伪娘
Young seekers uncover Japan's wartime past across borders; Precious historical archives delivered amid 8Ist anniversary of Japanese surrender_我的网站
一 | 图源：视觉中国 蓝鲸新闻7月20日讯（记者 翟智超）AI Agent能在几秒内给出实验假设、文献总结或实验方案，但真正把材料做出来，依然要靠人手——这是当前材料研发领域最核心的矛盾。 7月19日，小米材料CORE团队负责人俞之奡在接受蓝鲸科技等媒体采访时，首次系统阐述了小米在AI for Materials和前沿材料领域的布局逻辑，这个问题自他加入小米以来便被反复问起。
August 15 this year marks the 81st anniversary of Japan's unconditional surrender in World War II. As the main battlefield in the East for the war, China fought a 14-year resistance war at the cost of 35 million casualties. It pinned down the bulk of Japanese military forces and made indelible contributions to the victory of the global anti-fascist cause.
On this significant anniversary, French youth Marcus Detrez, his compatriot Bastien Rata, and Chinese youth Zhong Haosong will donate precious historical archives to the Memorial Hall of the Victims in Nanjing Massacre by Japanese Invaders.
Global Times reporters accompanied the trio throughout their cross-border archival retrieval mission, witnessing firsthand the handover of historical materials collected in Paris and Tokyo to archival institutions and memorial halls in Chongqing, Shanghai and Nanjing.
"Without justice, there can be no true peace," Marcus told the Global Times during his trip from Shanghai to Nanjing on Thursday.
History through multiple perspectives
A landmark handover of historical documents shedding new light on Japan's wartime bombing of China took place on August 10 at the Chongqing Municipal Archives, as Marcus and Zhong donated digital scans of 138 pages of rare materials related to the devastating Chongqing Bombing.
Compiled by Bastien from the Nantes Diplomatic Archives Centre in France, the mostly French-language files—accompanied by Chinese texts and maps—offer an unflinching third-party account of Japanese air raids on Chongqing, Chengdu and other cities between 1938 and 1941. The archives detail exact bombing dates, aircraft formations, ordnance types, civilian casualties and infrastructure destruction.
Of particular note within the archives are the Chinese and French translations of a so-called Statement of the Government of the Empire of Japan. Attached to a letter dated November 1938, the document brazenly demanded that China cease its resistance against Japanese aggression and submit to Japan's so-called "new order in East Asia", threatening military punishment if its demands were rejected.
Marking the 89th anniversary of the August 13 Battle of Songhu, also known as Battle of Shanghai, the historical preservation effort reached another milestone on August 13 in Shanghai, where youth activists presented 891 pages of invasion-era archives to Shanghai Songhu Memorial Hall for the War of Resistance Against Japanese Aggression for further research.
The archives, obtained from the Nantes Diplomatic Archives Centre, include records detailing the dates and locations of Japanese bombing raids on parts of Shanghai during the Battle of Shanghai, as well as the number of aircraft involved. The materials also cover Japan's aggression in Northeast China, the puppet state of "Manchukuo", Japanese atrocities and bombings, with some records also relating to the Soviet Union and the Korean Peninsula.
The third batch of materials, scheduled to be handed over to the Memorial Hall of the Victims in Nanjing Massacre by Japanese Invaders on August 15, was brought back from Japan last week by Marcus, Zhong and others. During their trip, they visited several WWII archival collections and secondhand bookstores in Tokyo, uncovering more than 50 items weighing about 20 kilograms, including pictorials, publications, books and photo albums dating mainly from 1932 to 1945. The materials cover the Nanjing Massacre and Japan's preparations for launching the July 7th Incident, among other topics.
Among the finds was a book titled The Nanjing Front, which documents events related to the Nanjing Massacre despite not mentioning the massacre in its title. They also found photos taken by Asahi Shimbun of Japanese air raids, as well as images of the Great Wall after it was damaged during the Japanese military's occupation.
"These archives should go where they belong," Zhong told the Global Times.
'Guarded' Japanese collectors
In France, archival research was a race against time. Bastien told the Global Times that they had only four days to work at the French diplomatic archives in Nantes. Each person could access no more than six boxes a day, or 24 boxes over four days. Some contained as many as 10,000 to 20,000 pages, while only 10 to 20 percent might relate to Japan's war of aggression against China. "We had to go through the pages one by one, quickly identify anything relevant and scan it immediately," Bastien said.
The team then spent nearly a month organizing the material, working about eight hours a day. They also developed an AI-powered system to translate and classify the records and extract information on dates, locations, people and events, while linking the French diplomatic archives with photos and notes left by Marcus' grandfather.
Finding wartime records in Japan posed a different challenge. Much of the material is not publicly catalogued online, so the team had to visit collectors one by one and keep asking specific questions.
"The biggest challenge is that many Japanese collectors are cautious about what we are looking for. Some know these materials exist but choose to avoid the issue, while others have clues but are unwilling to share them," Zhong said.
The team visited one Tokyo collecting institution twice. Despite explaining during their first visit that they were looking for wartime materials from the 1930s and 1940s, the collector initially showed them only ordinary photos of Japanese soldiers. The team then changed its approach, asking why many Japanese WWII records contain gaps concerning China. The collector said Japan generally refers to the conflict as the "Sino-Japanese War," rather than Japan's war of aggression against China. The team pressed further, asking whether there were records related to Shanghai, Nanjing or Chongqing. Only then did previously unseen materials gradually emerge.
The team also cross-checked promising finds with Chinese institutions, including the the Memorial Hall of the Victims in Nanjing Massacre by Japanese Invaders, and traced their provenance. One photo album, for instance, came from the family of a Japanese soldier and still contained the photographer's details and names of people pictured.
Anchoring hard-won peace
In early August, Marcus and Zhong stood at a busy intersection in Tokyo's Shibuya district, dressed in black and holding Japanese-language banners urging Japan to confront its history of aggression, including the Nanjing Massacre. They spent about three hours raising awareness on the street. Some foreign tourists stopped to ask what had happened, while some Japanese passersby tried to stop them.
There was also support. A Japanese female blogger livestreamed the event, saying she wanted to share the truth and promote peace. Japan's refusal to face its wartime history and apologize, was "irresponsible," she said.
"Our long-term goal is simple: to let the world see the truth of history," they told the Global Times.
Zhao Yongsheng, a professor at the Institute of Regional and International Studies at the University of International Business and Economics in Beijing, said the growing involvement of international friends in uncovering Japan's wartime history and preserving the truth carries unique value. It not only helps safeguard historical truth, but also challenges long-standing Western biases and neglect of the Eastern theater of World War II. More importantly, it demonstrates that exposing wartime atrocities and defending peace and justice are shared moral responsibilities of humanity.
"There are concerns that confronting Japan's wartime past could create new instability in Asia. In fact, the opposite is true," Zhao said. "Germany became a 'stabilizer' in Europe partly because it confronted its Nazi past. In Asia, only by fully confronting Japan's wartime crimes and curbing the ambitions of right-wing forces can the region escape the shadow of militarism and achieve lasting peace and development."
Regarding the newly recovered materials, an official with the Shanghai Songhu Memorial Hall for the War of Resistance against Japanese Aggression told the Global Times that experts will first authenticate the archives before their formal handover. Zeng Jin, deputy director of the Chongqing Municipal Archives, said the materials would complement existing collections and provide new evidence for research through third-party perspectives from foreign governments and societies.
。 今年6月，俞之奡正式组建小米Core团队（C-O-R-E，意为“核心”），主攻前沿材料和AI for Materials两个方向。此前，小米已通过泰坦合金和超强钢两个项目验证了AI辅助材料研发的可行性。 雷军也曾公开提及，“AI技术在小米的新材料研发中发挥着重要作用。在汽车大压铸材料的研发中，AI帮助他们从千万种配方中筛选出最佳组合，大大提高了研发效率。
二 | ” 俞之奡的履历恰好横跨AI+材料链条的两端：斯坦福博士期间，他做的是“拿瓶瓶罐罐”的材料湿实验；2020年，在GPT尚未问世、学界还在用小型机器学习模型计算材料参数时，他便开始接触AI for Science。 博士毕业后，他成为字节跳动AI4S早期核心成员，负责AI+材料和锂电材料等方面的研发，是国内最早将AI应用于储能材料并推动落地的技术骨干之一；此后他在美国创业四年，今年回国加入小米并组建团队。 团队定位非常明确：前沿材料是长期技术和战略布局，不求短期落地；AI for Materials是方法论，用AI加速材料从实验室到产品的全流程。为什么是小米？ 俞之奡的回答指向一个判断：材料科学最终的评价标准是“能不能用”——药物的唯一标准是能否治病，材料的唯一标准是性能是否提升、成本是否降低、量产是否可行。小米拥有从消费电子到汽车到家电的“人车家”全生态产品矩阵，这意味着足够多的场景去容纳新材料的尝试。用他的话说，“这个生态位或许是最合适的” 但行业仍有一个核心卡点没有解决。AI Agent可以在几秒内告诉你该做什么实验，但实验本身——真正做出看得见摸得着的材料——大部分仍靠人力在实验室完成。俞之奡称之为“闭环问题”：前面假设生成和文献搜索以及后面的审阅、校对、循环等已经做得很好，中间执行环节才是真正的决胜点。自动化和机器人的进展正在加速这一环节，但还需要时间。 另一个值得注意的信号是，俞之奡的加入意味着小米正在把AI材料研发从一个项目的成功，推向系统性能力的建设。
三 | 从金属到高分子、玻璃面板、电池，小米的意图是让“AI+材料”的范式覆盖整个生态链。
四 | “材料的性能要求没有上限，大家希望六边形战士的每一边都超过一百分。 以下为蓝鲸科技等媒体与俞之奡对话实录： 小米为何需要一个“材料连接者” 提问：为什么加入小米？选择小米的考量是什么？ 俞之奡：可以简单回溯一下。我大概2020年进入AI for Science这个领域，那时候还在读博，大家对AI for Science都没有概念，觉得材料计算就已经是所谓的AI for Science。
五 | GPT也不知道在哪，但大家就在探索怎么把一些很小的模型——不是现在的大模型，是专有的模型——一步一步应用起来，用一些很小的machine learning或者deep learning模型，那时候都已经做到头了。 从2020到2022年，这个方向越来越多的发展， 2022年末ChatGPT出来之后，以及2024年诺贝尔物理学奖和化学奖给了偏AI和AI for Science这两个方向，让这个领域有了非常大的爆发。对我来说，我本身的背景从本科到博士都是做化学材料实验科学的，真的是在实验室里拿着瓶瓶罐罐做材料。我非常清楚，无论是AI加材料还是AI加医药，最终的目的还是拿到那个真正能提升大家生活质量的材料或药物。 大家对药物来讲，这个药能不能治病是唯一的评价标准；对材料来讲，这个性能比起上一代有没有提升、价格有没有降低、量产性有没有提高、对整体产品有没有性能提升，这些是非常明确的。在这样的情况下，加入小米这样的企业生态位是非常正确的一件事情。
六 | 小米有那么多的人车家产品生态，从消费电子到家电可穿戴到车，整个都覆盖的情况下，我们有足够的场景去容纳新材料的尝试，真正做出有提升的东西。
七 | 这个生态位反而是最正确的。 提问：小米在AI加材料这个赛道的独特生态位是什么？ 俞之奡：我觉得小米非常独特的一个点在于，我们本身的基因——雷总也说了未来战略是人车家全生态——我们对产品一直以来就是最核心的一个基因。所以在落地方面，无论是人车家生态里面哪一个产品，我们都有非常大的力度。 在这个情况下天然就有场景，这是一个非常自然的事情。
八 | 我们有MiMo大模型的投入，有Agent的投入，然后有最适合的场景，整个生态链体系下的场景。我们希望把中间这一环接通，利用模型能力、Agent的能力，把更新的材料做出来，进入市场去服务客户。希望把模型和真正的产品做一个更好的连接，让更前沿的、性能更好的材料进入我们的产品。
九 | 提问：小米在泰坦合金、超强钢等项目上都用了AI，AI角色经历了哪些演变？ 俞之奡：泰坦合金是一个非常好的例子，就是珠玉在前。雷总在直播里也讲到。一方面AI本身这个范式在迭代——Agent、AI for Materials企业越来越多快速的迭代，把AI推到了一个之前不可想象的速度。 另一方面就是材料配方、工艺、生产制造的各种参数性能的综合优化。AI能够不停地加速这个过程，把更新的、更优化的材料提出来，我们去制造出来服务客户，进入产品让大家真正体会到产品质量和性能的提升。 提问：在之前已有的探索基础上，小米此时成立材料Core团队意味着什么？ 俞之奡：我更多的会把它形容成不是转变，而是一个进化。
十 | 咱们已经有AI+金属——无论超强钢还是泰坦合金——两个非常成功的例子。可以想象小米生态体系里肯定不止金属这一种材料，大家自然而然会去想怎么做到更多的材料。 尤其我们有消费电子产品、有车、有家电以及整个生态链里的所有产品。从玻璃到高分子到金属到电池会是一个非常大的、泛的材料领域。我们天然的想法就是能不能去泛化，把各种材料都涵盖进来，让AI加金属这个成功案例泛化到整个生态里，让产品整体有一个非常好的提升。 之前的话，AI的团队和业务的团队之间有一个协作，我来是希望能把这个方向打开以及泛化，更好地串联以及用到更多的材料里面。
十一 | 在我们整个生态链体系下的产品里面，在各个点都能有一些新的材料、更加前沿的材料，无论是从性能、成本还是制造的角度，各方面都能有更好的下一代材料用进产品里，让大家真正体验到产品性能的提升。 真正的决胜点在实验室，不在算法 提问：您认为AI加材料这个方向，目前需要攻克的难题是什么？ 俞之奡：这是整个领域目前的共识。
十二 | 大家看到尤其是Agent这一波从年初到现在爆发式增长，进入科学领域，无论中国还是美国比较头部的企业也都在布局。 整个工作流的搭建大家已经有了一些思路——怎么把Agent更好地用好，去提出假设、验证假设、搜索文献，到后面像实验室一样把出来的东西再去进一步优化迭代。但从材料的角度来讲，还是一个所谓闭环的问题——实验是真正做出看得见摸得着的材料，这一步现在是真正的决胜点。 Agent可能瞬间就能告诉你该做什么事情，但执行的时候，大部分还是靠人力在实验室做、在工厂做生产制造，这是一个瓶颈。怎么样把这个闭环快速打通、瓶颈快速解决，就能把AI for Materials整个循环大幅加速。现在已经有些希望了，大家都在不停地加速，比如自动化，包括机器人和具身大脑现在越来越泛化、能做的事情越来越多。 提问：AI相比传统材料研发方式，有哪些传统方式替代不了的？ 俞之奡：我们平时作为材料人，假如下游真的要做实验，大家可能会去想一些方向，但毕竟没有办法很快收敛到正确的方向，或者说要试的东西非常宽泛，会消耗很多时间和精力。AI能帮助我们快速收敛到一个正确的方向，然后进行快速的迭代，产生更好的配方。 提问：AI材料研发在周期、成本、良率这些维度上能带来多大的提升？ 俞之奡：目前来讲，AI for Materials大家还是把它作为Co-Scientist或者叫副驾驶，作为辅助角色帮助下游材料专家进行快速迭代，或者帮助我们想到一些人可能会忽略的东西。 整个大的图景里即使是最顶尖的科学家也，AI能够很好地帮我们查漏补缺，把整个图景画全。这个对下游真正的研发非常重要——一旦你漏掉的那个点正好是最好的那个点，就会花更长的时间才能发现这样的材料。
十三 | 成本、良率这些其实是综合优化的过程。人很难同时兼顾所有维度的性能指标，大家都希望一个材料是全维度优秀，也就是“六边形战士”。在这个过程当中AI能很好地帮我们补齐那些短板，尽量做出全维度均衡的材料——包括成本，包括量产时候的工艺、制程、制造，都能做统一的优化，而不是只限于某一个点说某一个性能很好但别的不在乎。
十四 | 提问：目前AI在材料研发哪个环节效果最好？哪些环节几乎没用到？ 俞之奡：基本上在今年年初之前，大家还是一些专有领域的小模型——金属有一些，别的方向各自有一些。
十五 | 但今年年初Agent这一波起来之后，范式发生了相当大的变化，大家发现基座大模型加上Agent Workflow——我们现在叫Harness——能把整个工作流真的搭起来。
十六 | 你告诉它一个目标，比如想让这个金属的强度提升多少，它可能自动去搜索文献、提出假设、想该做哪些实验、做哪些计划。然后下一步就是卡点——得去实施，真的做出来。做出来之后，有了配方的输入和性能的输出，再返回到前面，看到数据之后想下一轮该怎么去迭代材料、怎么进一步提升性能。 在这个范式里，前面几步——搜索文献、提出假设——已经做得非常好了，尤其是Agent和基座大模型的能力越来越强。但中间那个卡点——做实验和做真正的产品性能这一步，现在很多公司在做自动化的实验室，甚至自动化的中试线，怎么样去提速这一步，还需要一些时间。
十七 | 提问：现在基座大模型是通用的还是各材料领域有专用模型？小米会做材料通用大模型吗？ 俞之奡：大概是2026年年初之前，大家在做一些所谓的专有模型。
十八 | 目前大家都是以基座大模型加Agent的方式去做。但你提到的针对各种材料，甚至更加泛化的——不止材料，包括生物医药等科学领域的大模型，行业内也有一些人在探索。但我们可以随着这个领域的发展去看，到底哪一条范式或者哪一条路线最后收敛下来，这个确实需要一些时间。 期待“AI+材料ChatGPT时刻”到来 提问：超强钢接下来还会继续升级吗？AI加材料还有哪些更可为的作用？ 俞之奡：材料的性能要求其实是没有上限的，大家肯定希望越来越好，甚至不仅是六边形战士，六边都要一百分、超过一百分那样的产品。除了金属之外，还有高分子、有机化学这些领域，有些材料可能是人类历史上从未想象过的，远远超出简单材料的范畴。模型本身也在持续优化，配方也会有新的迭代，比如加入不同的元素成分，搜索范围和配方组合空间会更大。 提问：小米在AI加材料领域，看到了哪些可以通过新材料解决的具体问题？ 俞之奡：从产品布局的角度看，除了金属材料之外，高分子、玻璃面板和电池都是重要的方向。 电池在消费电子产品和汽车应用中都是核心要素。比如动力电池方面大家最关注的是安全性，以及电车能跑多远，充电速度、成本控制和量产能力也是关键考量。大家都希望在不改变充电速度和购车成本的前提下获得更长的续航体验。这些需求都会推动我们去探索新的电池材料方案。 提问：AI for Materials的ChatGPT时刻何时到来？ 俞之奡：这个问题是我今天的报告最后一页PPT标题。首先，要看怎么定义这个ChatGPT时刻。有一种观点认为它可能已经到来了——2026年Agent这一波起来，有基座大模型，有很好的Agent工作流，这一套已经能够帮助科学家或者下游真正做产品的人快速加速迭代，这件事情本身可能就被定义成ChatGPT时刻。
十九 | 如果说要再定义一个更进一步的时刻，我觉得有点像ChatGPT时刻之前，大家也不知道ChatGPT时刻是什么定义，可能还需要一些时间的发展去看终局是什么。
Current article:http://www.minshixuanhabing.bond/list_m2e/h9w.html
Published on:03:28:52
