools and product demos via vibe coding along the way.
设计剧情、人设与用户选择经 LLM 工具调用转成实时硬件指令的端到端框架;写了一层不依赖模型智能的 "Runway 过滤" 兜底机制。Designed the end-to-end frame turning narrative, persona and user choices into real-time hardware commands via LLM tool calls; built a post-generation "runway filter" safety layer that doesn't count on model intelligence.
站内创作者生态的数据监测与热点归因;和安全团队一起优化人审 + 机审的联动策略(HITL)。Creator-ecosystem data monitoring and trend attribution; refined human-in-the-loop moderation strategy with the trust & safety team.
从 SG 市场的内容质量管理做到 US 市场的画风巡查;深度参与内容质量评估模型的数据构造与效果评估。From content-quality ops in SG to style-level red-teaming in the US market; built training data for and evaluated the in-app content quality model.
百万粉丝 IP 矩阵从 0 到 1;直播、社群与跨平台流量闭环——19 岁的第一课:内容怎么长出来,流量怎么流动。Grew an IP matrix from zero past a million followers; ran livestreams, communities and cross-platform loops — my first lesson, at 19, in how content grows and traffic flows.
白天造 AI,晚上研究造 AI 的人。Building AI by day, studying the people who build it by night.
首篇把 Agent Harness 当作独立研究对象的系统综述:提出 H=(E,T,C,S,L,V) 形式化框架,综述 110+ 文献、23 个系统。从 idea 到 preprint,12 天。The first systematic survey treating the agent harness as its own research object: the H=(E,T,C,S,L,V) formalization, 110+ papers, 23 systems. Idea to preprint in 12 days.
文化趣味如何在"技术中立"的外衣下被制度化为审核规则,又通过模型的标准化执行反向规训创作者。这篇和我的日常工作互为镜像。How cultural taste gets institutionalized as moderation rules under the guise of "technical neutrality" — then disciplines creators back through standardized model execution. This paper and my day job mirror each other.
都能直接上手,不用想象。Everything here runs. No imagination required.
| 系统System | E | T | C | S | L | V | Σ |
|---|
23 个系统逐一打分。生产级系统六项俱全,研究原型只有两三项——排序一下,规律自己浮出来。All 23 systems scored. Production systems check all six; research prototypes barely three — sort it, and the pattern surfaces itself.
从竞品调研到 PRD 到 UI 概念稿的完整产品设计。这里是核心交互的手感,完整版是整套 UI 稿。A full product design from competitive research to PRD to UI concepts. This is the core interaction; the full version has the complete UI spec.
为对话产品设计的人格分类与玩法产品线:27 型用户分类、行为信号到人格维度的映射、敏感品类的合规化设计。Personality typing and gameplay for a conversational product: 27-type user taxonomy, behavior-to-trait mapping, compliance-aware design for a sensitive category.
人类学家的老本行:参与观察。只不过这次被观察的是你——纯前端、零上传、随时可清除。An anthropologist's oldest trick: participant observation. Except this time the subject is you — frontend only, zero upload, erasable anytime.
