太极张三丰

China crackdown on cyber scams nets thousands but leaves networks intact_我的网站

灭顶之灾

一 | 该图片使用了AI生成技术本文来自微信公众号：具身商业前哨，作者：何安这次去上海参加WAIC，让我印象最深的活动是“智启具身论坛2026——迎接物理AI智能涌现”。 BANGKOK -- Zhang Hongliang, a former restaurant manager in central China, took various gigs in and outside China to support his family after losing his job during the COVID-19 pandemic.In March, a job offer to teach Chinese cooking at a restaurant led him into a cyber scam compound in Myanmar, where he was instead ordered to lure Chinese into giving up their savings for fake investment schemes via social media platforms.Zhang is one of tens of thousands of people, mostly but not all Chinese, who have become ensnared in cyber scam networks run by powerful Chinese criminal syndicates in Southeast Asia. Regional and Chinese authorities have netted thousands of people in a crackdown, but experts say they are failing to root out the local elites and criminal networks that are bound to keep running the schemes.When scam operations are shut down in one place they often just resurface elsewhere. The problem is an embarrassment for Beijing and is discouraging ordinary Chinese from traveling to Southeast Asia out of fear they might be duped or kidnapped and caught up in a cyber scam operation. In recent years, media reports have uncovered instances of young people being lured to places in Cambodia or Myanmar for high-paying jobs, only to be forced to work as scammers. Rescue organizations say people are regularly beaten or face physical punishments such as being forced to run laps if they don't perform well. In August, China, Thailand, Laos and Myanmar agreed to set up a joint police operations center to tackle cyber scams in the region. On Oct. 10, China’s Ministry of Public Security announced that its “Summer Operation” had successfully brought back 2,317 scam suspects from northern Myanmar to China.China calls such people suspects, though experts say most of them are victims who were forced to work for the criminals. They question how they will be treated once back in China. The schemes based in countries like Myanmar, Laos and Cambodia are run by Chinese bosses hand-in-hand with local elites. Many are based in places where China has financed big construction projects through leader Xi Jinping’s signature Belt and Road Initiative.Myanmar's border regions long have been a magnet for criminals — historically including drug producers and traffickers — because of lax law enforcement. Such places are generally under the control of ethnic minority armed groups, either opposed to or allied with Myanmar’s central government. Some also cooperate with organized crime gangs.“From the vantage point of the Chinese government, it’s a source of extreme embarrassment that you have so many of these Chinese criminals operating all across Southeast Asia,” said Jason Tower, an expert on transnational crime with the United States Institute of Peace.The syndicates also are known for “pig butchering” cons, where scammers entice individuals, often halfway across the world, to invest their money in bogus schemes after duping them into digital romances.The scammers divide their targets into two categories: Chinese and non-Chinese. They use scripts, images of models and influencers and translation software to trick the people they contact by phone or online into parting with their money. Victims can be anywhere in the world.The criminals have “ridden on the shoulders of the Belt and Road Initiative,” said Tower, who outlined links between the criminals and Chinese state enterprises, think tanks and government officials in a 2020 report written for the United States Institute of Peace.Zhang was working in Thailand and on a visa run to Laos when he met the man who lured him to the scam compound in Myanmar. Giving what he said was his last name, Gao, he claimed to be a broker and travel agent for Chinese living in Thailand. Zhang and his wife wanted extra money to pay for in vitro fertilization to have another child. Gao suggested he go work in Myawaddy, in eastern Myanmar’s Kayin state, teaching a local chef how to cook Chinese dishes in Gao’s new restaurant. The pay would be double what Zhang made in China.Zhang was wary. Since a 2021 coup, military-controlled Myanmar has been embroiled in civil conflict. But Gao reassured him that he wouldn’t be doing anything illegal and said the restaurant would have plenty of customers since many cyber scam businesses were operating in the area.That might have raised a red flag but it was only once he got to Myanmar that Zhang realized his predicament. He asked to go back home, saying there was a family emergency. His family helped him scrape together some 40,000 yuan ($5,472) to pay off the debt Gao claimed he owed him, and he slipped away one night, swimming across the Moei River into Thailand, where he turned himself in to Thai police, who contacted the Chinese Embassy.Zhang showed the AP copies of his deportation notice from the Thai Immigration police and a temporary ID card. He returned to China in late June and was questioned by Chinese police but not detained. He has been sharing his story on Douyin, the Chinese version of TikTok, to alert others to the risks and says people often contact him about relatives trapped in cyber scam compounds.“We all went out with this wonderful sense of hope, but then reality slammed us in the face,” he said.In total, China has detained some 4,000 suspects and returned them back to China. The Ministry of Public Security has claimed “breakthrough results” through operations in coordination with Myanmar authorities. On Monday, they announced they had repatriated another 2,349 people. The ministry did not respond to a faxed request for comment.One 31-year old former chef who was smuggled into Myanmar’s Wa State earlier this year said he saw his company hand over four people to Chinese police with little fanfare in September. Other companies did the same, said the man, who was smuggled into Myanmar and later rescued by a non-profit organization. He declined to be named out of fear of government retribution, and The Associated Press could not independently verify his account.Overall, the enforcement actions don’t seem very comprehensive, experts say. The groups now based in Myanmar originally were located in Cambodia. When Cambodia cracked down on online gambling rings and illegal casinos in 2019, many of the groups just moved to less well policed places in Myanmar. Some were taken over by rival gangs.China's efforts to repair its image have so far not made much headway, said Thitinan Pongsudhirak, a professor of political science at Thailand’s Chulalongkorn University. “You can crack down on these symptoms and the manifestations … that you can see in the borderland areas," he said, "but they’ll come back unless you really have a sustained effort." -—-AP researcher Wanqing Chen in Beijing and AP writers Grant Peck and Jintamas Saksornchai in Bangkok contributed to this report.。这场论坛的含金量，不只在于智元、PI、Dyna、Sunday、腾讯和高校研究者坐在同一张桌子上，更在于他们讨论的是数据从哪里来、模型往哪走、本体要不要自己做这些行业尚未回答的硬问题。

二 | 论坛最后，主持人问了一个所有人都关心的问题：机器人的“ChatGPT时刻”，到底什么时候来？智元合伙人、觅蜂科技董事长兼CEO姚卯青说，两年； Sunday Robotics联合创始人兼CEO赵子豪说，三年以内；腾讯首席科学家、腾讯Robotics X实验室主任张正友说，三到五年； Dyna Robotics联合创始人兼首席科学家马也骋说，四年左右； Physical Intelligence研究科学家任至意说，四到五年；佐治亚理工学院徐丹飞教授给出的答案是五年。几位嘉宾所在团队的技术路线和产品落地方式并不相同，但他们给出的时间判断却出奇地接近：两到五年。这个共识，比我想象中更激进。不过，这些数字背后的口径未必相同。姚卯青给出的定义相对明确：机器人能够开箱即用，听懂开放式自然语言指令，并在常见任务上达到70%—80%左右的基础成功率。但其他嘉宾并没有逐一定义自己的标准。所以，大家对时间看似形成了共识，但对“什么才算机器人的ChatGPT时刻”，其实还没有共识。整场讨论中，有三个问题最受关注：机器人的数据飞轮怎么形成？“VLA已死”是不是一句伪命题？具身智能公司是不是正在走向软硬件全栈？这三个问题，正好构成了理解当下机器人行业的三条主线：数据决定机器人如何进化，模型决定智能如何涌现，软硬件路线则决定谁能把能力真正变成产品和生意。一、机器人的数据飞轮，不是采出来的过去几年，大语言模型的成功，给了机器人行业一个直接的想象：只要不断扩大数据、算力和模型规模，机器人是不是也会迎来一次能力涌现？可问题是，大语言模型需要的数据本来就存在于互联网上。机器人需要的物理经验，却没有统一的数据来源。姚卯青在现场做了一个估算：头部语言模型的预训练已经达到100万亿Token量级。如果要让机器人真正开箱即用，可能需要一亿小时级别的数据。这一亿小时从哪里来？真机数据质量高，但采集成本太贵；仿真容易扩大规模，却始终存在虚实差距；互联网视频和第一视角数据数量足够，却未必能看清最关键的手部动作和力觉信息。

三 | 讨论中，一个被反复提及的数据来源是UMI。你可以简单把它理解为一种低成本的人类操作数据采集方式：人拿着采集装置完成任务，系统记录视觉信息和动作轨迹，再把这些经验转成机器人可以学习的数据。赵子豪认为，短期来看，UMI可能是最高效的规模化数据来源。但他同时强调，长期真正重要的还是部署数据。

四 | 两句话放在一起，恰好勾勒出一条可能的数据路径： UMI解决“机器人还没有大规模上岗之前，去哪里获得经验”；部署数据则解决“机器人上岗以后，如何越干越好”。任至意更看重真机数据，同时主张通过多源数据和跨本体训练，提高模型的泛化能力。马也骋则把这些数据总结成一个“数据金字塔”。底层是规模最大、成本最低的互联网视频和人类第一视角数据，中间是真机多任务数据，塔尖则是数量最少、价值最高的部署数据，尤其是机器人犯错、停顿、恢复并最终完成任务的过程数据。它不是在告诉机器人“理想情况下应该怎么做”，而是在教机器人“真实世界不按剧本走的时候怎么办”。 Dyna在酒店部署机器人完成折毛巾、折餐巾等任务时，会把前几次部署的数据重新放回预训练。随着数据积累，进入下一家酒店需要重新适配的时间就会缩短。

五 | 表面看，嘉宾们都认同多种数据需要组合。

六 | 但真正到了资源投入上，各家公司押注的重点并不一样：有人优先扩大人类操作数据，有人坚持真机数据，有人则直接进入真实场景，用部署换数据。这里有一个典型的鸡生蛋问题。没有足够好的模型，客户不敢让机器人上岗；没有真实部署，机器人又拿不到最有价值的数据。说白了，机器人的数据飞轮不是采出来的，是在真实场景里干活干出来的。二、“VLA已死”，死的可能不是VLA 最近，具身智能行业很流行一句话：“VLA已死，世界模型当立。

七 | ” 但在这场圆桌上，这场争论被进一步落到了VLA与WAM（World Action Model，世界动作模型）两条路线上。 VLA是根据视觉信息和语言指令生成动作的模型；WAM则更强调世界如何变化，以及不同动作可能带来什么结果。 Dyna确实从早期的VLA路线，逐渐把更多精力转向WAM。马也骋提到，在一些强调鲁棒性、少量数据和高成功率的任务里，WAM可能比VLA更高效。那么，VLA真的要被淘汰了吗？任至意并不认同。

八 | 他认为，到目前为止，还没有看到哪家模型做出明显超过π0.7的演示效果。π0.7是PI推出的新一代通用机器人模型，也是PI对VLA路线的一次重要升级。更关键的是，任至意认为VLA与WAM并不冲突。

九 | VLA本身也可以吸收世界建模和未来预测能力。机器人最终既要理解语言和当前环境，也要预测未来结果。PI的做法，就是先生成未来的子目标画面，再用这个画面引导底层模型执行动作。智元也没有简单押注其中一条路线，而是在推动VLA与WAM融合，探索世界建模、推理和动作能力进一步结合的模型体系。

十 | 这场讨论的张力就在这里。

十一 | 任至意用π0.7说明VLA仍有很强的竞争力，马也骋则用真实任务中的效率和成功率说明WAM的价值。

十二 | 两个人没有直接否定对方，但公司的技术下注已经给出了不同答案。张正友又把问题往前推了一步。他认为，具身智能远没有形成类似大语言模型Transformer那样的统一架构。一个真正的机器人智能体，至少要包含认知系统、感知行动系统和底层反应系统。上层负责复杂推理和任务规划，中层负责感知与动作生成，底层则要以极高频率控制电机、处理碰撞和保持平衡。

十三 | 你不可能让一个数千亿参数的云端大模型直接控制每一个电机，也不能指望一个三四十亿参数的端侧模型解决所有问题。

十四 | 所以，“VLA已死”很可能从一开始就是个伪命题。未来形成主流的，大概率不是纯粹的VLA，也不是单独的WAM，而是一套融合语言理解、世界建模、未来预测、推理规划和实时控制的系统。 VLA没有死。真正可能死掉的，是“靠一个模型解决机器人所有问题”的想象。三、全栈能力会成为趋势，但全栈生意未必会第三个问题，是机器人公司究竟应该只做“大脑”，还是连身体一起做。 PI现阶段主要做通用基础模型，没有选择自己研发和销售本体。任至意坦率地说，没有自研硬件，确实可能导致部分任务无法完成。但PI目前更看重模型快速迭代，因此优先使用结构简单、维护成本低的机器人搭建研发基础设施。这是一条典型的“大脑路线”：先把通用模型能力做出来，把本体作为研发基础设施，而不是现阶段的商业重点。腾讯的选择也很明确。张正友说，腾讯只做大脑，不会卖本体。他甚至开玩笑说，腾讯过去做硬件基本没有成功过，所以要有自知之明。

十五 | 但腾讯的Robotics X实验室从2018年成立以来，却一直保留着硬件研发。因为如果没有真正做过硬件，就很难理解触觉、力觉、关节控制和端侧算力，究竟会给模型带来什么限制。这是一条很微妙的边界：商业模式上可以不卖本体，研发上却不能不懂本体。

十六 | 智元走的是另一条路。

十七 | 姚卯青认为，一旦机器人进入客户现场，面对的就不再是论文指标和Demo效果，而是可靠性、成本、成功率和一致性。

十八 | 这些问题没办法只靠模型解决。端侧芯片能运行多大的模型，电池能工作多长时间，关节如何控制，整机能不能稳定量产，都会反过来决定机器人的智能上限。

十九 | Sunday Robotics选择三指夹爪，也是同样的逻辑。

| 赵子豪认为，机器人的手要在成本与能力之间做平衡。三指夹爪追求的是用20%的成本获得80%的能力。今天，通用机器人迟迟没能大规模进入家庭和更多真实生产场景，未必是因为少了两根手指，更可能是因为智能不够、成本太高、稳定性不足。

| 这正是全栈路线的价值：不是为了什么都做，而是在模型、算力、本体、控制、成本和场景之间更好地做取舍。那么，软硬件全栈会不会成为新趋势？我的判断是：全栈能力会成为趋势，但全栈商业模式未必会成为唯一答案。未来，机器人行业很可能会出现类似苹果的公司，模型、本体、系统和产品都自己完成；也会出现类似安卓和Windows的平台，只做通用大脑，适配不同本体。但在今天，本体、传感器、灵巧手、芯片和控制系统都没有标准化。即使一家公司的最终目标只是卖“大脑”，也很难完全绕开身体。没有一家公司还能假装硬件不存在。四、两年还是五年，其实没那么重要这场论坛听下来，我最大的感受，是机器人行业远远没有收敛。哪种产品形态最适合家庭和工厂，没有答案；VLA、WAM、端到端模型和分层系统，谁会成为主流架构，也没有答案。数据配方仍在变化，商业模式同样分散。有人卖整机，有人做模型，有人做数据平台；有人押注工厂，也有人瞄准家庭。表面看，这是一种混乱。

| 换个角度看，这也是机会。在一个尚未收敛的行业里，大厂和创业公司、中国团队和美国团队，都没有提前拿到标准答案。所以，听完整场论坛，我反而不太在意答案究竟是两年还是五年。机器人的“ChatGPT时刻”，不会只由某一次模型发布宣布到来。它真正出现的标志，是数据、模型、本体与真实部署第一次形成自我加速：机器人部署得越多，积累的数据越多，能力提升越快，进入新场景的成本也越来越低。