Johannes Kepler Isaac Newton René Descartes Niccolò Machiavelli Baruch Spinoza Gottfried Leibniz Jeremy Bentham Charles de Montesquieu Voltaire Immanuel Kant Jean-Paul Sartre Greek Philosophers Newton Royal Society
~
udify.games / research laboratory

PROTEUS LAB

普罗透斯实验室

"万物皆流，无物常驻。在深邃的信息海洋中，定义下一个问题。"

"All things flow, nothing stays. In the deep ocean of information, define the next question."

我们基于 Udify 意图驱动内容演化系统，在游戏、AI、硬件与艺术的交叉疆域中，提出最前沿的问题。问题即方向，定义问题即引领变革。
Built on the Udify intent-driven content evolution system, we raise the frontier questions at the intersection of games, AI, hardware, and art. Questions are directions — defining the question leads the change.
Explore
I. Problems / 前沿问题

前沿问题

我们定义的、等待被解决的问题 / Questions we define, waiting to be solved

Udify 的核心洞见是：创作是变换，不是生成；意图比技术更重要。
The core insight of Udify: Creation is transformation, not generation; Intent matters more than technology.

Hardware × AI / 硬件 × AI
神经形态芯片能加速 MCTS 搜索中的价值评估吗？
Can neuromorphic chips accelerate value evaluation in MCTS search?
当前 MCTS 规划中 LLM 价值函数延迟高达数百毫秒。能否利用 脉冲神经网络（SNN）在 neuromorphic 芯片（如 Intel Loihi）上实现超低延迟的价值评估？这是否能将规划延迟压缩到游戏实时帧率以内？
Current LLM value functions in MCTS planning incur hundreds of milliseconds of latency. Can Spiking Neural Networks (SNN) on neuromorphic chips like Intel Loihi achieve ultra-low-latency value evaluation? Could this compress planning latency within real-time game framerates?
Game Host × Edge / 游戏主机 × 边缘
如何在游戏主机上运行轻量级内容感知模型？
How to run lightweight content perception models on game consoles?
PS5 / Xbox Series 的定制 APU 拥有被忽视的 AI 加速能力。能否将 感知引擎压缩为 <50MB 的量化模型，直接在游戏主机上实时提取内容图谱，而不依赖云端往返？
PS5/Xbox Series custom APUs have overlooked AI acceleration capabilities. Can the perception engine be compressed into a <50MB quantized model, running directly on game consoles to extract content graphs in real-time, eliminating cloud round-trips?
AI × Algorithms / AI × 算法
LLM 能否产生可验证的因果变换链？
Can LLMs produce verifiable causal transformation chains?
Udify 的 CDL Patch 需要可验证性。如何让 LLM 在生成变换计划时，同时输出形式化的因果推理链，使得每个 Patch 操作都可以被 SMT 求解器自动验证一致性？
Udify's CDL Patch requires verifiability. Can we make LLMs output formal causal reasoning chains alongside transformation plans, enabling each Patch operation to be automatically verified for consistency by SMT solvers?
RL × Personality / 强化学习 × 个性
强化学习如何从用户反馈中演化出个性化审美模型？
How can RL evolve personalized aesthetic models from user feedback?
当前审美模型是静态的。能否用 在线强化学习 + 贝叶斯偏好学习，让系统从每个用户的隐式反馈（停留时长、重试率、分叉行为）中实时演化出个性化审美函数？
Current aesthetic models are static. Can we use online RL + Bayesian preference learning to let the system evolve personalized aesthetic functions in real-time from each user's implicit feedback (dwell time, retry rate, fork behavior)?
Math × Content / 数学 × 内容
如何用范畴论形式化验证跨媒介变换的一致性？
How to formalize and verify cross-media transformation consistency using category theory?
游戏→小说→视频的跨媒介变换缺乏形式化保证。能否用 范畴论语义 定义变换函子的保结构性质，从而在变换前自动验证目标媒介的可达性？
Cross-media transformations (game→novel→video) lack formal guarantees. Can we use categorical semantics to define structure-preserving properties of transformation functors, automatically verifying target medium reachability before transformation?
Physics × Systems / 物理 × 系统
内容适应度景观中的相变点如何被算法自动识别？
How can phase transition points in content fitness landscapes be algorithmically detected?
类比统计物理中的相变检测。能否用 信息几何 + 拓扑数据分析，在内容的高维特征空间中自动定位"临界参数"，在相变点附近进行最经济但最有冲击力的变换？
Analogous to phase transition detection in statistical physics. Can we use information geometry + topological data analysis to automatically locate "critical parameters" in high-dimensional content feature space, executing the most economical yet impactful transformations near phase transition points?
Game Art × AI / 游戏艺术 × AI
如何从游戏引擎二进制中自动重建可编辑的内容图谱？
How to automatically reconstruct editable content graphs from game engine binaries?
Unity / Unreal 的运行时资产是黑盒。能否结合 逆向工程 + 程序分析 + 神经网络反编译，从打包的游戏中自动提取出带语义标注的内容图谱（Content Graph）？
Unity/Unreal runtime assets are black boxes. Can we combine reverse engineering + program analysis + neural decompilation to automatically extract semantically annotated Content Graphs from packaged games?
Rendering × Intent / 渲染 × 意图
如何在实时渲染管线中嵌入"意图驱动"的实时魔改？
How to embed "intent-driven" real-time modding into rendering pipelines?
当前魔改是离线的。能否在 Vulkan/Metal 着色器中嵌入意图解析器，让玩家在游戏过程中实时说出"更暗黑一点"，渲染管线即时调整光照、后处理和材质参数？
Current modding is offline. Can we embed intent parsers into Vulkan/Metal shaders, allowing players to say "make it darker" mid-gameplay, with the rendering pipeline instantly adjusting lighting, post-processing, and material parameters?
Cloud × Security / 云 × 安全
如何在沙箱中实现零信任的内容变换验证？
How to achieve zero-trust content transformation verification in sandboxes?
gVisor 等沙箱仍有逃逸风险。能否结合 形式化验证 + eBPF 沙箱 + 可信执行环境（TEE），构建一个可以运行不可信用户上传变换脚本的零信任执行环境？
Sandboxes like gVisor still carry escape risks. Can we combine formal verification + eBPF sandboxes + TEE (Trusted Execution Environments) to build a zero-trust execution environment for untrusted user-uploaded transformation scripts?
Distributed × Content / 分布式 × 内容
如何用 DHT 构建去中心化的内容图谱存储？
How to build decentralized content graph storage using DHT?
当前内容图谱集中在 Neo4j。能否用 分布式哈希表（如 IPFS + CRDT） 构建一个去中心化的内容图谱网络，让每个 Mod 创作者都成为图谱的一个节点，实现真正的 P2P 内容演化？
Current content graphs are centralized in Neo4j. Can we use Distributed Hash Tables (e.g., IPFS + CRDTs) to build a decentralized content graph network, where every Mod creator becomes a node in the graph, enabling true P2P content evolution?
◆

这些问题没有标准答案。我们寻找那些被问题本身点燃的人。
提出更好的问题，比给出平庸的答案更重要。

These questions have no standard answers. We seek those ignited by the questions themselves.
Raising better questions matters more than giving mediocre answers.

II. Domains / 研究疆域

研究疆域

Udify 思想体系下的交叉学科方向 / Cross-disciplinary directions under the Udify paradigm

Proteus Lab 不局限于单一技术栈。我们从 游戏 Mod 自动化 切入——最硬核、最可验证的领域——但目光始终投向更远的疆域。
We are not limited to a single tech stack. We enter through game mod automation — the hardest, most verifiable domain — but our gaze always reaches farther.

意图识别 / Intent MCTS 规划 / MCTS Planning 内容图谱 / Content Graph 游戏逆向 / Game RE LLM 编排 / LLM Orchestration Diff/Patch 审美模型 / Aesthetic Models 适应度景观 / Fitness Landscape 多智能体 / Multi-Agent 实时演化 / Real-Time Evolution 跨媒介转换 / Cross-Media 沙箱安全 / Sandbox Security 神经形态 / Neuromorphic 形式化验证 / Formal Verification 信息几何 / Information Geometry 分布式系统 / Distributed Systems 实时渲染 / Real-Time Rendering 创作者经济 / Creator Economy
III. Join / 加入我们

加入我们

寻找被问题点燃的人 / Seeking those ignited by questions

我们不招聘"职位"，我们寻找"解题者"。如果你对以上任何一个问题有执念，我们渴望与你对话。
We don't hire "positions"; we seek "problem solvers." If you're obsessed with any of these questions, we want to talk to you.

◈
架构炼金术士
你理解复杂系统的涌现性质，能将跨学科洞见转化为软件架构。你相信代码可以像有机体一样演化。
You understand emergent properties of complex systems, translating cross-disciplinary insights into software architecture. You believe code can evolve like organisms.
◉
AI 认知架构师
你痴迷于让机器"理解"内容的结构与意义。LLM 编排、RAG、多模态融合——无论你的武器是什么，你在寻找可验证的 AI 推理。
You're obsessed with making machines "understand" content structure and meaning. LLM orchestration, RAG, multimodal fusion — whatever your weapon, you seek verifiable AI reasoning.
◆
游戏系统逆向师
你能读懂二进制的语言，拆解过 Unity / Unreal / Godot 的资源格式，对文件格式、内存布局、渲染管线有直觉般的理解。
You read the language of binaries, have dissected Unity/Unreal/Godot resource formats, with intuitive understanding of file formats, memory layouts, and rendering pipelines.
◎
审美形式化者
你在寻找量化"美"的方法。信息论、拓扑、范畴论或认知科学——无论你的武器是什么，你相信审美可以被计算，但永远不会被穷尽。
You seek methods to quantify "beauty." Information theory, topology, category theory, or cognitive science — whatever your weapon, you believe aesthetics can be computed but never exhausted.
◇
系统铸造师
你关心如何让系统在生产环境中可靠运行。分布式系统、安全沙箱、形式化验证、性能优化——你是让理论落地的重力。
You care about making systems run reliably in production. Distributed systems, secure sandboxes, formal verification, performance optimization — you are the gravity that grounds theory into reality.
◐
跨媒介译者
你看见游戏与音乐、小说与视频之间的隐秘通道。你相信任何内容都可以被翻译为任何其他内容，关键是找到正确的中间表示。
You see the hidden passages between games and music, novels and videos. You believe any content can be translated into any other — the key is finding the right intermediate representation.
"我们不追问 LLM 是否真正理解，而追问：基于它的表示生成的改造计划，在实践中是否有效？"
"We don't ask whether LLMs truly understand; we ask: Are the transformation plans generated from their representations effective in practice?"
IV. Signal / 建立连接

建立连接

如果你听到了问题的召唤 / If you hear the call of the question

无论你是研究者、工程师、艺术家还是哲学家，
如果你相信 意图应当直接转化为形态，
如果你渴望在交叉学科的锋刃上行走——

Whether you are a researcher, engineer, artist, or philosopher,
if you believe intent should directly transform into form,
if you thirst to walk the razor's edge of cross-disciplinary frontiers—

发送信号。我们等待着与你共振。
Send the signal. We await resonance with you.

iamin@udify.games
◆

udify.games