Raskar 的创意六边形:六条路径,系统性产生新想法
Raskar's Idea Hexagon: Six Systematic Paths to Generate New Ideas
MIT Media Lab 的 Ramesh Raskar 是个连续创造者:飞秒摄影(femtophotography,“能看到光在运动”的相机)、NETRA 手机眼科检测、Camera Culture 研究组……在他 2012 年写给 Wired UK 的文章和 TEDxBeaconstreet 演讲里,他把自己产生新想法的方法总结成了一个六边形框架——Idea Hexagon。
这不是”多想想""保持好奇心”那类空洞建议。它是一套固定路径:从任意一个已知概念 X 出发,沿六个方向走,强迫你看到以前没注意到的角落。
框架全貌
Xd(维度泛化)
/ \
X+Y(异类融合) X→X̄(做反面)
| X |
锤子找钉子 X++(加形容词)
\ /
钉子找锤子
六条路径,没有先后顺序,任意一条都可以独立使用。
六条路径,逐一拆解
路径一:Xd — 维度泛化
核心问题:X 是在什么维度上运作的?如果换一个维度,它会变成什么?
不是简单地”扩大规模”,而是找到一个 X 当前不存在的坐标轴,然后把 X 沿着这条轴延伸过去。
经典例子:
- 2D 照片 → 3D 立体摄影 → 4D 时光场(light field)
- 单点测量 → 空间分布测量 → 时间序列分布
- 文字搜索 → 图片搜索 → 视频搜索 → 语音/音频搜索
AI 时代的版本:
- 文本补全(LLM)→ 代码补全 → 动作补全(Agent)→ 物理动作补全(机器人)
- 单模态理解 → 多模态理解 → 具身多模态(看 + 说 + 做)
Raskar 自己的案例:把”照明”从可见光波段泛化到微波,就得到了雷达;泛化到飞秒脉冲,就得到了能穿透不透明介质的飞秒摄影。
操作提示:
- 把 X 写在中间
- 列出 X 当前依赖的所有”参数轴”(时间、空间维度、频率/模态、规模、用户数……)
- 对每条轴问:更高?更低?方向反转?换成另一个变量?
路径二:X+Y — 异类融合
核心问题:把 X 和一个看起来完全不相关的 Y 强制合并,会产生什么?
关键词是异类——Y 越不像 X,越容易打开真正新的空间。把两个相似的东西合并只是功能叠加;把两个没关系的东西合并才会产生化学反应。
经典例子:
- CT(医学成像)+ 望远镜技术 → 便携 CT
- 音乐 + 数学 → 和声理论、序列音乐
- 货运 + 共享经济 → 拼车、货拉拉
AI 时代的版本:
- LLM + 代码解释器 → 可执行的推理(ChatGPT 代码执行)
- 搜索引擎 + LLM → RAG / Perplexity
- 机器人 + LLM → 语言指令机器人(RT-2, π0)
操作提示:
- 列出 20 个与 X 毫无关系的领域(随机越好)
- 强制想象”X + 这个领域”会产生什么
- 很多组合是垃圾——但这没关系,找到那 1-2 个有意思的
路径三:锤子找钉子(Hammers for Nails)
核心问题:X 作为一项技术/能力,所有可能的应用场景是什么?
这是从技术出发找应用,而不是从应用出发找技术。
经典例子(Raskar 自己的):
飞秒摄影能做什么?
- 看光在透明介质里的运动 → 医学成像(穿透皮肤)
- 看角落后面的物体 → 非视线(non-line-of-sight)成像
- 测量大气散射 → 气象感知
- 测量表面微振动 → 无接触声学麦克风
一项技术,十几个完全不同的行业应用。
AI 时代的版本:
Embedding 技术能做什么?
- 语义搜索
- 推荐系统
- 异常检测
- 代码克隆检测
- 跨语言文档对齐
操作提示:
- 把 X 的核心能力抽象成一个”超级能力”(不要用具体产品描述,要用物理/数学层面的描述)
- 枚举所有”哪些场景需要这种超级能力”
- 不要排除”太小”或”太奇怪”的场景,长尾应用往往是真正的蓝海
路径四:钉子找锤子(Nails for Hammers)
核心问题:对于一个特定的需求/问题 X,所有可能的解决方案是什么?
这是从需求出发找技术,强迫你不依赖第一个想到的方案。
经典例子(Raskar 的):
“数字重对焦”(照片拍完后才决定焦点)怎么实现?
- 光场相机(Lytro 路线)
- 编码孔径(coded aperture)
- 多张不同焦距合成
- 计算全息
- AI 深度估计 + 后处理重对焦(现代手机)
每种技术的成本、精度、适用场景完全不同——但它们都是同一个钉子的锤子。
操作提示:
- 把”解决 X”转化为”实现 Y 这个物理效果/信息转换”
- 跨领域枚举:有没有已经在其他领域解决了类似物理问题的技术?
- 不要停在第一个”显然的”解法——列到 10 个再做筛选
路径五:X++ — 加形容词
核心问题:在 X 前面加一个形容词,会产生什么新方向?
Raskar 推荐的形容词列表(不用全部,每次选 1-2 个):
- 自适应的(adaptive)
- 个性化的(personalized)
- 分布式的(distributed)
- 嵌入式的(embedded)
- 层次化的(hierarchical)
- 持续的(continuous / persistent)
- 社会化的(social)
- 实时的(real-time)
经典例子:
- 翻译 → 个性化翻译(用你的词汇库和风格习惯)
- 学习 → 自适应学习(根据你的进度动态调整难度)
- 监控 → 分布式监控(边缘设备协同,无中心节点)
AI 时代的版本:
- LLM → 持续学习的 LLM(在线 fine-tuning,不需要重新训练)
- Agent → 社会化的 Agent(Agent 之间协商、分工、合并记忆)
- 搜索 → 嵌入式搜索(在任意 App 内部,无需跳转)
操作提示: 把这个形容词列表打印出来贴在显示器旁边。每次做头脑风暴,机械地过一遍:这个形容词+我的产品=什么?
路径六:X→X̄ — 做反面
核心问题:如果 X 做的事情完全反转,会是什么?
Raskar 举的例子是 Fosbury Flop(福斯伯里翻滚)——跳高的传统姿势是正面跨越横杆;Dick Fosbury 在 1968 年奥运会上用背对横杆的姿势跳,打破世界纪录,此后这成了所有跳高运动员的标准姿势。
“反面”不是”坏的”,而是把假设颠倒。
经典例子:
- 搜索引擎(你找信息)→ 反向:信息找你(推荐算法/RSS/订阅推送)
- 教师讲、学生听 → 反向:学生讲、教师听(费曼学习法、学生主导的 PBL)
- 服务器渲染(SSR)→ 客户端渲染(CSR)→ 再反向:边缘渲染(Edge SSR)
- 模型越大越好 → 反向:模型越小越快(MobileNet、边端小模型)
AI 时代的版本:
- AI 辅助人类写代码 → 反向:人类辅助 AI 写代码(HITL,人类作为 verifier)
- 大模型中心化推理 → 反向:小模型本地推理(llama.cpp、Apple Intelligence)
- 模型学习数据 → 反向:数据学习模型(数据蒸馏 Data Distillation)
操作提示:
- 写下 X 的三个核心假设(X 假设谁发起?假设谁受益?假设什么是输入/输出?)
- 逐一颠倒这些假设
- 检查:颠倒后的世界里,谁得到了更大的价值?
实战:用六边形分析一个 AI 产品
以 RAG(检索增强生成) 为例,快速过一遍六条路径:
| 路径 | 操作 | 产出方向 |
|---|---|---|
| Xd | RAG 现在是文本→文本;维度泛化 | 多模态 RAG(图/视频/代码→答案),时序 RAG(记忆随时间演化) |
| X+Y | RAG + 知识图谱 | GraphRAG(微软),关系推理而非简单召回 |
| 锤子找钉子 | RAG 的核心能力是”外部记忆注入” | 代码库问答、文档审阅自动化、法律合规检查、药物数据库查询 |
| 钉子找锤子 | 需求:让 LLM 知道它不知道的事 | Fine-tuning / RAG / Tool use / 长上下文 / Continual learning |
| X++ | 自适应 RAG | 根据问题类型动态选择召回策略(密集/稀疏/知识图谱) |
| X→X̄ | RAG 是模型查外部 → 反转:外部系统主动推送给模型 | Push-based context(Proactive Context Injection),类似 RSS 但给 Agent |
六条路径,15 分钟内,出来 6 个方向,其中 2-3 个可能是没人做过的。
使用建议
当你被一个问题卡住:用”钉子找锤子”——强迫自己列出 10 种解法再评判。
当你有一项新技术/工具:用”锤子找钉子”——把核心能力抽象化,找所有潜在应用。
当你想做增量创新:用”X++“——选 2-3 个形容词,快速生成变体。
当你想做颠覆性创新:用”X→X̄”——找到领域最根深蒂固的假设,把它颠倒。
当你做产品规划:六条路径都过一遍,每条 5 分钟,作为 divergent thinking 阶段;然后选出 3 个最有潜力的方向做 convergent 评估。
一个有效的团队工作流:
- 每人独立过 6 条路径,写下至少 2 个想法/条(共 6×人数个)
- 贴在白板上不讨论,先聚类
- 对每个集群投票,选出前 5 个做深入分析
- 前 5 个用 Raskar 自己的评估标准过滤:Impact × Novelty × Feasibility
原始资料
这个框架的原始来源:
- Wired UK (2012):Raskar 亲自撰写的文章《Inventing a New Field in Vision》,收录在 Wired 杂志 2012 年 11 月刊 Start 专栏:http://www.wired.co.uk/magazine/archive/2012/11/start/inventing-a-new-field-in-vision
- TEDxBeaconstreet 演讲:Raskar 在台上完整讲解六边形的来由和例子:http://tedxbeaconstreet.com/rameshraskar/
- Raskar 的 MIT 主页:Camera Culture 研究组首页,列出他的代表作和飞秒摄影等项目:http://web.media.mit.edu/~raskar/
Raskar 提出这个框架,是为了回答一个他自己面对的真实问题:一个研究组怎么系统性地产生”下一个”项目的想法,而不是等待灵感降临?六边形是他的答案——一个可重复、可教学、可在团队里使用的流程。
MIT Media Lab professor Ramesh Raskar is a serial inventor: femtophotography (a camera that captures light in motion), NETRA mobile eye testing, the Camera Culture research group. In a 2012 article for Wired UK and a TEDxBeaconstreet talk, he distilled his method for generating new ideas into a single framework — the Idea Hexagon.
This is not “think more” or “stay curious.” It is a set of fixed paths: start from any known concept X, walk in six directions, and force yourself to see the corners you’ve been missing.
The Framework
Xd (Generalize dimension)
/ \
X+Y (Fusion) X→X̄ (Opposite)
| X |
Hammers for nails X++ (Add adjective)
\ /
Nails for hammers
Six paths, no fixed order. Any single path can be used independently.
Six Paths, One by One
Path 1: Xd — Generalize to a New Dimension
Core question: What dimension is X currently operating in? If you move it to a different dimension, what does it become?
This is not “scale up” — it’s finding a coordinate axis that X doesn’t currently occupy, then extending X along it.
Classic examples:
- 2D photo → 3D stereoscopy → 4D light field
- Point measurement → spatial distribution → time series distribution
- Text search → image search → video search → audio/voice search
AI era versions:
- Text completion (LLM) → code completion → action completion (Agent) → physical action completion (robotics)
- Single-modal understanding → multimodal → embodied multimodal (see + say + do)
Raskar’s own case: generalize “illumination” from visible light to microwaves → radar. Generalize to femtosecond pulses → imaging through opaque media.
Practical tool:
- Write X in the center
- List all the “parameter axes” X currently depends on (time, space dimensions, frequency/modality, scale, number of users…)
- For each axis: higher? lower? direction reversed? substitute another variable?
Path 2: X+Y — Fusion of Unlike Things
Core question: Force-merge X with something completely unrelated. What do you get?
Key word: unlike — the less Y resembles X, the more likely you’ll open genuinely new territory. Merging two similar things produces feature stacking; merging two unrelated things produces chemistry.
Classic examples:
- CT (medical imaging) + telescope optics → portable CT
- Music + mathematics → harmony theory, serial composition
- Shipping + sharing economy → rideshare, freight platforms
AI era versions:
- LLM + code interpreter → executable reasoning
- Search engine + LLM → RAG / Perplexity
- Robots + LLM → language-commanded robots (RT-2, π0)
Practical tool:
- List 20 domains with no obvious connection to X (the more random, the better)
- Force-imagine “X + this domain” for each
- Most combinations are garbage — that’s fine. Find the 1-2 interesting ones.
Path 3: Hammers for Nails
Core question: Given X as a technology or capability, what are all possible applications?
This is technology-first search for applications, not application-first search for technology.
Classic example (Raskar’s own):
What can femtophotography do?
- See light moving through transparent media → medical imaging (penetrate skin)
- See objects around corners → non-line-of-sight imaging
- Measure atmospheric scatter → meteorological sensing
- Measure surface microvibration → non-contact acoustic microphone
One technology, a dozen completely different industry applications.
AI era version:
What can embeddings do?
- Semantic search
- Recommendation systems
- Anomaly detection
- Code clone detection
- Cross-lingual document alignment
Practical tool:
- Abstract X’s core capability into a “superpower” (describe in physical/mathematical terms, not product terms)
- Enumerate all scenarios that require this superpower
- Don’t exclude “too small” or “too weird” scenarios — long-tail applications are often true blue oceans
Path 4: Nails for Hammers
Core question: For a specific need or problem X, what are all possible solutions?
Technology-agnostic. Forces you off the first solution that came to mind.
Classic example (Raskar’s):
“Digital refocusing” — choose your focus point after the photo is taken:
- Light field cameras (Lytro approach)
- Coded aperture
- Multi-exposure composite
- Computational holography
- AI depth estimation + post-processing (modern smartphones)
Each approach has completely different cost, precision, and applicability — but all are hammers for the same nail.
Practical tool:
- Translate “solve X” into “achieve Y physical effect / information transformation”
- Cross-domain enumeration: has another field already solved a similar physical problem?
- Don’t stop at the first “obvious” solution — list 10, then filter
Path 5: X++ — Add an Adjective
Core question: Put an adjective in front of X. What new direction does it open?
Raskar’s adjective list (pick 1-2 each time):
- Adaptive
- Personalized
- Distributed
- Embedded
- Hierarchical
- Persistent / Continuous
- Social
- Real-time
Classic examples:
- Translation → personalized translation (uses your vocabulary and style)
- Learning → adaptive learning (dynamically adjusts difficulty to your pace)
- Monitoring → distributed monitoring (edge devices coordinate, no central node)
AI era versions:
- LLM → persistent-learning LLM (online fine-tuning, no retraining)
- Agent → social agents (agents negotiate, divide tasks, merge memory)
- Search → embedded search (inside any app, no redirect)
Practical tool: Print this adjective list and tape it to your monitor. Each brainstorm session, mechanically run through the list: this adjective + my product = what?
Path 6: X→X̄ — Do the Opposite
Core question: Completely reverse what X does. What do you get?
Raskar’s example: the Fosbury Flop. Traditional high jump crosses the bar face-first. Dick Fosbury at the 1968 Olympics jumped backward. Broke the world record. Every high jumper since uses this technique.
“Opposite” doesn’t mean “worse” — it means inverting the assumption.
Classic examples:
- Search engine (you find information) → reverse: information finds you (recommendation algorithms, RSS, push subscriptions)
- Teacher lectures, students listen → reverse: students lecture, teacher listens (Feynman method, student-led PBL)
- Server-side rendering → client-side rendering → back again: edge rendering
- Bigger models = better → reverse: smaller models = faster (MobileNet, edge inference)
AI era versions:
- AI assists humans to write code → reverse: humans assist AI to verify code (HITL, human as verifier)
- Large model centralized inference → reverse: small model local inference (llama.cpp, Apple Intelligence)
- Model learns from data → reverse: data learns from model (Data Distillation)
Practical tool:
- Write down three core assumptions of X (who initiates? who benefits? what is input/output?)
- Invert each assumption one by one
- Check: in the inverted world, who gets more value?
Live Example: Six Paths on RAG
Apply the hexagon to Retrieval-Augmented Generation:
| Path | Operation | Resulting direction |
|---|---|---|
| Xd | RAG is text→text; generalize the dimension | Multimodal RAG (image/video/code → answer), temporal RAG (memory evolves over time) |
| X+Y | RAG + knowledge graph | GraphRAG (Microsoft), relational reasoning rather than flat retrieval |
| Hammers for nails | Core capability: “inject external memory into LLM” | Codebase Q&A, document review automation, legal compliance checking, drug database queries |
| Nails for hammers | Need: let LLM know what it doesn’t know | Fine-tuning / RAG / Tool use / Long context / Continual learning |
| X++ | Adaptive RAG | Dynamically selects retrieval strategy per question type (dense / sparse / graph) |
| X→X̄ | RAG: model queries external → reverse: external pushes to model | Push-based context (Proactive Context Injection) — like RSS but for agents |
Six paths. Fifteen minutes. Six directions, 2-3 of which may be genuinely unexplored.
Usage Recommendations
When you’re stuck on a problem: Use “Nails for Hammers” — force yourself to list 10 solutions before evaluating.
When you have a new technology or tool: Use “Hammers for Nails” — abstract the core capability, find all potential applications.
When you want incremental innovation: Use “X++” — pick 2-3 adjectives and quickly generate variants.
When you want disruptive innovation: Use “X→X̄” — find the domain’s most entrenched assumption and invert it.
For product planning: Run all six paths (5 minutes each) as your divergent thinking phase, then pick 3 most promising directions for convergent evaluation.
Effective team workflow:
- Each person independently runs all 6 paths, writing at least 2 ideas per path (produces 6 × team size ideas)
- Post all ideas on a whiteboard without discussion; cluster silently first
- Vote on each cluster; take top 5 for deeper analysis
- Filter top 5 through Raskar’s own criteria: Impact × Novelty × Feasibility
Original Sources
- Wired UK (2012): Raskar’s own article, “Inventing a New Field in Vision,” published in Wired magazine’s November 2012 Start column: http://www.wired.co.uk/magazine/archive/2012/11/start/inventing-a-new-field-in-vision
- TEDxBeaconstreet talk: Raskar presents the Hexagon in full, with live examples: http://tedxbeaconstreet.com/rameshraskar/
- Raskar’s MIT page: Camera Culture research group, representative projects, and femtophotography: http://web.media.mit.edu/~raskar/
Raskar built this framework to answer a real problem he faced: how does a research group systematically generate the next project idea, rather than waiting for inspiration? The Hexagon is his answer — a repeatable, teachable, team-usable process.
关于本站 · 免责声明
🍄 Mushroom Research Blog 是非营利、免费公开的个人科技观察博客与公众号 XStack18,不接受商业合作、不代表任何企业或机构立场,也不谋求商业利益。我们以个人视角客观中立地记录和分析 AI、Web3 等领域的最新模型发布与技术动态——不止转述新闻标题或二手信息,而是给出有独立思考的深入分析,希望帮更多人获得有价值的一手科技认知。
⚠️ 文中介绍的开源代码与模型,仅供学习交流与技术借鉴。它们大多仍处于早期阶段,有待进一步研究和验证,请勿直接用于工作或生产环境;如需采用,请先自行充分测试,并核实其许可证与安全性。
Open-source code and models featured here are shared for learning and reference only. Most are early-stage and still need further study and verification — please don't use them directly in your work or in production. Test them thoroughly and check their licenses and security first.
- 本站文章均为作者基于公开信息的个人研究与观点整理,不代表文中提及的任何公司、产品、模型的官方立场,未与其构成商业关联或合作关系。
- 科技行业信息更新极快,我们尽力保证内容准确、及时,但不对完整性、实时性做绝对保证,具体请以相关企业/项目官方公告为准。
- 文中引用的第三方商标、产品名称、图片、数据等版权归原权利人所有,我们会尽量注明来源;如你认为存在版权疑问或侵权,请通过下方邮箱联系我们,收到通知后会尽快核实处理(更正、加注来源或删除)。
- 文章内容仅为技术科普与个人观点,不构成投资、法律或其他专业建议,据此进行任何决策的后果需自行判断和承担。
📮 侵权 / 勘误 / 合作咨询:hello@mushroom.cv
💬 评论与讨论
使用 GitHub 账号登录后发表评论