您正在使用IE低版浏览器，为了您的雷峰网账号安全和更好的产品体验，强烈建议使用更快更安全的浏览器
此为临时链接，仅用于文章预览，将在时失效
0
导语：GPT-6之后，机器人公司还有价值吗？智元用2万台出货量背后的“Worken”算力账本回应。
作者丨齐铖湧
编辑丨岑 峰
2026年，具身智能赛道正在陷入一场前所未有的情绪焦虑。
GPT-6 Astra展示出的具身控制能力，让"物理AI即将被AI击穿"的论调甚嚣尘上；负责感知与决策的S2层，正在以肉眼可见的速度被通用智能接管。行业里甚至出现了明确的路线分化：一批做模型的公司开始转型卖数据，"卖铲子"成了新的安全区。
而在烈火烹油的显示场景中，哪怕今年中国机器人整机产量突破10万台，能真正在流水线上替代现实工作、替人类扛起KPI的确寥寥无几。
所有从业者都在等一个答案：当通用大模型不断吞噬智能的上层，做机器人的公司，价值锚点到底在哪里？
10月29日，上海世博中心，智元创新日（A.I. DAY），稚晖君、姚卯青、熊彦、王闯和他们的智元军团用两个多小时的发布会，给出了一套完整的回应。现场超1000名科研伙伴、开发者、商业伙伴与媒体参会，线上超50家媒体直播。
AI科技评论认为，这场分享里，有非常多的干货和行业痛点，值得跟大家剖析：
01
Worken：
稚晖现场造了一个具身新名词
发布会开场，稚晖君先泼了一盆冷水。
“中国今年机器人整机产量预计突破10万台，但我们也必须清醒：生产了多少台机器人，并不等于真实世界就多了多少生产力。”
这句话说完，现场一片沉默。
这也是AI科技评论在一些行业调研中发现的“行业幻觉”：量产的新闻满天飞，但真正持续投入生产力的寥寥无几。对此，智元给出的解法是一个全新的度量衡：Worken（有效智能工时）：
Worken = 活跃机器人数量 × 自主工作时长 × 产出效能比
三个乘数，各自藏着具身行业的三个痛点问题：宕机、离线、维护中的机器不算活跃；遥操作时段不计入自主工作时长。镜头前忙得团团转、镜头后站着三个手柄操作员的机器人，在这套算法下工时直接清零；产出效能比以一个标准工人的单位时间合格产出为1.0，干得比人慢就不达标。
为解释这个概念，稚晖君串起了一段度量衡史：蒸汽时代看"马力"，电力时代看"千瓦时"，数字AI时代则是用"Token"度量思考量。他总结道：“token衡量机器思考了多少，Worken衡量机器行动了多少。”
在AI科技评论看来，这个概念的提出，远比发布某个产品重要，它代表着资深从业者对这个行业的深度思考。
大模型产业之所以能形成今天按量计费、按效果定价的商业闭环，token居功至伟。它让"智能"首次成为可定价交易的商品。而物理AI行业苦无度量衡久矣：买卖双方无法定价，资本算不清楚投入产出，Worken正是解开这个死结的关键。
稚晖君没有仅仅停留在造新词上，现场他直接切出了实时工时看板：龙旗产线的上下料机器人、香格里拉酒店的服务机器人、长隆乐园的文旅机器人，每台机器当月贡献的有效工时实时滚动。他调侃了道：“上过班的朋友都填过工时，未来对机器人的服务，也要按Worken来计费。”
同场，由中国信通院等权威机构联合编制的《2026具身智能产业创新报告》正式发布，Worken及相关部署态实例成功入选。度量衡刚发布，就已经有了权威背书。
在AI科技评论看来，“从卖机器人到卖工时”，是理解智元全部商业逻辑的钥匙。
当机器人按Worken计费，客户买的就不再是一台出厂即巅峰、逐年折旧的设备，而是一个持续进化的生产力单元。稚晖君的另一句话点明了资产属性的变化：“硬件会陈旧，但智能可以升值。”
02
六个月前出的题，这次交卷了
对Worken的构思最早源于六个月前的APC2026智元合作伙伴大会。当时，智元创始人邓泰华断言：2026年是具身智能从"开发态"迈入"部署态"的元年。稚晖君则定调：“行业真正的分水岭，是从’卖机器人’转向’交付结果’。”
彼时支撑这个判断的，是龙旗工厂一场8小时无剪辑直播：精灵G2以"正式员工"身份连续作业，成功率100%。那是单点实证，证明"能行"，但还证明不了"能规模化地行"。
六个月过去，具身行业和智元的落地细度已经完全不同。
在经营面上：22026年上半年出货9700台，以超43%的份额稳居全球第一；9月第2万台机器人下线，从万台到两万台只用了半年；上半年收入接近20亿元。
万台量产是人形机器人行业的“生死线”，考验的是供应链管理、良品率控制和成本摊薄的综合功夫。智元合伙人王闯给出了跨越这道坎的底牌：100%本土化A链、半小时供应链圈、92项车规级量产测试。
"量产不是把机器人造出来就完了，是要保证造出来的每一万台，都像第一台一样可靠。"王闯直言，“这背后没有捷径，全是产线上一个一个坑踩出来的。”
场景面上，4月在龙旗产线打工的"老演员"，至今稳定运行超过7个月；6天直播累计作业超60小时，产出17625台电脑，机器人累计作业64828次，单件节拍22.5秒，与人类工人相当。"能不能扛住一整年的产能考核"的悬念，也有了答案。
工业纵深也在拓宽：精灵G2已在龙旗工厂实现平板量产质检工段全覆盖，进入芯片封测与CNC；精灵G2 Max落地食品、快消与电商搬运；横琴长隆部署近500台机器人，带动乐园客流提升近40%；广州地铁超120台机器人上岗巡检；成都宽窄巷子落地了全球首个VLA模型驱动的双足人形零售作业，准确率98%。
还有一个容易被忽略却极为关键的数字：人机运维比。
王闯公布的这个数字，已经从最初的1:1（1台机器人配1个调试人员）降到了10:1，目标直指30:1。在AI科技评论看来，这条曲线的斜率，才是"部署态"是一门真正的生意，还是一堆赔钱定制项目的试金石。
“实验室是温室，真实世界是旷野。行业不缺技术难点，缺的是能批量上岗的生产力。”
4月的APC回答"部署态是什么"，10月的A.I. DAY回答"部署态怎么度量、怎么算账、怎么进化"。从提出概念到交付度量衡，智元用了六个月。
03
技术全景：
给Worken的三个乘数上保险
这场发布会的技术清单密度极高：第三代具身基座模型GO-3 Preview、四层架构的AGILE运动引擎、全模态交互大模型WITA-Omni、第三代硬件架构AHA3以及重载本体G2 Max等。
但在AI科技评论看来，与其逐条罗列参数，不如看清这些发布共同指向的结构：它们分别在为Worken公式的三个乘数上保险。（雷峰网(公众号：雷峰网)）
GO-3 Preview保的是"产出效能比"。 姚卯青现场宣布，GO-3已在RoboDojo等近期评估GPT-6具身能力的三大权威榜单全部登顶。
而比榜单更重要的，是部署后的进化能力：与清华联合研发的群体经验学习模块CEL，机器人专拿失败案例“练手”，实测100台机器人集群训练的任务成功率从50%量级拉到98%；在真实便利店场景遇到"挂钩遮挡"的意外时、它能衍生新策略，把单次成功率从20%拉到80%以上。
正如姚卯青一阵见血的总结：“具身模型的成长之路，不只是在数据里训练出来的，更是在真实场景里’干’出来的。”
AGILE和WITA-Omni保的是"自主工作时长"。 机器一旦宕机或求助人工，工时即刻清零。为此，AGILE 运动引擎通过四层架构赋予机器人“闭环恢复”能力。稚晖君直言，这彻底告别了当下许多跳舞机器人“闭眼盲目模仿”的行业通病。
在交互侧由另一位智元合伙人熊彦分享，也是全场最有"人味"的环节。WITA-Omni将视、听、说、动同步建模，实现"边听边想边说"，响应速度压到最快400毫秒，已经进入了人类日常对话的自然节奏区间。
AHA3与G2 Max保的是"活跃机器人数量"。 三智在硬件层面物理合一，控制器体积降50%、延时降50%；全球首款皮下式关节液冷系统，散热能力达到传统风冷的7倍，让关节持续扭矩输出达到风冷的3倍；G2 Max双臂峰值负载50公斤、双电池热插拔撑住8小时全天候作业。
同步发布的还有新一代灵巧手家族OmniHand 3系列，其中OmniHand 3是首款已在产线连续作业的工业级灵巧手。
这条产品线的成熟度，直接决定精细操作的上限。还有那个110台远征A3创下的"全球最多人形机器人群舞"吉尼斯世界纪录。稚晖君称之为"行业迈向规模化的重要起点"。
在软硬配置拉满之外，智元还留下了一个更具杀伤力的后手：发布业界首个具身原生操作系统灵渠 OS 1.0，并承诺年底全面开源。配合半年联动超 8 万名开发者的“元苼计划”，智元正在试图垄断物理世界的基础设施。
一台台拆看是技术，合起来看，是一套"让有效工时持续变多"的精密装置。
04
正面回应"击穿论"：
大模型是弹药库，而非掘墓人
这场发布会最值得记录的，是智元没有回避那个悬在行业头顶阴影：大模型会彻底“击穿”具身赛道吗？做机器人的本体公司还有价值吗？
面对这一焦虑，智元合伙人姚卯青给出了本场最锋利的一句回应：“如果行业总将被AI击穿，那希望我们站在AI这一侧。”
他进一步解释：通用大模型单次推理耗时极长、算力成本极高，“在要求低延时、高精度真实工业场景里难以落地部署”。智元的选择是立足本体、饱和投入大模型自研，同时把能力沉淀成"生产级技术"，并且宣布GO-3和整套具身AI技术框架年内开源，把架构变成生态。
稚晖君则从针对整个行业关心的问题做了两种未来的可能性推演：
第一种，通用基础模型越来越强且趋于开放，谁都能获得先进智能。那竞争焦点就变成谁能以更低成本、更高可靠性把智能接入真实世界；
第二种，模型依然高度依赖不同本体、不同场景的数据训练，不存在万能模型。那“谁拥有大规模真实部署、能稳定获取高质量数据”，谁就赢。
而这两种未来，都是本体厂商的主场。
“两条技术路径看似不同，最终都指向同一个结论：规模化的本体优势会越来越重要。”
这是一个相当漂亮的论证。在AI科技评论看来，"把智能接入物理世界"从来不是软件问题，而是可靠性、成本、供应链、部署网络、运维体系的综合工程问题，这恰恰是互联网基因的大模型公司最难补课的部分。
智元如今2万台出货量、99.99%稳定性、92 项车规级测试、10:1人机比，就是阻挡纯算法公司的护城河。通用智能不是部署态的掘墓人，而是弹药库；思考交给大模型，但行动的壁垒，永远属于贴着物理世界生长的公司完成。
05
一个现场“接话”的细节，
和具身行业真正的成熟
最后想说一个本场最容易被忽略的瞬间：发布会现场，智元合伙人熊彦在演示的时候，问远征A3一句"智元谁最帅"，机器人毫不犹豫说：“我可是智元的颜值担当。”接着追问“除了你呢”，远征A3没有机械回答，回答而是接了一句“当然你了，你的蓝色POLO衫很有精神。”
现场演示里，远征A3在嘈杂会场中不插话、等时机，能自动分辨谁在跟它说话；更难的是语言的弹性：熊彦现场带着它从普通话无缝切到粤语、四川话、闽南话，甚至完成了同声传译。
熊彦的解释很务实："方言识别不是炫技，是上岗资格证。"WITA-Omni支持超过100种语言和29种方言，并已登顶Daily-Omni等四项具身交互权威榜单榜首，12月将在国内完成全面部署。这些能力落在北大人民医院的导诊台、宽窄巷子、长隆的咨询点位上，就是实打实的在岗工时。
这个瞬间之所以值得记录，是因为它背后站着的，是本场发布会最"有人味"的一个模型：WITA-Omni全模态交互大模型。它不是把语音、视觉、动作几个模块简单拼接，而是把所有输入输出放在同一个模型里同步建模，才能做到夸人夸得如此具体，有情商。
除了能观察，有情商外，还能做到让机器人"边听边想边说"，响应速度压到最快400毫秒、平均600毫秒，已经进入了人类日常对话的自然节奏区间。
熊彦在现场的演示，更像一场突击考试：带着远征A3从普通话无缝切到粤语、四川话、闽南话，甚至完成了同声传译；在嘈杂的会场里，机器人不插话、等时机，能分辨出谁在跟它说话、谁只是在和旁人闲聊。WITA-Omni支持超过100种语言和29种方言，并已登顶Daily-Omni等四项具身交互权威榜单榜首，12月将在国内完成全面部署。
类似的质感遍布全场：龙旗产线22.5秒的节拍、便利店挂钩遮挡的失败案例、长隆"人机网络物理隔离"的5G方案、运维平台"新增点位一键复用"的设计。这些都不是实验室里能规划出来的，是旷野里长出来的。
六年前，这个行业比拼谁的机器人能翻跟头；三年前，比拼谁的Demo视频更流畅；今年4月，智元说比赛换成了"谁的机器人能连续干活"；这个10月，它干脆给比赛发了计时器。
正如稚晖君在开场说：“智元第一次参赛就拿到了第一，但支撑这份成绩更有意义的，是商业化部署。我们的机器人既能上赛场，也能进工厂、去商场。”
46枚运动会金牌是面子，实时滚动的工时看板是里子。Worken能否成为行业公认的度量衡，最终取决于买卖双方的共同采用，这需要时间验证。但至少在这个被"击穿论"笼罩的秋天，智元做了一件行业早该有人做的事：
当所有人都在争论机器人够不够聪明时，它第一个站出来，给机器人记了考勤，并且告诉具身行业：下半场比的不是谁更会表演智能，而是谁更能交付工时。（雷峰网）
雷峰网原创文章，未经授权禁止转载。详情见转载须知。
为了您的账户安全，请验证邮箱
请验证您的邮箱
本文由 AI宝库自动抓取并翻译，版权归原作者所有；如需查看原文请点下方链接。
You are using a low-end IE browser and strongly recommend a faster and safer browser for your Thunder Peak account security and better product experience.
This is a temporary link, only for article preview, which will expire at the time
0
Introduction: Does robotics have value after GPT-6? Ji-won responded with a “Worken” account that was behind 20,000 deliveries.
Author
Edit Peak
In 2026, a smart track was being plunged into an unprecedented level of emotional anxiety.
GPT-6 Astra has demonstrated a physical control capability that makes the argument that "physical AI is about to be punched through "; the S2 floor, which is responsible for perception and decision-making, is being taken over by universal intelligence at an eye-visible rate. There's even a clear division of routes in the industry: a group of modeled companies started re-selling data and "selling shovels" became a new safe zone.
And even if China’s whole machine produced more than 100,000 units this year, there are very few people who can actually replace real work on the waterline and carry KPI for humans.
All the practitioners are waiting for the answer: where are the value anchors when General LLM continues to devour the upper layers of intelligence, the companies that are robots?
On 29 October, the Shanghai World Wide Centre, the Day of Innovation (AI. DAY), Jeong Yi, Yao Zhiqing, Yong-hyan, Wang Ban and their group of Jijun gave a complete set of responses with publish meetings over two hours. Over 1,000 scientific partners, developers, business partners and the media were present and over 50 media outlets were broadcast live online.
According to AI, there are a lot of dry goods and trade pains in this sharing, and it is worth an analysis:
01
Worken：
It's a new term.
publish's meeting started with a cold-water squirt.
“China’s total robotic output this year is expected to exceed 100,000, but we must also be sober: the number of robots produced is not equal to the productivity of the real world.”
After that, there was silence.
This is also the “industry illusion” found in some of the industry studies by the AI-Technology Review: the news is full of volume, but there are few real and sustained investments in productivity. The solution given by Ji-won is a new measure: Worken (effective smart working hours):
Worken = number of active robots x duration of autonomous work x output efficiency ratio
Three multipliers, each of which contains three pains in the body industry: the machine under maintenance is not active; the remote operating time does not count for the time of the autonomous work. The robots of the shot who were busy before, and the shots who were standing behind three handle operators were cleared directly at the end of the algorithm; the output was 1.0 more effective than the unit time-qualified output of a standard worker and did less than humans.
In order to explain this concept, Teenage set up a history of measures and measures: "Maritime" in the steam age, "Kilowatt-hour" in the power age and "Token" in the digital AI age. He concluded: "Token measured how much the machine thought, and Worken measured how much the machine acted."
According to the AI Science and Technology Review, this concept is far more important than a product of publish, which represents a deep reflection of the industry by senior practitioners.
The LLM industry has made it possible to create today ' s business closed loops that are measured in volume and priced in effect. It makes smart a priceable commodity for the first time. The physical AI industry has been struggling for a long time: buyers and sellers are unable to price, capital is unclear about input output, and Worken is the key to breaking this knot.
He did not stop on the new word, and he cut the board of real-time working hours directly at the scene: the top and bottom robot of the flag line, the service robot of the Shangri-la Hotel, the Manga robot of Longlong Park, and the real-time rolling of the effective working hours each machine contributed during the month. He said, "Friends who have been in work have filled out their working hours, and future services to robots are billed according to Worken."
At the same venue, the report on innovation in the 2026 Smart Industries, which was prepared jointly by authorities such as the Chinese Institute of Information and Communication Technology, was successfully selected as an official publish, Worken and an example of the relevant deployment. publish is the only measure that has an authoritative endorsement.
According to AI-Technology comments, “from selling robots to selling time”, is the key to understanding the full business logic of the resource.
When robots charge Worken, the customer no longer buys a plant, or peak, depreciated year by year, but rather a continuously evolving productivity unit. The other sentence of pharaoh indicates a change in asset attributes: “The hardware is old, but the intelligence can appreciate.”
02
It was six months ago. This time it's on.
The concept of Worken was first conceived by the SAP 2026 Partnership Congress six months ago. At that time, Deng Taehua, the founder of Ji-won, asserted that 2026 was a year of intelligence from "development to deployment." "The real watershed in the industry is the shift from selling robots to delivering results."
This was supported by an eight-hour, uncut broadcast at the Dragon Flag Factory: Elf G2 works continuously as a "regular employee", 100% of success rate. It's a single piece of evidence to prove that it works, but it doesn't.
Over the past six months, the level of detail of the trade and the intellectual capital has been completely different.
On the operational front: 9,700 in the first half of 22026, with a steady global share of over 43 per cent; 20,000 robots down in September, from 10,000 to 20,000 for only six months; and close to $2 billion in the first half of the year.
The mass product is the “life line” of the human robotics industry, testing a combination of supply chain management, good quality control and low cost. The Ji-won partner Wang gave a bottom card for crossing the fence: 100 per cent of the localized A-chain, half an hour of supply chain circles, and 92 car standard mass tests.
"The mass production is not the end of the robot, but the guarantee that every 10,000 machines built are as reliable as the first. "The king goes in, "There's no shortcut behind it, it's all a pit on the line."
On the scene, the "old actor" who worked on the Dragon Flag Line in April and has been operating steadily for more than seven months; six days of live broadcasting over 60 hours of cumulative operation, with a total of 17625 computers, 64828 robotics, 22.5 seconds of single-synthesis, comparable to human workers. There's also an answer to the suspense of "can we hold a whole year's performance test."
Industry is also expanding: the Elf G2 has fully covered its flat mass inspection section at the Dragon Flag Plant, entering the chip sealing and CNC; the Elf G2 Max landed food, fast-down and electrician handling; the Quixin Longlong deployed nearly 500 robots to boost the visitors to the park by nearly 40 per cent; the Guangzhou Metro Metro has over 120 robots on duty patrols; and Cheng has landed in the first VLA-driven double-foot retailing operation around the world with a 98 per cent accuracy rate.
There is also an easily ignored but crucial number: human wibi.
The number published by Wang Jun has dropped from the original 1:1 (1 robot with 1 debug person) to 10:1 and the target is directed to 30:1. According to AI, the slope of this curve is whether the "deployance" is a real business or a litmus test of a custom-made project.
“The laboratory is the greenhouse, the real world is the wilderness. There is no shortage of technical difficulties in the industry, but of the productivity that allows them to go in bulk.”
APC in April said, "What's the deployment? It took six months to move from a concept to a delivery measure.
03
Technical Panorama:
Insurance for three multipliers for Worken.
The list of technologies for this publish meeting is extremely dense: the 3rd generation base-base model GO-3 Preview, the AGILE motor engine of the 4th layer, the full-modular interactive Z43ZWTA-Omni, the 3rd generation hardware architecture AHA3 and the reloading body G2 Max.
However, according to AI-Technology, rather than listing the parameters by article, it would be better to see the structures to which publish is referred: They are insured on the three multipliers of the Worken formula. (Ray Peak Network (WeChat Official Account:Ray Peak Network))
GO-3 Preview protects "output efficiency ratio". On the spot, Yao Zhiqing announced that GO-3 had come to the top of three authoritative lists that had recently assessed GPT-6's ability, such as in Robo Dojo.
What is more important than the list is the ability to evolve after deployment: CEL, a group experience learning module developed jointly with Chinghua, robotics take “trainers” for failure cases, and the task of measuring 100 robot cluster trainings success rate is pulled from 50% to 98%; it can generate new strategies to pull a single success rate from 20% to more than 80% in the event of a real convenience store scene.
As Yao Duk-chung summed up with blood: "The path to growth with a body model, not only trained in data, but also `dry' in real scenes."
AGILE and WITA-Omni guarantee "free working hours". Once the machine is defunct or assisted, the hours of work are emptied. To this end, the AGILE motor engine gives robotics a "closure recovery" capability through a four-tiered structure. This is a complete departure from the current industry of many dancing robots, "close blind copycats."
It's also the most "human" part of the scene. WITA-Omni is able to visualize, listen, talk, move in sync, achieve "hear and think" and reach the speed of response to the fastest 400 milliseconds, and has entered the natural rhythm of human daily dialogue.
AHA3 and G2 Max secures the number of active robots. The three minds are physically integrated at the hardware level, reducing the volume of the controller by 50 per cent and the delay by 50 per cent; the global primary cortex-coated liquid cooling system, which has a heat-dispersion capacity of seven times the conventional cold and a continuous cortex output of three times the cold; and the G2 Max arms peak, which carries 50 kg and double-battery heat-plugs for eight hours a day.
Synchronization of publish is also carried out by the new generation of smart-hand family Omnihand 3 series, of which Omnihand 3 is the first industrial-class smarter who has been working continuously on the production line.
The maturity of the product line directly determines the upper limit of fine operation. And the world record of the 110 expedition A3 "the world's largest human-shaped machine dance." It's called "the important starting point for the industry to scale."
In addition to the soft and hard configuration, Ji-won left behind a more lethal backhand: The first publish industry canal with a physical operating system OS 1.0, with a commitment to full access by year end. As part of a six-month-connected project with 80,000 developers, Ji-won is trying to monopolize the physical world's infrastructure.
One table looks technical, combined, and it's a precision set of "to keep working hours running."
04
Positive response to the "Strike Up":
LLM is an ammunition depot, not a gravedigger.
This publish will be the most valuable record of the fact that Ji-won didn't avoid the shadow hanging over the head of the industry: would LLM be able to “blow through” the track? What's the value of robotics?
Faced with this anxiety, Jiwon's partner Yao Zhiqing responded with one of the sharpest words: “If the industry is always going to be hit by AI, hopefully we will be on the side of AI.”
He further explained that the generic LLM single-debate reasoning was extremely time-consuming and very costly to calculate, “deployed in a real, high-precision, industrial landscape that required low delay”. The choice is to be self-studyed, saturated and LLM, while at the same time reducing capacity to "production-level" technology, and to declare GO-3 and the full-body AI technology framework an open source for the year, turning the architecture into an ecological one.
The young man, however, has developed two possibilities for the future:
The first is that universal basic models are becoming stronger and more open to anyone with advanced intelligence. The focus of the competition then becomes who can access the real world with a lower cost and higher reliability.
Second, the model remains highly dependent on data training from different ingenues and scenarios and does not have a universal model. Then “whoever has a large, real deployment and has stable access to high-quality data” wins.
And both of these futures are home to their own manufacturers.
“The two technical pathways appear to be different and ultimately point to the same conclusion: the intrinsic advantages of scale will become increasingly important.”
It's a pretty good argument. According to the AI-Technology Review, "Intelligence into the physical world" is never a software issue, but rather an integrated engineering issue of reliability, cost, supply chain, deployment network, transport systems, which is precisely the hardest part of LLM, an Internet gene.
Today, Ji-won's 20,000 deliveries, 99.99 per cent stability, 92 car-level tests, 10:1 man-to-man ratio stand in the way of pure algorithms' moat. General intelligence is not a deployed gravedigger, but a ammunition depot; thinking is given to LLM, but the barriers to action are always done by companies that grow in the physical world.
05
The details of a live “interception” are:
And the real maturity of the trade.
Finally, I want to say one of the most easily ignored moments of the scene: publish was at the scene, and during the demonstration, Jimoto's partner, Yuhiko, asked the excursion A3 "who's the best-looking" and the robot didn't hesitate to say, "I am the face of Jimoto." Then I asked "other than you," and the excursion A3 didn't answer mechanically, but instead said, "Of course you, your blue POLO is very spiritual."
In a live demonstration, excursion A3 can automatically tell who's talking to it when it's in the middle of a loudspeaker. On the spot, he carried it from Mandarin to Mandarin, Sichuan, Tonnan and even completed simultaneous interpretation.
The explanation is practical: "Voice recognition is not a showdown, it's an entry certificate. " WITA-Omni supports more than 100 languages and 29 dialects, and has been on the top of four interactive lists of authority, such as Daily-Omni, which will be fully deployed in the country by December. These capacities fall on the North Great People ' s Hospital ' s reference desk, in the open alley and at the advice point in Longlong, which is the time worked.
This moment is worth recording because it stands behind it and is one of the most "humanist" models of publish on this field: WITA-Omni Total Modular Interactive LLM. It does not simply collide several modules of voice, vision and action, but it uses all input outputs in the same model for simultaneous modelling to be so specific and emotional.
In addition to being able to observe and be emotional, it is possible to get robots to "hear and think" that the speed of response is at the speed of 400 milliseconds at an average of 600 milliseconds and has entered the natural rhythm of human daily dialogue.
The demonstration at the scene was more like a surprise test: with excursion A3 cutting seamlessly from Mandarin to Mandarin, Sichuan, Tonan, or even completing simultaneous interpretation; in the noisy arena, robots don't interject, waiting for time to tell who's talking to it, who's just talking to others. WITA-Omni supports more than 100 languages and 29 dialects and has been placed at the top of four interactive lists of authority, such as Daily-Omni, with full deployment in the country to be completed in December.
Similar senses are present throughout the scenes: the 22.5-second beat of the flag line, the failure of the convenience store connection, the 5G programme of the Long Rong "Physical Separation of Human Networks" and the design of the "New Point One-Pilot" platform. None of this was planned in the lab. It grew in the field.
Six years ago, the industry was better than the robots that were slamming it; three years ago, it was smoother than the Demo video that was slamming it; in April this year, Ji-won said that the game had been replaced by "Who's robots can work in a row"; this October, it simply gave the timer to the game.
As Shao-jun said at the opening: “Though Ji-won's first race was first, but it was commercial deployment that supported this achievement. Our robots can enter the race, the factory, the mall."
Forty-six sports fair gold medals are face and the real-time rolling working-time panels are Leeko. Worken ' s ability to become an industry-recognized measure ultimately depends on the joint adoption of the seller and buyer, which takes time to verify. But at least in the fall, when Ji-won was surrounded by the "smuggling theory", he did something that should have been done in the industry:
When everyone was arguing whether the robot was smart enough, it was the first one to come out, remembering the robot's attendance, and telling the trade: The second half is more intelligent than who is more capable of delivering time. (Ray Peak Network)
Original Ray Peak article, unauthorized reproduction. For further information, see the information below.
For your account security, please verify the mailbox.
Please verify your mailbox.
Auto-fetched and translated by AI Gem Box. All rights belong to the original author.