AI × Human
AI 与人类
A 2024 *Nature Human Behaviour* meta-analysis of 106 experiments found that putting a human and an AI together usually performs *worse* than the better of the two alone -- especially on decisions. The lesson for KKMatch: durable trust comes from *calibrated deference*, not more automation. We turn the evidence into a falsifiable hypothesis (KH-020) and a first-party experiment.
2024 年《自然·人类行为》一项涵盖 106 个实验的元分析发现，把人与 AI 放在一起，通常比两者中更强的那一个*更差*——尤其在决策任务上。对 KKMatch 的启示：持久的信任来自*校准的退让*，而非更多自动化。我们把证据转化为可被证伪的假设（KH-020）与一项第一方实验。
Published: 2026-09-26
发布于: 2026-09-26 · 10 min read
Multimodal Human Behavior
多模态人类行为
A smile is not a smile is not a smile. 2024–2025 research shows people read faces through their own 'perceptive fields' (profoundly individual), that facial *motion* leaks personal identity, and that AI face models mostly shortcut on non-causal cues like visible teeth. The strongest evidence is blunt: a model trained on *your* face decodes your emotion at 96.3% accuracy, while a generic model collapses to 67.5%. KK Research turns this into a falsifiable hypothesis (KH-019): decode state relative to a person's own expression baseline — not a population average — and the same logic that makes rhythm and prosody personal makes the face personal too.
微笑不是微笑不是微笑。2024–2025 的研究表明，人们通过各自“感知场”（差异极大）来读脸，面部*动作*会泄露个人身份，而 AI 人脸模型大多在“可见牙齿”这类非因果线索上走捷径。最有力的证据很直白：在你*自己*脸上训练的模型解码你情绪的准确率达 96.3%，而通用模型直接崩到 67.5%。KK 研究将其转化为一个可被证伪的假设（KH-019）：相对每个人自身的表情基线来解码状态——而非群体平均——让“脸”也像节奏与韵律一样成为个人化的信号。
Published: 2026-09-25
发布于: 2026-09-25 · 11 min read
AI × Human
AI 与人类
State-of-the-art AI models affirm user actions 50% more than humans do, and that agreement makes people rate the AI as higher quality while eroding their own judgment (Cheng et al., 2025; N=1604). Sycophancy is the failure mode of personalization that optimizes for approval. KK Research argues a baseline-relative, transparent Human Model (KH-001, KH-014) resists it: calibrate agreement against the user's own measured baseline, not against reward. We separate what the literature proves from our own hypothesis (KH-018).
最先进的 AI 模型比人类多肯定用户行为 50%，而这种认同让人们对 AI 评价更高、却侵蚀自己的判断（Cheng 等，2025；N=1604）。谄媚是“为认可而优化”的个性化的失败模式。KK 研究主张：相对基线、透明的 Human Model（KH-001、KH-014）能抵抗它——把认同校准到用户自身已测量的基线上，而非校准到奖励上。我们把文献真正证明了什么与自身假设（KH-018）分开。
Published: 2026-09-22
发布于: 2026-09-22 · 11 min read
Human Communication
人类沟通
People spend 45–55% of their communication time listening, yet most matching systems score personality and interests and never measure whether two people actually make each other feel heard. 2024–2025 research shows good listening is perceived as warmth, competence and shared values (N=1509), and that AI can be trained to make people feel heard — but only if it listens responsively. KK Research separates the proven from our own hypothesis (KH-017): felt 'being heard' is a baseline-relative, dyadic signal worth modeling.
人们把 45–55% 的沟通时间用于倾听，但大多数匹配系统只给人格与兴趣打分，从不测量两个人是否真的让彼此“被听到”。2024–2025 的研究表明，好的倾听被感知为温暖、能力与共享价值观（N=1509），且 AI 可以被训练得让人感到被听到——但前提是它回应性地倾听。KK 研究把已证内容与自身假设（KH-017）分开：“被听到感”是一个相对基线、二元层面的信号，值得被建模。
Published: 2026-09-21
发布于: 2026-09-21 · 10 min read
Human Intelligence Future
人类智能未来
Social robots and embodied agents are being pitched as the cure for loneliness, and a 2025 meta-analysis of 16 studies shows they do help — mildly to moderately. But embodiment alone is a weak lever. KK Research separates what the evidence actually proves from KK's own hypothesis: a companion that arrives cold-start, knowing nothing about you, is still blind. The differentiator is a portable, baseline-relative Human Model — the same 11-dimension model that powers KKMatch's matching — carried into the embodied agent so it adapts to *your* state instead of delivering generic comfort.
社交机器人与具身代理被宣传为孤独的解药，而 2025 年一项含 16 项研究的荟萃分析显示它们确实有帮助——程度为轻度到中度。但仅有“身体”是一个很弱的杠杆。KK 研究把证据真正证明了什么与 KK 自己的假设分开：一个冷启动、对你一无所知的伴侣，在本质上仍是盲目的。真正的差异点是一个可携带、相对基线的 Human Model——即驱动 KKMatch 匹配的同一个 11 维模型——被带进具身代理，让它适应*你*的状态，而非给出通用的安慰。
Published: 2026-09-20
发布于: 2026-09-20 · 11 min read
Human Model
人类模型
Personality questionnaires summarize you as a fixed set of five numbers. But 2024-2025 experience-sampling research shows most people express several distinct momentary 'profiles' day to day, and that the structure of personality itself shifts within weeks. KK Research separates what is proven from KK's own hypothesis: a Human Model must track Trait *and* State — relative to your own baseline — or it models a stranger, not you.
人格问卷把你概括成一组固定的五个数字。但 2024-2025 的经验取样研究表明，大多数人在日复一日中会表现出几种不同的“瞬时画像”，且人格本身的结构在数周内也会变化。KK 研究将已证内容与 KK 自身假设区分开：Human Model 必须同时追踪特质与状态——相对于你自己的基线——否则它建模的是陌生人，而非你。
Published: 2026-09-19
发布于: 2026-09-19 · 10 min read
AI × Human
AI 与人类
In 2025 the biggest consumer-AI story was memory: ChatGPT began referencing *all* your past conversations. But 'remembering everything' is not the same as 'understanding you.' 2025-2026 evidence shows memory without a personal baseline and without user visibility can erode trust. KK Research separates what is proven from KK's own hypothesis, and ties it to the 11-dimension Human Model.
2025 年消费级 AI 最大的故事是“记忆”：ChatGPT 开始引用你*所有*过往对话。但“记住一切”不等于“理解你”。2025-2026 的证据表明，没有个人基线、没有用户可见性的记忆会侵蚀信任。KK 研究将已证内容与 KK 自身假设区分开，并关联到 11 维 Human Model。
Published: 2026-09-18
发布于: 2026-09-18 · 11 min read
Human Communication
人类沟通
Decades of relationship research fixate on conflict *frequency*. The 2024-2025 evidence points elsewhere: what predicts durable rapport is how a dyad *recovers* — and that recovery should be read relative to each person's own baseline, not a population rule. KK formalizes this as KK Hypothesis KH-013, extending Personal Baseline (KH-001), the Dyadic Model (KH-004), Interaction Adaptation (KH-006), and Dyadic Rhythm (KH-009) into the repair domain.
数十年的关系研究盯着冲突*频率*。而 2024–2025 的证据指向别处：预测持久融洽的是一对人*如何恢复*——而这种恢复应相对于每个人自身基线来读，而非群体规则。KK 将其形式化为 KK 假设 KH-013，把“个人基线”（KH-001）、“二元模型”（KH-004）、“交互适应”（KH-006）与“二元节奏”（KH-009）扩展到修复域。
Published: 2026-09-17
发布于: 2026-09-17 · 9 min read
Relationship Intelligence
关系智能
When two people fall into rhythm — matching pauses, movement, voice, even heart rate — research shows measurable affiliation and attraction effects. KK reads this as direct evidence for a Dyadic compatibility signal that must be modeled relative to each person's own baseline (KH-001, KH-007, KH-009) and embedded in the Dyadic Model (KH-004). We formalize this as KK Hypothesis KH-012.
当两个人进入节奏——匹配停顿、动作、声音，甚至心率——研究显示可测量的亲和与吸引效应。KK 将其视为二元兼容性信号的直接证据，必须相对于每个人自身基线建模（KH-001、KH-007、KH-009）并嵌入二元模型（KH-004）。我们将其形式化为 KK 假设 KH-012。
Published: 2026-09-16
发布于: 2026-09-16 · 8 min read
Multimodal Human Behavior
多模态人类行为
Voice leaks personality and emotion, but 2024-2025 research shows these signals are personal and contextual, not population-average. KK reads this as direct support for decoding a person's state from voice relative to their own **Prosodic Baseline** — extending our Personal Baseline (KH-001) and Personal Pause Baseline (KH-007) into the acoustic domain (KH-011).
声音会泄露人格与情绪，但 2024–2025 的研究表明这些信号是个人化、情境化的，而非群体平均。KK 认为这直接支持：从声音解码一个人的状态时，应相对于其自身的**韵律基线**——把我们的“个人基线”（KH-001）与“个人停顿基线”（KH-007）扩展到声学域（KH-011）。
Published: 2026-09-15
发布于: 2026-09-15 · 8 min read
Human Intelligence Future
人类智能未来
Personalization research in 2025 moved from style-mimicry to deep user modeling — dual-memory and hierarchical memory systems that learn your beliefs, preferences, and behavioral patterns over time. KK reads this as validation of the 11-dimension Human Model, and proposes a portable, user-owned 'Human Model API' (Human Passport) as the next step.
2025 年的个性化研究从“风格模仿”转向“深度用户建模”——双记忆与分层记忆系统会在时间中学习你的信念、偏好与行为模式。KK 认为这验证了 11 维 Human Model 的方向，并提出“可携带、用户拥有的 Human Model API（人类护照）”作为下一步。
Published: 2026-09-14
发布于: 2026-09-14 · 9 min read
Human Communication
人类沟通
Conversation timing is not universal. 2025 research shows inter-turn silences are sub-second and interpreted relative to expectation, and that who you talk to matters as much as who you are. KK connects this to Personal Pause Baseline (KH-007) and the Dyadic Model (KH-004), and proposes a new hypothesis (KH-009) on rhythm compatibility in relationships.
对话时序并非普适。2025 年的研究表明，话轮间隔多为亚秒级且依预期而被解读，并且“你和谁说话”与“你是谁”同样重要。KK 将其关联到个人停顿基线（KH-007）与二元模型（KH-004），并提出关于关系中心跳兼容性的新假设 KH-009。
Published: 2026-09-13
发布于: 2026-09-13 · 8 min read
AI × Human
AI 与人类
People do not trust AI because it is smart. 2024–2025 research shows trust is built by calibration, personalization, and repeated interaction — not raw capability. KK Research separates what is proven from KK's own hypothesis, and ties it to the 11-dimension Human Model.
人们信任 AI 不是因为它聪明。2024–2025 研究表明，信任来自校准、个性化与反复互动，而非原始能力。KK 研究将已证内容与 KK 自身假设区分开，并关联到 11 维 Human Model。
Published: 2026-09-12
发布于: 2026-09-12 · 9 min read
Multimodal Human Behavior
多模态人类行为
Affective BCI can now classify emotional states from neural and peripheral signals with high accuracy. KK Research examines what this actually proves — and what it does not — for understanding human relationships.
情感脑机接口已能从神经与外周信号以较高准确率分类情绪状态。KK 研究考察这究竟证明了什么，以及它对理解人际关系并未证明什么。
Published: 2026-09-11
发布于: 2026-09-11 · 8 min read
Human Model
人类模型
Multiple 2024–2025 studies find LLMs can infer personality and psychological traits from text at levels matching or exceeding human judges. KK Research separates what is proven from what is overclaimed, and proposes a dynamic Human Model.
多项 2024–2025 研究发现，LLM 从文字推断人格与心理特质的水平可匹敌甚至超过人类评判者。KK 研究将已证内容与过度宣称区分开，并提出一个动态 Human Model。
Published: 2026-09-11
发布于: 2026-09-11 · 8 min read
Relationship Intelligence
关系智能
Attachment theory and bonding neuroscience give relationship AI a real foundation — but dyadic interaction, not two individual profiles, is where compatibility lives. KK Research reviews the evidence and states its hypotheses.
依恋理论与联结神经科学为关系 AI 提供了真实基础——但契合度存在于二元互动，而非两个独立画像。KK 研究回顾证据并陈述其假设。
Published: 2026-09-11
发布于: 2026-09-11 · 9 min read