ABOUT / SHAWN XU

关于我About

我是徐葛神戎，一半工程师，一半设计师。我在做 AI Agent，也在意一个产品落到人身上是什么感觉。模型越来越强，我更想做一点有人味、不会被模型进步取代的东西：技术托底，审美定形，以人为本。I'm Shawn Xu — half engineer, half designer. I build AI agents, and I care how a product feels when it reaches a person. As models get stronger, I want to make things with a human touch that model progress won't replace: engineering as the base, taste as the form, people at the center.

联系我Contact 能力图谱Capabilities GitHub
现在NOW
Eazo.ai · AI 架构师Eazo.ai · AI Architect2026.08 至今Since Aug 2026
在读STUDY
国际关系学院 · 网络空间安全University of International Relations · Cybersecurity本科 · 2027 届B.Eng · Class of 2027
方向FOCUS
Agent 工程 · 后训练与评测Agent engineering · post-training & evaluationAI 安全 · 3D 交互AI security · 3D interaction
所在BASED
北京Beijing寻找全职机会Open to full-time roles
徐葛神戎的黑白肖像 / Black-and-white portrait of Shawn Xu

01 / HALF & HALF

一半工程师，一半设计师Half engineer, half designer

在刻度上滑动Slide the scale

现在的我：一半工程师，一半设计师。Today: half engineer, half designer.

我以前其实是个纯粹的工程师。后来越来越像设计师，是因为有了 AI，技术正处在一种平权的状态：你很难再比别人建立起更高的技术壁垒。I started out a pure engineer. I became more of a designer because AI has made technology an equalizer: it is hard to build a technical moat above everyone else.

在这个节骨眼上，拿什么去战胜别人、取得优势？我的答案是设计。好产品往往并不取决于底层能力有多强。So what gives you an edge now? My answer is design. A good product rarely hinges on how strong the underlying capability is.

CASE 01

豆包Doubao

大家都觉得它模型弱，甚至嘲笑它，可它偏偏用户量最大。People call its model weak, even mock it — yet it has the largest user base.

它能胜出，核心在于体验。It wins on experience, above all.
CASE 02

WorkBuddy

模型绝不比 Claude、Codex 占优，却是 Agent 领域用户量最大的产品。Its model is in no way ahead of Claude or Codex, yet it is the most-used product in the agent space.

同样，赢在体验。Again: it wins on experience.

02 / HUMAN

做有人味的东西Make things with a human touch

勾选你也会问的问题Tick the questions you would ask too

随着模型能力越来越强，我发现以前做的很多工具都没用了。As models grew stronger, many tools I had built before became useless.

所以现在更想做能真正落到人身上、能留存下来、不会轻易被模型进步取代的产品。我要把技术和审美结合起来。做产品的思维是以人为本：必须充分考虑用户的使用场景和体验。So now I want to build products that genuinely reach people, that last, and that model progress will not easily replace. I want to combine engineering with taste. Product thinking is people-first: the user's context and experience have to be fully considered.

动手之前，我会先问Before I build, I ask

03 / DATA

数据决定上限Data sets the ceiling

拖动，调整训练数据的质量Drag to change training-data quality

质量差的数据，一定训练不出好模型。这是模型世界的基本规则。Bad data never yields a good model. That is the basic rule of the model world.

在模型层面，我认为数据才是决定能力的根本。不管是现在的大语言模型，还是之后的多模态、工业模型，乃至各种垂直场景模型，都高度依赖数据。有些数据容易获取，有些很难；未来的具身智能同样需要海量数据。所以数据行业里，存在巨大的利润空间和机会。At the model level, I believe data is what ultimately determines capability. Today's LLMs, tomorrow's multimodal and industrial models, and every vertical model all depend heavily on data. Some data is easy to obtain, some is very hard; future embodied intelligence will also need massive real-world data. That leaves a huge margin and opportunity in the data industry.

MODEL OUTPUT模型上限 · 38%Ceiling · 38%

数据的质量，决定了模型的上限。Data quality sets the ceiling of the model.

04 / GOLD

三座金矿Three gold mines

对我而言，AI 行业未来最大的金矿在这三层For me, the industry's biggest opportunities sit in these three layers

01 / APPLICATION

应用落地Application

模型能力要变成某个人每天真的会打开的东西，价值才算落了地。越往后，越是这一层决定谁能留下来。Model capability only counts once it becomes something a person actually opens every day. Further out, this layer decides who stays.

05 / FRONTIER

模型层面的探索Exploration at the model layer

拖动、点击或用方向键切换Drag, click, or use arrow keys

体验之外，我同样非常在意模型本身的新技术。这是我做过、在研究、在关注的方向。Beyond experience, I care deeply about new technology in the models themselves. These are directions I have built, am researching, or am following.

01 / POST-TRAINING实践过Hands-on

模型训练 · 后训练Training & post-training

做过 SFT 与 DPO 的对比实验；在混元 HY4 的评测里制定 Rubric、构造高难样本。这让我确信：奖励信号和评测本身，就是数据问题。接下来关注 GRPO、可验证奖励（RLVR），以及在 Agent 轨迹上做强化学习。I ran SFT vs. DPO comparison experiments, and authored rubrics and hard samples for the Hunyuan HY4 evaluation. That convinced me that reward signals and evaluation are themselves data problems. Next: GRPO, verifiable rewards (RLVR), and reinforcement learning on agent trajectories.

  • SFTSFT
  • DPODPO
  • GRPOGRPO
  • RLVRRLVR
  • Rubric 奖励Rubric rewards
02 / INFERENCE实践过Hands-on

推理架构Inference architecture

在 Eazo 参与 Prompt Caching 与模型调用链路优化，从这里摸到推理侧真实的成本结构。关注 KV Cache 复用与前缀共享、投机解码、Prefill / Decode 分离，以及长上下文下显存与延迟的取舍。At Eazo I worked on prompt caching and model-call path optimization, which showed me the real cost structure on the inference side. I follow KV cache reuse and prefix sharing, speculative decoding, prefill/decode disaggregation, and the memory vs. latency trade-off at long context.

  • Prompt CachePrompt cache
  • KV CacheKV cache
  • 投机解码Speculative decoding
  • PD 分离PD disaggregation
  • 长上下文Long context
03 / ALGORITHMS在研究Researching

算法Algorithms

关注那些改变成本曲线的结构：MLA、稀疏注意力与线性注意力，MoE 的路由与负载均衡；还有 test-time scaling，怎样把推理时的算力换成真正的推理能力。Structures that bend the cost curve: MLA, sparse and linear attention, MoE routing and load balancing; plus test-time scaling — how to convert inference compute into genuine reasoning ability.

  • MLAMLA
  • 稀疏注意力Sparse attention
  • MoEMoE
  • Test-time ScalingTest-time scaling
04 / AI FOR SCIENCE在关注Following

AI for ScienceAI for Science

科学数据昂贵又稀缺，是「数据决定上限」最极端的场景。关注蛋白质与分子设计、材料发现、天气与气候模型，看模型怎样在实验数据很少的领域里被训练和验证。Scientific data is expensive and scarce — the most extreme case of "data sets the ceiling." I follow protein and molecule design, materials discovery, and weather/climate models, looking at how models are trained and validated in fields with little experimental data.

  • 蛋白设计Protein design
  • 材料发现Materials discovery
  • 气象模型Weather models
  • 科学数据Scientific data
05 / AI INFRA实践过Hands-on

AI InfraAI Infra

把 Agent 执行从后端 Worker 迁到 E2B 沙箱、拆分控制面与执行面，是我接触 Infra 的起点。再往下是训练与推理的集群调度、数据管线和评测平台，模型能力最终都落在这些地基上。Moving agent execution from a backend worker into E2B sandboxes and splitting control from execution was my entry into infrastructure. Deeper down lie cluster scheduling for training and inference, data pipelines and evaluation platforms — the ground model capability ultimately stands on.

  • E2B 沙箱E2B sandbox
  • 控制面 / 执行面Control / execution plane
  • 数据管线Data pipeline
  • 评测平台Evaluation platform

06 / TRAJECTORY

经历Trajectory

悬停或点击展开Hover or click to expand
01Eazo.aiEazo.ai2026.08 — 至今Aug 2026 — now

2026.08 — 至今Aug 2026 — now

ASI X Inc.（Eazo.ai）ASI X Inc. (Eazo.ai)

AI 架构师AI Architect

  • 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。Contributed to the V4→V5 upgrade, moving agent execution from a backend worker into E2B sandboxes, splitting control plane from execution plane; the frontend connects to sandboxes directly over WebSocket.
  • 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。Worked on prompt caching and model-call path optimization: connection reuse, parallel config initialization, incremental sync for extensions.
  • 参与构建 Skills 驱动的编码、设计与 3D Agent 体系，支持子 Agent 并行与能力按需加载。Helped build the Skills-driven coding, design and 3D agent system, with parallel sub-agents and on-demand capability loading.

07 / UPSTREAM

开源与贡献Open source & contributions

读源码、改服务，把改进交回上游Read the source, patch the service, send it upstream

08 / HOW I WORK

工作方式How I work

  1. Ⅰ

    对交付负责Own the delivery

    独立推进律所系统上线，也参与政务平台的完整交付。习惯把需求、架构、部署和真实使用放在一起考虑，关注功能在实际环境里能否走通。I took the law-firm system live on my own and took part in the full delivery of a government platform. I consider requirements, architecture, deployment and real use together, and care whether a feature holds up in the actual environment.

  2. Ⅱ

    愿意进入技术内部Go inside the technology

    遇到工具限制，会读源码、改服务和交互，并向上游提交改进。希望把解决眼前问题的工作，积累成之后仍然可用的能力。When a tool hits a limit, I read the source, change the service and the interaction, and send the improvement upstream — so fixing today's problem also builds capability that lasts.

  3. Ⅲ

    工程与表达并重Engineering and expression

    做 Agent 时关注执行、权限与结果依据；做作品时亲手打磨建模、材质和动效，让技术进入具体体验。With agents I focus on execution, permissions and evidence for results; with creative work I hand-tune models, materials and motion so the technology becomes a concrete experience.

A CONVERSATION STARTS HERE

聊聊下一件
作品。Let's talk about
the next one.

一个想法，一次合作，
或只是聊聊技术与设计。An idea, a collaboration,
or just a chat about technology and design.

AI 与智能体AI & agentsAgent 安全评估Agent security evaluation开发工具与工作流Developer tools & workflows产品与设计Product & design开源与技术写作Open source & technical writing一起造物Building together
正在形成共振…Finding resonance…
3D CHLADNI ↗拖动旋转 · 方向键也可以Drag to rotate · arrow keys work too

也可以，从 AI 造物社认识我Or meet me through the AI Makers Club ↗

Shawn Xu

engineer & designer

01 / EMAIL

写一封邮件Send an email

xugeshenrong888@gmail.com

02 / WECHAT

微信WeChat

Xu_gsr2365

03 / XIAOHONGSHU

小红书RED

ShawnXu · 42224190089

04 / GITHUB

从代码开始Start from the code

@ShawnRyan2365

LET’S MAKE SOMETHING.很高兴认识你。Nice to meet you.