Notes

A line or two, written down before it gets away.

2026
  1. 13 Nov

    有一个事情我在做的，或者是一个定位吧 - 通过记录真实工厂，以及我对机器人应用的观察、判断和验证，逐渐建立“具身落地的现场实践者”这个印象。让工厂愿意向我开放问题和场景，也让技术伙伴愿意和我一起尝试

  2. 12 Nov

    Just noticed an interesting trend for me
    While AI gets better at writing, the more I write without it. Influenced by Paul Graham I am very aware of that writing is thinking, and no one can do your thinking for you.
    Would be very scary if you lose the ability to critically think.

  3. 12 Nov

    When I started out in robotics I was SO stoked about humanoids
    but the more I learn about the space & the more factories and warehouses, the more actual industrial sites I visited, the more I am leaning towards specialized robots versus humanoids for industrial AI
    From intuition I believe use cases with the following characteristics would be best suited for humanoids:
    - open-ended
    - general
    - low volume & low-scale
    - emphasis on emotion & interaction
    But when it comes to industrial sectors, you want smart, reliable, robust BUT specialized robots for the job. Not throwing humanoids at everything.
    Sure bitter lesson is a real thing but I don't think it applies that well to hardware.
    An interesting case to look at in home robots:
    If you have dish washers already, it doesn't mean humanoids will replace them, it's smartest to treat dishwasher as a tool for humanoids versus trying to make a humanoid learn how to wash dishes with their bare hands.

  4. 12 Nov

    Be a man with honor.

  5. 12 Nov

    这两天很可能要养一只豹猫了！！！

  6. 11 Nov

    Three missions / motivations that fuel my work in robotics, daily:
    1) Liberating humans from working dangerous, laborious, harmful jobs
    2) Unlocking more SMBs to use affordable, intuitive robots. They have been blocked by traditional automations (due to complexity, lack of flexibility, expensiveness, and needing experts to program them)
    3) Figuring out a way (with policy & regulation) to create new opportunities in society for workers displaced by robots in 8-10 years. Create new policies, incentives, programs, jobs, roles, anything...

  7. 10 Nov

    Robot learning & general robotics is so sexy because it allows people to dream about deploying robots without retrofitting the deployment environment
    unlike current / traditional automation where you need to change the env dramatically for the robots
    Basically humans are the ones absorbing all the entropy of the system and carefully program the robot so it can repeat motions. Its a delicate system that breaks even only when there's a slight variability
    But even if general robotics advances dramatically, does it really make sense to just toss the robots into unstructured environment and hope data & scale would make the robot learn and adapt quickly?
    The answer is probably no, and it would be more practical (and clever) to reduce entropy by the means of structure as scaffolding, removed layer by layer as the models get better
    What I mean by structure is things like
    - Using fixtures so parts show up at known position & orientation
    - Using form-fit finger-tips to make sure the object self-aligns in the gripper
    - Standardizing the feeding
    Then you remove one layer at a time to let the model absorbs more errors without disrupting production at the deployment's site
    This is basically the same principle @Waymo adopts when expanding its ODD (aka operational design domain) - they started in sunny, wide-road Phoneix suburbs, build guardrail and then expand city by city as the system earns the right to do so.
    If you can design geometry to absorb errors, why not use it instead of perception? Use it well, use it to speed up your game.

  8. 05 Nov

    Everyone is all about policy and robot learning because it is believed to be an efficient & scalable way to teach robot new skills
    However is this true?
    One often doesn't consider the time and resource spent during the data curation and data ops phase when thinking about learning
    I would lose my mind if I have to collect tens of hours of data (or just a few hours of data even) to train or fine-tune a model
    I only care about fastest time to value

  9. 04 Nov

    我觉得我这一个月最大的问题是没有搭建 example scene in the office... 我现在要搭一些工作台，物流分拣，上下料，插拔，抛光的工位

  10. 04 Nov

    找到那些工人又讨厌，然后企业主又需要的工序，这个才是商业闭环的关键

  11. 04 Nov

    做机器人生产力部署的，为终端用户负责的 - 你需要承担的不仅是技术上的不确定，还有market risk

  12. 04 Nov

    昨天去参观 HS precision 之后，我有一个感受：人并不是他们最大的痛点，而是扩张产能。
    产能扩张后，老板需要掏钱去买机械，这是重资产的投入。投入之后，他不知道来年的产能是否能够让他 justify 这个 investment，所以机器的投入才是大头，而不是人力。这就是做制造业一个比较尴尬的点。
    Unfortunately, may not be a good case. 所以有两个事情：
    1. 找另外一个 sector；
    2. 把它当成一个试点，先积累真实部署的经验再说。
    但我又觉得有点纳闷，你问他们招工和员工的留存率是不是一个 concern？他们当然会说：“哇，招工人很难很麻烦，流失率很高。”可是我就是感觉这不是他们最重要的痛点，而且人员很便宜，这就是最大的问题。
    但如果是在美国那边的话，这个痛点就很真实了。就像他昨天给我一个很确切的数字：你搞一个 10 个机器的产线，去那边招工人，一个月的成本要 50 万人民币。而且他们还有各种各样的劳动法保护，你不能让他们去三班倒，也不能让他们加班。
    劳工确实是一个痛点，这对美国来说确实是一个很大的痛点，只不过在中国并不是。
    所以我的另外一个结论也是 - 如果这个工序的成本主要是靠人工的话，那么机器人落地会更现实（而不是 opex）

  13. 03 Nov

    Fiverr for physical work, done by robots :)

  14. 29 Oct

    创业始终要想怎么滚雪球，幂次定律。软件创业的滚雪球以前很明显，就是靠产品做出来后的 network effect（用户量），我觉得这个的前提是做出这个产品的工程 efforts 本身有够大。但现在不一样了，生成产品已经被自动化，所以 where value acrues will shift. 模型是这个时代的能够滚雪球的东西，因为它把 events 都给吃进去了自己的权重里。

  15. 28 Oct

    Intelligence is not about what you can already do. It’s about how quickly you can learn to do something new.

  16. 28 Oct

    今天看到了 Machenta emerged from stealth - 前段时间，我一直在关注创始人 Edward Johns 的 research。听了 柯丽一鸣的 podcast 之后，我发现脉络其实已经很清晰了：以最快的速度把机器人送到工位上，让它自己开始作业起来，之后再使用真机数据去迭代。我觉得这个应该就是 deployment 范式的 holy grail

  17. 26 Oct

    原来「把隔开的水塘引到一起」有个名字：tertius iungens，联结的第三方（Obstfeld, 2005）。它的反面是 Burt 的 tertius gaudens，坐收渔利的第三方，靠让两边一直隔开来获利。Obstfeld 还发现，前一种人更多地参与到创新里。

  18. 26 Oct

    一定要做资源分配者。我最近的一个感悟就是我似乎越来越能够嗅到对方的需求，或者多方的需求，并且相互拼接起来。我的起心动念是想要把原本想隔开的水塘给引到一起，让大家能够一起加进来把蛋糕做大，都从中获益。我发现这种能力是根植于我善于倾听且喜欢去了解对方的这个特点，又追溯到了我十几二十岁时环游中国时锻炼出来的社牛能力

  19. 26 Oct

    机器人会需要一个 OS 层，而这件事情已经被很多人前赴后继地尝试过，他们都失败了。因为 OS 层在智能，而不是另外一套软件或框架。

  20. 23 Oct

    如果基础模型已经会简单任务，剩下的都是很难的任务。可机器人本来就做不了这些难任务，你怎么遥操作它产生成功示范？

  21. 23 Oct

    等基础模型已经具备很多通用能力之后，让机器人学会一个有经济价值的新任务，最缺的那部分经验是什么？我现在采集的专家数据（视频、轨迹、触觉，etc），能不能提供它？这是个很重要的技术问题，也是个核心的商业问题

  22. 21 Oct

    我想象中的产品形态不是机器人，而是「如何用机器人」的这套系统，一种新的使用机器人的方式.. 这套系统的 input 是数据，输出是完整的工作流

  23. 21 Oct

    我感觉我现在闻到了钱的味道，像鲨鱼一样，我很兴奋，我感觉我看到了机会，但我现在还没有足够多的启动资金、人、资源