冷暖人生（全文在线阅读）>
冷暖人生世界人工智能大会上，顶尖专家都在讨论什么？_我的网站
一 |  A massive new geological map of the moon, measuring about 2.8 meters long and 1.2 meters high, has recently made its debut in China, offering one of the most comprehensive portraits yet of the lunar surface and its billions of years of geological evolution. Compiled by the Institute of Geology under the Chinese Academy of Geological Sciences (CAGS), together with the Institute of Geochemistry under the Chinese Academy of Sciences, Jilin University, Shandong University and China University of Geosciences (Beijing), the 1:5,000,000-scale Geological Map of the moon marks 13,519 impact craters, 81 impact basins, 14 types of geological structures and 17 rock types, and includes detailed inset maps of the lunar north and south poles. More than simply a new rendering of the lunar surface, the map incorporates the latest findings from China's Chang'e lunar exploration program. Scientists involved in the project told the Global Times in exclusive interviews that the new map recalibrates key boundaries in the moon's geological chronology, updates the distribution of important lunar rock types and establishes a more systematic framework for interpreting the moon's evolution. "The Chang'e-5 and Chang'e-6 missions have successively returned samples from the moon's near and far side, producing findings that have significantly changed our understanding of lunar evolution," Ding Xiaozhong, a researcher at the Institute of Geology under CAGS and a core member of the mapping team, told the Global Times. Samples showed that volcanic activity on the moon continued until about 2 billion years ago - roughly one billion years later than previously thought under the conventional framework. Meanwhile, observations and samples from the lunar far side have provided new information about the composition and magmatic history of the moon's deep interior, Ding said. "These new discoveries created an urgent need for a new and systematic geological map that could integrate the latest data and research results and establish an updated framework for understanding lunar geology," he noted. The 1:5,000,000 scale was chosen with practical applications in mind. Compared with China's more detailed 1:2,500,000 lunar geological atlas released in 2024, the new map provides a more condensed, global view of the Moon, enabling researchers to quickly identify major geological relationships while also making it suitable for education, popular science and the early-stage planning of exploration missions. Jin Ming, an associate researcher at the Institute of Geology under CAGS, told the Global Times that the project also represents an important step toward establishing a mature Chinese standard for planetary geological mapping. Its newly designed symbols, legends and color system have been certified as a group standard by the Geological Society of China, filling a gap in China's standardized rules for planetary geological mapping. Although the latest project formally began in 2022 and took about four years to complete, Ding and Jin said its scientific foundations stretch back decades. Data accumulation and methodology development began with China's early lunar exploration missions, particularly after Chang'e-1 started returning scientific data. Several generations of Chinese planetary geologists have since developed a comprehensive system for lunar geological interpretation, remote-sensing data processing and map compilation. One of the toughest challenges was bringing together enormous quantities of remote-sensing data obtained by different instruments, Jin said. Different payloads aboard the Chang'e probes produced data with varying resolutions and spectral characteristics, requiring specially developed algorithms as well as extensive manual verification to place them under a unified mapping standard. "We had to make the different datasets compatible under the same system while striking a balance between scientific completeness and visual clarity - avoiding information overload without losing critical geological clues," Jin said. China previously released the world's first 1:2,500,000-scale high-precision lunar geological atlas in 2024. The newly completed 1:5,000,000 map is therefore not simply a reduced-scale copy, the scientists stressed, but a scientific update incorporating newly available data and revised interpretations of lunar chronology, lithology and geological evolution. China's new lunar geological map has achieved an all-round improvement over previous versions and is now at the international forefront, Ding said. We are the only country to have exclusive access to directly measured sampling data from the Chang'e-5 and Chang'e-6 missions. By using age anchors from lunar samples independently obtained by China to calibrate the geological chronology, this new data source and calibration system represents a core advantage of China." Like Earth, which is divided into geological periods such as the Cambrian and Jurassic, the Moon also has a geological timescale. One of the most significant tasks behind the new map was to recalibrate that "lunar clock." According to Jin, the recalibration draws on two major sources - the latest international isotopic dating results, including updated measurements of Apollo samples and lunar meteorites, and the precise age obtained from Chang'e-5 samples. The latter provided an especially important chronological anchor. Radiometric dating of Chang'e-5 basalt has placed its age at roughly 2 billion years, demonstrating that lunar volcanism persisted far later than had once been assumed and substantially changing scientists' picture of the moon's late-stage thermal evolution. Based on the updated evidence, the team refined the starting ages of several ancient lunar periods, including the Aitkenian, Nectarian and Imbrian periods, to approximately 4.33 billion, 4.17 billion and 3.92 billion years ago, respectively. More significantly, the researchers used the 2-billion-year age boundary to divide the Eratosthenian into an early and a late epoch. The team said this marks the first time an age boundary derived from samples independently collected by China has been incorporated into such a detailed Chinese lunar geological chronology. "The moon's story itself has not changed, but its timeline has," Jin said. "We once thought that major lunar volcanism had essentially cooled down about 3 billion years ago. Now we know that the Moon remained volcanically active until at least about 2 billion years ago. That means the thermal evolution of the lunar interior and many of the causal relationships in its later history need to be re-examined." Ding highlighted a newly mapped rock type in the South Pole-Aitken Basin - gabbronorite - identified with the help of the latest Chang'e-4 and Chang'e-6 exploration results. The vast basin on the lunar far side is one of the moon's largest and oldest impact crater. "This is like being able to directly see material from deeper inside the moon," Ding said. "Previously, much of this could only be inferred from remote-sensing observations. With Chang'e-4 in-situ exploration and Chang'e-6's returned samples, we now have much stronger evidence for identifying its existence and distribution." Another important revision concerns KREEP, a lunar geochemical component enriched in potassium, rare-earth elements and phosphorus and commonly associated with elevated concentrations of radioactive heat-producing elements such as thorium and uranium. According to Ding, earlier models suggested that KREEP-rich material was concentrated in a relatively limited area. Higher-resolution data used in the new mapping project indicate a broader and more continuous distribution than previously thought, suggesting that its abundance may have been underestimated. Ding said the finding is scientifically important not only for assessing potentially resource-rich lunar materials, but also because the distribution of KREEP provides clues to the crystallization of the early lunar magma ocean, the thickness of the lunar crust and the moon's internal thermal evolution. The map's colors also carry geological meaning. The team follows a principle of "lighter for younger, darker for older": relatively young Copernican impact materials are depicted in bright pale yellow, ancient Aitkenian impact materials in deep purplish red, while mare basalts are represented in varying shades of blue. "The color system combines the logic of geological evolution with elements of traditional Chinese color aesthetics, giving China's lunar geological maps a distinctive scientific and visual identity," Ding said. Beyond basic science, the researchers said the new geological map can serve as a strategic base map for China's future lunar exploration, including subsequent Chang'e missions and planning related to the International Lunar Research Station (ILRS). A refined geological chronology can help scientists understand the age and modification history of candidate regions; updated rock mapping can identify areas of particular scientific or resource interest; and the global-scale representation allows mission planners to compare broad regions before moving to higher-resolution site studies. Asked whether the map could directly support the Chang'e-7 and Chang'e-8 missions, Jin stressed that its role is primarily strategic rather than tactical. "The 1:5,000,000 geological map is a global-scale foundational map. It can be used for broad comparisons of potential landing regions and for evaluating their overall geological context," Jin said. "Precise landing-site selection and detailed rover-route planning, however, require higher-resolution regional geological maps. The two work together, moving progressively from the global scale to finer regional scales." The scientists also stressed that the map is intended to be shared with the international scientific community. Digital versions are expected to be released through professional publications and public scientific databases, while the standardized symbols and mapping rules could facilitate exchanges with other planetary geological mapping systems. Ding said significant room remains for international cooperation, from comparing and translating different geological mapping standards to joint studies of the evolution of the lunar far side and the origin of the ancient South Pole-Aitken Basin. "China possesses unique exploration data and returned lunar samples, which provide a strong foundation for cooperation," Ding said, adding that planetary science, scientific training and deep-space exploration technologies all offer potential areas for expanded exchanges. "A precise geological map of the moon is an important foundation for humanity to explore deep space and understand the universe," Jin said. The latest map, he noted, is not an endpoint. As China proceeds with subsequent lunar exploration missions and acquires new observations and samples, the geological picture of the Moon will continue to be refined. "Future exploration will allow us to keep updating our understanding of lunar geology," Jin said. "At the same time, we hope these results can contribute to broader international scientific cooperation and to humanity's exploration of the moon." 。 文 /巴九灵 过去几年，AI似乎越来越像人。 它会写代码、做PPT、分析财报、生成视频，也开始进入工厂、医院和家庭。
二 | 许多人相信，我们距离通用人工智能（AGI）已经越来越近了。 但就在前两天，图灵奖得主、强化学习之父理查德·萨顿泼出了一盆冷水：“现在的大模型，谈不上原生智能。” 过去四天，2026世界人工智能大会（WAIC）在上海如期召开，这是过去9年规格最高的一届，也是全球AI领域最重要的年度盛会之一。 有人在现场逛展，亲身体验最先进的AI产品；有人盯着龙头企业，关注最新发布的大模型、应用、成果；而小巴在关注，全球AI行业的领军人物齐聚一堂，他们究竟在讨论些什么。 2026世界人工智能大会 本次大会汇聚了来自全球的上千位顶尖科学家、专家、企业家、投资者，除了理查德·萨顿，还有“深度学习三巨头”之一的约书亚·本吉奥、图灵奖得主吉勒斯·布拉萨德、图灵奖得主姚期智、诺贝尔奖得主奥马尔·M·亚吉、“硅谷精神之父”凯文·凯利，等等。 小巴梳理了大会四天的演讲、论坛后发现，专家们似乎把目光聚焦在了以下五点上：我们是否高估了今天的AI、AI如何获得经验、AI落地为什么难、智能体的后果谁来负责、发展AI的终点是什么？ 要点一：我们高估了今天的AI 过去几年，我们一直在见证AI的进步，它替代了许多低级岗位，给人最直观的印象就是它不断刷新大家对智能的想象。然而，在一些顶尖科学家看来，我们可能高估了今天AI的智能程度。 答案藏在一个长期被混淆的问题里：知道很多事、会做很多事，是否就等于拥有了真正的智能？ 图灵奖得主、中国科学院院士姚期智表示，AI虽然看起来很强大，但未必就是智能的终点。萨顿则讲的更明白——大语言模型虽然看了海量文字、图片和视频，但它还是在学习和重新组织人类已有的知识，而不是自己创造知识，计算能力的突破，不是智能的突破。 人类拥有智能，不只是因为读过书、记住了知识，更是因为我们会行动、会犯错，能够从现实世界得到反馈，并在一次次反馈中修正自己，获得经验。 举例来说，一个人没拿稳杯子，于是知道玻璃杯掉在地上会碎，而AI没有自己砸碎过一只杯子，它是从人类的文字中得知杯子掉在地上会碎，这是它记住的知识，不是它自己的经验。 复旦大学浩清特聘教授苏昊认为，这也是大语言模型产生“幻觉”的深层原因之一：语言只是世界的投影，不是世界本身。 因此，在萨顿看来，AI需要从“人类数据时代”走向“经验时代”。 要点二：AI缺少的是经验 在这样的背景下，“智能体”和“物理智能”，成为了今年大会上被反复提及的两个方向。 智能体让AI从“回答问题”走向“执行任务”，它可以围绕一个目标制定计划、调用工具、采取行动，并根据结果调整自己的行动；物理智能则让AI走出屏幕，进入由重力、摩擦、材料、空间等构成的真实世界。
三 | 而这也引出了今年大会上被反复提及的一个概念——世界模型（World Model）。 简单来说，世界模型是AI在理解现实世界运行规律的基础上，在自己的“大脑”中对现实世界进行模拟：如果我这样做，接下来会发生什么。 比如，当人准备去拿一个杯子时，大脑其实已经提前完成了一次预测：杯子会不会滑？需要用多大力？如果手一松，会不会掉到地上？ 这种预测能力，无法仅靠阅读、记住更多知识获得，它正需要AI走进现实世界，通过智能体或者物理智能与环境互动，最终获得属于自己的经验。 2026世界人工智能大会上的人形机器人 换句话说，智能体或者物理智能不完全相同，但都能让AI在行动中获得反馈，在反馈中积累经验，不断完善自己的世界模型，逐渐形成对真实世界运行规律的理解。 借用北京银河通用机器人联合创始人张直政的总结，世界模型应当拥有三个不可或缺的能力：建模/预测、规划/策略、评价。简单来说，就是AI能预测自己行动会给环境带来什么变化，利用这些预测指导自己的决策和学习，并判断自己的行动是否达到目标。 要点三：AI落地难点重重 然而，帮助AI获得经验、发展世界模型的这条大道虽然光明，却依旧险阻——大模型已经发展了很多年，却迟迟没有大规模走进真实世界，这背后，是因为存在很多难点。
四 | 智能体发展的难点之一，在于智能体操作系统的完善。 如果说模型是AI的“大脑”，那么操作系统决定了它能够调用哪些工具、连接哪些设备、完成哪些任务。再聪明的大脑，没有手，也很难真正工作。 因此，阶跃星辰董事长印奇表示，智能体能做什么，不只取决于模型有多聪明，也取决于操作系统为它打开了多大的能力边界。他提出了一个公式：智能体能力=模型能力×智能体操作系统能力。
五 | 在此之外，智能体发展还存在许多问题，比如，如果未来有上万个智能体同时工作，它们如何协作？如何共享能力？不同公司的系统能否互相兼容？ 对此，上海润科具能科技有限公司CEO周斌期望，行业能够加速建立统一标准与开放接口，让智能体真正“聊起来”、能力可共享，形成“智能伙伴网络”，而非各自为战的工具集合。 对于物理智能，苏昊认为，难点同样也是突破口，在于知识的聚合。 视频告诉AI世界长什么样，物理方程告诉AI世界为什么会这样运转，机器人不断提供新的实验数据……只有把这些不同来源、不同领域的知识聚合进同一个模型，互相校准补盲，AI才有可能真正理解现实世界。 而聚合这件事，注定不是一家企业能够独立完成的，而是这一代人要完成的科学工程——苏昊表示，它需要整个行业乃至全社会的通力协作：共建数据，共立标准，共享仿真、评测等基础设施。 不过，这并不意味着物理智能必须等到“什么都会”才能落地。 在张直政看来，物理AI未必要先实现全面通用，而是可以先在仓储、物流等相对固定的场景中实现稳定工作。
六 | 未来，物理智能或许会像当年的电气化一样，先点亮工厂和仓库，再走进商店和医院，最后才走进千家万户。 2026世界人工智能大会的全球首个AI鉴别机器人 要点四：AI治理规则尚未完备 然而，当智能体和物理智能真的加速走进千家万户，安全问题也变得越来越重要、无可回避。 印奇提出了三个问题：智能体可以代表谁行动？行动产生的后果，由谁来负责？当它连接数据、设备和真实环境，怎样确保身份可信、权限可控、行为可追溯？ 上海人工智能实验室主任周伯文提出，AI可以代表人行动，但不能代替人负责，让渡代理权的速度，不应超过我们验证它的能力。 纽约科学院院长兼首席执行官杜宁凯表示，要采取行动开展人工智能治理，必须要创造出新的问责方式。
七 | AI能力越强，人类越需要建立起一套与之相匹配的规则，而如今，关于AI的代理权、问责权、治理能力，显然还未完备。 要点五：AI的意义是放大人 不过，治理并不是AI发展的终点。治理的目的，是让AI能够更安全、更可靠地融入人类社会。
八 | 回看今年的WAIC，一个变化已经十分明显：大家讨论的重点，正在从“模型有多强”，转向“AI如何获得经验、如何进入真实世界、如何安全地与人协作”。
九 | 这背后，其实回到了最初、也最根本的问题：AI发展的目的究竟是什么？ 苏昊说，“物理智能的使命，是把人还给人”；印奇说，“智能体时代最大的意义是放大人”；上海绘话智能科技有限公司CEO贾学锋说，“AI大发展的终极价值应当回归于人”；商汤科技董事长兼首席执行官徐立说，“AI真正意义上走向大众并不是通过替代，而是通过放大每个人的能力”…… 图灵奖得主约翰・爱德华・霍普克罗夫特则提醒，面对信息革命带来的种种变革，各国的最佳战略战略，或许是着力培养能够应对这些变化的人才。 今天，大模型可能依旧谈不上真正的智能，但如果AI能够在现实世界中不断学习、积累经验，并始终与人协作，为个人、社会乃至整个世界创造更多可能性,那么，它距离真正的智能，或许就又近了一步。
十 | 作者|蒋紫涵 主编|何梦飞 图源|VCG、网络。 Current article:http://40z9al.gamajionghuanmuniuzhuanguchu.cyou/news/20260826_24696.html Published on:12:14:11 |
