爆款重要，让 AI 在更多品类里 scale 也一样重要。
过去两年多，在智能硬件领域，投资人和创业者都热衷于为 AI 寻找一个个物理世界里的新「容器」，各种新形态的 AI 原生硬件成了投资市场的热门项目。
但其实，很多成熟形态的硬件同样有 AI 化的需求。这也许没有「AI Native」的概念那么性感，但市场一点不小，复杂度丝毫不低。
以家电行业为例：供应链长，改一次硬件就要重走研发、测试、认证；销售链条长，从厂家、经销商到上门安装的电工，每一环都得接得住；成本又极度敏感，多一块钱都要反复掂量。
好在，家电已经为 AI 打好了地基。过去十几年，大量家电已经具备联网能力——这意味着灵活升级的 OTA「管道」已经铺好，行业缺的是让设备真正被用起来的解决方案。
那么，硬件到底需要什么样的 AI？9 月，TypeSafe 推出的 Jev 在开发者圈引发了不少讨论，它提供了一个参照：Jev 不生成文字，只输出带置信度的判断。在硬件乃至具身领域的不少场景里，判断比生成更重要。
作为一名在硬件和芯片行业里从业近 20 年的创业者，博联创始人刘宗孺近期和我们谈了谈博联最新的技术以及商业思考。在博联看来，「不让通用模型挑大梁」也许才是硬件领域的最优解。他们推出的寰思中枢由一系列参数很小的专家模型组成，大模型则退到后台，充当「阅卷老师」，复盘错误并自我迭代。
据刘宗孺介绍，目前寰思中枢已经在多个品类里和 B 端客户实现落地，AI 成本可以被压缩到「单台产品每年 1 元人民币」的级别。
以下是 Hardwire 和博联智能董事长兼 CEO 刘宗孺的访谈，经编辑整理。
访谈 | siqi、郑玄
作者 | siqi
编辑 | 郑玄
智能家电头 10 年：联网了，但不够智能
Hardwire：很多读者可能没听说过博联，但可能家里很多产品里都有博联的模组。先请您介绍一下，博联是一家什么样的公司？
刘宗孺： 我们是做 ToB 的，所以对外宣传比较少。博联主要做两块业务：
第一块是帮家电、照明、电工这些厂家做智能化方案：从芯片、模组，到通信组网协议；厂家没有能力做云服务和软件的，我们就打包提供，也帮他们对接外面的各个生态。到了 AI 时代，又多了一块 AI 方案。
第二块是空间场景的智能化，比如全屋智能、智慧酒店、智慧办公。我们自己不做家电，而是把客户的硬件拿过来，再叠加系统和软件。
所以我们其实是 To B to C，通过厂家和这些大客户，把智能化送进用户家里。十几年下来，我们最核心的积累就是连接：芯片是我们自己的，组网协议是从头写的，整个通信这一层都在我们自己手里。
Hardwire：智能家电这个品类，目前走到什么阶段了？
刘宗孺： 这个品类其实已经出现 10 多年了。IoT 行业大概从 2013 年开始，就开始给各种设备接上 Wi-Fi 模块，家电是其中智能化比例最高的，现在的空调基本都是 Wi-Fi 空调了。但其实家电行业的第一波智能化并不成功，大部分 C 端用户还没有为智能买单。
Hardwire：问题出在哪里？
刘宗孺： 根本问题是联网率太低。以厨房电器为例，之前有创业者尝试过 Wi-Fi 烤箱、Wi-Fi 电饭煲等品类，卖点是联网之后可以下载「云菜单」。但这些产品并没有解决动手操作这个问题。如果用户最终还是要走到设备面前亲自动手，掏手机联网就没什么必要。
相比之下，空调和照明是目前联网激活率最高的两个品类。我觉得很大原因是用户在这两个场景里有「遥控」需求，但联网率大概也就是 20% 左右。
之前有很多厂商尝试过让智能家电做语音交互，都没有成功，根源也是联网率太低。
Hardwire：这一轮 AI 爆发之后，有助于解决这个问题吗？
刘宗孺：在我看来，AI 要真正进入家电，需要满足 3 个条件：确定性、低延迟、低成本。这些要求其实是通用大模型不擅长的。
Hardwire：听起来有点反直觉。模型能力正在越来越强，怎么会不擅长呢？
刘宗孺：举个例子：当我跟一个通用模型说「我有点冷」的时候，它会生成一些问候，可能会建议你加件衣服。但在家电场景里，系统更该意识到这是在和设备说话，最正确的反馈是对应到具体的控制指令上。前一段时间很火的 Jev，其实也是这个思路，不生成文字，只输出判断。
至于更重要的成本问题，如果每次交互都要用通用模型推理、都要消耗 token，那这笔钱谁出？用户很难为了一台更好的 AI 电器，每个月交 9.9 元的订阅费，所以这笔成本还得由厂家承担。在这么卷的市场里，厂家的每一块钱成本都很重要。所以，如果做一个 AI 空调、AI 烤箱要增加几十上百元的 token 成本，那这个产品就只能停留在 demo，没法大规模商用。
Hardwire：行业里还有一条思路，既然老产品背着包袱，不如为 AI 重新设计一代硬件。这两年，关于 AI 原生硬件的探索也很多，您怎么看这条路？
刘宗孺：如果我是一家创业公司的 CEO，我可能会去做一个新硬件，直接面向消费者。对新品牌来说，这可能是更好的机会。
但新形态产品需要比较高的教育成本。其次，对硬件创业者来说，更大的挑战是卖完第一个产品之后，怎么持续创新。
作为「创业老炮」，我的心态肯定和年轻创业者们不一样。博联今天主要服务的是家电领域的 B 端客户，所以一定会优先选择成熟且大规模出货的品类，用低成本、稳定的方式交付 AI，让它们的体验大幅提升。你可以把我们理解成卖铲子的，帮厂商完成快速转型。
博联智能董事长兼 CEO 刘宗孺 | 图片来源：BroadLink 博联智能
「一年仅需一块钱」的 AI 账单
Hardwire：如何理解「快速转型」？
刘宗孺：基本操作方式就是，硬件部分基本不动，只给它的联网模组做一次 OTA 升级。
因为对大厂来说，动一下硬件是伤筋动骨的。要经历立项、研发、测试等等很多环节。而且每改一次硬件，都要重新认证，成本非常高。
好在现在大部分智能家电用的是 Wi-Fi +蓝牙的组合模组。以前蓝牙只用来辅助配网，帮用户用手机把设备连上路由器。也就是说，这些硬件本身已经有通信能力。我们在模组上做一次 OTA 升级之后，用户拿微信扫一下产品上的二维码，就能通过小程序里的 AI 助手跟它对话。
这个思路其实也是受了 OpenClaw 这类产品的启发：它们可以用社交软件直接跟家里的设备交互，还有长记忆、主动服务。在中国，社交软件就是微信。
Hardwire：为什么不是各个厂商自己来做 AI ？
刘宗孺：其实所有厂家都有 AI 焦虑，都觉得这是必须拥抱的技术浪潮，但我觉得很容易卡在人才、成本这两件事上。
第一个问题是人才，家电行业很难吸引到真正顶尖的模型人才，也开不起那么高的薪资。
第二个问题是成本，也就是 token。前面说过，如果接通用模型，那成本会很高；也有厂家想过另一条思路：在家里装一个中心网关，跑一个微调过的 OpenClaw 当大脑。但那条路的阻力也很大，还需要用户多花几千块装一台主机。
Hardwire：所以博联理解的家庭 AI，是不需要在本地部署一个带算力的中枢（Hub）的？
刘宗孺：从我们的角度来看，在家里部署一个 GPU（本地算力中心）这件事成本太高，很难规模化。
博联在做的寰思中枢，主要跑在云端，包括一个 IoT 专用模型。它由一组很多个非常小的专家模型组成——每个小模型的参数量可能不到 10B，所以如果用户愿意，也完全可以部署在本地。
Hardwire：这个 IoT 专用模型的架构是怎样的？核心要解决什么问题？
刘宗孺：核心还是解决前面提到的减少幻觉、降低 token 消耗、低延迟三个问题。我可以稍微展开说一下它的架构。整体来看，你可以把它理解成一个多 Agent 的架构，包括系统一、系统二和一个「自进化」的过程。
其中，大部分用户交互的指令我们都在系统一里解决掉，系统一不生成、只判断和执行，目标是更高的命中率和更快速的响应。系统二负责推理和解决系统一解决不了的问题。
我们现在的命中率在 95% 以上，目标是做到 99%的命中率。
Hardwire：这个「命中率」怎么理解？
刘宗孺：就是系统一自己能解决的问题，不需要走大模型、消耗 token。
Hardwire：通用大模型在这个架构里提供的核心帮助，主要体现在什么地方？以我自己的体会为例，在家里操作一些智能设备的时候，会发现通过短链路识别意图时，错误率其实挺高的。
刘宗孺：这个问题其实是反直觉的。我们观察下来的结论，跟您提出的这个感受恰好相反。大模型对于理解意图的帮助，对我们来说可能用处更小一些。更大的帮助，其实体现在我前面提到的「自进化」过程。
这个过程大概的路径是：每天晚上，会有一个真正的大模型——我们现在用的是 DeepSeek 的最新模型——来判卷。在前一天里，我们的系统后台记录了很多交互中的行为和数据，大模型批改完，相当于把它的判断「蒸馏」给小模型，用来重新训练。也就是说，大模型对我们最核心的作用是：我们理解错的、执行错的、或者理解不了的，每天都会更新，每天晚上它都会重新训练我们的模型。
博联的智能体矩阵 | 图片来源：BroadLink 博联智能
Hardwire：这部分训练模型的成本，是由谁买单呢？
刘宗孺：这一部分的自进化、自演进，是我们公司的成本。也就是说，把我自己的 IoT 模型越训越好、数据越来越全、扩展的领域越来越多，这部分不需要厂商做任何事情。
Hardwire：那产品使用过程中的 token 消耗由谁买单呢？您前面有提到 C 端订阅制很难跑通。
刘宗孺：我觉得现阶段，硬件厂商基本没有能力规模化向用户收订阅费，所以这笔钱只能厂家自己承担。对于我们来说，这就要求我们提供的服务价格必须低到让厂商能接受，目前单台设备每年的成本可以做到「1 元级别」，而且这个价格博联是有利润的。
Hardwire：厂商怎么给博联付费？要付多少年？
刘宗孺：质保期多长，就一次性付给我们对应时长的费用。大部分产品是 3 年，我们现在有几个客户都是一次性付 3 年费用的。
其实，如果真了解家电行业就知道，厂家对成本抠得非常细，每年都要和供应商讨论成本结构。如果一共只多一块钱，他们二话不说就上了；但如果是每年一块钱，一个用 10 年的产品，负担还是很重的。
家电领域的安卓
Hardwire：现在是厂家替用户买单。但用户的入口正在变：越来越多的人习惯了跟豆包、ChatGPT 这样的个人智能体说话。如果将来用户都通过个人智能体来控制家电，控制权和钱会不会都流到它们那边？它们会不会直接绕过博联这样的中间层？毕竟很多设备用的都是蓝牙这样的标准协议。
刘宗孺： 基本绕不过去，因为厂家对控制接口几乎是不开放的。像烤箱这样的产品，不可能简单对接一下就让你操控，对厂家来说这是安全底线。厂家不会把电控逻辑开放出来。
Hardwire：如果有一天 Agent 发疯，把家里的煤气都打开了，那也挺恐怖的。
刘宗孺：是的。更现实的问题是，IoT 行业产业链太长了，涉及的工厂很多，不是所有工厂都有能力提供一个 API 和 Agent 开放接口。可能像海尔、美的、格力这样的巨头，愿意对接接口，但也不会是把它的硬件能力完全开放，核心还是安全考量，安全是底线。
Hardwire：您认同用户入口会改变吗？如果改变，商业模式会怎么变化？
刘宗孺：我高度认同这一点。将来不管是 Personal Agent 也好，还是 C 端用户的入口这一端，大概率是社交巨头的天下，毕竟用户在他们那里。
从我们的角度也很简单：我们的 Agent 把能力给他们，用 Agent to Agent 的方式合作。如果这种模式成立，小米、华为现在相对封闭的人车家生态，对大家也就开放了。其他第三方也都可以有一个自己的「人车家」。汽车、手机厂商也不需要对接各种各样的产品，都是一个 Agent 去对接其他 Agent 就好了。
Hardwire：从您的视角看，假设未来几年，让设备智能化、上网这件事变成一个竞争更激烈、大家都要去卷的赛道，博联最核心的优势是什么？
刘宗孺：我们希望帮助商业世界构建一个「第三方」。小米、华为可以讲人车家的故事，别的硬件厂商、手机厂商、汽车厂商怎么讲呢？但他们也要发展。
Hardwire：是不是可以理解成 iOS 和安卓的关系？
刘宗孺：（笑）你完全可以这么理解。
Hardwire：长远看，家里的 AI 该由谁来付费呢？
刘宗孺： 这个问题真的很难判断。我觉得未来如果是 Agent to Agent 的模式，比如手机的智能体调用了我们的智能体，费用怎么分，都可以谈。在现在这个阶段，我们要做的是让厂商接受成本规模，让更多人先在家里体验上 AI 家电。
厨房里将走出爆款产品
Hardwire：具体到「智能家居」这个场景，以前更多选择是前装，装修的时候就要提前规划，是一个整体方案，现在似乎产品的落地方式也在改变。接下来，智能会以什么方式进入用户家里？
刘宗孺： 我觉得以后可能在前装阶段，「智能家居」这个概念会退场，接下来很明确要往后装走。
Hardwire：相比前装，后装的优势在哪里？
刘宗孺：智能家居做了这么多年，最大的问题是沟通成本太高：到底什么是智能家居，要先跟用户讲半天，然后上门勘探、布点、施工、安装、调试，人力成本也高。这条路很难做大，用户也嫌烦。
我们这几年做的所有事情，包括把组网做到傻瓜、稳定，再加上现在 AI 的加持，就是想让智能产品真正像传统产品一样：用户买回设备，扫一下码，剩下的交给 AI 就行。相比前装动辄就是一套复杂的系统，单个设备的升级要灵活得多。
Hardwire：您看好哪些品类？
刘宗孺：从走量的维度，前面提到的具有「遥控属性」的设备，联网率相对高一些，量产应该更快。
以空调为例：现在设置空调 25 度，反馈的是回风口的温度，跟人的体感没关系，所以我们经常会觉得热了或者冷了。但如果加入 AI 和组网，在家里再配一个 10 块钱左右的温湿度传感器，就可以做到真正的「体感恒温恒湿」，半夜不用再起来给孩子盖被子了。对厂家来说，这就是卖点。
Hardwire：那其他过往智能化程度没那么高的品类呢？有没有可能因为 AI 实现体验上大的跨越？
刘宗孺： 越复杂的生活电器，AI 化之后用户就越喜欢。我个人目前最期待的是 AI 蒸烤箱，还没有正式发布，但是所有人来我们这里看了样品，都会问「在哪能买到」。
对于厨房场景来说，过去的产品功能已经很丰富了。但问题是交互复杂，很多人并不知道怎么操作——先选蒸还是烤；怎么调温度、曲线、菜单；时间多少合适……往往做饭之前，得先去小红书、抖音找博主视频，体验很割裂。
AI 化之后，直接说「我要烤一块牛排，外焦里嫩，七分熟」，就结束了。如果烤箱本来就内置了这道菜的菜单，就直接调用它自己的程序；如果用户的需求烤箱里没有预设，就用我们的菜单模型，把「七分熟的牛排该怎么烤」这个问题转化成机器电控能执行的方案。
再或者，水族箱这个品类，以前需要很多人工经验判断。接入 AI 之后，拍张照，系统就能识别养的什么鱼，然后调节氧气、温度这种参数，体验提升也很明显。
AI 蒸烤箱使用场景示例 | 图片来源：BroadLink 博联智能
Hardwire：这种体验的提升，也不需要动硬件吗？
刘宗孺：有厂家试过在烤箱里装摄像头，但无论是耐高温摄像头的 BOM 成本，还是在线实时识别的推理成本，都会带来挺大的压力。我们的方案，不增加任何硬件成本，只加了一年一块钱的费用。至少年轻人会喜欢，小孩也会用了，烤箱变活了、变好用了，这就是价值。
Hardwire：烤箱这样的产品火了以后，怎么从一台烤箱扩展到全家？毕竟烤箱和空调可能来自不同的厂家，跟消费者讲这个故事并不容易。
刘宗孺： 没错，这是难点。最大的难点还是市场教育，所以需要一个 killer 产品。难也难在还没有形成共识：烤箱的提升也许是革命性的，空调的提升没那么大。而且最终得是用户觉得好，产品才能真正卖起来，只是厂家觉得好没有用。
Hardwire：开头我们聊到，上一波智能化做了十几年，用户也没有真正买单。极客公园今年年底的大会，有一个主题叫「未来 500 天」。站在今天看，未来 500 天，您希望在这个赛道上看到什么？
刘宗孺： 第一，智能家居一定会 AI 化。用户可以随心所欲地往家里加产品，再通过 AI 去交互，这在我们这里已经实现了。只是我们不直接面向消费者，C 端用户还没有感知，但我的客户会去讲这个故事。
第二，AI 家电会来得比上一波快得多。上一波智能家电，大家花了 10 年去搞联网；这一波我估计用不了两年。现在所有厂家都迫不及待地在讲自己的家电怎么 AI 化。不用等 500 天，大概率明年 3 月的 AWE 展会上，大家都会讲 AI 家电。
第三，家真正被 AI 管起来，马上就能实现。不用去幻想具身智能进入家庭，我不需要具身智能的机器人进来，但一个真正的 AI 管家，用不了 500 天就能做到。
对市场的期待，就是我们努力把成本降下来，让用户无感地进入 AI 时代，而不是为了体验 AI 一定要付一笔订阅费。对厂家来说，如果一台烤箱能提升这么多体验，多花 3 块钱绝对没问题，他愿意替用户买单。我个人认为，明年 AI 厨电一定会风靡起来。
Hardwire：其实这两年，硬件行业里最热的话题往往都是「爆款」产品。但听下来，博联更看重的是让智能在硬件中 scale。在您看来，爆款和规模化是什么关系？
刘宗孺： 我觉得两件事都重要，只是解决的问题不一样。爆款是最快提升用户感知的事。市场教育是最难的，所以一定需要 killer 产品，让用户用了觉得「AI 真能帮我」，他才会去找下一个、再下一个产品。
但对整个家电乃至硬件行业来说，更难的是规模化：这么多已经在卖、还在源源不断生产的产品，怎么让它们都用上 AI，而且足够便宜、足够稳定。每个品类的电控不一样，每家厂商的情况也不一样，不可能一个一个靠人去做。所以我们这些年一直在做底座：组网、专用模型、开发平台，让厂商自己用自然语言就能开发智能体。
这件事我们正在一边做一边提升，品类在一个个接进来，成本也在往下压。我觉得，很多年以后可能「AI 家电」也不是新词了，所有的产品都可以具备 AI 的能力。
本文由 AI图鉴自动抓取并翻译，版权归原作者所有；如需查看原文请点下方链接。
Battery is important, making AI in more class scale is also important.
Over the past two years, in the area of smart hardware, investors and entrepreneurs have been keen to find a new "container" in a physical world for AI, and new forms of AI raw hardware have become a popular project in the investment market.
But in fact, many mature forms of hardware also have an AI-based demand. This may not be as sexy as the "AI Native" concept, but the market is not so small and complex.
For example, in the domestic electricity sector: the supply chain chief, who changes hardware once again, has to re-engineer, test, certify; the chain of sale, from the manufacturer, the distributor, to the electrician who installs it at the door, has to be followed every ring; and the cost is extremely sensitive, and the additional dollar is subject to double-checking.
Well, home electricity has been set up for AI. Over the past decade or so, large amounts of household electricity have been connected – this means that the flexible upgrade of the OTA “pipe” has been prepared, and what is missing in the industry is a solution that will actually put the equipment to use.
So, what exactly does hardware need, AI? In September, TypeSafe launched Jev, which generated a lot of discussion in the developer's circle, provided a reference: Jev does not generate text, but only gives confidence. In many contexts, both hardware and physical, judgment is more important than generation.
As an entrepreneur in the hardware and chip industry for almost 20 years, the founder of the Federation, Liu Zong and Ying, recently spoke to us about the latest technology and business thinking of the Federation. In the view of BOB, it is probably the best solution in the hardware field to "learn the general model to pick the beam." The thought center they launched consisted of an expert model with a small range of parameters, and LLM withdrew to the backstage to act as a "roll-reading teacher" with error and inversion.
According to Liu Zong Zong and the children, the center of thought has now landed in a variety of goods and B-end customers, and AI costs can be reduced to the level of 1 yuan a year for single-line products.
Here's an interview, edited, with Hardwire and BoA's Chairman of the Smart Board and CEO Liu Zong and his children.
Interviews
Author
Editor Zheng Hyun
Smarthouse head 10 years: connected, but not smart enough.
Hardwire: Many readers may not have heard of BOB, but many of the products in the home may have BOB models. I'd like to introduce you. What kind of company is BOU?
We're doing ToB, so there's less publicity. BOBU has two main lines of business:
The first is to help domestic electricians, lighting, electricians and so on: from chips, models, to communication group protocols; if the factory is unable to do cloud services and software, we pack them up and help them connect with the surrounding ecology. By the time the AI age, another AI program was added.
The second is the intellectualization of space scenarios, such as house-wide intelligence, smart hotels, smart offices. We don't do home electricity on our own, but we bring in the client's hardware, superimpose systems and software.
So we're actually To B to C, sending smarts into the user's home through the factory and these big clients. For more than a decade, our core accumulation is connectivity: the chips are ours, the network protocols are from the beginning, and the entire layer of communication is in our hands.
Hardwire: What stage is this?
It's been there for more than 10 years. The IOT industry has been connecting the Wi-Fi module to various devices since about 2013, with the highest rate of intelligent domestic electricity, and now basically Wi-Fi air conditioners. But the first wave of smartness in the domestic electricity industry has not been successful, and most C-end users have not yet paid for intelligence.
Hardwire: What's the problem?
The fundamental problem is the low rate of connectivity. In the case of kitchen appliances, for example, entrepreneurs have tried the Wi-Fi ovens, Wi-Fi cooks, etc., selling points that can be downloaded from the Cloud menu after the Internet. But these products do not solve the problem of hands-on operations. If the user is finally going to go to the device and do it himself, there is no need to get a cell phone online.
In contrast, air conditioning and lighting are the two categories with the highest rate of network activation. I think it's because users need remote control in both scenarios, but the connection rate is about 20 percent.
There have been unsuccessful attempts by many manufacturers to interact intelligent household electricity with voice, and the root cause is the low rate of connectivity.
Hardwire: Does this round of AI help solve this problem?
Liu Zong Chong and children: In my opinion, there are three conditions that AI needs to meet for real access to electricity: certainty, low delay, low cost. These requirements are not really good at General LLM.
Hardwire: Sounds counterintuitive. Modeling is getting stronger. How come it's not good?
Liu Zong and children: For example, when I say "I'm a little cold" to a generic model, it creates some greetings that might suggest that you add clothes. However, in the home electric scene, the system should be more aware that this is speaking in and with equipment, and that the most accurate feedback is on specific control instructions. Jev, it was the same idea, not the text, just the judgment.
As for the more important issue of cost, if each interaction is based on generic model reasoning and consumes token, then who pays the money? It is difficult for users to pay a subscription of $9.9 per month for a better AI appliance, so the cost is borne by the manufacturer. In such a roll of markets, every dollar a manufacturer costs is important. So, if an AI air conditioner and an AI oven cost hundreds of extra tokens, the product can only stay in demo and not be commercial.
Hardwire: There's another line of thought in the industry, since the old product carries the burden, it's better to redesign a generation of hardware for AI. Over the last two years, there's been a lot of exploration of AI raw hardware. What do you think of this road?
Liu Zong Cong and Children: If I was the CEO of an entrepreneurial company, I would probably make a new hardware and direct it to consumers. This may be a better opportunity for the new brand.
However, new products require higher education costs. Second, for hardware entrepreneurs, the greater challenge is how to continue innovation after selling the first product.
I'm sure my mind is different from that of young entrepreneurs. BOB's main service today is the B-end customer in the field of household electricity, so it is certain that it will give priority to mature and large-scale deliveries, delivering AI in a low-cost, stable way, so that their experience will increase dramatically. You can interpret us as selling shovels, helping manufacturers to make rapid transformations.
BOR's head of the board of directors and CEO Liu Zong Zong Zhu Zhou Zhou Zhou Zhou Zhou Zhou
"One dollar a year" bill.
Hardwire: How do you understand "quick transition"?
Liu Zong Cong and children: The basic method of operation is that the hardware component is almost intact and only an OTA upgrade is made to its network module.
Because, for a large plant, it's hard work. There are many links to project development, research and development, testing, etc. It is also costly to re-certify every hardware change.
Fortunately, most of the smarts now use Wi-Fi+Bluetooth combination models. Previously, Bluetooth was used only to assist the network by connecting the equipment to routers with a mobile phone. In other words, the hardware itself already has a communication capability. After we did an OTA upgrade on the module, the user sweeps the two-dimensional code of the product with WeChat, and talks to it through the AI assistant in the applet.
This idea is inspired by OpenClaw: They can interact directly with home-based equipment through social software, as well as long memory and proactive services. In China, social software is WeChat.
Hardwire: Why don't the manufacturers do AI?
Liu Zong Chong: Actually, all the manufacturers have an AI anxiety that they think is a technological wave that they have to embrace, but I think it's easy to get caught up in talent and cost.
The first issue is talent, and it is very difficult for the household electricity industry to attract truly top model talent and to afford such high salaries.
The second problem is the cost, which is token. As I said earlier, it would be very expensive to connect to a generic model; there is another idea that manufacturers think about: putting a central gateway at home and running a slightly modified OpenClaw as their brain. But that road is also very resistant, requiring users to spend thousands more on a mainframe.
Hardwire: So the BFU understands that AI does not need to deploy a computer hub locally?
Liu Zong Chong and children: From our point of view, it is too costly to deploy a GPU at home, which is difficult to scale.
BOU is doing a thought center, mainly in the clouds, including an IOT dedicated model. It consists of a group of very small expert models - parameter count for each small model may be less than 10B, so it can be deployed locally if users so wish.
Hardwire: What is the structure of this IOT dedicated model? What's the core to solve?
Liu Zong Chong and children: The core also addresses the three above-mentioned problems of reducing hallucinations, token consumption and delay. I can expand it a little. Overall, you can understand it as a multi-Agent architecture, including systems I, II and a "self-evolution" process.
Of these, most interactive user directives are addressed in the first system, which is not generated, judged and implemented, with the goal of a higher hit rate and faster response. Systems II is responsible for reasoning and solving problems that the system cannot solve.
We now have a hit rate of over 95%, and the goal is to achieve 99%.
Hardwire: How does this hit rate work?
Liu Zong and children: The system can solve its problems without LLM, consumption token.
Hardwire: Where is the core help provided by General LLM in this structure, mainly? In my own experience, for example, when running some smart devices at home, you find that there's actually a high error rate when you recognize intent through a short link.
Liu Zong Cong and children: The problem is anti-intuitive. The conclusion that we have observed is exactly the opposite of what you have suggested. LLM may be of less use to us for the purpose of understanding intentions. More help is in the "self-evolution" process I mentioned earlier.
It's about the path: every night, there's a real LLM. Now we're using DeepSeek's latest model -- to judge. In the day before, many interactive behaviors and data were recorded on the back of our system, and LLM was batched, amounting to "distilling" its judgement to a small model for re-training. That is to say, LLM's role at the heart of us is that we understand wrong, implement wrong, or fail to understand, is updated every day, and every night it retrains our models.
AI agent Matrix of BOB
Hardwire: Who pays for this training model?
Liu Zong Cong and Children: This part of evolution and evolution is the cost of our company. That is to say, the better my own IOT model gets, the more complete, the more extensive, and the more extensive, part of it does not require anything from the manufacturer.
Hardwire: Who pays for the token consumption in the product? You mentioned earlier that C-end subscription is hard to run.
Liu Zong Cong and children: I do not feel that at this stage the hardware manufacturer has the capacity to scale up subscription fees for users, so that the money can only be borne by the producers themselves. For us, this requires that the price of the services we provide be as low as it is acceptable to the manufacturer, that the cost of a single unit of equipment can be “one dollar” per year and that the price chain is profitable.
Hardwire: How does the manufacturer give BOB paid? How many years?
Liu Zong Cong and Children: The length of the Quality Assurance period is a one-time payment to us for the duration of our response. Most of the products are three years, and we now have a few clients who pay a one-time fee of three years.
Indeed, if one really knew about the household electricity industry, one would know that the plant was very expensive and would discuss the cost structure with the supplier each year. If there's only one more dollar, they'll say nothing, but if it's one dollar a year, one product for 10 years, it's a heavy burden.
Andre in the field of electricity.
Hardwire: Now the company pays for the user. But the user portal is changing: more and more people are used to talking to individuals like Doubao, ChatGPT, AI agent. If, in the future, all users control their power through personal AI agent, will control and money flow to them? Do they go straight through the middle of the BFC? After all, many equipment uses standard protocols like Bluetooth.
Liu Zong and children: It's almost impossible to go around because the plant is almost closed to the control interface. A product like an oven can't be easily connected to you, it's a safety line for the factory. The plant won't open up the power logic.
Hardwire: If one day Agent went mad and turned on all the gas in the house, it would be scary.
Liu Zong Chong and children: Yes. More realistically, the IOT industry industrial chain is too long, involving a large number of factories, not all of which are capable of providing an API and Agent open interface. A giant like Haile, America, Gwig may be willing to interface, but it will not be completely open to its hardware capability, core or security considerations, and security is the bottom line.
Hardwire: Do you agree that the user portal will change? How would business models change if they changed?
Liu Zong Cong and children: I highly agree with this. In the future, whether it's personal access or C-end users, it's probably social giants, after all.
From our point of view, it's simple: our Agent gives them the ability to work together in the same way as Agent to Agent. If this model is established, it will be open to all to Mi and Hua, who are now relatively closed human beings. The other third parties can have their own "car house". Cars and mobile phone manufacturers don't have to pick up all kinds of products, but it's just one Agent to pick up the others.
Hardwire: From your point of view, what is the strength of BOB's core, assuming that in the coming years it will become a more competitive track for equipment to be intelligent, Internet access, and everyone to roll?
We want to help the business world build a third party. What about Mi and Hua, who can tell the story of a man's car, and other hardware, mobile and automobile manufacturers? But they also have to develop.
Hardwire: Does it understand the relationship between iOS and Andre?
Liu Zong and children: You can totally understand that.
Hardwire: In the long run, who's supposed to be at home with AI?
It's really hard to judge. I think the future can be talked about if it's the same model as Agent to Agent, like the AI agent on the phone that called our AI agent. At this stage, what we're going to do is get manufacturers to accept the scale of the cost, get more people to experience AI electricity first at home.
The kitchen will be out of the explosive product.
Hardwire: With the "smart home" scenario, which used to be more front-loading, pre-planned at the time of renovation, it was a holistic approach, and now it seems that the way the product landed is changing. So, how does intelligence get into the user's house?
I think the concept of "intellectual home" is going to get out of the front stage.
Hardwire: What's the advantage of the rear over the front?
Liu Zong Chong and children: The biggest problem is that the cost of communication is too high: what is a smart home is a half-day talk to the user and then go to the door for exploration, dotage, construction, installation, commissioning and labour costs. It's a hard road, and users are tired.
All the things we've done over the years, including making the network stupid, stable, and now the AI plus, are trying to make smart products really like traditional products: users buy back the equipment, sweep the code, and leave the rest to AI. The upgrading of individual equipment is much more flexible than the previous installation of a complex system.
Hardwire: What are you looking at?
Liu Zong Cong and children: The dimensions of walking, the equipment with the "controlled properties" referred to earlier, have a relatively high rate of connectivity and production should be faster.
Take air conditioning as an example: we're setting up an air conditioner at 25 degrees, and the feedback is that the temperature of the vent is not related to the human body, so we're often hot or cold. But if you join AI and the network, and you have a temperature sensor at home of about $10, you can get real "smooth and humid" and you don't have to get up to cover the kids in the middle of the night. For the factory, that's the point of sale.
Hardwire: What about other things that were not as intelligent as they used to be? Is it possible to achieve a big leap in experience because of AI?
The more complex life appliance, the more users like it. What I'm looking forward to personally is the AI steambox, which has not yet been officially publish, but everyone who comes to us and looks at the samples asks, "Where can I get it?"
For the kitchen scene, the product function of the past has been very rich. But the problem is complex, and many people don't know how to do it — whether to steam or bake; how to adjust the temperature, curve, menu; how much time is appropriate... Before cooking, you have to go to REDnote, Douyin to find the blogger's video and experience is very fragmented.
After AI went down and said, "I'm going to roast a steak, and I'm going to curry out, and I'm going to be fine." If the oven already contains a menu for this dish, it calls its own program; if there is no preset in the user's needs oven, we use our menu model to turn the question of how to bake a seven-square steak into a machine electro-control program.
Alternatively, the aquator category requires a lot of experience to judge. With access to AI, the system can identify the fish that it maintains, then regulate the parameters of oxygen, temperature, and experience increase.
Image source: BroadLink
Hardwire: Do you need to move hardware for this experience?
Liu Zong Cong and children: While some manufacturers have tried to install cameras in ovens, the cost of BOM for high temperature cameras, or the cost of reasoning for real-time online recognition, can be quite stressful. Our programme, without any increase in the cost of hardware, is only one dollar a year. At least the young people will like it, the kids will use it, the oven will be alive, the oven will be useful, that's the value.
Hardwire: How can products like ovens be expanded from one oven to the whole family? After all, ovens and air conditioning can come from different manufacturers, and it's not easy to tell the story to consumers.
Yes, it's difficult. The biggest difficulty is market education, so a killer product is needed. It is difficult, and it is difficult, to reach a consensus: the upgrade of ovens may be revolutionary, and the upgrade of air conditioners is not so great. And ultimately, the product will be sold only if the user feels good, except that the manufacturer feels good and useless.
Hardwire: At the beginning, we talked about how the last wave of smartness had been for over a decade, and the users had not actually paid. At the convention at the end of this year, there was a theme called "The Future 500 Days." What do you want to see on this track?
First of all, there's a certain altruism of intelligent family. Users can go home and add products as they please, and then interact through AI, which is done here. It's just that we don't go directly to consumers, and C-end users don't know yet, but my clients will tell the story.
Second, AI will come much faster than the last wave. The last wave of smart home appliances, it took us 10 years to get connected; I don't think it'll take me two years. Now all the manufacturers are dying to talk about how their electricity is AI. You don't have to wait 500 days for the AWE exhibition next March, and everybody's gonna talk about AI.
Third, home is really under AI control, and it's about to happen. I don't have to fantasize about being smart, I don't need a smart robot to come in, but a real AI housekeeper can do it in 500 days.
The expectation for the market is that we are trying to lower the costs and let the user enter the AI era unsurely, rather than having to pay a subscription to experience AI. For a manufacturer, if a oven can enhance so much experience, it's absolutely no problem to spend an extra $3, and he's willing to pay for the user. Personally, I think next year AI's cooking power's gonna be awesome.
Hardwire: In fact, the most hot topic in the hardware industry in both years is the "explosive" product. But listen, the BoC is more interested in getting intelligence in hardware. In your opinion, what's the relationship between explosive money and size?
I think both things are important, but the problem is different. It's the fastest way to raise user awareness. Market education is the hardest thing to do, so there's got to be a killer product that allows users to think that "AI really helps me," so he's going to look for the next one.
But what's harder for the entire electricity and hardware industry is to scale: so many products that are already sold and are still being produced, how can they be used for AI and cheap enough to be stable enough? Electricity controls vary from one product to another, and the situation is different from one manufacturer to another. So we've been doing the bottom seat for years: group networks, specialized models, development of platforms that allow manufacturers to develop AI agent in their own languages.
We're doing this while we're doing it, and the class comes in one by one, and the cost is going down. I don't think that "AI home electricity" is a new word for many years, and all products can have AI capabilities.
Auto-fetched and translated by AI Atlas. All rights belong to the original author.