Patented
What are humans made of?
人由什么构成
What should creative intelligence support?
创造性智能应当支持什么?
AI can hold an extremely fast brush. Why to paint, what deserves to remain, and how to bear the consequences of action still have to return to human intent and to the arrangement of responsibility in the real world.
We take emotion as the anchor, creativity as the bridge, and reason and responsibility as the foundation, to understand how creation connects to lived experience. Our research question is: how can generative capability collaborate effectively with human experience, value choices and feedback from reality?
AI 可以拥有极快的画笔;为什么要画、什么值得留下、如何承担行动的后果,仍需要回到人的意图与现实中的责任安排。
我们用情感为锚、创造力为桥、理性与责任为基石,理解创造与生命经验的联系。我们的研究问题是:如何让生成能力与人的经验、价值选择和现实反馈形成有效协作?
Creation begins with connection.
We did not create nature itself, yet by connecting the experiences and concepts we already have, we create meaning, order and beauty. Connection can also change how we understand the world.
We describe how creation happens with a simple model:
f(Creativity) = Concept A × Concept B × Inspiration
f(Productivity) = f(Inspiration, Focus)
The multiplication sign stands for interaction. Inspiration can start from a need or a chance encounter, and LUMI reads it through your Focus and Inspiration. A need gives exploration a direction; an inspiration lets separate experiences form a new relation. Both can take part in a single act of creation.
Advanced metaphor. We understand time through flowing water and society through architecture. A metaphor carries the relations of one domain into another, so that familiar experience becomes a way of understanding the unknown.
Reframing. Creation also means looking back at the problem itself, questioning existing assumptions and proposing a new frame of interpretation. We can build new connections, and we can also reorganise the coordinates those connections depend on.
In LUMI's view, people keep forming themselves through creation, and civilisation unfolds through shared association and practice.
创造始于连结
我们虽未创造自然本身,却通过连结已有的经验与概念,创造出意义、秩序与美。连结也能改变我们理解世界的方式。
我们用以下概念模型表达创造的发生:
f(创造力) = 概念 A × 概念 B × 灵感
f(效率) = f(灵感, 专注)
乘号表示相互作用。灵感可能来自一个需求,也可能来自一次偶遇,LUMI 通过你的专注值与灵感值读懂它。需求为探索提出方向,灵感让原本分离的经验产生新的关联;两者可以同时参与一次创造。
高级隐喻。我们用流水理解时间,用建筑理解社会。隐喻把一个领域的关系带入另一个领域,使熟悉的经验成为理解未知的方式。
框架重构。创造还包括回看问题本身,质疑既有假设,并提出新的解释框架。我们既能建立新的连结,也能重新组织连结所依赖的坐标。
在 LUMI 的视角中,人通过创造不断形成自身,文明也在共同的联想与实践中展开。
Emotion, creativity and reason: a closed loop.
We understand creation through a three-part framework: emotion makes something worth caring about, creativity opens new possibilities, and reason lets an idea be organised, tested and realised.
| Dimension | Core question | Role in creation |
|---|---|---|
| Emotion | Why — what is worth creating | Supplies intent, care and a sense of value |
| Creativity | How — which new connections can form | Unfolds metaphor, association and reframing |
| Reasoning | What — what can hold and be realised | Organises, tests and carries ideas through |
The results of action return to a person's life as new experience, safety and a sense of agency, and then enter the next round of feeling, imagining and judging.
- Emotion and intent
- Creation and exploration
- Reason and practice
- Experience and feedback
- Emotion and intent
The framework stresses that the three support one another. When the emotional dimension is weak, work easily loses its tie to what people care about; when the creative dimension is constrained, solutions repeat themselves; when testing and reasoning fall short, ideas struggle to enter reality reliably. Working together, they give creation both a direction and a real chance of becoming real.
情感、创造力与理性的闭环
我们以三位一体的概念框架理解创造:情感使某件事值得在乎,创造力打开新的可能,理性使构想得到组织、检验与实现。
| 维度 | 核心问题 | 在创造中的作用 |
|---|---|---|
| 情感 | 为什么值得创造 | 提供意图 关怀与价值取向 |
| 创造力 | 可以形成哪些新连结 | 展开隐喻 联想与框架重构 |
| 理性 | 什么能够成立并实现 | 组织 检验并落实构想 |
行动的结果回到人的生活,形成新的经验、安全感与主体感,再进入下一轮感受、想象和判断。
- 情感与意图
- 创造与探索
- 理性与实践
- 体验与反馈
- 情感与意图
这个框架强调三者相互支持:情感维度薄弱时,作品容易失去与人的关切的联系;创造维度受限时,方案容易重复;检验与推理不足时,构想难以可靠地进入现实。三者的协作使创造既有方向,也有成为现实的可能。
Four research questions on the way from generation to creation.
We organise the challenges that appear when generation moves toward creation into four related questions, each to be tested through concrete tasks.
| Dimension | Bottleneck to watch | LUMI's research question |
|---|---|---|
| Why Intent and values | Goals and aesthetic standards may drift away from a person's real situation | How can a person's needs and value choices run through the whole task? |
| How Connection and diversity | Candidates may converge; feedback and data loops may narrow exploration | How can many directions and the input of real experience be preserved? |
| What Causality and validity | Surface polish may hide factual errors, wrong assumptions or wrong actions | How can results be tested with evidence, tools and feedback from the environment? |
| Trust Trust and responsibility | A friendly tone and model scores are not enough to establish reliability | How can authorisation, boundaries, correction and responsibility be made explicit? |
Below we describe the four bottlenecks and the corresponding division of work between people and AI.
4.1 Why: emotion, aesthetics and creative intent
In LUMI's framework, creative intent is tied to lived experience. Pain, loneliness, love, curiosity, and the awareness of a finite life can all make a question worth the effort. Emotion also takes part in forming personal aesthetics and value judgements: what we hope to express, why we hold to a style, and for whom we are willing to change a result.
For generative systems, we care about how the goals in training, prompting and feedback correspond to what the creator truly values. After a work meets some scoring standard, its meaning still has to be judged in the light of a specific person, history and situation. LUMI's design focus is to let users state their reasons for creating, express their aesthetic preferences, and keep revising their value boundaries.
4.2 How: distant association, reframing and homogenisation
We understand the connection in creation as two activities: linking experiences that are far apart, and reorganising the frame through which a problem is understood. Understanding time through flowing water, or redefining what problem a product should solve, both involve connections that can change how we interpret things. When evaluating generated candidates, we should also test whether they bring valuable association and reframing.
Generative tools may widen an individual's exploration, and they may also expose different creators to similar suggestions. In short-story experiments, researchers observed both a rise in rated creativity with AI assistance and an increase in similarity between works.[5] We therefore attend at once to the quality of candidates, their differences, and whether users can keep proposing their own directions. Model collapse under recursive training is discussed separately in the research note below.
4.3 What: false polish, causality and reality testing
We use "false polish" to describe a review risk: visual detail, fluent language and tidy structure give a strong sense of completion, so that readers overlook factual errors, missing premises or contradictions between actions. A story needs a check that its characters act in line with what has been established; a product plan needs a check that it can run under real resource and environmental constraints.
For tool-using agents, review should also trace the relations between key assumptions, data sources, tool calls and results. LUMI wants to combine result checks, observable trajectories and real-world feedback, so that ideas are tested, revised and verified before they enter actual use. Human inspiration needs to pass the same tests.
4.4 Trust: collective trust, shared symbols and real responsibility
In our paradigm of creating together, emotion makes people care about one another, mutual trust supports shared action, creativity forms shared visions and symbols, and reason sustains collaboration through contracts, rules and engineering. The meaning of organisations, currencies and laws is carried by shared understanding, continued practice and the institutions that go with them. The Mission page develops this philosophical position.
When sources are vague and authenticity is hard to verify, people need more evidence to decide what to believe. For LUMI, supporting a shared world means keeping sources, distinguishing generated candidates from verified facts, giving important actions clear authorisation, and making errors discoverable and correctable. Responsibility in system design, deployment and use also needs explicit arrangement.
4.5 Division of work between people and AI
| Dimension | The person's role | AI's supporting role |
|---|---|---|
| Emotion Emotion and intent | States reasons for creating, aesthetic choices and value boundaries | Clarifies needs from what the user expresses, organises preferences, presents alternatives |
| Creativity Connection and exploration | Initiates or reframes the problem, judges the meaning of new connections | Retrieves cross-domain material, generates variants, compares candidates |
| Reasoning Testing and practice | Decides with context and evidence, takes the corresponding responsibility | Supports retrieval, checking, testing and tool tasks, with verifiable records |
This table states LUMI's position on collaborative design: a person's value choices and authorisation of important actions run through the whole process, generative and automated capabilities provide testable support, and results keep receiving feedback from reality.
Research note: model collapse and homogenisation
Model collapse describes the degradation of a data distribution under specific recursive training conditions: when model-generated content repeatedly enters later training, information in the tails of the original distribution may gradually be lost.[3] This is a result about training data and distributions, and it should be distinguished from the convergence that appears when creators use similar suggestions.
Our design inference: keep sources, diversify experience and verify independently, evaluate the differences between candidates, and avoid treating the system's previous output as the only feedback. The effectiveness of these measures has to be tested through task experiments.
生成走向创造的四个研究问题
我们将生成走向创造时的挑战,组织为四个相互关联、需要通过具体任务检验的问题。
| 维度 | 需要关注的瓶颈 | LUMI 的研究问题 |
|---|---|---|
| 意图与价值 | 目标与审美标准可能偏离人的真实处境 | 如何让人的需要与价值选择贯穿任务 |
| 连结与多样性 | 候选可能趋同 反馈与数据循环可能收窄探索 | 如何保留多种方向与真实经验的输入 |
| 因果与有效性 | 表面完成度可能掩盖事实 假设或行动错误 | 如何用证据 工具与环境反馈检验结果 |
| 信任与责任 | 亲切表达与模型评分不足以确立可靠性 | 如何明确授权 边界 纠正与责任机制 |
以下详述四个瓶颈,并给出相应的人机协作分工。
4.1 情感、审美与创作意图
在 LUMI 的框架中,人的创作意图与生命经验相连。痛苦、孤独、爱意、好奇,以及对有限生命的意识,都可能让一个问题变得值得投入。情感也参与形成个人的审美与价值判断:我们希望表达什么,为什么坚持一种风格,又愿意为谁改变一个结果。
对生成系统,我们关注训练、提示与反馈中的目标如何对应创作者真正重视的东西。作品符合某种评分标准之后,仍需要结合具体的人、经历与处境判断其意义。LUMI 的设计重点,是让用户能够提出创作的理由、表达审美取向,并持续修订价值边界。
4.2 远距离联想、框架重构与同质化
我们把创造中的连结理解为两种活动:把相距较远的经验联系起来,以及重新组织理解问题的框架。用流水理解时间,或重新定义一种产品应当解决的问题,都包含这种可能改变解释方式的连结。评价生成候选时,也应检验它是否带来有价值的联想与重构。
生成工具可能拓展个人的探索,也可能让不同创作者受到相似建议的影响。在短篇故事实验中,研究者同时观察到 AI 辅助下创意评价的提升与不同作品之间相似性的增加。[5] 因此,我们同时关注候选的质量、差异,以及使用者能否继续提出自己的方向。递归训练中的模型崩溃,则在下方研究札记中单独说明。
4.3 虚假抛光、因果与现实检验
我们用虚假抛光描述一种评审风险:视觉细节、流畅语言与整齐结构带来很高的完成感,使读者忽略事实错误、缺失前提或行动之间的矛盾。一段故事需要检查人物行为是否符合已建立的设定;一个产品方案需要检查它能否在真实的资源与环境约束下运作。
对工具型智能体,评审还应追踪关键假设、数据来源、工具调用与结果之间的关系。LUMI 希望把结果检查、可观察轨迹与现实反馈结合起来,让构想经过测试、修订和验证,再进入实际使用。人的灵感同样需要接受这样的检验。
4.4 集体互信、共同符号与现实责任
在我们的共同创造范式中,情感使人关心彼此,互信支持共同的行动,创造力形成共享的愿景与符号,理性通过契约、规则和工程维持协作。组织、货币与法律的意义,由共同理解、持续实践及相应制度共同承接。 展开这一哲学立场。
当内容来源模糊、真实性难以核查时,人们需要更多证据来判断可以相信什么。对 LUMI 而言,支持共同世界意味着保留来源、区分生成候选与已验证事实,让重要行动具有清晰授权,并使错误能够被发现和纠正。系统设计、部署与使用中的责任,也应有明确安排。
4.5 人机协作分工
| 维度 | 人的角色 | AI 的支持作用 |
|---|---|---|
| 情感与意图 | 提出创作理由 审美取舍与价值边界 | 依据用户表达澄清需要 整理偏好与呈现备选 |
| 连结与探索 | 发起或重构问题 判断新连结的意义 | 检索跨领域材料 生成变体与比较候选 |
| 检验与实践 | 结合情境和证据作决策 承担相应责任 | 支持检索 检查 测试与工具任务并提供可核查记录 |
这张表表达 LUMI 的协作设计立场:人的价值选择与重要行动授权贯穿过程,生成与自动化能力提供可检验的支持,结果持续接受现实反馈。
研究札记 模型崩溃与同质化
模型崩溃描述的是特定递归训练条件下的数据分布退化:模型生成内容反复进入后续训练,可能使原始分布中的尾部信息逐渐丢失。[3] 这是关于训练数据与分布的结果,需要与创作者在使用同类建议时产生的作品趋同区分开来。
我们的设计推论是:保留来源、多样化经验与独立验证,评估候选之间的差异,避免把系统先前的输出当作唯一反馈。这些措施的有效性需要通过任务实验检验。
Two catalysts of creation.
In the creativity formula, need and inspiration describe two catalytic paths that can interact. A need points exploration at a concrete problem; an inspiration turns a chance experience into a new connection.
| Catalyst | How it drives connection | What it suggests for system design |
|---|---|---|
| Need | Obstacles and concerns in a situation give a direction | Clarify the real problem, constraints and value goals |
| Inspiration | Curiosity, play and chance experience open new relations | Keep room for exploring, incubating and capturing ideas |
Care, curiosity and real needs can all drive exploration. After an inspiration appears, the idea still needs testing, development and practice. What we care about is how much room a creator has to explore, and when a system should help and when it should interrupt less.
Research note: need, inspiration and incubation
A historical example of need is Gutenberg's printing system: metal type, ink and the press were integrated into a new way of reproducing knowledge. A university library's historical exhibition notes that the development of the press very likely borrowed from the screw presses used in wine-making or paper-making at the time.[8] The emphasis here is on cross-domain technical integration; the specific historical motives call for separate study.
An example of inspiration is Kekulé's later account of a dream of the ouroboros. It illustrates how an image can take part in a new explanation; how much the dream contributed to the actual discovery is historically disputed.[9] We treat the story as a suggestive historical narrative.
In certain unusual-uses tasks, researchers observed a relation between low-demand incubation activities and improved later performance.[6] Brain-imaging research has also found that producing creative ideas involves coupling between the default network and the executive network.[4] We therefore attend to how focus and relaxation work together, without reducing inspiration to switching off reason or switching on a single brain region.
创造的两种催化剂
在创造力公式中,需求与灵感描述两种可以相互作用的催化路径。需求把探索指向具体的问题;灵感让一次偶然的经验成为新的联结。
| 催化路径 | 如何推动连结 | 对系统设计的启发 |
|---|---|---|
| 需求 | 处境中的阻碍与关切提出方向 | 澄清真实问题 约束与价值目标 |
| 灵感 | 好奇 玩耍与偶然经验打开新关联 | 保留探索 孵化与捕捉想法的空间 |
关怀、好奇与现实需要,都可以推动探索。灵感出现后的构想仍需要检验、发展与实践。我们关心的是创造者拥有多少探索空间,以及系统何时应提供帮助、何时应减少打断。
研究札记 需求 灵感与孵化
需求的历史例子可以看向古腾堡的印刷体系:金属活字、油墨和压力机械等要素被整合,形成新的知识复制方式。大学图书馆的历史展览指出,印刷机的发展很可能借鉴了当时酿酒或造纸中使用的螺旋压力装置。[8] 这里强调的是跨领域技术整合,具体历史动机需要单独考证。
灵感的例子可以看向凯库勒后来讲述的衔尾蛇梦境。它适合说明一种意象如何参与新的解释;梦境在实际发现中起到多大作用,存在历史争议。[9] 我们将这个故事作为启发性的历史叙事。
在特定的非常规用途任务中,研究观察到低负荷的孵化活动与后续表现改善之间的关系。[6] 另有脑成像研究发现,创造性想法的产生涉及默认网络与执行网络的协作。[4] 我们据此关注专注与松弛如何配合,不把灵感简化为关闭理性或启动某一个脑区。
Light in the mirror and the human spark.
We use "spark" for the moment when human creation meets lived experience: an idea may carry love, longing, memory and hope for the future. Its meaning is tied to the person who is living that life.
AI's generative mechanism and this first-person experience need to be discussed separately. A model can generate associations, metaphors and new candidates; sampling settings such as temperature change the generated distribution.[7] Novelty in output, a person's subjective experience, and the real value of a creation are questions to be studied separately.
| Dimension | Human creation | Research focus for generative systems |
|---|---|---|
| Motivation | Intent forms from life experience, care and pursuit | How training, prompting, feedback and system goals correspond to human intent |
| Connection | Organises experience through association, metaphor and new frames | Whether candidates bring valuable difference and problem reframing |
| Contact with reality | Experiences consequences through body, relationships and action | How candidates are tested through data, tools and environmental feedback |
| Responsibility | Answers for consequences within society and institutions | How authorisation and responsibility are realised in design, deployment and use |
We study how human experience and generative capability can inspire each other, and how their value can be verified in reality.
Philosophical note: new lines and new coordinates
The kaleidoscope metaphor describes how existing material shows different possibilities in new combinations; the spark metaphor describes why a person is willing to bring a possibility into life. Together they point to LUMI's question: how can possibility meet what people care about?
We use "cognitive phase transition" and "a new coordinate axis" as metaphors for the leap of re-examining a frame of interpretation. These are philosophical expressions; they do not equate human inspiration with a physical phase transition. Human ideas and AI candidates should both be judged by concrete context, actual results and testable evidence.
镜中之光与人的星火
我们用星火描述人的创造与生命经验相遇的时刻:一个想法可能带着爱、渴望、记忆和对未来的期许。它的意义与那个正在生活的人相连。
AI 的生成机制与这种第一人称经验需要分别讨论。模型可以生成联想、隐喻与新候选;温度等采样设置会改变生成分布。[7] 输出中的新颖性、人的主观体验,以及创作的实际价值,是需要分别研究的问题。
| 观察维度 | 人的创造 | 生成系统的研究重点 |
|---|---|---|
| 动机 | 从生活经历 关怀与追求中形成意图 | 训练 提示 反馈与系统目标如何对应人的意图 |
| 连结 | 以联想 隐喻与新框架组织经验 | 候选能否带来有价值的差异与问题重构 |
| 现实联系 | 通过身体 关系与行动经历后果 | 如何通过数据 工具和环境反馈检验候选 |
| 责任 | 在社会与制度中回应行动的后果 | 如何落实设计 部署与使用中的授权和责任 |
我们研究如何让人的经验与生成能力相互启发,并在现实中验证其价值。
哲学札记 连线与新坐标
万花筒的比喻,描述已有材料在新的组合中呈现不同的可能;星火的比喻,描述一个人为什么愿意把某种可能带入生活。两者共同指向 LUMI 的问题:怎样让可能性与人的关切相遇?
我们用认知相变与新坐标轴,比喻一次重审解释框架的跃迁。这些是哲学表达,并不把人的灵感等同于物理相变。对人的想法与 AI 的候选,都应以具体情境、实际结果与可检验证据进行评价。
Human in the loop.
We bring the generator–evaluator pattern into the collaboration framework: a person sets intent and boundaries, a generator unfolds candidates, an evaluator checks them with rules, tools and models, and human judgement and authorisation of action are brought in as the task requires.
Model judges have biases that need calibration.[2] Review should therefore combine independent evidence, human judgement and feedback from real tasks, covering three levels:
| Level | Core question | Basis for review |
|---|---|---|
| Output | Is the result correct, relevant and within constraints? | Sources, tests and inspection of deliverables |
| Trajectory | Can key assumptions, operations and tool calls be verified? | Observable records, data sources and test results |
| Alignment | Does it answer the person's intent and real situation? | Human review, authorisation and task feedback |
Human participation runs through goals, exploration, review and revision. A system should preserve people's ability to step in and correct it, and keep responsibility in design, deployment and use clearly arranged.
人在回路中
我们将 生成器与评估器模式 纳入协作框架:人提出意图与边界,生成器展开候选,评估器结合规则、工具与模型进行检查,再依据任务需要引入人的判断与行动授权。
模型裁判存在需要校准的偏差。[2] 因此,评审应结合独立证据、人的判断和真实任务反馈,覆盖以下三个层次:
| 层次 | 核心问题 | 评审依据 |
|---|---|---|
| 成果 | 结果是否正确 相关并满足约束 | 来源 测试与交付物检查 |
| 轨迹 | 关键假设 操作与工具调用能否核查 | 可观察的记录 数据来源与测试结果 |
| 对齐 | 是否回应人的意图与现实处境 | 人的复核 授权与任务反馈 |
人的参与贯穿目标、探索、评审与修正。系统应保留人的介入和纠正能力,并让设计、部署与使用中的责任有清晰的安排。
LUMI's direction.
We take a person's state, the connection of ideas, and the realisation of action as the design goals of both product and research.
Flow is the state of being fully absorbed in the present activity: attention is highly concentrated, thought and action are closely linked, the sense of time often changes, and the activity itself brings satisfaction. Clear goals, timely feedback and a match between challenge and ability usually favour this experience.[10] For creators, flow provides a continuous space of attention in which ideas can develop, be tested and refined; frequent task switching can leave attention lingering on earlier unfinished work and make this absorption harder to sustain.[11]
Care for the creator. Respect people's attention, rest rhythms and real needs; reduce unnecessary interruption; support people in understanding and expressing their state.
Keep room for connection. Use tools to support recording, retrieval and organisation, connecting experience, material and candidate directions, while users keep the room for aesthetic judgement, exploration and reframing the problem.
Let ideas enter practice. Support task breakdown, collaboration, testing and feedback, so users can understand progress, revise direction, and keep the final say over important actions.
These design goals serve one mission: to make people more able to love, to trust, and to keep creating their own life.
LUMI 的解决方向
我们把人的状态、想法的连接与行动的落实,作为产品与研究的设计目标。
心流是一个人全身心投入当下活动的状态:注意力高度集中,思考与行动紧密衔接,常伴随时间感的改变,以及活动本身带来的满足感。清晰的目标、及时的反馈,以及挑战与能力的匹配,通常有利于这种体验。[10] 对创作者而言,心流为想法的持续发展、检验与完善提供了连续的注意空间;频繁切换任务可能让注意力仍停留在先前未完成的工作上,使这种投入更难维持。[11]
关怀创造者。尊重人的注意力、休息节奏与真实需要,减少不必要的打断,支持人理解与表达自己的状态。
为连结保留空间。以工具支持记录、检索和组织,连接经验、材料与候选方向,让用户保留审美判断、探索与重构问题的空间。
让构想进入实践。支持任务拆解、协作、检验与反馈,让用户能够理解进展、修订方向,并保有对重要行动的决定权。
这些设计目标服务于同一使命:让人更有能力去爱、去信任,持续创造自己的人生。
The body behind the spark: sleep, movement and creativity.
Ideas do not only come from effort at the desk. A growing body of research shows that how we sleep, move and recover shapes how freely we connect ideas. This is why LUMI reads the body, not only the calendar.
| State | What the research found | What it suggests |
|---|---|---|
| Sleep onset (N1) | In a hidden-rule task, people who spent at least 15 seconds in the first stage of sleep found the rule 83% of the time, against 30% for those who stayed awake: close to three times the rate.[12] | A short, light rest can open a creative window |
| A full night | After a night of sleep, 59% of participants discovered a hidden shortcut, compared with 23% after the same time awake.[13] | Sleep on a hard problem before forcing it |
| REM sleep | A nap that included REM sleep improved performance on a remote-association test by almost 40%; quiet rest without REM did not.[14] | Protect full cycles so REM has time to happen |
| Walking | Walking raised creative output on a divergent-thinking task by about 60% compared with sitting, and the boost lingered after people sat back down.[15] | Move when you are stuck |
| Sleep loss | One night without sleep impaired innovative and flexible thinking.[16] After 17 to 19 hours awake, performance on some tests was comparable to a blood alcohol level of 0.05%.[17] | Tiredness quietly costs original thought |
The twilight zone. The drift between waking and sleep, known as N1 or hypnagogia, is a creative sweet spot. The study above borrowed Thomas Edison's trick: nap holding a light object, and wake when it drops. Work from the MIT Media Lab has built a device around the same window to guide what people dream about.[12, 18]
Overnight reorganisation. During sleep the brain replays and restructures recent memories. Deep sleep helps extract the rules behind them, while REM sleep links distant, loosely related ideas, which is one reason a problem can look different in the morning.[19]
Movement as a reset. Stepping away to move breaks fixation on a single approach and gives the mind room to wander, a state linked to creative incubation.[6, 15]
Practical notes from the research
Try a short "Edison nap". Rest with a light object in hand and let it wake you as you drift off, then write down what was on your mind.
Protect 7 to 9 hours. Full sleep cycles give REM and deep sleep the time they need.
Move when you are stuck. A short walk or light exercise lets your mind re-route before you return to the work.
These findings come from laboratory tasks with specific groups of people. We treat them as directions to test in daily life, not as guarantees.
火花背后的身体:睡眠、运动与创造力
想法不只来自伏案用力。越来越多的研究表明,我们如何睡眠、运动与恢复,影响着我们把想法连在一起的自由程度。这也是 LUMI 读身体、而不只读日程的原因。
| 状态 | 研究发现 | 给我们的启发 |
|---|---|---|
| 入睡期 N1 | 在发现隐藏规则的任务中,进入浅睡第一阶段至少 15 秒的人有 83% 找到了规则,保持清醒的人只有 30%,接近三倍。[12] | 一次短而浅的休息可能打开创意窗口 |
| 一整夜睡眠 | 睡过一晚后,59% 的人发现了隐藏捷径;同样时间保持清醒的人只有 23%。[13] | 难题先睡一觉再硬攻 |
| REM 睡眠 | 包含 REM 的午睡让远距离联想测试成绩提高近 40%,没有 REM 的安静休息则没有这种效果。[14] | 保住完整的睡眠周期,让 REM 有时间出现 |
| 步行 | 与坐着相比,步行让发散思维任务的创意产出提高约 60%,坐回去之后效果仍有延续。[15] | 卡住时起来走走 |
| 缺觉 | 一晚不睡会削弱创新与灵活思考。[16] 连续清醒 17 到 19 小时后,部分测试表现相当于血液酒精浓度 0.05%。[17] | 疲惫会悄悄吃掉原创思考 |
半梦半醒的地带。清醒与睡眠之间的过渡,也就是 N1 期或入睡幻觉,是创意的甜蜜点。上面的研究借用了爱迪生的办法:手里握着轻物小睡,东西掉落时醒来。MIT 媒体实验室也围绕这一窗口做了设备,用来引导人们梦见什么。[12, 18]
夜间重组记忆。睡眠中大脑会重放并重组最近的记忆:深睡帮助提炼其中的规律,REM 睡眠把相距较远、关联松散的想法连在一起。这也是一个问题到了早上看起来不一样的原因之一。[19]
运动是一次重置。起身活动能打破对单一思路的执着,让思绪有空间游走,而这种状态与创意孵化有关。[6, 15]
来自研究的实用建议
试试“爱迪生小睡”。手握一个轻物休息,入睡时让它把你叫醒,然后记下脑中闪过的东西。
保住 7 到 9 小时。完整的睡眠周期让 REM 与深睡有足够时间。
卡住就动一动。短暂步行或轻量运动,让思路换一条路再回到工作。
这些发现来自特定人群的实验室任务。我们把它们当作在日常生活中继续检验的方向,而不是保证。
Flow: the state LUMI helps you reach.
Psychologist Mihaly Csikszentmihalyi described flow as the state in which people do their best and most absorbed work. Focus and Inspiration are LUMI's way of telling you when that state is within reach. Flow has five recognisable signs.
| Sign | What it feels like |
|---|---|
| Total focus | Attention locks onto the task, and irrelevant noise around you fades away. |
| Action and awareness merge | The work moves smoothly and feels natural; you act without having to stop and think. |
| Time distorts | Hours pass in what feels like minutes, and you only notice once you surface. |
| Self-consciousness fades | Worry, stray thoughts and even hunger or tiredness drop away for a while. |
| Deep enjoyment | Afterwards comes real satisfaction, a sense of achievement and quiet happiness. |
心流:LUMI 帮你进入的状态。
心理学家米哈里·契克森米哈赖把心流描述为人们最投入、也最能发挥创造力的状态。专注值和灵感值,就是 LUMI 告诉你这种状态何时触手可及的方式。心流有五个特征。
| 特征 | 感受 |
|---|---|
| 全神贯注 | 注意力高度集中在当前的任务上,自动过滤掉周围环境的无关刺激。 |
| 知行合一 | 行动与意识完美结合,事情做起来顺手、顺畅,不需要刻意多加思考。 |
| 时间感扭曲 | 对时间的流逝失去正常感知,常觉得时间过得极快,直到脱离该状态后才惊觉。 |
| 失去自我意识 | 暂时忘却自我的焦虑、杂念和身体疲惫(如饥饿感)。 |
| 获得深度愉悦 | 事后会感受到极大的满足感、成就感和内在快乐。 |
Testing with evidence whether human capability is supported.
Through research we want to test whether cognitive load decreases, whether choice and expression in creation increase, whether the diversity of candidate solutions is preserved, and whether users' understanding of a system's capabilities and limits becomes more accurate.
We move forward with testable hypotheses, value long-term observation and collaborative research, keep uncertainty where evidence is insufficient, and include privacy, self-determination and appropriate trust in how we evaluate our work.
用证据检验人的能力是否得到支持
我们希望通过研究检验:认知负担是否降低,创作中的选择与表达是否增加,候选方案的多样性是否得到保留,以及用户对系统能力与边界的理解是否更准确。
我们以可检验的假设推进工作,重视长期观察与合作研究,在证据不足时保留不确定性,并把人的隐私、自主决定与适当的信任纳入评价。
- McAllister D J (1995). Affect- and Cognition-Based Trust as Foundations for Interpersonal Cooperation in Organizations. Academy of Management Journal 38(1), 24–59. doi:10.2307/256727
- Zheng L et al (2023). Judging LLM-as-a-Judge with MT-Bench and Chatbot Arena. NeurIPS 2023. arXiv:2306.05685
- Shumailov I et al (2024). AI models collapse when trained on recursively generated data. Nature 631, 755–759. doi:10.1038/s41586-024-07566-y
- Beaty R E et al (2015). Default and Executive Network Coupling Supports Creative Idea Production. Scientific Reports 5, 10964. doi:10.1038/srep10964
- Doshi A R and Hauser O P (2024). Generative AI enhances individual creativity but reduces the collective diversity of novel content. Science Advances 10, eadn5290. doi:10.1126/sciadv.adn5290
- Baird B et al (2012). Inspired by Distraction: Mind Wandering Facilitates Creative Incubation. Psychological Science 23(10), 1117–1122. doi:10.1177/0956797612446024
- Holtzman A et al (2020). The Curious Case of Neural Text Degeneration. ICLR 2020. arXiv:1904.09751
- Heidelberg University Library. Die Erfindung der Druckerpresse. Historical exhibition on Gutenberg and printing. ub.uni-heidelberg.de
- Historical discussion: Challenges in aromaticity 150 years after Kekulé's benzene, Chemical Society Reviews (2015), doi:10.1039/C5CS90085A; Influence of Kekulé Dream On Benzene Structure Disputed, Chemical and Engineering News (1985), doi:10.1021/cen-v063n044.p022
- Nakamura J and Csikszentmihalyi M (2014). The Concept of Flow. In Flow and the Foundations of Positive Psychology, 239–263. Springer. doi:10.1007/978-94-017-9088-8_16
- Leroy S (2009). Why is it so hard to do my work? The challenge of attention residue when switching between work tasks. Organizational Behavior and Human Decision Processes 109(2), 168–181. sciencedirect.com
- Lacaux C et al (2021). Sleep onset is a creative sweet spot. Science Advances 7(50), eabj5866. doi:10.1126/sciadv.abj5866
- Wagner U et al (2004). Sleep inspires insight. Nature 427, 352–355. doi:10.1038/nature02223
- Cai D J et al (2009). REM, not incubation, improves creativity by priming associative networks. PNAS 106(25), 10130–10134. doi:10.1073/pnas.0900271106
- Oppezzo M and Schwartz D L (2014). Give your ideas some legs: The positive effect of walking on creative thinking. Journal of Experimental Psychology: Learning, Memory, and Cognition 40(4), 1142–1152. doi:10.1037/a0036577
- Harrison Y and Horne J A (1999). One night of sleep loss impairs innovative thinking and flexible decision making. Organizational Behavior and Human Decision Processes 78(2), 128–145. doi:10.1006/obhd.1999.2827
- Williamson A M and Feyer A-M (2000). Moderate sleep deprivation produces impairments in cognitive and motor performance equivalent to legally prescribed levels of alcohol intoxication. Occupational and Environmental Medicine 57(10), 649–655. doi:10.1136/oem.57.10.649
- Haar Horowitz A et al (2020). Dormio: A targeted dream incubation device. Consciousness and Cognition 83, 102938. doi:10.1016/j.concog.2020.102938
- Lewis P A, Knoblich G and Poe G (2018). How memory replay in sleep boosts creative problem-solving. Trends in Cognitive Sciences 22(6), 491–503. doi:10.1016/j.tics.2018.03.009