Silicon Valley Walks the Jagged Edge

发布时间    来源
Episode 设置


登录已过期或未登录,无法修改。请先登录后再试。

摘要

Links: - Patreon (Support the channel directly!): https://www.patreon.com/Asianometry - X: https://twitter.com/asianometry - Newsletter & Podcast (available through Stratechery Plus): https://asianometry.passport.online/ - LinkedIn (feel free to connect): www.linkedin.com/in/jon-y-asianometry

GPT-4正在为你翻译摘要中......

中英文字稿     

来自波士顿的问候!我在这里待上几天,见见朋友和一些有趣的人,然后再飞回家。在这之前,我刚刚去过旧金山。所以这一次,我又录制了一期“Vibes”视频,分享我学到和讨论的一些事情。刚离开美国没多久,我就再次踏上旅程,原因我稍后会解释。即便是在硅谷待一个月也是很充实的,我总是满脑子新话题、关心的事情以及即将面临的问题。在今天的视频中,我做了一次小型旧金山之旅,并分享了一些新的想法。我此次旅程的原因是受邀在Semicon West的测试视野研讨会上发表主题演讲。演讲题为《左移与右移,ATE的演变》,涵盖了自动化测试行业从20世纪50年代创立至今的历史。
▶ 英文原文 ⏱
Greetings from Boston, where I'm spending a few days meeting friends and interesting people before flying back home. Prior to this, I was in San Francisco. So here we got another Vibes video report of things I learned and talked about. I know, so soon. This US trip comes just a few weeks after my last one, and I will explain why in a bit. A month in Silicon Valley is still a lot, and as always, I come away thinking on new topics, concerns, and looming issues. In today's video, I made a small trip to San Francisco and share a few new thoughts. The reason I made this sojourn was to give a keynote at the Test Vision Symposium at Semicon West. The talk was titled Shifting Left and Right, the Evolution of ATE, and it covered the history of the automated test industry from its creation in the 1950s until today.

这段话翻译成中文如下: 这有点像我之前做的关于ATE行业的视频,只不过加入了2000年后行业发展的一些趣闻,涉及移动和人工智能行业的测试发展。遗憾的是,这次讲座没有录制下来。但如果时间允许,我将来会把它制作成新的视频。这次演讲对我来说是一个极大的荣誉,是让我记忆深刻的事情之一。感谢Test Vision的主办方邀请我。接下来谈谈展会。我想先说一下,我错过了2026年Semicon West展会的第一天开放日。我是在10月14日上午到达的。刚下飞机,在展厅漫步时有些疲惫,所以这可能会影响我对展会的印象。
▶ 英文原文 ⏱
It was a little similar to the video that I did on the ATE industry with additional anecdotes on the industry after the 2000s, spanning test developments for the mobile and AI industries. Sadly, it was not recorded. But if time permits, then I will turn it into a new video in the future. It was a huge honor for me to give this talk. One of those things that I will remember for a long time. Thank you to the Test Vision organizers for the invitation. Okay, the show. I want to preface this by stating that I missed the first open floor day of Semicon West 2026, arriving October 14th in the morning. Having just come off the plane, I was a bit tired when walking the floor. So maybe that is affecting my impressions of the show.

说完这些,我觉得2026年的展会就像一个时代静静落幕的夕阳。在去年的Semicon West展会上,有人告诉我,他们计划今后将在亚利桑那和旧金山之间轮流举办。但现在看来,这个计划不会实现。有很多人跟我说,这将是旧金山的最后一场展会。Semicon West 2027将会在2027年3月于亚利桑那举办。是的,我知道,这距离上一次展会只有六个月。我猜他们想找一个空闲时间举办。因为Semicon West和另一场大型展会时间重叠了,那就是2026年在圣何塞举办的OCP全球峰会,而显然OCP抢走了风头。半导体和晶圆制造设备行业已有70多年的历史,这些领域的竞争早已尘埃落定,格局已定。
▶ 英文原文 ⏱
Having said all that, I get the sense that the 2026 show felt something like the quiet sun setting of a different era. During last year's Semicon West, I was told that they were going to alternate between Arizona and San Francisco going forward. But it seems like that plan will not happen. More than a few people have told me that this is the last SF show. Semicon West 2027 will be in March 2027 in Arizona. Yeah, I know. That is like six months since the last show. But I'm guessing they want a time that is more free. You see, Semicon West overlap with another major show. The 2026 OCP Global Summit in San Jose. And OCP clearly stole its thunder. The semiconductor and wafer manufacturing equipment industries are 70 plus years old. The battles have been fought and won. It is stratified.

另一方面,OCP 涉及现代战争,比如推理硬件等。而在这些领域,比如冷却和网络,SPECT 的关键发明者尚未决定出来。该展会门票售罄,所以我没有去。但有访客告诉我,现场气氛十分火爆。所以,最好换到一个更宽敞的地方。此外,美国半导体行业已经从过于以圣克拉拉的英特尔为中心,转移到美国内陆的新生产基地。我完全预计这个展会不久后会移到德克萨斯州。尽管距离我最近一次参加 Hot Chips 大会才几周时间,但气氛已经发生了变化。硅谷内外的情绪之间出现了有趣的分化。
▶ 英文原文 ⏱
OCP, on the other hand, covers modern era battles like inference hardware. And there, the SPECT inventors for essential areas like cooling and networking have yet to be decided. The show was sold out. So I didn't go. But a few visitors told me that it was popping. So, better to move somewhere with a bit more legroom. Moreover, the American semiconductor industry has shifted from an overly Intel-centric axis in Santa Clara to new production sites in the American hinterlands. I fully expect the show to make its way to Texas before long. Even though it has been just a few weeks since my most recent trip to Hot Chips, the mood has changed. There is an intriguing split between the moods inside and outside Silicon Valley.

外面的情况一直不太稳定。金融界的人感到不安。在我拜访的那一周,《金融时报》报道了Anthropics的财务数据。这些数据对大多数企业来说都很不错:160亿美元的收入和优秀的毛利率。但市场认为这些还不够,这导致AI股票价格陷入了动荡。我记得上次我谈到了2027年下半年可能出现的计算过剩问题。但是对于供应链来说,现在是时候开始考虑2028年了。是的,我知道,我之前说过计划在2027年底实现50吉瓦的计算能力。这很可能在现有电力基础设施下实现。
▶ 英文原文 ⏱
Outside, things have been unsteady. The financial people are feeling antsy. The week that I visited, the Financial Times had reported Anthropics numbers. The numbers would be great for most any business. $16 billion revenue with excellent gross margins. But they were deemed insufficient by the market. Throwing AI trade stock prices into turmoil. I know last time I talked about the possibility of a compute glut in the second half of 2027. But for the supply chain, it's now time to start thinking about 2028. Yeah, I know already. I said earlier that the idea is to have 50 GW of compute by the end of 2027. This is probably possible with existing power infrastructure.

他们打算如何使用这些计算能力,我稍后再说。总之,这就是2027年的计划。至于2028年的新增产能,估算数值有所波动。但目前给我的感觉是还会增加40-50吉瓦。不过这只是我听到的数字,别引用,因为这个数字还不确定。由于供应链需要大约一年到一年半的时间来生产所有这些设备,所以最终决定可能要到2027年年中到年初才会做出。也就是大概六到八个月之后。我们是否会在接下来的六个月内有足够的信息来说服供应链,让我们在2027年新增的40-50吉瓦基础上再增加50吉瓦呢?
▶ 英文原文 ⏱
What they are going to use all that compute for, let me get back to in a bit. Anyway, so that is 2027. For 2028, the incremental capacity estimates fluctuate. But the target vibe I get so far is like 40-50 GW additional. Don't quote me on that, it's a number that I heard. The number also remains in flux. Since the supply chain needs about a year or year and a half to produce all that equipment, final decisions are probably not going to be made until mid to early-ish 2027. Which is like six to eight months away. Will we see six months away from now enough to convince the supply chain that we should have another 50 GW on top of the 40-50 GW we get in 2027?

2028年的50 GW预期是否太低?这意味着从2027年开始,增长将会平缓甚至停滞。然而,半导体方面的支出预期增长速度要快得多。拉姆研究公司在其2026年10月的财报电话会议上表示,他们预计2026年用于晶圆制造设备的支出将达到1500亿美元,高于今年年初时的1400亿美元。到2027年,他们预计这一支出数字将超过2000亿美元。高盛则预计2027年的支出为2180亿美元,2028年为2810亿美元。而且,一些公司的内部预测数字甚至更高,达到3000亿美元或更多。
▶ 英文原文 ⏱
Is the expectation of 50 GW in 2028 too low? That would be flat or mild growth from 2027. Expectations for semiconductor spending are rising far faster than that. Lamb Research recently said in their October 2026 earnings call that they expect wafer fabrication equipment spending in 2026 to end at like $150 billion, up from $140 billion earlier this year. In 2027, they expect a spend number that is over $200 billion. Goldman Sachs, for their part, estimates 2027 spending to be $218 billion and $281 billion in 2028. The whisper numbers of other firms go even higher, to $300 billion or more.

我在Semicon West会议上听到了些传言。我不断听到有人提到晶圆制造设备公司的招聘冻结。虽然这些信息不太连贯,但我确实听到了一些。牛鞭效应的一个关键特点是,需求信号随着在供应链中传递,会被放大。我觉得这些WFE的预测显示出一个正在形成的牛鞭效应。而公司们正在为此做准备。不管怎样,这些都是比我薪水更高的人需要解决的问题。对于硅谷的建设者来说,生活仍在继续。
▶ 英文原文 ⏱
Now, there is a rumbling that I've heard at Semicon West. I keep hearing passing mentions of hiring freezes at the wafer fabrication equipment companies. It's a patchy signal, but I hear it. A key characteristic of the bullwhip effect is that a demand signal tends to amplify as it travels up the various levels of the supply chain. I think these WFE projections show a bullwhip effect in the making. And the companies are bracing for it. Anyway, those are questions for people with much higher salaries than mine. For the builders inside Silicon Valley, life continues.

我想问题是,为什么你要这样做?我从许多人那里反复听到的一个短语是“锯齿边缘”。我可能无法用这个解释来完全表达它的意思。但大意是,今天的语言模型(LLMs)在某些方面非常出色,但在其他方面表现不佳,这就是“锯齿边缘”概念。哪些方面表现突出并不是直观的。而正是因为这个“锯齿边缘”,创业公司才应该开发自己的特殊模型,比如世界模型或者用于处理特殊情况的模型。
▶ 英文原文 ⏱
But I guess the question is, why are you doing this? A phrase that I have heard repeatedly from various people as part of their answer is the jagged edge. I'm probably not doing it justice with this explanation. But the concept is that today's LLMs have this jagged edge where they're much better at certain things than they are at others. And what those things are, are not intuitive. The jagged edge is why a startup should build their own special models. Things like world models or models for handling special cases.

人们普遍认为,像Anthropic或OpenAI的大型语言模型在编程、解决Navier-Stokes方程或各种数学问题上表现出色,甚至可以破解政府网站。然而,如果让它们偏离核心任务,它们的表现就可能大打折扣。这种现象被称为"边缘效应"。我听说这个术语也被用于描述中国的开源模型。它们在某些方面似乎很强,但在其他方面则明显不足。此外,还有一种普遍观点认为,并非所有的那些数以千兆瓦计的电力都能按时进入市场。由于电力生产严重短缺、美国的劳动力问题、数据中心建设的阻力、政治风向的变化等多种因素,似乎不太可能按计划实现目标。
▶ 英文原文 ⏱
The anthropic or open AI LLMs might be whizzes at coding, be able to beat Navier Stokes or every math problem out there, or hack a government website. But put them outside of their golden path and they supposedly fall far short. That is the jagged edge. I've heard the same term be applied to the Chinese open source models. They seem to be quite good at a few certain things, but sorely lacking in others. Moreover, there is an outstanding belief that not all of those gigawatts will actually hit the market on time. Between profound power generation shortages, American labor, data center nimbyism, shifting political winds, and the like, it feels unlikely that things will actually hit on schedule.

出于这些原因,模型训练初创公司仍然在大力投资计算能力。我遇到的一些最聪明的人也在大力关注计算能力。他们相信每增加一分的计算能力都会比现在显得更加珍贵。但最合理的做法是通过筹集股本资金来尽量减少债务,因为无法偿还债务是公司倒闭的原因之一。虽然也有人指出,如果初创公司没有购买足够的计算能力,就可能会因为规模不如竞争对手而失去竞争力,也可能因此失败。
▶ 英文原文 ⏱
For these reasons, model training startups are still going long on compute. Some of the smartest people that I can meet and talk to are going long on compute. They believe that every incremental ounce of compute will become even more valuable than it already is today. But the most sensible way to do that is to raise equity and minimize debt because missing debt payments is how companies die. Though others have pointed out to me that a startup can also die if it does not buy enough compute, thus allowing competitors to out-compete them on scale.

所以,也许仅仅筹集1亿美元或其他金额来购买芯片是不够的。你还应该在这些芯片的基础上借更多的钱,这样你才能买到更多的芯片。如果不这样做,那你实际上不是在真正的做多,而是“结构性做空”。听到这话时,我笑了,我觉得这本来是个幽默的说法。通过这次旅行,我觉得我应该和一些在Frontier Labs基础上发展的较小公司交流。他们是生态系统中的关键部分——是Anthropic和其他大型实验室瞄准的商业用户。
▶ 英文原文 ⏱
So maybe it is not enough to raise $100 million or whatever to buy chips. You should then be taking out debt on top of those chips so you can buy even more chips. If not, then you are not truly long. You are in fact "structurally short". I laughed when I heard that. I think it is meant to be humorous. With this trip, I felt that I should be talking to some of the smaller companies building on top of the Frontier Labs. They are a critical part of the ecosystem - the business users that Anthropic and the other big labs are targeting.

他们是如何经营业务的呢?让我感到惊讶的是,他们非常依赖前沿实验室通过订阅提供的补贴。早在2026年年中,一些研讨会发布了一项发人深省的研究,发现ChatGPT和Claude的订阅相比API定价提供了更多的价值。花200美元的订阅计划,你可以得到大约2000到14000美元的API价值。资金紧张的初创公司知道这一点。因此,他们选择购买多个订阅,并在它们之间切换,以完成所需的开发工作。这些是充分利用每一个提供给他们的代币的高级用户。
▶ 英文原文 ⏱
How are they running their businesses? What strikes me is the dependence on subsidies doled out by the Frontier Labs through subscriptions. Back in mid-2026, seminars published some thought-provoking research finding that the ChatGPT and Claude subscriptions offer way more value than API pricing. Buy a $200 plan and you get something like between $2,000 and $14,000 in API value. Startups low on cash know this. And so load up on multiple subscriptions, switching between them to do whatever development work they need to do. These are the power users that leverage every token that is offered to them.

当重置限制被宣布时,使用代币尤其困难。我想我不应该感到惊讶。毕竟,这些都是试图快速成长以求稳定的初创公司,他们会利用每一个机会。不过,我觉得,如果他们突然或逐渐失去这些订阅补贴,将会受到严重影响。我也听说一些初创公司甚至将面向客户的产品与这些订阅绑定,依靠补贴的代币来运营。
▶ 英文原文 ⏱
Burning tokens especially hard when limit resets are announced. I guess I should not be surprised. After all, these are startups trying to hack their way to stability and they will take every advantage they can. But I get the sense that if they were to be cut off from those subscription subsidies, suddenly or even gradually, they will be severely impacted. I have also heard of some startups have even wired up their customer-facing products to these subscriptions. Running off subsidized tokens.

这听起来像是一种极其糟糕且严重扭曲的经营方式。我认为任何人都能看出这种方式是不可持续的。但我知道它确实存在。当我想到补贴使用时,我会想到优步、Lyft 和其他拼车公司。我记得在2015年,优步试图占领全球市场,他们和Lyft在乘客补贴上花费了数亿美元。2016年,优步泄露的财务数据显示,顾客仅支付了实际乘车成本的41%。我认为这是一个综合数据。不过,说实话,这确实是综合数据,其中包括当时中国大陆市场的荒唐情况。在中国,优步给司机的奖金高达车费的三倍。去掉这一部分,我想情况可能更合理一些。总之,我不认为这些AI订阅服务有那么大的补贴。
▶ 英文原文 ⏱
That sounds like an absolutely horrible and egregiously distorted way to run a business. I think anyone can see how unsustainable that is. But I am aware that it happens. When I think about subsidized usage, I think about Uber, Lyft, and the rideshare companies. I remember back in 2015 when Uber was trying to take over the world, they and Lyft were spending hundreds of millions of dollars in rider subsidies. In 2016, Uber's leaked financials showed that its customers were only paying 41% of the actual cost of their trips. I think it was a blended number. But to be honest, that is a blended number. That included the ridiculousness that was then going on inside mainland China. Where Uber China paid out driver bonuses up to three times higher than its fares. Take that out, and I think maybe things are a bit more economical. Anyway, I don't think these AI subscriptions are subsidized to that extent.

在这里可以进行一些价格歧视,特别是针对高价计划,就像OpenAI最近推出的Pro 500计划一样。这些AI初创公司似乎正在为这种情况做准备,它们的产品经济基于API定价,并探索编排机制。比如,基于更智能的专有模型之间的代理和基于开源模型的子代理之间的互动。也许子代理甚至可以本地运行?我们拭目以待。最近,我一直在思考数学方面的事情。即使在旧金山的科技圈之外,AI对数学领域的影响也是众所周知的。数学从业者正在认真考虑他们的世界可能在短短几个月内被彻底改变。在某些方面,这让人联想到AI对围棋或国际象棋选手所造成的影响。
▶ 英文原文 ⏱
And there can be some price discrimination to be done here with higher price plans, as OpenAI has already started with their recent Pro 500 plan. And these AI startups seem to be preparing for such an eventuality by basing their product economics on API pricing and exploring orchestration. Like between agents based on smarter proprietary models and sub-agents based on open source models. And perhaps maybe sub-agents run locally? We shall see. In recent days, I have been thinking about the mathematics thing. Even outside of the San Francisco bubble, what AI has done to the math field is well known. Math practitioners are reckoning with the possibility of their world being thoroughly dynamited over the span of just a few months. In some ways, it reminds of what happened to the chess or Go players when the AIs came for their games too.

幸运的是,那些行业顺利转型为体育项目。运动员们成功转型为粉丝感兴趣并愿意支持的运动明星和公众人物。或许同样的事情也会发生在数学家身上。数学竞赛可能会成为一个小众领域。我有点怀疑,但谁知道呢,历史上更奇怪的事情也发生过。 我的猜测是数学领域本身可能会发生变化,对不同的人会有不同的影响。像陶哲轩这样的资深教授应该会安然无恙。年轻的博士们应该也会适应,因为他们有时间去调整。最受影响的可能是那些处于中间阶段的人。他们已经承诺将数学作为职业,但还未建立起自己的声誉。
▶ 英文原文 ⏱
Fortunately for those guys, their industries pivoted to becoming sports. The players smoothly transitioned into athletes and personalities that fans are interested in following and supporting. And maybe that becomes the case for the mathematicians too. Mathletes might become a niche thing. I kind of doubt it, but who knows. Weirder things have happened in history. My best guess is that the math field itself probably changes, and it will affect people differently. The older professors, like Tao, will be fine. The younger PhDs, I think, will be fine because they have the time to adjust. The ones probably most affected are those in the middle. Those who have committed to math as a career, they haven't built a reputation yet.

无论如何,抛开那些内容不谈,我一直在思考的问题是为什么会这样。OpenAI 和 Anthropics 似乎在这方面投入了相当多的开支。据估算,Navier-Stokes 数学问题花费了大约 650 万美元,按 API 令牌价格计算,可能相当于三个 200 美元的订阅。我知道 Frontier Labs 拥有大量的计算资源。但即便如此,这么多钱仅仅用于一篇营销博客文章似乎也太多了。这一切的经济动力是什么呢?我想 Irrational 以前也问过同样的问题。我一部分觉得这可能仅仅是出于营销或者对游戏的热爱。 如果演示模型的能力能够让估值提高哪怕1%,理论上讲,这就可以收回所有的 API 成本。有时候,你确实会因为对游戏的热爱而去做些事情。
▶ 英文原文 ⏱
Anyway, setting that stuff aside, the question that I have been pondering is the why. OpenAI and Anthropics seem to be throwing quite a lot of spending at these. The Navier-Stokes math problems cost about $6.5 million estimated at API token prices, which is maybe three $200 subscriptions. I know that the Frontier Labs have an embarrassment of compute. But even this feels like a lot just for a marketing blog post. What is the economic impetus for all this? I think Irrational has asked the same thing before. Part of me thinks this really is just marketing and/or the love of the game. If demonstrating the model's capabilities helps boost the valuation just 1%, then that, in theory, pays back all the API costs. And sometimes you really just do things for the love of the game.

有几个人建议说,Anthropic和OpenAI如此专注于数学,因为这与RSI(递归自我改进)有关。这与风险条件无关。也不是指Relational Software Inc.,即拉里·埃里森的公司在改名为甲骨文(Oracle)之前的名字。RSI指的是递归自我改进。提到RSI时,你可能会想到一个自动运行的循环,不断进行改进并提出新想法。就像是数据中心内的一个无人工厂。虽然最终目标似乎是那样,但目前的目标似乎比较谦逊。这个谦逊的起点似乎是让大型语言模型(LLM)来证明一个假设的正确与否,就像数学证明一样。一旦这个工作可以稳定完成,LLM可以被用来自动化地处理大量由人类生成的假设,证明其真伪,并将学习到的知识反馈到模型中。
▶ 英文原文 ⏱
But a few people suggested that Anthropic and OpenAI are focusing so much on math because it ties to RSI. Again, not the risk condition. Or Relational Software Inc., the name of Larry Ellison's company before change to Oracle. Recursive self-improvement. When you think of RSI, you might conjure up some image of an autonomously running loop that is constantly making improvements and coming up with new ideas. A lights-out factory inside a data center. And while that does seem to be the eventual goal, the current goal post seems to be more humble. This humble start appears to be getting the LLM to prove a hypothesis right or wrong. Like a math proof. Once this is consistently done, the LLM can then be employed to autonomously run through a big pile of hypotheses, which for now are human-generated, proving things true or not, and letting the learning be put back into the model.

根据这一理论,RSI(研究科学智能)不仅仅与大型语言模型(LLM)有关,还包括进行LLM训练实验的自动化。我喜欢这个想法。毕竟,这些LLM最初都是基于数学的,对我来说,这些公司更加投入于解决数学问题是合理的,因为这对他们未来的研发计划至关重要,而不是为了炫耀或营销。尤其是在营销没能在人们中间引起积极反响的时候。我还在思考另一个附带理论,那就是这些公司正在利用这项工作,通过Lean工具进行芯片设计的形式化验证。这听起来很有趣,但我在这个领域的知识还不足,因此无法对其真实性做出明确判断。
▶ 英文原文 ⏱
According to this theory, RSI is not only about LLM. But automation for doing LLM training experiments. I like the idea. Those LLMs are all math in the first place, right? It fits for me that these companies invest in solving math problems more because it is a critical part of their future R&D plans, and not for flexing and/or marketing. Particularly when the marketing isn't delivering warm, fuzzy feelings amongst the people. There is a side theory that I am also pondering that they are leveraging this work in Lean for the formal verification of chip designs. That sounds interesting, but I am not yet educated enough in this field to make a solid judgment on whether it is real.

无论如何,如果RSI的猜测是真的,那么这可以解释Anthropic和OpenAI如何在未来一年预期使用他们大量的计算资源。他们会用来进行RSI。在我之前的视频中,我在餐巾纸上做了一个大致的计算,假设训练和推理的比例为50/50。由于推理无法经济地吸收那么多千兆瓦,我感觉某些地方一定有问题。解决这个难题的方法是,训练和推理的比例实际上过多偏向于推理。结果表明,其实训练和推理的比例更像是75/25。所有这些训练之所以可以进行,是因为自动化实验,即RSI。
▶ 英文原文 ⏱
Anyway, if the RSI speculation is true, then that can explain how Anthropic and OpenAI will expect to use their tsunami wave of compute in the coming year. They are going to use it to do RSI. In my earlier video, I did a back of the envelope on the napkin that assumed a split between training and inference of 50/50. Since inference cannot economically absorb that many gigawatts, I felt that something had to be wrong somewhere. The way out of the dilemma is that the training-inference split puts too much compute onto inference. Rather than an even 50/50, it turns out to be more like 75/25 training to inference. All that training can be done because of this automated experimentation, i.e. RSI.

我一直明确表示,我不提供投资建议,你可以去别处寻找这样的服务。但我想对此发表一些看法,并以此结束。最近,专业投资者普遍感到市场的东西越来越贵。我不是第一个告诉你,人工智能交易是美国屈指可数的仍然有效的经济趋势之一。除了技术领域,特别是人工智能交易之外,还有什么值得投资的呢?各个领域似乎都有问题。几乎所有行业,几乎所有国家,你都能发现令人担忧的事情。各方面似乎都有问题。
▶ 英文原文 ⏱
I have always made it clear that I do not deliver investing advice. You can go somewhere else for that. But I do want to talk on this and close this way. There has been a feeling amongst professional investors as of late that things are getting expensive. I am not the first person to tell you that the AI trade is one of the few headlining economic trends in the US that still works. Outside of technology, and the AI trade in particular, what is there that is good to invest? Everything out there has warts. Almost any sector. Almost any country. You can go and find something concerning. Things just seem broken.

同时,AI行业的投资已不再像过去那么容易。像芯片、功率、光学和内存这样的股票,回头看都显得理所当然,但它们的股价已经飙升,现在处于高估值状态。接下来会发生什么呢?在私营领域也是如此。这些初创企业在计算上进行长期投资所需的资金费用不菲。尽管投资兴趣似乎仍在,但窗口能持续多久并不完全明朗。我依然持乐观态度,我宁愿保持这种乐观,而不是悲观。
▶ 英文原文 ⏱
Yet at the same time, the AI trade isn't nowhere as easy as it used to be. Plays like chips, power, optics, and memory - all of which now look obvious in hindsight - have done their stock surges and now sit at elevated valuations. What is next? The same goes in the private space. Funding these startups as they go long on compute is costing a pretty penny. And while the appetite still seems to be there, it is not entirely clear how long the window stays open. I continue to be optimistic. I prefer to be as such rather than being a downer.

因为我一想到如果一切崩溃,我们将要面对什么样的事情,就感到不寒而栗。不过,是的,氛围确实变了。硅谷仍然沿着险峻的边缘前进,但现在掉下去似乎比以前更容易了。请订阅频道,注册Patreon,我下次再见大家。
▶ 英文原文 ⏱
Because I shudder to ponder what kind of things we have to look forward to if it all comes crashing down. But yes, the mood has changed. Silicon Valley still marches forward along the jagged edge. But falling off now somehow feels easier than before. Subscribe to the channel, sign up for the Patreon, and I'll see you guys next time.