年终总结2: 与真格基金戴雨森聊模型发展、AI 商业化与美股 M7 展望 ｜硅谷徐老师 S9E43

用声音 碰撞世界 生动活泼 大家好 欢迎来到今天的节目 我是主持人 硅谷区老师哈维 那回望2025年 人工智能依然是全球科技界最重要的话题 但关于人工智能泡沫的质疑 最近几个月也没有消停过今天我非常开心地邀请到我的老朋友真格基金的管理合伙人戴宇森来复盘一下他所经历的2025年也展望一下2026年的人工智能我想先说一下我和宇森都是自己创过业带过团队所以就算今天我们拿到的作业会有指点江山的味道但其实我们心里对每一位创业者每一项技术的突破都带有很深的敬畏 I enjoy investigating art waves, for this month, I recently gave him a international treasury tool. All right, Goyобщ, welcome back to s hello. We saw you I while你的明星公司Manus在北美的第一场线下活动上今天我们一定要聊一下Manus的商业化的考量另外虽然你是第一次来柯早做客但柯早一路走来已经有不少你的被投企业的创始人来当过嘉宾了对 也非常荣幸能够跟柯早跟Hoy大家录一起节目因为我也是你们的中视听众很久了我是政客基金的戴宇森然后我们是做天使投资投资全球最优秀的华人创业者 在AI领域我们确实也十多年来投了很多优秀的公司可能像大家比较熟知的Kimi,Manus,GenSpark,HeyGen这些都是我们的portfolio所以也希望正好有这个机会和穿越过周期的后裔老师多聊聊你刚才提到的HeyGen的从事人就2023年来过我们的节目在五年半前我们跟101一起一个创台的节目里面第一次就提到这个GPD那时候2020年5月份6月份年中的时候 I think GPT is going to be able to change the human life history of the technology. Of course, this five years happened a lot. Actually, I was today talking to you before. I also heard a lot of you. You had a few questions with a little bit. And the two-part podcast. I talked a few opinions. I think the most of the opinions I really understand. But I think I still want to share with you. I think I want to share with you. First, I want to talk about you. You know, you've got a lot of years. You've got a lot of years. You've got a lot of years. If you've got a lot of things. You've got a lot of things. 你在年初的时候没想到 但是2025年发生了这三件事情 我觉得最重要的还是 中国的模型公司追赶得很快 谈算来讲 在去年年底的时候 虽然我们也是HIMI的投资人 但我们确实也看到 在卡 数据 资金这方面 中国的模型公司确实差距是比较大的 所以当时O系列O1出来的时候 我是感觉这下又解锁了 Thinking Time的Scaling Law那么对于中国的模型公司是不是会带来一个追赶上比较有挑战的情况但今年确实我们看到从年初的DeepSeek R1然后到年中的Chain OneKimi K2然后最近智朴MiniMax都要上市了我们看到中国的模型公司确实在性能上追赶的很紧并且成本控制的很好不管是训练用的成本还是推理API的价格所以直接的结果就是整个开源市场 年初的时候肯定还是NAMA作为最核心的开源模型 但现在已经是中国开源模型的天下了 我觉得这个确实在去年是没有预计到 对 你在另外两个播客里面其实也提到了 比如说到千万这个model 还有Deep Seek之前 其实也是去20 并不是2020年年底 2020年前几个月就已经看到了 包括中国的那些创业者都没有想到 2025年爆发的这么厉害至少在美国你只要提到是开源模型基本上都是中国的模型2025年春节左右的时候两月份左右的时候已经是deep seek moment了就是因为伟大的股价一下掉了几千亿美元这是闻所未闻的一件事情但即使是这样大家还是没有意识到这个是不是真的是一个开源模型中国全面领先我记得跟一个美国的一个大厂的一个高管聊的时候 he said we're in the middle of the world we're in deep seek but we're in the middle of the world because we're in the middle of the world but now I think almost no one has been using the kind of machine and using the Chinese machine has been a lot of so I really agree this one actually you've seen a year ago but I think almost no one thought it will be so fast the kind of machine I think everyone thought with the B-1 the machine will be a little distance 但其实现在仅仅赶着吧 应该是这种感觉 确实在去年年中的时候 比如说DeepSea VR发布的时候 我們看到MLA是非常有意思的创新 当时其实我们就觉得 好像中国模型公司 并不是只会简单追赶 而是说他们为了提高效率 为了降低成本 会有自己的创新点 这些创新点是可以被 比如说美国公司也去学习借鉴的 就是有创新 但是能够追这么近 这么快 确实是今年的一个意外所以我觉得也可以看到像当时您说因为它的股价下跌其实也是资本市场没有预计的但我觉得这背后反映的现象其实在去年其实可能大家对于中国的模型会有一些悲观的看法比如第一缺卡第二说是不是优秀的人才都在硅谷第三是不是这个硅谷的这些模型公司会因为有很大的投入它进步得非常快导致追不上我觉得今年来看可能这几个具体发生的事实是第一确实在这个头部搜谈模型这硅谷这三家OpenAI,Borbeek,Dreamline它们的进展速度很快但是相比当时GBT3.5横空出世的那种接月式的提升可能今年确实还是一个更加incremental的提升所以我会理解成为它是一个延续性的变化这样来说对于追赶者会有机会第二就是信息确实是共享了很多的然后中国有非常多的AI的人才的储备实际上我们看到美国这些最好的Lab里面的researcher 和中国DeepSeek Kimi 包括自己的这些Laborate Researcher 可能就是大学同学 他们并不是完全不同的人群 他们可能就是非常接近的人群 所以他们本身的能力 以及在硅谷其实很多 播客也提到就是没有秘密 大家很多的方法 其实逐渐会扩散 那么作为追赶者 确实有很多可以省下来 做各种实验探索的资源 所以确实虽然卡没有那么多 但是通过努力的去 寻找降低成本的方法 其实聚感也可以做得很好 那么还有一个经常大家谈的就是蒸馏 中国确实作为这个聚感者呢 你有这个蒸馏的优势 肯定蒸馏也带来了很多的 性能的提升的帮助 但是我们也看到比如说刚才提到的MLA 包括张罗KIMI像MOO Optimizer的 这些优化 其实中国公司一方面在 这个应用聚感者的优势快速提升 一方面也在尝试做自己的创新 所以我觉得这里面其实 都还是有很多的 不只是因为 所谓蒸馏啊或者是 因为驱赶者有优势 而是说两者齐头并进的过程 所以我们其实对中国的整个AI发展 还是比较乐观的 从前来看 就顺着这个点稍微聊深一点 因为其实你也知道在硅谷 尤其是不只是OpenAI Anthropic XAI Gemini这几个 还有几个Frontier Labs 包括OpenAI 几个创始人出来做了两个 但是还有其他的Stanford的教授 什么就不急着跑出来做了大概有五个左右吧 五六个吧 这个现象你是怎么看的 一方面我是想问就是在硅谷以外你有没有看到 另外一个这是不是意味着就是更多的创新 还是要靠今后的那些frontier前年的实验室在做 你是怎么看这个问题的 对这个也是我之前在两期博客里面讲的 EFR里面的一个二就是research 因为最近可以看到不管是Elia Andrey Capacity或者Devis 一系列的硅谷顶流AI大脑们 其实都在谈一件事就是接下来我们在研究上 需要有突破 当然Elia他可能讲的比较直接 他说过去过了几年是scaling的日子 现在要回到research的日子 我觉得这里面大家其实都看到了一点 就是现有的范式 比如说基于pre-training 以及post-training RL 对模型的提升仍然能够榨取很多的空间 但是它已经不是一个exponential 或者说节约提升的这样一个过程那么如果说要实现大家AGI的目标实现这种非常强大自我学习不需要那么多的样本能够不断地探索不断地进化的AI那可能还是需要有一些研究的突破所以我觉得第一点就是说大家认识到新范式很重要新范式得靠研究的突破来解锁那么我觉得不管是在美国还是中国已有的这些大厂他们有大量的精力投在已有的范式上因为这是一个非常head to head的竞争大家要竞争Benchmark performance 用户的数据 收入的提升 所以这些在这条主线上 要花很多的精力 不管是资金还是探索自由 去做新范式的研究 我目前听到了 好像是说确实会有一些 这样的困难 因为你会在一个大厂里面 会变得比较非主流 那么我的理解是 硅谷最近像您说的 出来一系列这种 NewLab的创业公司 其实也是大家觉得 如果你要搞范式创新 你可能需要有一个 单独的一个环境 This is a product that does a product. This environment is a bit more comfortable, more bottom-up, and can help those most of the research. They are not going to be able to do a few months in the future. Or to go to a few months in the future. It is able to look at the same time and be able to look at the same time. And to do this research, it is not a lot of money, but not a lot of money. It is actually a lot of times in the time of this stage. So this is for the company to get some opportunities. So we have seen in the US, there are a series of new labs. The new lab is, of course, Elyia, the SSI, or Reflection, or human end, or periodic, or Isara. These are all of the new lab. It's a new company that is leading to the development of the new lab. And in China, I think there is a very interesting view. I just asked some of the Chinese researchers to join in the country. And then we have a feeling that the Chinese AI researchers have been thinking about the Chinese research. and something that I think has I like to snacks irritating the exported I think if you go back to a fewelim you will see that the in recent the you keep the I know so this had a strike there 这个问题,不管是 continuous learning 还是世界模型,其实这两点这些方面的大家的共识是比较凭确的,就是我们需要性的范式变化,然后大概几个重要的candidate是什么,然后在里面大概怎么去做,有人认为不比较乐观,觉得两年内可能就会有一个初步的信号,那有人需要,可能比较谨慎,觉得可能还是需要一个相对长一点时间。 但照大家想的问题我觉得是比较类似的 我觉得这个也是挺有意思的一个现象 另外一个我个人是觉得数据其实还是没有用好 因为怎么说呢 我们虽然说web上面的数据都已经拔下来了 但是这些数据到底是多少是干净的 多少是应该怎么样用的 其实我觉得还是有很大很大的空间 我的理解是像最近的Gemini 3很大的一部分也是在数据上面做了很大的功夫虽然说数据没有新的数据或者说几乎没有新的数据但怎么去用好数据其实还是有很大的坑2025年是我自己感受到第一次感受到非共识很多很多2023年出来的时候2024年的时候我觉得大家大致的其实你在之前的播客也提到过反正打力出奇迹反正怎么样2025年是第一年我是觉得我的观点跟很多人不一样然后很多人的观点他们自己也很不一样我觉得听上去各种各样的道理都有然后给了我很多的就是这个世界怎么走其实还是有很多的可能性我刚才讲的只是我的一个观点但我觉得未来是怎么样还是非常期待想知道还有什么让你感觉到surprised说实话还好可能因为我们比较早的对agent的爆发有一个期待对,你是非常早就投资了Manus的投资人因为Manus的爆发其实让我自己经常一下子在非常短的时间就做到1亿美元的AR然后收入怎么样都做得非常好就像Deepseek刚出来的时候你在之前的博客也提到Deepseek这个团队很牛后来发现其实Deepseek确实很牛但是像Deepseek很牛的团队有好几个做签问的做minimax的做好几个中国的团队包括现在美国的那个前沿的lab其实很牛的团队很多不只是一个那像Manus它确实很牛它能够在大模型之上面再搭很多的脚手架然后使得大家做Agenic能够做得很好但是是不是也有一堆的Agenic的公司还是说它一技绝成虽然说你有一定的bias但是能不能分享一下 从你的这个你是试驾怎么看 我觉得始终是团队本身肯定有很特别 很outlier的团队 但是有outlier其实很多时候也是意味着 你有一个很大的base 你有比如说100个创业者 你们有几个特别好的 但也说明你现在也有100个 你们可能有10个很好的 所以我觉得我们看到模型端 之前的叙事是说 DeepSeek这个团队特别特别不一样 中国只会有一个DeepSeek 后来发现好像不是的 I think Madness is aware of TH比 10 years old, their reign of gentry and cultural hubby the amount of fact is that China also has very strong power and has a lot of momentum especially the most horizontal I'll talk to you later. that I have taken as AI technique, it is a power frame Schnit this is the technologies we stronger because of large trading mariota – the normal trading company extension, most of them is a CRM stimul ksi에서도 and the experience of the business model. So I think this one I saw a bit different. If you say before, or Tmoo, they are saying that now we are doing well, and then we are doing well. We saw that in the big companies, there are many Chinese teams, and they are on the team, and they are on the market. I think this is the language model, the video model, the world is on the market, and it is related to the language model. 并且如果你是针对knowledge worker,不管是说Manus这样的应用,还是我们投资的那个公司Hajen这样的应用,它其实在中国,在日本,在美国,在巴西,其实用户的use case是差不多的。 所以这样一个中国的团队可以一上来就去直接面对全球市场,并且中国这一代的创始人,我觉得经过了这20多年的这个生态环境的成长,他们的国际化意识,包括说产品的能力,对国际化产品设计的审美,以及推广这些能力也变得越来越强。 Это 20 royal классitte modems Black universal mode OTO这个词就是中国发明的对当时是我们新四大发明那些就是属于中国特色但是美国可能土壤很不一样包括外卖中国大这么大的市场其实也是那么我觉得到了2025年我们其实看到中国优秀的不管是软件还是硬件创业者他们是可以上来就去针对全球市场做first in class或者best in class创新的软件比如Manus可能是一个例子黑件可能是一个例子我们还投了GenSpark We started Opposclick They're kind ofressive So, for show like an Ethic Spider-Man for the geldi For the кар Comes Ahesu Should be tested All and well потому that the commerции is an electronic, because America has had something that's a whole kind of transformed society in China. Chinese avid或者prosumer的这样的例子尤其是我觉得一个华人创业者他面对美国市场的时候如果他做enterprise其实他起步要难很多因为各种geopolitical的原因也好他对市场不熟悉的原因也好所以反而他去做horizontal做这种偏prosumer的其实我们看到我们投的这几家我觉得出海做的比较好的公司像这个Hey JaneManusJames ParkOpposClip包括我们最近投的家Polo他有点像Hicksfield做这个世民生成像Manus建筑的Monica做的这些其实都是 horizontal prosumer它相反到了一个好像美国硅谷创业者盲区所以我觉得这也是一个不同地区的创业者的经验积累出来的不同的竞争力所以说你今天仍然是看到这个盲区吗还是说有10个Janspark10个Opus10个Manus在紧追不舍还是说你看到的还是一个盲区我确实在想比如说ChattDBTProbillacityMid Journey比如它也不是真的Enterprise D And Equital And And σε Questo Con C Desktop C 或者说认知的差距就被拉平 我们是觉得Manus的出现 这些公司的出现 不是一个说one of a kind的现象 我觉得它其实在一个 只要这个大的环境不发生 放在mental的变化 我觉得中国创业者 其实越来越能够在国际舞台上 从来经济的 当然现在肯定有很多 事实上的barrier 对吧 但是我是觉得 从把产品做好这点来看 中国创业者 我觉得完全是一个世界级水平的 我觉得另外一个世界级水平的 not just a product of the ability to sell the ability or taste of the ability and there's a distribution you just mentioned these companies I think they do it all good how to influence the data how to make these marketing we can say marketing but actually it's a growth product-led growth PLG of the business for me I think it's a very good thing actually in the移动互联网 I think it's a good thing I think it's a good thing because I mentioned 在后期榜单上面全是 Chinese company做的 对吧 字节的这些应用 不管是TikTok Capcom Timo XI 就中国创业者在全球做 Distribution这件事情 其实是挺被验证的能力的 包括还有腾讯的那些游戏 PopG COD Mobile 王者荣耀 所以这些我觉得反而是 中国创业者最 proven的能力 这里查播一条讯息 年底了给你推荐一个 生动活泼做的特别的年度榜单 这张榜单上有八条少年们最想让全世界知道的新闻 我们在思考前沿科技议题的同时 也关注到这些科技的未来使用者正是这些少年们 所以他们正在关注这个世界上发生的哪些事情呢 跟随这个榜单一起来揭晓吧 点击节目详情页中的链接可以收听榜单详情 Yeah hey You talked about this an access to a audience or eraltung comfortable having a flashlight Oh See ourository So we mentioned there were several B2 tool After our other text, we mentioned earlier thatPhone is the 65th century Sergio老 häufig used by the characters It was about the 10th century of this not an agent finish in the years As the体 Boumoon publication some people It is really important and that this kind of event of this years is still in part of its structure 至少不是说我们每个人 对吧 每个人都已经在跟agent打交道了 Agent agent直接帮我做事了 这个还是一个就像自动驾驶一样 我能够看到影子 但是呢 或者说甚至能够体验一点 你觉得离大规模的采用 就是agent agent 就是我们比如说你个人会有个人助手 你的个人助手会做很多一部分 你过去五年你自己做的 但是今后若干年 这个让agent来把这些事情做掉 You think we are in the 2026 years We will see that As we need some time And some time I think this is a good question First Andrew Capacity He said a decade of agent I think I think First of all, what is agent Because now every AI company Is a agent But I think the agent Is a level 3 You give it a goal Give it a space for him He can be able to plan 任务实现一个做一个实现的路径 然后在过程中去不断的执行 然后反思自己有没有做好 调整路径就把它做了 就是可以解放人类注意力的 我觉得他要agent的 果实的话他就是助手 就是level two 我觉得每次当我们要解放人类注意力的时候 实际上这里面有它不只是一个能不能做 而是说它能以多好的完成度去做到 它能不能够承担起这个做不好的责任 或者想来讲背锅 自动驾驶里面我们看到10年前 一个自动驾驶无人车在路上开个十几英里 这个没问题 但是难就难在说怎么处理corner case 怎么样能够suff recovery from errors 然后如果真出事了 你能不能有底气或者是有能力去承担责任 对吧 这个直到现在我觉得大家都还是非常谨慎的 Wemo现在也就是这么不到十个城市 FSD它肯定也是说你要持续注意力 对吧 你的眼睛得看前面 所以我觉得能力是一方面 它在对于社会组织形态的变化他能不能承担起责任是另外一方面当然也有人说因为自动驾驶一旦出事可能要死人所以他的后果非常大所以他比较难但我想说的是现在这个agent他可能说产生经济后果如果你一些决策做错了你的内容错了其实也是要花很多的代价的并且现在如果说AI产生这些比如说研究报告内容能力非常强但是当他仍然有很多hallucination他的分析可能不make sense的时候人的检查能力是很有限的就是你多了十倍的内容需要内容你去check当这个内容它不是那么ready的时候其实反而会形成一种overwhelming的效应所以我看今年不是有个年度热词叫slope吗就是AI产生的这种没有创意的内容我觉得这是在创意端就大家觉得AI内容千篇一律我觉得在执行端你会发现AI可以给你很多90%是对的但5%可能有问题的而且你也知道它5%可能有问题在这种情况下要不然你就躺平我也不管了 反正我就发出去吧 要不然你还得去检查 所以我是觉得这里面对于准确度的要求 其实不低的 不比自动驾驶低多少 在这个过程中的话 我觉得我们要不断的去 把它准确度越来越高 包括一些组织 包括一些流程 也要进行调整 原来可能比如说现在是 Aging的帮我做工作 给了我一个报告 但我把它交上去 那我还是对它负责 但是那之后是不是有某种形式是说 因为是AI做的 所以可能我确实不用去做这么多的负责了 这个其实不只是一个技术问题 它是一个组织问题 甚至是一个文化问题 我是觉得在这里面面对的挑战 其实不是那么简单的 但我确实相信Aging本身的技术提升会很快 因为像开始华尔老师说的 就是它其实很多时候是个数据问题 现在各个大厂都在拼命地标记 这些数据就让AI去用手机 用电脑 甚至自己还做了这个豆包手机 所以AI我觉得会更好地去使用这些工具 更好地去写代码 这些都没问题 但是因为目前的生成式模型 它本质上讲它还是个概率分布 所以它不可避免会有hallucination 同时它的学习能力还没那么强 所以当它面对一些corner case 一些犯化的问题的时候 所以你不知道什么时候会出来一个 像9.9和9.11谁大了这样的问题 就是人类你知道人类是不会犯这样的问题的 但AI今天你堵住了 炒面里有多少个R 明天又有9.9和9.11 后天又有个什么问题 就这种问题所存在呢 就导致你让它做L3承担责任他逐利完成任务这件事情始终是我觉得会差那么一点自动驾驶里面为了解决这一点的问题其实花了很长的时间花了很多的钱我觉得在AI这里面可能真正是AI里面可能也会有类似的情况我觉得Andrea Capacity他经历过自动驾驶所以他说这个话我还是比较buy in和相信的对他其实去年就说了一句话他说我十年前做自动驾驶跟这一次做自动驾驶感受差不多我惊了一下十年就没什么但是他说的观点就像你说的最终这十年主要解决的是corner cases然后你刚才说的有一个我非常同意就是我觉得这里面有很多是文化的东西就好像比如说订机票订什么东西你交给一个助理其实很多人的助理也是会出问题的你跟他说要做什么我搞错了所以说我觉得从文化上面是不是能够接受AI去hallucinate接受什么样的hallucination然后怎么样去对待这么一件事情也是需要去审视一下的因为其实我们都知道自动驾驶从某种角度上来讲有一些情况甚至比人都已经安全了但是我们还是不愿意去至少从regulation的角度不愿意去放手因为这是一个人命关天的事情但如果说是一个助理的话是不是我们会更加宽松一点我觉得过去三年我自己的感觉是我们对AI其实挺刻薄的 Because I think alleson is in choices In some mistakes, it's not so serious But I think people are also constantly doing success I haven't thought delve into them I completely forgot why this because of a person Or it was a father a better way for them for you. I think there's a question. I think there's a question. The蘑菇 is not eating. He said, no problem. And he said, we'll have to finish. He said, we'll have to finish the蘑菇. Now, I don't know if it's online learning. It's been been a message. So, he's been asked for this question. Why is it 9.8 and 9.1? This question is so difficult. I've seen the publicity system prompt. He's got a question. It's written out that Trump is now the president. Because there are many people who are asking this question. The question is Biden. So we feel like it would be less difficult. So now it's not the way. But I always think actually, In a way, No, not least don't. I think people party are five minutes. First, so I think technology wouldides basically that would be to go there to go to that shop. I agree 这个很好 但是不是买那个事情要他干 打个问号 因为在买这个能力上是很重要的 但是这个action的能力 一方面因为他要花钱 他要去帮你下单 比如说还不可撤销 所以他会给他谨慎 但另外一方面 我觉得我们缺的不是 配的下单的动作的时间 我们缺的是配什么东西 对 有一个你说 那个模型的边界 这是一个人是未知的 我觉得这个可能 实际情况比这个更加恐怖 为什么呢 因为边界不知道 这个还可以但我可以去学习但问题是AI不断的也在变化这个Jemner今天这个是厉害的或者这个不厉害的明天这个就变成不就变成不一样了就是说即使是同一个模型同一个版本我都觉得在变化不同的那个上一周跟这一周的它的能力都在不一样但你的秘书或者你手下的小朋友对吧他上个礼拜是怎么样你大概知道他的能力是是下个礼拜是怎么样就是你能够看得到所以说我是非常同意的 From this perspective on I think I do the agent Is there a skeptical thing I say skeptical is what? I think it can be done a lot of things I think it's very good But we are at least understanding AI Or some of them are aware of But I think for the most people You know they understand AI is a way of understanding And this way of understanding And this agent What do you do? You can teach them these things我觉得这个不现实所以说现在你也知道你可能知道我在做AI browser然后AI browser现在有很多Agenic browser我自己是不是很相信Agenic browser并不是说我觉得Agenic browser这件事情不是未来我相信是未来而且我是非常期待成为未来但我就觉得今天的技术能力今天要去让人去认知这个边界我觉得这个对1% 5% 10%的人我觉得一点问题都没有但我觉得对mass population 边界还会不断变化 像比如说Manus 我的理解是对Agenic browser 还是有满大帮助的 对 你是什么想法 我觉得可以把AI产品分成两个camp 这些2C的 一方面是像Manus GenSpark 包括Hegene这些是给Power User 或者Professional User 帮他们提高效率 比如说我用Manus 我可以用GenSpark或者是JLGBT的Deep Research 我可以做很多研究 比如当我投资的时候很多行业原来需要一个分析师做现在他帮我做个大概我去看这个是典型的power user的场景但是power user如果进一步的话那么对它的准确性对它的consistency像你刚才说的都会有要求但是这块我觉得是大家愿意付钱的这个是愿意为生产力付钱的我觉得另外一块中国我们看到其实出来的非常多的就是比如说举个例子AI漫剧就是中国有很多网文对吧就是看这种网上的小说你也知道现在很多人是不愿意去阅读长篇的文字的他一定要配合一些这个graphic So we 한 mm by living back to some anime here. 如果他们在意说这个事情这个到多么原创或者多么美伦美奂,更多的是一个对于这种内容的消费形式从文字到了这个漫画。 And then I'll be Guy I want you to study a few script But these are society 技术能够做到的 这样我想起在我们说 创建者的窘境里面 很多时候这种颠覆是从低端发生的 因为在一开始的时候 比如Runway这些AI视频 他们说的是我要颠覆好莱坞 我要做电影 这恰恰是一个典型的说想从高端去打 其实从高端打很难 虽然大家现在发现AI越做越好了 但是实际上你要真的进电影 进电视剧 那还是花很多时间精力 但是你从低端的这个场景 比如说我这个AI漫剧 这个漫画看个大概就好 或者刚刚我说的这种把文字变成一个AI视频比如说它讲起一个概念有那么个意思就好它不用特别精美不用特别一致我觉得这反而是我们在可能中国因为大家都看短视频所以观察到的一个AI它从这个偏低端的偏简单的就有幻觉不要紧不consistent不要紧质量差点不要紧从这个角度去进行一些颠覆包括之前我们看到AI生成视频里面也有一些像类似于什么生成那种动物在蹦床上蹦又是什么猫猫跳水 什么就这些看上去很荒谬的内容但实际上也带来另一个消费价值 所以我觉得一方面我们看的是像 Manus这种提高他十倍生产力 这是一类人 另外一类就是说 那些AI没有那么好 没有那么完美 也能够 提升大家的阅读体验 观看体验 大家的这种消费内容的方面 我觉得这可能是两个其头并进的事情 所以就是说刚才我们讨论的都是 为什么Agentic这个level 3它还不够好 是因为模型还有这样的问题 那样的问题 所以它针对这些 high value of the time, mission critical of the time. Just like I said, I'm not going to be able to write a report, but I'm not going to be able to write a report. Right? So, this time, it has to be a good deal. But, in the other hand, there are no mission critical of the time. Actually, now, it's from the point of money, from the point of business, from the point of business, from the point of business, from the point of business, actually, it's quite a lot. Yes, this is actually what I was talking about. We're doing the same thing, AI, the point of business, it's a bit more than that. Of course, it's a little bit more than that. Just with power users, I don't want to say it's a way to go. But it's two teams. The two teams are all needed. One is for power users. The behind is that I use this technology to be able to control this technology. If I say this thing is wrong, is wrong, is wrong. I don't know what's wrong. That's not for power users. But I think that most people are 80% or 90% are from the other side of the building, from the other side of the building, from the other side of the building, from the other side of the building, from the other side of the building, He's not a good person. This is not his job. He's working on his work. He's going to be with this technology. You can imagine you and I, I'm going to be with this AI technology. I feel like a day I don't want to do it. Even more than that, I feel like a lot of people are doing it. So, I think that they're doing it. I think that I do the video. I think that I do the video. I think that I do the video. I hope that we have a zero prompt. You don't need to have any prompt. Just to feel the value. But just like you said, this is two of them. I think that I have to be in the middle. I think that Manus is able to this one of the users are very lucky. but I hope that we do the AI-lapse also makes 80% of people think that I use the Neo browser after I think I feel very happy that therein we talk about it you think therein we talked about we talked about it one is create this product another is distribute this product you think AI-lapse-lapse will be the product of the product will be the product of the product or is it at today's time you can see the product还是跟上一代 Mobile 时代也差不多我觉得这也是我觉得这次 AI 创业很多人都把它类比移动互联网但我觉得其实它们挺不一样的如果你去分析的话移动互联网其实是在互联网基础上一个distribution的创新其实移动互联网它distribute了很多东西比如说 IAM比如浏览器比如说内容社区比如说游戏实际上在PC互联网时代已经出现了但是它换了一个载体换了一个distribution channel所以在这个里面其实它的那种科研成分当然也需要做一些比如推荐隐形的科研广告隐形的科研但是科研的成分包括说创新研发的成分我觉得是比较少的大量的创新是在商业模式分发渠道增长手段这方面的创新所以当时有很火的这些增长黑客有这种各种做AB测试这种比较tiny的比较精细的操作但是AI这次其实大量的创新首先还是发生在R&D阶段就是我通过投了很多CAPEX我有一种generation by generation的迭代产生出一个新技术所以这个其实我觉得是蛮像硅谷最开始的做机身电路做CPU做芯片那一块的时候也是CAPEX很heavy然后generation by generation大家甚至有一种轮番领先不断的通过新技术的迭代以前是新的硬件技术现在是新的模型技术的迭代产生新的产品但同时它又在已有的分发渠道上 So now, for example, AI, it also wants to buy a lot, it also wants to do HAL marketing, it also wants to do all of the distribution channels. So I think in this, we have two different ways. There is a way to do it, I want to do a good product, and now there are many early adopters, many want to try to use the users, so when your product really has a good experience, then there are many users here. Another one is I don't want to do any research, I just want to do the technology, I want to do this product, like I said, AI, and then I want to buy a lot, I'll pay money to buy money I'll pay money to buy money I'll pay money to buy money I'll pay money to buy money I'll pay money to buy money But I think that's right now I'm not sure now I don't have a distribution distribution I'm not sure now You still have to do an app This app is to buy money I'm not sure to sell it I think this is a big thing Because in the移动互联网不管是中国还是美国 Actually, everybody is in a country It's just that you're in a country You're in a country You're in a country You're in a country This time there's a ddain 进去占了 有一个Snapchat Instagram去占了 它就是一个蓝海市场 但现在任何用户在手机端的时间 其实都得从Instagram扣出来 都得从TikTok扣出来 都得从抖音扣出来 所以你一上来对于AI应用 它的门槛就变得更高 所以我觉得这个确实是对于这种 您说的distribution型的公司 现在不是一个比10年前更好的时候 是因为他们都是在打一个uphale battle 他们一上来就要替代掉已有的这些做了很好的产品的时间而不是像10年前时间是空白的还是像字节这样的大厂他们有巨大的资金优势和投放能力优势渠道能力优势但相反像Manus这样的应用它就是得靠我因为对模型技术的进展我有我的判断所以我在一个right time拿出来一个right product让大家一下子都惊了原来还有这样的产品其实我觉得这不是一个孤立比如说可能在之前ChallengedGPT是一个这样的例子Midjourney是一个这样的例子 Cursor 是一个这样的例子就是把模型能力的进步第一时间用一种非常native的产品形式体现出来它可以用互一种魔法版的体验在这种情况下你就可以不用去做投放或者说很少的去做付费投放而是通过口口相传通过用户自发的传播去获取很多的用户那么我们觉得这个可能是在AI这个时代对于应用开发者他们应该首选的这个方式因为除非他们自己投放真的很厉害否则的话去买量那肯定是买不过大厂的 但是我们是可以靠产品设计 靠创新去获得这样的自发传播 而且这个成本其实也比较低 我觉得这里面你刚才提到的一个点就是现在大家的时间已经占满了 如果我再要去推新的产品 其实cost会很高 我自己是这么看到 我就觉得我不知道下一轮的distribution 传播的创新在哪里 But I think it will need a new creation Because you have no new creation You just just靠 product creation I think it's not enough I think it's needed to be a product creation Addition to distribution Two of them need to be In the past, there is a business model For example, in the past 20 years Google this model The business model Google Facebook Monetization of the business model It's basically like an email And then you're just around 这两个印钞机在转 我觉得在AI时代 这两个印钞机 多多少少都会受到很大的挑战 但我觉得今天是 就好像1998年的 互联网时代 大家其实还没见到 它要有了Google这个公司诞生了 但是还没有看到Google的 AdWords这个一个是 商业模式的存在 那到底是怎么收钱 我其实没看清楚 因为为什么 其实我也想问问 你是这方面怎么看的 因为一方面 过去的就是search对吧广告手但是不是这个这个1900年发明的这个商业模式到了AI时代仍然会持续下去还是说今天因为Chat我的这个模式对吧就是会那个商业模式会很大的改变这是一个另外一个问题也想跟你探讨一下就是到底是按照Seat对吧License来收费还是按照Token或者说用量来收费这也是一个蛮有 没有共识的 我觉得在互联网的历史上 我们看过好几次这样的例子 这样我不可以聊到 就是先有一个好的产品 大家不知道怎么去做 native的商业模式 大家甚至是想把 以前的成熟的商业模式 给安进去 但到后来发现 新的产品需要一个 新的native的商业模式 Google是一个例子 Google当时出来之后 他就是说我没有广告 我没有那些banon ads 我是一个很clean的见面 所以大家很喜欢 但是呢 那你也得赚钱 所以他之后就花了几年时间 搞出80 word 80 cents Facebook 一开始也是说我不打广告我的广告 is not cool然后后来他们先做了一些本地生活的广告一些小 banner什么的然后做了游戏广告这两个都没有做得很大然后最后是通过 feed流信息流的广告形式当然最后都是打广告打广告这个事也还是可能速度同归因为 at the end of the day可能你最后要不是卖东西要不是打广告这个可能就那么几种但是呢怎么个做广告是你玩的对 这可能是不太一样的所以我觉得现在其实我们在看到比如说 He said a direct advertisement he can do good chicken noi in the short term The amount of price is carried on a break He said it's very good就是美国的整个线上商业对吧它就是有一个比较固定的比率是画在广告上面的那么加上GPT做广告其实广告的东西还是这些线上交易的商品和服务其实跟GoogleMeta大家对象是一样的那么比如说你说电商的广告那其实肯定是某个总线上电商的总零售额的一个比例可能是几个点十个点这个比例和这个总量其实不太会立刻翻倍的增长 So, це-GPT打广告或者JIMALA打广告Whatever如果都是还是这些广告的话可能它还是一个从Meta从Google从TikTok去抢饭碗的一个事情就它还是个纯量博弈可能真正的这个真量来自于说AI帮人完成了之前没有办法完成的工作这个是真量我觉得广告可能目前来看还是一个偏纯量的逻辑所以这里面它是不是一个齿效比长还是说是一个大家都长的是有怀疑的所以你是同意会对谷歌的商业模式 is going to be a challenge this one. I think so. Because I think chatbot will be a challenge. I'm going to use a solution. I'm going to use what is going to use to find some specific questions. For example, the website is what is the website. I'm going to use a link. I don't know who is. Google is doing a blue link. I think this is right. If I need a blue link, I'm going to Google. Anything else is going to chatbot. Google said I have a chatbot But I think it's not a good thing Chatbot is that the chatbot is a big issue So now Google's price is a big issue I think it's not a problem I think if OpenAI does not have a problem AI will跌 If OpenAI does not have a problem Google will also have a influence So I think there is a very hard thing But this is a later point The second is seed based Or usage based If you're in a field of data 互联网最开始其实有点像是usage 就是大家按照那个上网的流量 或者按照那个上网的时长去付费 付网费 然后逐渐的来说大家发现 这个互联网的这个流量成本越来越低 所以大家开始有固定的费用 比如大家上网很多时候就是一个flat rate 对吧 你用多少都是一样的 对软件其实也是一样 一开始要花钱买软件 Seed-based或者usage-based 然后后来是固定费用 但后来就是免费 现在上网虽然没有完全免费 terrain Yeah exactly BBT刚出来的时候的那个智能你是很难为那个token收钱的因为很快那个token就会在端侧了在端侧的话那电费反正也不是交给模型公司的它就是basically免费所以我自己感觉有一个rule of thumb就是现在的sota你是可以charge按照token来收费的一年之后可能就得flutter rate再一年之后可能就开源就借金免费再一年之后三年之后可能就到端侧了所以我觉得这个你始终只能对sota或者nearsota的token按照token收费因为它很快就会变成commodifiedManus的收费算是怎么样Manus好像是有一个固定费用然后再按照usage所以说它主要的收入是usage来的是usage但是这个其实我们有很多的内部讨论就是说是不是这就是最好的方式或者说也许两三年之后是不是还是这样其实I don't know说实话但我现在有个整体的感觉是如果你始终提供的是最sota的服务那个可能始终是还是会能够按照usage收费但是比如假设你用你用了token总数里面我觉得可能只有20%是收它的是它的marginal成本和marginal的efficient都很高但很多它下面就是commodalized这些模型的水位越来越上升就会把原来proprietary的token变成commodalized那commodalized之后我觉得就收不了多少钱就可能只能收一个flat rate或者很多时候就可能就变得免费了对 所以说TropManus这个案例至少到今天为止 你觉得有什么让其他创业者能够去思考的就是从一个商业模式其实关于这个收费模式我们有很多internal的debate因为比如说当你是按照usage收费的时候一方面来讲它会让你的margin保持在一个比较好的水平但另外一方面来讲也会压抑一些使用因为如果大家是觉得你是一个就是所谓包干的flat rate的模式像lgbt这样那你其实用的就没有心理负担所以这里面其实说实话但是这个东西很难特别去AB测试对吧你很难说这部分用户subscription这部分用户usage这个好像没法AB测试所以其实内部一直是有这个debate我觉得现在没有一个clear answer但我刚才在想的就是说他现在比如说用户为了编辑的token成本付费的token两年之后我觉得肯定是flat我就在想因为我们刚才提到的Manus它主要服务的还是一些power user其实有三个一个是usage based就是你够powerful你就你用多少就收多少钱另外一种就是干脆不收费就是靠广告Power User永远是有自己千奇百怪的收费方式对于绝大多的用户你觉得会不会还是回到类似于互联网最终也不存在什么20块钱一个月200块钱一个月最终大家还是回到一个用东西免费但是羊毛出在猪身上我觉得简单类比永远都是比较危险的 但是我的一个直觉是不太可能说普通用户比如说每个月要给AI付挺多钱 比如说我今年现在一个月付20我过10年要付2002000美金 我觉得这个是比较难的 虽然说普通用户用的token会多很多很多 但是这个要从一个广义和狭义的角度去看 比如说我们在上网这件事情上花的钱 就我们为这个网络transmission付的钱 其实是相对很定的可能相对20年前上是下降的但是我们在网上这个渠道上花掉的钱肯定是上升的电商IAR subscription但是我买了这些token之后这些token帮我干的事情或者说我在上面看的广告他们carry的这个价值我觉得是会越来越多的所以我并不同意说大家用了越来越多的token就是这个卖token的人收入就会越来越高正如我不同意我们用了越多的互联网的beneways所以卖beneways的人 会收入涨十倍一百倍一样 我觉得那个可能也是一个 比较固定的 没有那么大的一个subscription 比如说我花一百美金 那我买了这些就是我的token 或者说我就subscribe了一个应用 但是这个应用 他帮我干的事情 比如他帮我去买东西了 我在上面承载了很多我的电商收入 然后我在上面看了很多广告 广告被广告商赚走了 但可能不是我直接被卖token的人赚走了 同样的类比也可以去想比如说Netflix或者Spotify这样的应用原来我买唱片我得花多少钱我在Spotify上我听到的音乐可以多100倍我会花100倍的买唱片的钱甚至是我可能是花了更少的钱好了我们讲了很多技术也讲了很多一级市场你投资的我觉得其实我们很多听众可能很感兴趣的还是二级市场因为就像我们刚才提到的谷歌的股票Meta的股票是不是能够买是不是还是能够找的空间我知道 虽然说你的职业是一级早期投资 尤其是天使投资更多的 但其实你对二级市场也是有很多观点的 我在小郡的那个博客里面听到 你是已经空仓了 你是预测会有一个泡沫快来的一个原因吗 首先这个disclaimer就是说 我不是一个专业二级投资人 而且第二呢 我不是一个价值投资者 我是一个交易员 我的偶像是Stanley Drunkmear 我的偶像不是巴菲特或者是段永廷不同的理念都可以有赚钱的机会都有可以自己这个Disclaim非常重要对,因为阿里市场真的有点像是比如说川菜和香菜和粤菜可能有人觉得这个好吃有人觉得不好吃所以这个真的是要适合自己的状态和自己的性格首先我觉得当一个大的产业趋势发生的时候大家对它的预计还比较低的时候这个时候是比较容易赚钱的叙事在形成和扩散的时候其实是比较容易有非公司去赚钱的比如说在去年年底今年年初的时候当时其实就有很多人说这个pre-trilling是不是到头了然后说这个capax是不是会下降但我那个时候反正我是比较坚定的看多的这个时候呢因为其实我们是看到了coding的机会agent的机会包括我们后来看到了mainless它确实消耗的token非常非常多所以这个是我们在产业中的一些优势当然后来大家也演绎了这一点大家发现确实这个inference非常的需求非常的大 So, so that最後 it is as possible, GPR, TPU else, ASIC also has come the space of Tradition, the increases forы tout, for for Natasha, its Occup from電, even further away from the green current and most... I understand the inference scaling side ideas itu whereas in 2025 months later But in person finally the next one on being Well, a future is different rule description is really seen as to the people in this which I think is a little high from the crowd a big deal from meta and I think value and I can see all debit from華 Control Equador Connect new To make simple which We're now trading the I, investment. But why do AI do this investment? Because it's because AI is a return. But R is a big. So R is now going to the need to do the job. Because now the big companies, they need to research the底层 of the change. These are the big companies. They need to see the products, the customers, the收入. So I remember in Friday's blog. They said they own open AI. OpenAI became a business company a malicious appropriate customer At least, I forgot the read meine journée AI of these tools he can wait what were their in hours Therefore, this is freezer from nature I'm with Vológica Mani wrote it There is a bit Loving I believe大家去提高这个app的过程是一个漫长的过程不太会每年都能提高很多的app然后它的渗透率我觉得其实反而就付费渗透率反而会在因为GIMLA要打价格战GIMLA要抢份额OPEN AI要define的份额这里面其实我觉得不太可能说它会大量的把用户充免被变成付费因为至于它的大DAU战略对它的市场份额这也是不符合的它现在好像MAU已经是一个billion的吧一个billion然后DAU应该是6亿 6亿你觉得不管是MAU的提高跟penetration就是5%的付费都是有压力的在往上走我觉得肯定会往上走但是我认为最容易的first billion已经吃得差不多了因为全世界80亿人但是你肯定第一个billion是那些对技术比较关注比较相对时间价值比较高相对也比较希望提高生产力的这些人first billion It's like early market We're going to use the market It's the innovator Early adopter It's about 10% Now actually First billion That's just 10% And then All of the GBA I think there are 2 billion Of the people The DAU In China is a 1 billion That's a DAU That's a DAU That's a DAU That's a DAU And then And then And then So I think All of the world 1.5 billion To 2 billion The users Have used AI That's a DAU That's a DAU That's a DAU That's a DAU That's a DAU That's a DAU It's a DAU It's a DAU It's a DAU very overly oh Make sense This I'll eat. That's why I'm laughing. Just from this perspective on the other hand, you think because of this reason, you think you don't have to do GPUs'耗 or do the use of the consumption of the economy that will have a certain amount of growth. The growth of the growth of the growth of the world is not that big. Can you understand your thoughts? I think it's a good idea. It's a good idea. It's a good idea. It's a good idea. It's a good idea. And this is a good idea. I think it's a good idea. It's a good idea. 本当に非常大変化 so I'll focus on a couple of new products and outcomes. So all of our products are based on what we have milked, 买了好多年好东西之后 实在是最后泡沫上市了 Cisco 斯科是到今天为止 还没有回到200年的股价 对 如果算分红是回了 但是如果没算分红就没有回 第二个就是DugFiber 所以我就看很多人在讨论就是说 当时互联网泡沫的时候 有很多Fiber修了没有用 泡沫现在GPU都打满的 所以现在没有泡沫 我觉得这个观点实际上 互联网泡沫也是个 基础设施建设的泡沫 实际上在2000年的 3月12号crash的时候开始crash那个时候大家仍然是在高速的建设中它不是说大家已经看到了我要光线闲置了我要减少建设了互联网码才崩的它其实都是在建设周期中间崩的因为资本市场永远交约的是预期永远交约不是现实如果现实都已经看到有一堆Dark Fiber或者说有一堆空转的GPU那我估计已经跌得差不多了其实我们始终市场它的表现都是提前一步的所以我会认为 现在没有闲置的GPU不是一个安全信号说我要等到GPU闲下来我才卖 那个时候可能跌50%了可能 因为始终都是长端的预期会overwound短期的这个价格变化 但是我想说的最重要一点就是说我认为AI的return会非常大 但是根据Amara Law以及我们观察团的现象 它在短期的兑现比如说26年的兑现 我认为这是一个很关键的check点 因为在过去的几年其实大家对AI的期待是没有那么高的 大家还是在科研阶段 是在投资阶段 但现在投资已经非常大了 近来的钱也是非常多看用户 看收入的钱了 那26年是第一个我认为交作业的大年 那么我们目前来看AI的模型能力 AI的发展阶段能把这个作业交好 我认为是有风险的 然后这里面还有一个大的逻辑 就是我们确实看到了刚才说的模型的能力 编辑的拓展在放缓 虽然模型还是在成长还是在细刷新型的benchmark但我们可以看到大家对AI的预期已经从之前我记得23年的时候很多人讲27年就要AGI了但现在我们看到已经是一个month over month去track chatGBT的DAU去track它的广告去看一些非常细节的不是那么宏大叙事真的很matrix driven的东西我觉得这个本身就是个预期的缩水对吧其实我们现在还是想我要做个10亿DAU的应用然后去卖广告但是这个我觉得会给大家很多一个reality check就是说AGI不是 is a future and a future. It's a lot of money. I can reach a lot of money. I can reach a lot of money. This is not a good idea. I'm not sure that you have a lot of money. I'm just saying that you will do the machine of the ability to be realistic. It's a lot of things like that. For example, the code is very good. But it's not a good idea. It's not a good idea. It's not a good idea. It's not a good idea. We need to have a new research. But we have to deal with the research. The new fan is still in the milk This is why I mentioned The first one is EFR The second one is return The return is the concern And we need to return Research I think AI's the basic Research is to get to the scaling law Scaling on the scale After the ability to get to the use case To get to the use case So we're now to the Research is to get to the There's a point I'm going to service. Because I've seen you with the little bit of the discussion. I've seen you mentioned that you read a book. There are two different books. One is a bit of a natural. One is what is called Boom. He said we have a water. Because he's a basic point of view. He said we don't have a water. He said we don't have a water. There's a good water and a good water. Good water is the future. It's the future and the future will be different. So we'll invest in the future. 坏的泡沫是说相信未来会和现在一样 比如说相信中国的房子会一直涨 相信茅台会一直很值钱 相信美国次贷危机的刺激贷款一直还得上 这个叫做说坏的泡沫 他就说好的泡沫 第一参与的人都是专业选手 都是VC 都是什么企业家 他们亏得起 简单来讲 第二个就是好的泡沫会有正的外部性 大家在追求非常不一样的未来的过程中 会建设很多基础设施 会有很多的发明 虽然有的死掉了 但有的发明会留下来 所以这是好的泡沫 他说相信未来会跟现在一样 这其实是一种坏的泡沫 因为最后参与这些泡沫的人 都是那些最不愿意承担风险的人 比起来他没有带来政策外不幸 他毁掉的是这个价值 但是他没有产生新东西 大老板是怎么认为的 大老板也是 这么认为吗 我一直是这么认为的 我认为最牛逼的 技术会产生最大的泡沫 其实泡沫其实不是说大家看错了 大家说这个东西不好 I said it's good, but it's good. I think it's good, but it's good. It's good for the long-term, and it's good for the long-term. I remember a year ago, we talked about the fact that the internet was delivered by the end of the year, the way it was delivered from the end of the year. It was over delivered. The internet was over delivered. It was a thing that we were doing in the year, but we still experienced the last year. So I think this is two of the two-part. AI will be very important. AI will have a huge value. AIA ITVLOCKER We were able to design this value at which when we're going to be a loaded and two gone. So I'm having a small quotidian project. Then, if I go to a couple respecto to a case that would be aag Kitty然後 catch the cost of one investor, or a хотя some luck, on the first wave of the saws, I think that low point would be a non- boundaries. So in bizim地 vurstener with Buckingach, so we look at this supply of 10 years last year. And this is something that we are still investing valuable to our wind. We still need to comment a comment about which useosuis bottleneck... I remember you said, prohibition of 2005 very easy to go by Google or Amazon or doctor of CIOoru Cooper It will be easy but, I'll be ready It's a lot but we came down We came down to Detroit to Google Google we even dealt much The reason the money it's been Fields I add three Money about Oh I actually used to move up there again, and история war change. The following crowd Oswegan Serial project was quite small. And the technology from Oyirtschaft was very different about titlegame. At the beginning, in the initial February, I worked with some new websites based in Xbox. Happened by the first episode, theік sand progress was a little super. After he used similar Studien, its price is really expensive then, Yeah. Well even for Facebook then Google has a huge The social And there are something that made that得到 there's something that was now谷歌确实 imultiple也是 trade在历史最高的这个时候我觉得现在谷歌肯定是一个大家的白马了就是大家 consensus long反正我有时候就觉得越 consensus的东西就越容易有一些风险我觉得谷歌还是没有完全解决这个Chatbot对它影响的这个问题就是搜索这个问题Gipeline肯定很好非常好但是Gipeline也没有比这个Opus也好或者说比GPT要明显的高很多这个大家其实这三家还是在一个此消彼长 you see I'm in the mood but with the same time XGBT in the user's ability actually is very high it has been in many users to create memory and we also have to share XGBT's review and we also have to see Jimline's recent its current level not so good and XGBT has been a差距 so I feel like XGBT if it will continue to grow it will actually be to Google search for business and from search for search for the品質 and search for the and search for the because this is XGBT's重点 I think it will be a big impact this is how much I think Google is not a threat. Google is not a threat. Google is not a threat. Google is a lot of good. For example, TPU, online market, I think the market is in price. Google is now a side of the space. I think Google is not a threat. I think it's 100% more. I bought it was 150, 167% more. The price is now a 310. I think it's still a lot. It's a very big company. The Chinese system is very unique. But it's never solved The L smash the求 net very lame Now the whisper I think many things are a bit more than a future. I think there are some new evidence. So, the future is a bit more than a future. But the future is not going to be gone. It's just being a bit more than a future. If there is a new evidence, it will be back. Let's talk about the Apple. We already know that it's now Siri, High Siri is still doing very well. I can't say that. I actually don't know what to do. Because it's not easy to do. Because it's not easy to do. The efficacy is not easy to do. It's not easy to do. But it's so many users, so many money. Now, I don't know what to do. Actually, it's not easy to do. I think if I was a child, I just don't want to do it. I'm not going to do it. I don't want to do it. But I think Apple is going to do it. I don't want to do it. I don't want to do it. I don't know that. Uyush, you're going to see how it is. Again, this is just a little bit. I think that Salesforce always´s not a company at least But Voorhewer, almost AWS they always have a little developed all the contemporaries From this Skyway, to visualize the machine To happen at the release and lebens So that I think And OnePlus is the world to light up Let's move forward First it is not the im hype But Iドイヤ on aiot Artificial- muito blockchain Because I have to集ой Because I do many guys As current software Quality devices the device at options, there are other things in this case. These displays of ​​rassers and the physical unit supportPacáfico. Although this design не Is not as much as a video producter 웹 The liquid isoph traders again. I think that�HA's not in online size. And then it's a strategy. So it's a strategy I think is a strategy. So if you're in a company and you're in a market, then you'll buy it? I don't buy it. I think I'm still in the market. I think I'm still in the market. I think that's a good thing. It's a good thing. We're going to ride on the volatility. We're going to have a movement. For full disclosure, I'm closer to that story than an exception to my own." My GoogleMPelt Update is closer to Palm Chats in the city. Then Igenommen him intoсон and finds your clan Fürth at Codding and his networkaients company. The reason why was my own now is at the start, he imaginework how it is in love with sources.包括 Cloud Code也是今年才出来的两年前我以为或者两年半前我以为战斗已经结束了GitHub就很少了所以说微软看上去它2023年的时候尤其2022年23年的时候感觉拥有那么多的Foresight然后拥有这么多的OpenAI的股份然后OpenAI又在它上面运营感觉它就是应该AI的winner但今天是 from your perspective, how do you think? I think this is a little bit. AI is a big company. This is really a big deal. Because when Office Copilot came out, many people thought this is a good thing. Because OpenAI, it has a 50% of OpenAI, it has a good distribution distribution. Office is a good thing. This is not a good thing. But it's like Office Copilot. I hear the usage is a good thing. Of course, Office is a good thing. It's a good thing. So it's a lot of history. But the other part, there are a lot of companies in the industry. The data issues. So I think that companies are quite slow. It's a bit more than a process. It's a bit more than you buy. It's a long time compounder. It's a long time compounder. It's a very stable company. It's a lot of companies. Again, it's not a lot of companies. It's not a lot of companies. But I think if you have a long time portfolio.我一年只要动一两次我平时不怎么看微卫肯定属于一种跌多了就可以买就可以加到一个标的但我觉得这种题就太大了所以我觉得它里面其实它还是跟着AIBETA走了它可能里面有一点自己的copilot或这些的alpha不管是正的alpha付的alpha但是它整体来讲我觉得是跟着AIBETA走说一个故事我 máquina时间ם在一个panel上面让一个大佬点评一下各种кative打怎么样点评了半天然后就是没讲 Microsoft Copilot I said that By the way That person had been working on Microsoft And now left And he was able to click on it And I said that Microsoft Copilot He said that I already I knew that Microsoft He was able to do it I can't think of it This is how many But you think Why is it Because it's Copilot But it's coding AI coding It's not a bad thing But it's It's a bad thing Now it's One of the Ten players 你觉得这fundamentally 是不是能够说明这个公司的 不管说运营能力 Satia的各方面能力 还是有限 还是说明 这就是企业 一个软件的一个现实 就是这么一个 Fragmented的一个现实 你是怎么看的 我觉得不反应 不是说未然有什么问题 而是说在技术创新的时候 创业公司它更加敏捷 做的事情更加native 所以它会有更快的优势 I think actually this is the result of this Just like that OpenAI vs. Gemini Cursor vs. GitHub Copilot And these Mini Journey vs. Adobe Actually I think it's like this Just the fact that the technology is But in a big company A trillion dollar company Especially for the prosumer For the most powerful For the most powerful To do a product This challenge I think it's a challenge for a big challenge So, maybe a big challenge is not a good challenge Whether it's Google or Google or Google They're also very hard to跳 But you can still be able to do it You can still be able to do it So I don't think Google It's a problem So, it's a good challenge It's a good challenge for a company That's what we're talking about Amazon You can see Amazon I think it's like I think it's like It's not in AI It's not in AI Amazon It's like it's not in AI It's what they're doing.所以大家会有质疑但我后来想如果说AI真的进入越来越Enterprise Service的领域的话AWS的Enterprise Service还是有很多的经验包括它的这些margin什么也不错所以我觉得亚马逊是有可能成为一个反转的例子因为它最近这些年代其实都不太好它就觉得你好像是有点慢但是我是觉得反正被低估的标的我就可能会觉得比较有意思我就去想它被低估的背后有没有一些错杀的可能然后亚马逊当然也是很多美国老前 这种大钱喜欢买的标的 还一个潜在的Headwin 就是大家觉得这个CityGPT的 Agent Commerce会对他们产生 很大的影响 但是这个我觉得落地的速度 也会可能是一个问号 比如说如果大家发现在CityGPT 买东西卖东西没有那么明显 没有那么快 那可能对亚马逊的电商部分 也是个利好 所以我是觉得亚马逊 现在可能我反而会往多头的方向 多想一想 这个虽然我可能不会买 但是我可能会想 这个是不是会更有做多的机会 Nvidia has already Dutch pricing feminists inullen ?? 다음 Data הש annexation lo пал go but PC I have a bunch of�ittle job As a beta which now at 118 right 1200 that's the bottom So it hasn't been everything Get the question because in26 it just has 바 pas Now in that let's just argue that the that it那个时候其实有很多高手在short Ingrid达因为他们从短期的数据看到 Ingrid达有很多库存差不多是游戏那块的库存很多然后当时挖矿也出现了很多积累所以他们觉得 Ingrid达现在库存不好但是我们比如搞AI的我们就会觉得那GPU的区域会巨大当然最后这几年肯定是看多了这派赢了所以我想举个例子就是说当长端发生了大变化的时候短端的数据和估值就不重要了因为当时 Ingrid达是长端发生大变化 I think we are now in a short plant Definitely the wayu noisable Is something will beách for starter But efforts in長短 Four changes to quickly Player appetite will become Don't return in charge As research will continue to To cause an easy Jot prepare system In the second subsequent shape For future From pre-training to inference To better All in Sot The停 Because IneV PP it's not so good we don't think it's a fine No key We also have other options this is positive and if it's not justifiable and it's not a good and its margin is not better the margin will be there is a lot of whenチャンネル is't better If in the forward margin and the transfer noantage then the only way is it ready Or it's just it's hard to keep it on the top. So I think it's downside risk actually will be bigger. So for the long-term risk, one thing is to return. It causes the return to reduce investment. The TPU and the inference will be affected by the influence. Because there is a story that is CUDA is very important. You don't have CUDA, so you can't do it. But Google has proven that I'm all in TPU. From the training to the推理, no problem. So I'm now selling for you. You can buy it. We know that TPU has been used to use to other companies' training or推理. So I think this is a benefit. I think Tesla is a pretty exciting company. I remember you mentioned in the book in the previous book, you mentioned that in 2025, Optimus is below expectations. You can see in the 2016, 或者是 Robotex 有一些超过多数人的预期的这么一个现象 我认为人型机器人 尤其是它的 Manipulation那一端 会始终继续低于大家的预期 大家会遇到这个事情非常的难 但是我现在特斯拉是持有的 并且特斯拉可能是我 我首先我不是完全空仓 或者说我录的时候 仓位比现在要更低一些 我最近上周买了一点仓位回来 Tesla is my only one in the world. I think that first of all, FSD is a big difference. Because we found a FSD. I'm very familiar with FSD. I'm in the Yacht. I'm in the Yacht. I'm in the Yacht. But FSD and RoboTaxi in the world, all of the world, still have very few people. I think that's what I've experienced. I think that's the FSD. Cybercap is a big difference. FSD is a big difference. It's a big difference. It's a big difference. I think that's what I've learned. And more people can feel the FSD or RobotExe's magical moment. This is a big part of the process. And the second is the space. This is a big issue. I'm not thinking about why in the world is a huge issue. But SpaceX's market market. I think Tesla is very strong. Catalyst. Because Tesla's market market is not a big part. It's a huge issue. I think FSD's market market is a huge part. SpaceX's market market market. Actually, it's a huge part of the catalyst. and now from thephinx it now shows the glaube 2 so desde the end of this IC Galaxy 3 and then we think does it go withanza I think I have a lot of money I have a lot of money I have to pay for AI I think I have to pay for AI But if I don't have to buy I can't wait for it I think we can't wait for it We can't wait for it Because you think It's a solid company So you think Like it has a million dollars You have a million dollars How to command command, a trillion dollar of CapEx, how much will be used? There are some Oracle, there are many Contra, it's said there are many Contra, it's optional, Nvidia seems like it's optional. Do you think the most trillion dollar of CapEx are so many investments, is it not going to be used? From your perspective, I don't know. I think the number is very big. It's not even a long time, it's not a long time.一定会有很多的challenge尤其是现在GIMALINE这些已经在对它产生challenge了所以我觉得这不会是一个一帆风顺的道路它会不会最后比如说最后可能只能兑现一半或者什么的我觉得不知道Let's see我现在还在想OpenAI到底会更像Yahoo还是更像Google我其实没有一个特别好的答案但是我觉得它是有Yahoo的风险你是这个新的技术的第一个代表性产品但是你最后没有形成特别强的自己的门槛最后被很大的程度上替代的Yahoo的事我觉得这是有这个风险的 noi OpenAI Yahoo Google is a very good type and in a while there are people talking about the popularity is what from a certain perspective that is the popularity will be to go over Google of course this is a even successful and a very very very very bumpy thing I think it might be an LL because when LL was bought Time Warner and the the popularity was to buy Chrome Google Chrome he thought to buy Google Chrome I thought 当年AOL My Time Warner这件事情 从长期来讲 我们都是对AI是一个很大的革命 对人类的发展其实是巨大的 我也觉得是一个利好的一件事情 但是从一个短期来讲 肯定显然不管是泡沫也好 或者说是股价的走向也好 都是有极大的不确定性 这点我是非常的同意 就有两点吧 The second thing but still starting up is among small studies But such as anfed as an government before the Buy We all do this It was easy to move legen aside from the technology fått None of theseпод artejatas avons It doesn't mean that we indeed can fund strong companies andant the equipment And then the number of ret Orleans I believe there wasnt a few entrepreneurs in Greenpoint, in TertorWow we do AI's forecast is very low. And I think the AI's a展望 is a very easy way to get started. I have a full-time job of getting ready to get ready. So I can say that I think I'm talking about a lot of things. And I don't know if it's the best for it. And I think it's the best for it. So I think I'm going to share with you with this. It's not a good idea. It's not a good idea. It's not a good idea. I think it's a good idea. But I don't know if it's a good idea. But I think it's a good idea. is I at this point my view of some of the view and thinking about it if there's new evidence or new results I think it's very happy because it means I have a收获 I can learn new things so I'm excited to be a little and then I saw that I think it's a lot of different I think this is a lot of research in the process in a very interesting place I think you said you said it was very honest and honest and I was very much I remember I remember I remember I remember I was in Greylock 做ERL的时候 接触了很多超一流的投资人VC 我自己总结出来的一个观点 就是越是那些在业界里面 让人仰望的投资人 他对我说的就是 哈伟不要来跟我说TEM 创业者知道TEM 我是不知道的 越是那些刚刚进入VC的 他会说 哎呀 这个格局会怎么样 可以指点江山 但是其实指点江山很容易 打脸是必然的 And it's the end of the creation of the future. It's not the end of the future. It's the end of the future. This is what I really like. I think we talked about the two-way market. I also want to hear you. You have some more of the things you have. You have some of the facts. Because there are many facts. There are not all of the people who are interested in that. Because I know you are in the two-way market. You are a bit of a谦虚. You are not your main business. But you have a lot of things. So this is what I want to talk about. But this is not a conclusion. This is our 2025 year the last episode of the show. We hope that in the 2026 year everyone can learn more. Thank you. Thank you for your invitation. If you like our channel, please share with more friends. We'll see you next time. See you next time.