RSI trends are all about models. Could it be part of you?
自进化主体和对象都是模型，能不能是人自己？
Machine RSI ends at benchmark and scale. Your best state does not.
机器RSI 受限目标和规模。追寻最佳状态，是人的无限游戏。
Loop your seed improvers: body, context, wearable, model, brain.
种子优化循环：身体、上下文、可穿戴，模型，大脑。
Recursive从未成为用户体验
Opus 4.6 wrote wisely. 5.0 only codes, while Americans still mourn for GPT-4o.
Opus 4.6 写作绝佳，5.0只能代码。至今怀念GPT4o是美区热门话题。
Open Oura or Whoop. Is that screen actually you?
打开 Oura /Whoop App，它展示的到底是不是你？
Add Plaud or Looki, pipe it into ChatGPT Health or Claude Loops?
以上+Plaud/Looki接进 ChatGPT Health/Claude Loops，你会吗？
AI has never remembered or absorbed your taste or state. It only reads input.
AI 从未觉察和习得你的品味和状态，都是读取你的输入。
Self-improvement: SOTA model → personal reward loop → recursive persona model.
自进化会从众人模型，到个人 reward 循环，最后递归个人模型。
Data volume sets tunable parameter. So it has to be always-on.
数据量决定了可调节参数量：一定是Always-on。
Personalization sets the reward. So it has to be No Setup. Zero Skill.
个性化决定了Reward：一定是 No Setup. Zero Skill.
Health, focus, creation — if AI's gain is first-order, you become the overseer.
无论健康、效率、创作，如果AI只实现一阶加速，人就会沦为监工。
为什么不是Mem、PEFT、LoRA或各种SecondMe？
SOTA models outperform human experts — or at least scale skills they've shown.
SOTA Model 超过人类专家水平，至少可以规模化他们展现过的能力。
Dense models come close on domain skills, at a fraction of the cost.
Dense model在特定能力上接近，人均后训练成本显著降低。
MCP collections like OpenConnector already wire up internet context.
类似OpenConnector等MCP集合已经拉通互联网上下文。
Open memory projects like Mem0 can't carry multi-modal alignment.
Mem0等开源记忆项目，无法承载多模态人类数据对齐。
Hazy Cartridges, Transformer² — no consumer product path to scale.
Hazy Cartridge、Transformer²，根本没有消费产品和规模化工程可行性。
SOTA model steers, follows expert recipes, patrols loops.
最强模型发现和创造Skill，遵循专家意见和巡视问题。
Dense models move from a cohort to a person.
Dense Model 会从人群、领域，走向个人化。
Reward comes from real trajectories — and preference alignment.
reward 来自真实的工具轨迹，也来自真实偏好对齐。
One month of tuning parameters is 100x what academia and open source hold.
1个月可调节参数是现有学术和开源项目的100x。
Privacy unlocks the full stack, and a recurring revenue stream.
隐私诉求会激活full stack更持续的收入流。
可穿戴的切入形态
Ugliest wearable of the year. 7-day test, trending on Bilibili & Douyin.
Not health, not productivity. What audiences want is being aware and awake.
观众的共同兴趣不是健康，不是提效，是觉察自己，觉察自我模式。
Everything shipping through Q1 2027 trades away the AI agency we can ship.
已发售和预众筹至27年1季度发售的产品，主动智能落差很大。
Multimodal and all-week wear — still no design that does both.
多模态 和 持久戴 还没有两全方案。
Persistent: lab direction + expected gain.
续航：合作方向和可能效果。
Light weight: lab direction + expected gain.
轻量：实验室合作方向和可能效果。
Software works the moment you reserve. Rewards on context, open to any agent.
软件追求即刻可用，开放嵌入各种agents。
大家经历的三条线，现在撞到一起
1bn USD Rev Management
10 亿美元营收盘子
Huawei India — the only full-stack unit outside China.
华为印度：消费者业务在中国以外，唯一产研销一体的国家公司。
Multi-token prediction
Multi-token prediction
Now core to DeepSeek V4 and Qwen3 / 4-Next inference.
后来成了 DeepSeek V4、Qwen3/4-Next 的核心推理架构。
M6 & PLUG, first adopter
M6 & PLUG 首个业务方
First Alibaba keynote as head of Intelligent Connectivity.
第一次在阿里发布会上，以智能互联总裁身份出场。
Visual ChatGPT · 34k stars
Visual ChatGPT · 34k star
Core dev at Microsoft. First LLM tool-orchestration stack.
为微软做核心开发，当时第一个 LLM 多模态编排工具。
Qwen's first personalized LLM
Qwen 第一个个性化大模型
For games and hardware. Its alignment set still ranks.
面向游戏与硬件；配套的对齐数据集至今在 ModelScope 前列。
Ring and band — leading product experience
戒指和手环，都有领先产品经验
Software lead from RingConn,
hardware lead from Amazfit.
软件产品负责人来自 RingConn，
硬件负责人来自华米 Amazfit。
The three threads converge
三条线合流
Scroll the chart sideways →图可以左右拖 →