基于音频「半导体的未来：从新材料到新芯片，AI能否推动自身进化？｜硅谷101年度线下大会」完整转写整理（音频1784秒，约29.7分钟）。本场为英文Keynote（开头一段为节目中文旁白），与天文学的AI革命篇同属硅谷101 Alignment 2026年度线下大会——本场讲者开场提到「之前关于AI与金融的panel」，中途又说「昨天有一场关于望远镜的精彩演讲」，所指正是孙凤梧那一场天文演讲。讲者自述、并经Duke大学电子与计算机工程系官网核实为Haozhe “Harry” Wang：Duke ECE（Pierre R. Lamond电子与计算机工程系）助理教授（2023年8月入职建组），上海交通大学本科、MIT博士（2021年）、加州理工学院博士后。文中标注「经核实」的事实共五组：讲者身份、ATOMIC框架（ACS Nano 2025论文）、与Google DeepMind合作的co-scientist论文（arXiv 2608.26701，2026-08-27）、「螃蟹化」（carcinization，Borradaile 1916年提出）、石墨烯获2010年诺贝尔物理学奖；其余事实与数字均为讲者或节目旁白口述口径，未做外部核验。转写疑似错误文末统一标注。

先说结论

这场演讲回答了一个所有AI从业者都会脱口而出、却很少往下追问的问题：AI的算力从哪来，往哪去？Haozhe “Harry” Wang的答案分三层。第一层是物理：硅基晶体管缩到5纳米以下时电子迁移率会崩塌，行业宣传的「2纳米节点」「1纳米节点」其实是纳米片堆叠后的名义数字——继续缩小，得换沟道材料，候选就是石墨烯、二硫化钼（MoS₂）这类只有一个分子层厚（约0.65纳米）的二维半导体，台积电路线图已把它排在纳米片、CFET之后。第二层是产业：新材料十年前就在学术界的单器件测试中赢了硅，却始终进不了手机，原因讲者只用一个词概括——「良率，句号」，而良率管理至今仍靠人眼看缺陷。第三层是他的实验室正在做的事：用AI吃掉「看材料、长材料、造芯片」全链条——ATOMIC自主显微镜零训练把材料识别做到99.7%（ACS Nano 2025论文，经核实）；与Google DeepMind合作的co-scientist系统读文献、出配方、直接控制CVD设备，一次长成单层MoS₂（arXiv 2026-08，经核实）；以及未发表的下一步：在65.6%良率、被产业伙伴称作「废料」的材料上，让AI挑出好区域造出数千晶体管的芯片，良率提升约30个百分点。终点是他给这个闭环起的名字：autogenic intelligence（自生智能）——AI设计并制造运行自己的芯片，「心智与大脑从不分离」，并以1916年动物学的「螃蟹化」现象向听众发问：AI加芯片的演化，会发散、收敛，还是也长成一只「螃蟹」？

一、水龙头、漏电与「名义制程」：硅为什么撑不下去了

演讲前半段是一场诚实的「Silicon 101」（讲者特意玩了个梗：今天是硅谷101，那我先讲「硅101」）。他把晶体管比作家里的水龙头：栅极是开关，关上它，电子就无法从源极流向漏极；打开，电流通过，这就是0和1。「所有大语言模型和现代计算再复杂，最底层一切都是0和1。」scaling law也不是AI的发明——硅的世界里它已运行五十年：单颗芯片塞进越多晶体管就越强，讲者举的例子是英伟达Blackwell B300 AI处理器，按他的口径单颗芯片有约2000亿个晶体管（公开口径Blackwell为约2080亿、双芯粒封装，见文末标注）。戈登·摩尔的摩尔定律本质上就是「把每个元件做得尽可能小，就能在同样面积塞进更多」，这正是「硅谷」名字的来历。

但缩小有两条物理账。第一条是缩小路线本身：五十年里晶体管从微米级缩到纳米级，结构从平面（planar）到鳍式（FinFET）再到纳米片（nanosheet）——本质是堆叠。堆叠带来新问题：界面处发热、散热吃紧。第二条是失效模式：沟道太厚时，关掉「水龙头」电子仍会从底下渗过去，这就是漏电（leakage），白白耗电并发热——所以沟道越薄越好。而硅的死穴在这里：讲者强调，硅缩到5纳米以下，电子迁移率（电子流动的能力）会明显变慢，开态电流出不来；二维半导体——石墨烯（2010年诺贝尔物理学奖材料，经核实）、MoS₂（一个钼原子夹两个硫原子，单层厚度约0.65纳米）——即使薄到5纳米以下仍能保持电子迁移率。学术界的对比测试过去十年反复显示：单层MoS₂器件的性能优于同尺度硅器件（讲者口径）。

那为什么新闻里已经有「2纳米节点」？讲者拆穿了行业命名术：现在的制程节点号都是「名义节点」（nominal nodes）。纳米片结构是把多层硅堆起来，比如总厚约7纳米堆3层，均摊每层约2.3纳米，于是营销上叫「2纳米」——「7除以3，大约2纳米」。他用英特尔做了个类比：从奔腾单核到今天12核、14核，不是因为单核做强了，而是因为单个元素缩不动了，行业选择「堆」。结论很直白：硅在物理上做不进二维材料那个尺度，这是他想让听众带走的第一个判断。而产业界自己也认：讲者展示了台积电的路线图——从纳米片到CFET（互补场效应晶体管，把n型和p型器件上下堆叠），再到「very near future」的二维过渡金属硫族化合物（TMD，MoS₂即其代表）。新材料不是学术幻想，它已经在最保守的路线图上。

这一段还藏着一个可爱的插曲：二维材料不止能做芯片。讲者实验室把它印在软基底上，材料见光会像尺蠖一样定向爬行——他顺嘴开个玩笑：这就像博士生，导师找谈话就弯一下腰，不谈就躺平，「但我们总在动」。Duke工学院的公开报道也印证这个实验室确实在做MXene二维材料软机器人（经核实）。

二、一个词决定命运：良率

如果材料这么好，为什么还没进你的手机？讲者的回答斩钉截铁：「就一个词：良率。句号（yield, period）。」台积电之所以是台积电，是因为它是全世界最好的良率管理者；英特尔、三星也很好。一颗芯片动辄上千亿晶体管，你怎么保证每一颗都工作？这对硅已经是极致工程：硅用物理方法提纯到「10个9、12个9」（99.99999999%以上）的纯度——讲者的说法是，在12英寸晶圆上逐原子检查，只能找到一两个层错（stacking fault）。而二维半导体靠化学气相沉积（CVD）生长：气相前驱体把半导体「长」在衬底上，单层性能极好，但完美度远达不到硅的水准——如果材料本身不如硅完美，现有工业流程就得改写。良率管理对二维材料「需要一套新的逻辑」，而这个逻辑至今仍是行业里最依赖人类经验的环节：看材料、看缺陷、判断好坏，主要还是人眼。

人眼看不过来，恰恰是讲者两年前的出发点。这里还有一层学术界的痛点：可复现性。不同实验室的炉子、管路、几何都不一样，「我的实验室有这个炉子、那根管，但论文里的 setup 是另一套」，所以学术结果出了名地难重复——配方换台设备就失灵，新材料的工艺知识极度依赖隐性经验。这两件事——人眼质检不可扩展、工艺知识不可迁移——构成了后面所有AI工作的靶子。

三、AI的三级跳：看材料、长材料、造芯片

**第一跳：让AI看材料。**大约两年前（讲者口径，当时「智能体」概念都还很早，他们只敢叫「工作流」），实验室开始用大语言模型直接控制显微镜：LLM当大脑指挥设备，图像分割交给Meta的Segment Anything Model，再用无监督聚类把区域分类——「别忘了，机器学习是AI的开始」。成果是ATOMIC框架（Autonomous Technology for Optical Microscopy & Intelligent Characterization，经ACS Nano 2025论文核实）：零训练、不需要数据中心，纯提示工程驱动，就能判断视野里是单层、多层、裸硅还是什么都没有。关键数字：单层识别准确率99.7%，与人类专家相当（论文口径，经核实）；讲者补充延迟约6秒、正在压低，而且某些任务AI比人强——人眼凭直觉只圈出整片 flakes，模型逐像素分析，还能标出人眼忽略的「畴界」（domain boundaries），那本身就是缺陷。现场演示还出了个小事故：屏幕上的系统卡死（「I think we’re frozen」），讲者笑笑继续——自主显微这条路的日常。

**第二跳：让AI长材料。**讲者判断行业正从「AI for Science」转向「AI for Engineering」：AI已经证明了假说能力——Google的co-scientist（2025年发布）、DeepMind的AlphaEvolve都能产出新想法——短板在物理世界的验证，正如前一天AI4Science panel反复说的。他的实验室和Google DeepMind合作（论文arXiv 2608.26701，2026-08-27挂出，讲者称「四周前发布」，经核实）：把co-scientist接进自家半自动CVD设备，从读文献、提假说、给执行计划，到生成协议、控制软件面板，全链路跑通。目前人还在环里——称前驱体、装管、封口仍靠手，讲者畅Soon用机械臂替掉这一步。效果（经论文核实）：进化搜索模式优化的配方一次长出大尺寸单层MoS₂；用Gemini 3 Deep Think做快速推理时，几分钟出配方、约1小时完成实验，MoS₂、MoSe₂、WS₂三种单层半导体全部一次长成——后两种材料他的实验室此前零经验。针对MXene新前驱体路线，论文产出了与Ti₃C₂Tₓ晶格具有关键结构相似性的层状材料（论文同时注明尚需进一步实验确认原子结构——讲者现场说得比论文乐观：他认为理论上预测过但从未长出的材料，这条路线够得着）。

更贴近学术日常的是「定制化配方」：把本实验室的几何约束（炉子配置、可用化学品、衬底）交给AI，让它对着文献生成适配自家设备的配方。讲者讲了个故事：他催学生长一种实验室从没长过的新材料，学生说「王博士，我得先看几周文献」；后来索性让大模型直接出配方——一次成。「我开玩笑说，以后你连文献都不用看了，照模型说的做就行。」这省下的是几个月。

**第三跳：让AI造芯片（未发表工作）。**显微镜看到一片故意选的「不理想」材料——黑点密布全是缺陷，「如果交给产业伙伴，他们会说这是垃圾」；材料本身良率只有65.6%。但让AI接管制造：它会在缺陷图上挑出好区域，传统流程追求数十亿个整齐划一的晶体管，AI流程则是「每个晶体管都不一样」，用良率图反推器件布局。讲者称已把AI连进洁净室制造设备，造出了数千晶体管级别的芯片——「不是十亿级，但我们认为这是很好的开始」，并在低良率材料上得到高良率晶体管阵列，良率提升约30个百分点（讲者口径，未发表，统计口径未说明）。「我认为这会在未来改变、颠覆整个半导体制造，这是我的信念。」

四、AI4Engineering、初创公司与钱的方向

基于这些结果，讲者「大约一个月前」创办了初创公司（讲者口径），方向三个候选：良率智能（yield intelligence）、实验室到工厂（lab-to-fab）方案、或者为二维半导体建一条小型中试线（pilot line）。他的商业逻辑藏在一句玩笑里：「你们所有人的钱最后都流向英伟达，而英伟达最后得付给台积电——怎么把芯片造出来，才是最花钱的环节。」与其在数据中心侧卷，不如在制造侧给出一条更便宜的替代路线。

五、「自生智能」与螃蟹之问

演讲的最后五分钟，讲者把叙事拔到了AI行业最热的话题上：RSI（递归自我改进）。「大家都在谈RSI，但缺的那一环是：怎么让AI制造它自己的大脑。」他画了一个飞轮：AI设计芯片→AI制造芯片→芯片训练AI（数据在中间）→更高效的飞轮。他给的类比是生物学的：心智与大脑从不分离，而今天的模型和芯片是「分开的」——用模型造芯片、用芯片跑模型，两者融合，「这是一种新的生命形态，你可以叫它硅基生命」。顺着最近「SI（超级智能）比AI更受宠」的命名风潮，他给制造侧闭环起了个名字：autogenic intelligence（自生智能）——「让它回到AI，闭合这个环」。

那这种硅基生命会演化成什么样？讲者借来一个一百多年前的动物学概念：1916年动物学家Lancelot Alexander Borradaile提出的「螃蟹化」（carcinization，经核实）——甲壳动物不管出身如何，独立演化多次都会长成螃蟹的样子，「大自然多次尝试演化出一只螃蟹」。鱼过几百万年成了帝王蟹，乌龟（他笑着说）成了蓝蟹。「所以我的问题是：AI加芯片，会发散，会收敛，还是也会长出一个『螃蟹形态』的硅基生命？」散场前的最后一个梗：「下次你吃螃蟹的时候想一想——也许你正在吃一种非常有智慧的动物。」

把整场演讲压成几条可带走的链条：

  • 良率瓶颈链条：新材料单点性能超越硅（十年学术验证）→ 大面积CVD生长缺陷密布（65.6%良率级）→ 人眼质检不可扩展、良率管理是行业最后的人类堡垒 → 零训练AI质检（99.7%）+ AI按良率图挑区布局器件 → 低良率材料上长出高良率阵列（约+30个百分点，讲者口径）→ 二维半导体产业化时间表可能被提前。启发：当一项技术的瓶颈从「性能」转移到「良率/质检」，行业杠杆就从研发端移到测量与自动化端——识别这个转折点，比追性能参数更值钱。
  • 实验闭环链条：学术工艺不可复现（隐性设备知识）→ 把几何约束显式交给co-scientist → 读文献+出配方+控设备，分钟级配方、一次长成 → 材料研发周期从「数周看文献」到「约1小时实验」→ 「AI for Science」转向「AI for Engineering」。启发：隐性know-how的显式化与仪器接口，是AI进入物理世界真正的中间件；论文自己也承认跨实验室可复现性尚待验证——这是判断此类工作成色的关键缺口。
  • 自举飞轮链条：AI设计芯片→AI制造芯片→芯片训练AI → 模型与硬件从分离走向「心智/大脑」式融合 → autogenic intelligence，制造侧自举成为RSI叙事之外第二条递归自我改进路径 → 钱的结构佐证：所有AI开支终点是英伟达→台积电，制造侧 cheaper alternative 空间巨大。启发：算法侧RSI叙事火热的同时，盯住制造侧闭环的每一环（EDA、fab自动化、先进封装、2D中试线），那里可能是更「物理」的自举路径。
  • 命名脱钩链条：制程节点号与物理尺寸早已脱钩（纳米片堆叠的名义制程数学）→ 营销数字掩盖了单元素缩不动的物理极限 → 真正的下一步是换沟道材料（2D TMD，已上台积电路线图）。启发：读行业基准数字先看定义——「2纳米」不是2纳米。

接下来值得盯的观察指标：台积电路线图上2D TMD沟道何时进入量产节点（A16/ A14之后的实际节奏）；这家刚满月的初创公司选哪条产品线落地（良率智能、lab-to-fab还是中试线）；co-scientist类系统跨实验室配方的可复现性验证（论文明确留作future work）；二维材料芯片从数千晶体管向百万级阵列的爬坡；以及英伟达—台积电资本开支周期对「制造侧替代路线」窗口的影响。当AI开始制造自己的大脑，硅谷这个姓氏，可能比任何时候都更名副其实。

ASR转写校正与存疑说明

本场演讲为英文（Qwen3-ASR整体识别质量良好），开场一段为节目中文旁白。以下第1—13项为已确认校正（按上下文、讲者自述或外部来源核实），已同步应用到文末附录；第14—20项存疑或按讲者口径保留原文。逐项如下：

  1. 「nano sheet」→ nanosheet（纳米片，全篇统一）。
  2. 「Nvidia Blackwell three hundred」→ NVIDIA Blackwell B300（Blackwell Ultra系列AI处理器）。
  3. 「the MOS2」→ MoS₂；「two dimensional transition metal that chalcogenides」→ two-dimensional transition metal dichalcogenides（二维过渡金属硫族化合物，TMD）。
  4. 「you stack the silicones」→ you stack the silicon（指纳米片结构堆叠硅层）。
  5. 「about point six five sixty five nanometer」→ about 0.65 nanometer（单层MoS₂厚度，去掉重复口误）。
  6. 「meta segment anything model」→ Meta Segment Anything Model（SAM，分割一切模型）。
  7. 「supervised clustering」→ unsupervised clustering（无监督聚类；经ATOMIC论文原文核实为unsupervised）。
  8. 「we call it atomic」→ ATOMIC（Autonomous Technology for Optical Microscopy & Intelligent Characterization；经ACS Nano 2025论文核实，单层识别准确率99.7%）。
  9. 「Google’s co scientists I all actually published」→ Google co-scientist actually published（Google AI co-scientist，2025年发布的多智能体科研系统）。
  10. 「alpha for alpha evolve」→ AlphaEvolve（Google DeepMind进化编码智能体）。
  11. 「large angle model」→ large language model（大语言模型，[1320s-1440s]段）。
  12. 「sixty five point six sixty five point six percent」→ 65.6%（材料良率，去掉重复口误）。
  13. 「constantization」→ carcinization（螃蟹化；Lancelot Alexander Borradaile于1916年提出，经Wikipedia核实，全篇2处）。
  14. 「Valley 101」：讲者将「Silicon Valley 101（硅谷101）」口头缩略并顺势引出「silicon 101」双关，保留原文。
  15. 「two hundred billions switches or transistors in a single chip」：讲者口径约2000亿晶体管；公开口径Blackwell单封装约2080亿晶体管（双芯粒），按讲者口述保留。
  16. 「Intel’s original」：疑为「Intel’s history/era」（指英特尔历代处理器演进），未确认，保留原文。
  17. 「which material won the Nobel physics prize Nobel Prize in two thousand ten」：按上下文指石墨烯获2010年诺贝尔物理学奖（经核实），口语结构保留原文。
  18. 「autogenic intelligence」：讲者自创术语（可译「自生智能」），保留原文。
  19. 螃蟹化举例（鱼→帝王蟹、龟→蓝蟹）：讲者玩笑式类比，非严格演化关系，保留原文。
  20. 「sixty five point six percent … thirty percent around thirty percent increase」：良率提升幅度讲者未说明统计口径（30%还是30个百分点），正文按「约30个百分点」表述并标注讲者口径。

转写原文（修正版）

以下为整份转写稿全文，已按上方校正表中已确认项（第1—13条）修正（nanosheet、Blackwell B300、MoS2/TMD、硅层堆叠、0.65纳米、Meta SAM、无监督聚类、ATOMIC、Google co-scientist、AlphaEvolve、大语言模型、65.6%、carcinization），存疑项保留原文；[起-止] 时间戳保留原样。原转写由 Qwen3-ASR（Intel Arc GPU，120秒分段）生成。

[0s - 120s] 过去几十年，芯片性能的提升很大程度上依赖于晶体管不断缩小。但随着传统硅基芯片逐渐逼近物理极限，半导体行业正在寻找下一代材料和制造技术。而有意思的是，AI不仅是这场技术变革的需求推动者，也可能成为下一代芯片的设计者。那么，当AI开始参与制造下一代AI芯片，智能的发展会走向什么方向？接下来，让我们回到Alignment 2026大会现场，观看这场Keynote的完整演讲。So, um, in the next maybe 30 minutes, I want to introduce, like, uh, maybe how chips are built from raw materials and what’s the future of the of the silicon chip after or beyond silicon. And then after that, I will introduce our in so so our lab. We try to use AI to boost the silicon or or semiconductor discovery and manufacturing. So again, introduce myself a little bit. I’m Harry Wang. I’m currently assistant professor at Duke ECE department, and I finished my undergraduate in Shanghai Jiao Tong University, and then a PhD at MIT. I spent a few years at Caltech postdoc, and I started my lab in Duke about three years ago. So I want to start with today is Valley 101. Okay, I want to start with something like like I will call it a silicon 101, and and basically give you some basic knowledge about how how the chips actually working. Okay. So firstly, I want to say I really like this slide. Actually corresponds to the previous panel, like like uh, previously I will say like AI definitely needs chips to run it and train inference everything. But now in last two three years, you notice like silicon or the chips related industry booms in the stocks at at least right. So in that case, like I will say like we also see some startups are trying to use AI or to

[120s - 240s] To to boost the design of chips and everything like that. So I will say this can become a loop. The AI needs chips and chips needs AI. But what is what are the chips and what are the basic elements of chips? So basically everything is binary. So so even though like all the large language models and the computing become very complex in a modern era, but still everything goes to zero and one at the very bottom. So if you see like how does chips work? The basic element is called transistor, which is more like like like a faucet. Okay. So if we turn off the gate, so the electrons cannot flow from the source to the drain. So if we turn it on and faucet on, you can see here the electrons can flow through the source to drain. And this is the very basic element. It turn from zero to one, one to zero, and realize everything now based on the chips and based on the AI. All right. So so again like we we talk about the scaling law, but the scaling law is already there in in in silicon or in the chips. Okay. So in semiconductor, so we do believe like like if we stack more transistor in a single chip, it’s just be more powerful and has higher performance. Okay. Here shows like from a single transistor, if we accumulate to billions of transistors, it can become a chip and become very functional. And showing here if we see this example like if we we see the Nvidia Blackwell B300 AI processor, it actually has two hundred billions switches or transistors in a single chip. Okay. So so in that case like basically if you have you want to have better chips, you need to have more transistors in a single device. So in that case, that comes a very natural question like how we can fill in so many transistor in a

[240s - 360s] Single chip, so so basically this is basically the the the the the beginning of the success of the Silicon Valley, right? We are the Silicon Valley, and that’s the name come from. So basically, so so Gordon Moore basically realized the Moore’s law, actually saying that if you have smaller transistor, each element as small as possible, you can fill in more transistor in a single in a certain footprint. Okay, so in that case, here shows just a look back of how transistor develop in last fifty years. So originally from pretty large transistor, and then shrink it down to from a micron scale to nano scale, and then from kind of like planary to the the we have FinFET, and then we have like nanosheet structure. Basically, we stack everything together. So however, recently there are some problems. So if you so so currently we have FinFET, and then we stack it, but after stacking it, there will be other problems such as the the thermal problems, the heat generated at these interfaces generate some problems. Okay, so in that case, like so I want to introduce a little bit a common common failure mode of the transistor. So transistor when it shrink it into small scale, so sometimes there’s a problem as it also happens in your in the faucet in your home. Okay, so if if it’s very thick, so basically the the electrons sometimes not just flowing. If you turn it off, it’s not flowing directly from here, but also can flow from the underneath. So we call it leakage. So these leakage actually that’s why like consume a lot of power and waste a lot of power and generate the heat. Okay, so in that case, like if you use a thick material like very thick like this, sometimes like like like like like the the electrons can go underneath it. In that case, you have to make the material as thin as possible.

[360s - 480s] So in that case, if it’s staying, then we can close the faucet. In the off state, no electrons can flow here. In the on state, the electrons can flow well. Okay. So in that case, like it’s actually, I try to give you some common maybe concept in in silicon one one or semiconductor one one. So smaller is better for each element, and thinner is better for each element or semiconductor materials. So we already see this trend, but I want to convince you that currently the silicon has big problem. Okay. I I really value like in the previous panel we we talk about the finance maybe some problem or is still going right. But I want to convince you like the silicon currently has some problem. The reason is if we we we basically cannot shrink silicon infinitely. So if silicon goes from from like pretty large scale to shrink it to below five nanometer, so electrons become the the mobility and speed of electron become slow. Or finally like like you can see it’s not flowing well. In that case, in the on state, so no electrons can flow if you basically have really small scale of silicon. So in that case, like as a scientist, so we think it’s possible to have next generation semiconductor. We call it two dimensional semiconductor represented by graphene, molybdenum disulfide, like which material won won the Nobel physics prize Nobel Prize in in two two thousand ten, right? So so basically so so but comparing with the two dimensional semiconductor, so two dimensional semiconductor can keep its electron mobility or the capability of flowing electrons even it’s below five nanometers. Okay. So so you you might be feel confused. You you you look at the news.

[480s - 600s] Actually, the semiconductor industry always tell you, oh, we now now have a two nanometer node. In the future, we have one nanometer node. Why it’s not like what I just said is five nanometer is a limit, right? So I want to to say like every like the node numbers showing here is is currently is all nominal nodes. That basically means like if we check like like the the the the the structure and the history of semiconductor and transistors. So currently we are using the nanosheet structure. Nanosheet structure, you stack the silicones. So in that case, even though you have seven nanometer, you have three layers, then you have seven over three. That’s around two nanometer. Okay. So similar thing, we actually already seen in in the Intel’s original. We have Pentium, Pentium one, Pentium two, Pentium three, and now we have like twelve cores or like fourteen cores. So because it’s it’s hard to shrink a single element, so so basically industry decided to stack more. Okay. So in that case, like I want to convince you like like silicon still fundamentally has a problem. You can make it. You cannot make it extremely small like like other two dimensional materials. So but this materials, as I mentioned, it only has one molecular layer. Here shows two layers. So you have one molecular uh uh molybdenum uh atom and two sulfur atoms on, and in total this material has only the thickness below one nanometer is about point six five nanometer. So so so but still as a semiconductor, it has been showed in academia in last ten years, a single device built upon this material is much better than a single device of using the silicon to build it. Okay. So if if we see like uh when when this material enter your phone, into your computer, become really electronics in product. So this shows

[600s - 720s] The roadmap by TSMC, it basically shows like so. So now we are at between the nanosheet structure moving to the CFET as I just showed to you. But in the very near future, this this will go to the two dimensional transition metal dichalcogenides, or I just showed you as MoS2. Okay. So in that case, so so you might think like like this material just the material like I I tell you is very thin and can perform as a channel material of the transistor. So basically, so the problem is like how to build it efficiently. So even though we have some small flakes and we demonstrate single devices great, but we always need a lot of like as mentioned like hundreds of billions of transistors in a single chip. So by building this material, basically currently in the university lab and I I I I know there’s some industry partners are working on these similar things. We use a method called chemical vapor deposition. Have some two precursors. The precursor can carry with the gas and grow this extremely thin semiconductor on top of substrate. Okay. So the problem is like this one layer can perform really well as a channel in the in in the transistor, but it’s it’s not like silicon. Silicon is we use a physical way to build it, so have really high purity like 10 nines or 12 nines. That means it’s on a 12 inch wafer. If you check every atom, so only one or two stacking fault on on 12 inch wafer. However, this material still struggling about like gate is like perfect as silicon. So if it’s not as perfect as silicon, basically the current industrial flow should be changed. So in that case, like so so I want to also to mention that this material not just can be built a semiconductor. In my lab, we also develop something. We found.

[720s - 840s] If we print this material on top of some soft substrate, and this material kind of responds to to the light, you can see it’s like inchworm can just move in one direction. I’m joking with my friends. This is doing PhD. When the advisor talking with student, they just bend a little bit. If I don’t talk with them, just laying down, right? So, but but we are moving. We are always moving. So on one, you cannot always bend, right? So so so so so maybe the manager is laughing. So you know what’s going on. So so so so so then you are questioning like this material seems very magical, but why is still not is not in your phone, not in the daily life? The problem is like I will say that just one word is yield. Period. So TSMC is great because you can manage yield the best in the world. Okay, and Intel and Samsung also very good on yield management. That means like like any semiconductor industry, if you want to like like like get a share in the industry, so the yield management very important. That means you have to build a chip about hundreds of a billion transistor on one chip. But but but you how how do you make sure everyone is working, everything is working, or every transistor is working? So TSMC is doing great, but for this new two D semiconductor yield management needs some new logic, but still something like you need to solve. Okay, so basically I will say like in in the whole go back a little bit like to in the whole semiconductor industry, so we still rely on human to do the yield management and at least to see these materials, to see the defects, see the see see it’s good or see it’s not good. Okay, so in that case, then like about two years ago, so my lab we started to to look at like whether we can use AI to to see these materials or to see the defects of these.

[840s - 960s] Materials. So, so, so by by doing this, yesterday there was a really good talk about telescope, like to see the stars and everything. That goes to a macro scale of of seeing things. So in our lab, if we want to engineer in nano scale, we have to to see these small things as as as small as possible. So therefore, like we published paper last year, we call it ATOMIC. We use AI, basically large language model, to control a microscope and let the microscope to see these nano materials and characterize these meta nano materials and understand these material micro materials. So it’s it’s it’s the work is done two years ago. At that time, even agents as a concept was still early. So we call it a workflow, and basically we can use the large language model to control the microscope and to segment the the the the the the image captured, and we use a Meta Segment Anything Model, and then we can use some unsupervised clustering, which is a very basic machine learning. Don’t forget machine learning is is the beginning of AI. Okay, so so so so so so we can teach the AI to understand what materials are there. Is that clean? Is there defects, or is there something just nothing there? Okay, so in that case, like we we we can see it’s very interesting. This is not responding now. Maybe we should replace it to not silicon based thing. So so I think we’re frozen. But but but again, like like basically what we did is like we can let the microscope to do the task, research task by itself. So by doing this, basically it’s it’s it’s possibly to to we can see this is more like in the category of AI for science, or we can say like it’s actually uh uh like like in in in a more like a structure of like how to use AI to actually uh see.

[960s - 1080s] The manufacturing of the semiconductors. Okay, so and and most importantly, we notice it’s possible to so so so so at that time, like people still questioning is ChatGPT or or still have a lot of hallucination. Okay, this word recently not heard by by many of the research, but at that time, the hallucination is a problem. But we we validated that it’s possibly like to use AI to identify the material and also to finally classify the results by just seeing the material. What are single layers? What are multi layers? What are bare silicon or or what are the nothing over there? Okay, so so basically, we found it’s this model works really well and it’s working now. Yeah, so so this this model works really well. So we can see like comparing the human eyes, this the yield of the model can reach to ninety nine point seven. Even though the latency is still a problem, that human just look at it and and then you can get result, but we still need six seconds. So recently, we are we are lowering the latency is working pretty well, but again, like like like in some tasks, it’s even better than human. So if I show the left image showing here, basically I’m telling you like asking you to to label all the possible like like like I’m I’m asked to flakes. It sometimes the human will just label like this in the red. Okay, but the model cannot just see the the the whole flakes, but also identify some. We call it like domain boundaries. It’s a boundary of different domains, and these boundaries are actually a kind of defect. So so there’s not no tricks here, but because the human we just use intuition by our eyes, but but this like model they they have pixel by pixel analysis. So in some certain tasks, like I’m I try to convince you that AI can work definitely better than human. Okay, so we also try different kinds of like materials in semiconductor. For example.

[1080s - 1200s] Showing here like different kinds of, for example, graphene, which is Nobel Prize-winning material, and all kinds of 2D semiconductors. This model works really well, and most importantly, we even don’t need to train it. Can you believe it? We don’t need data center. We just we we can just build a agentic system and and and and do the prompt, and we can we can manage it to to harness it to to for our for our application. All right. So um so so so so then like still even though we teach AI to see this material, still like we are far from actually how to manufacture it, how to get it, right? So I will say like in last two years, AI for science is very important, and I will say like soon we’ll change to I will I’m thinking like maybe the AI for engineering will be very important. So in that case, like we think like for this novel semiconductors, it’s possible to use AI to speed up the discovery and speed up the manufacturing. Okay. So so actually we already see the AI is very good on hypothesizing ideas. So so Google’s co-scientist actually published last year and and and the the realize they are using some some some structure like like AlphaEvolve. So AI can give more ideas. Okay. So and also the problem is like as as mentioned yesterday in the AI for science panel, actual physical experiment and also the actual real world validation is really difficult. Okay. So so in that case like what we did this paper just released a lot uh four four weeks ago. So so my lab collaborated with Google DeepMind, and we try to use this co scientists structure and added into the physical world in physical experiments to do the end to end validation and discoveries. Okay. So basically the the idea is I want to introduce the work a little bit.

[1200s - 1320s] So the co-scientist now cannot just do the the the literature review. It can give ideas, and also can do the give the experimental or the execution plans. And also we connect our our instrument and a self-built experimental setup, and let AI to control it. And now it can do the experiment by itself. Okay. So in that case, like in in about one year ago, people are thinking in the future, Harry, you don’t need a PhD student. Okay. Then after I saw this work, I’m saying like maybe you don’t need a professor by itself because because AI can give ideas better than I. So why need a professor? And and also it can write paper by itself, right? So um, so maybe we should all work on work on data center. Yeah. So so basically this work can do the the the read literature, design experiment, validate end to end, and more importantly here this video shows like how it works. Is the co-scientist geometry can can can predict the precursor and also generate a protocol to actually execute it, and also like controlling some panel in the software. And human still in the loop right now. So so so we have to use hand to weigh the precursor, put it into the to the tube, and and and and also like seal it. Okay. But maybe soon like we can use a robot arm in that that hallway to replace this step. Ah, but but again like it’s generate a really good recipe and can even like in the paper we we we highly, truly believe like it’s possible to grow some new semiconductor never grown before. Okay. Even theoretical predicted, but but it’s very hard to grow. And for the conventional 2D semiconductors, as I showed you, MoS2 and everything. So it also shows really good results. We just let AI to control the instrument. It looks at the the literature.

[1320s - 1440s] And get the results, and most importantly, we found like I will say like in academia, many of the results are very hard to repeat because different labs have different geometry. So my lab have this furnace, have that tube, but but the published result has a different setup. So in that case, like we found, like AI can work really good on how to repeat the results. That means that if we have the geometry of our lab and give it to AI and let it look at the literature and try to see like can you have a customized recipe using my customer’s instruments, it works really well. So it’s it’s actually showing here like the the the in is always prompt and the task description and out is can be the directly get from the customer’s customized instruments. So basically, here shows a demo like I try to push my student to grow my my lab is growing this material for for a long time, but this new material I try to push my student to to can you grow it? Can you grow it? And my my student say, oh, Dr. Wang, I have to look at literature. It takes a few weeks or something. Then one day I just say, can you just try use large language model and try a recipe and let’s see whether it’s work. Then it’s it’s working. Okay. Then then it’s I joke with my student like like like in the future like like you don’t need to even like look at literature. You can just follow what models tells you. Okay. So that saves me a few months. So I I try to push my my my students. There’s a lot of mental mental torture for myself, right? So but but again, here shows like it’s AI can do the try the the industrial scale manufacturing just with one shot or is we say there’s just first trial. Okay. This is also something if you want to repeat something and you have just a new setup and AI can save your time and speed up the the discovery and your research. So.

[1440s - 1560s] Basically, here shows if we go one step beyond, we already see the material, and also we can grow the material. How about we directly build the actual chip? So here image showing like not published yet. So we can combine this microscope. You can see the microscope seeing some materials we grown, which is not ideal. We on purposely choose an non ideal material. You can see a lot of black dots are actually defects. If I give this material to the industry partner, they will say this is crap. Okay, you can never grow any chips on top of that. But if we use our AI and let AI to build it, so what AI will do is just choosing the good area. So you can see conventional way, if we have uniform transistors, billions. Okay, but if we use AI, we can see every transistor are different. And now we are connecting AI with all the clean room manufacturing instruments, and we can build like it’s actually like actually build the chip. So even though the chip is not billion scale, okay, so we can have still have thousands of transistors. But we think it’s a very good beginning. So basically, like we also see like by taking the let AI control the whole semiconductor manufacturing, it’s increase the yield. So originally this this material can I only have sixty five point six percent, which is crap. But built upon this low yield material, we can have a high yield like transistor arrays, which is a very basic chips. So this thirty percent around thirty percent increase, we think it will wholly like like will will change or disrupt the whole manufacturing semiconductor manufacturing in the future. That that’s my belief. Okay. So so in that case, so in that case, basically we recently built a startup about a month ago, and we think like using AI to to to interrupt the the semiconductor industry will be very interesting. We think.

[1560s - 1680s] AI native semiconductor manufacturing will be more and more important. We think we will try the yield intelligence or lab to fab solutions, or we think of like like maybe we want to build a small pilot line for the new 2D semiconductors. But again, like like for everyone talking about today, so all your money goes to Nvidia in the end, and Nvidia have to pay to TSMC in the end, right? So in the end, is how to build these chips actually takes the most, right? Even though we talk about AI, we talk data center and everything. Now I give you an alternative solution which is cheaper. Okay, so so so so so again, like I will say, like AI has already write some models. Recently, a lot of people talk about RSI, recursive self improvement, and then like like we think like building the missing link, how to use AI to manufacture its own brain is very important. So this shows like a flywheel, like if AI can design the chip, can build the chip, can train the AI chips, can train the AI with the data in the middle, so it’s actually become a more efficient like flywheel. Or I will say like like like the model and chips can finally combine to each other, similar to our biological brain. So for example, our minds and our brain are never separated, right? So but currently the model and the chips are kind of separated, right? So if we combine these chips and model by using model to build chips and using chip to run the model, so I will say this is a new format of life. Maybe we can say it’s like a new new format of the silicon life format, whatever you call it. Okay, and I will say like this could be the next step if we have a conventional AI and then we have RSI and then this RSI maybe we can call it a super intelligence. And recently, like it seems like SI is more favorable than AI, but but

[1680s - 1784s] But again, if we realize put the manufacturing into the the building the chips, and I will say this could be new intelligence, and because no nobody has has given name of that, maybe I can try to give a name to bring it back to AI. I will call it autogenic intelligence. Let’s go back to AI again to finish this closed loop. Okay, so so again, like we think, if we combine the minds and the brain or the chips and models together, it’s a new life format, and and the question is where it goes, right? As a professor, I always think. So so so basically, I want to borrow some concept. So discovered in 1916, about or one more than 100 years ago, this this concept is called carcinization. So what does it mean? Is like people find, no matter what kind of ocean animals are, after millions of years, it’s become a crab. So what does it mean? Like if you have a fish, millions of years crab, maybe a like king crab. So if you have a turtle, millions of years it become like a blue crab, for example, right? So it’s carcinization is observed in our nature, mother nature for for already there for billions millions of years. Okay, so in that case, I’m thinking like I give you a question for you. So do you think the AI plus chips will be diverge or converge, or will that be a crab shape of this silicon life format? All right. So but but maybe next time when you eat eat the crab, you should remind think think think about it, and maybe you are eating a very intelligent animal, right? So with that, I conclude my talk. Thank you very much.