State of AI 2025 with Nathan Benaich: Power Deals, Reasoning Breakthroughs, Real Revenue
发布时间 2025-10-30 12:01:46 来源
这份文字稿记录了FirstMark的Matt Turk对Astrid Capital创始人Nathan Benneish关于其“2025年AI状况”报告的采访。该报告可在stateof.ai免费获取，是对AI领域的一份全面概述。
他们首先讨论了AI研究的进展，尤其是在**推理**方面。Benneish强调，2025年标志着一个重大飞跃，他指出，大约12个月前，他们只看到了系统展现推理的早期迹象。现在，进展令人震惊，尤其是在可验证的领域，例如数学。他提到了AI在国际数学奥林匹克竞赛中获得金牌，以及模型被用作生物学和科学领域的AI共同科学家，帮助解读疾病新靶点等例子。他指出，这种进步使AI能够应对即使是聪明的人类也无法解决的挑战，不再仅仅是随机的。
对话转向了**机器人技术和“行动链”**，这是一种机器人行动前规划步骤的推理过程。Benneish指出，机器人技术正在经历一场寒武纪大爆发，语言模型正在指导机器人行动。他以CERIAC为例，并表示它确实有效，不仅仅是研究成果。他认为机器人技术的辉煌时刻已经到来，尤其是在工业环境、物流和仓储领域。虽然人形机器人吸引了人们的注意力，但他预测其发展路径将类似于自动驾驶，会出现零星的成功，但在长尾效应方面面临挑战。
谈到**AI的商业化**，Benneish断言它终于赶上了炒作的步伐。他强调了顶级AI公司的收入增长，现在已经达到数百亿美元，以及小型AI公司的快速增长。他引用了Ramp的数据，显示AI订阅的续订率有所提高，客户在AI产品上的支出也大幅增加。他指出，现在有44%的美国企业为AI工具付费，个人使用率更高（95%），反映了组织内部的“影子AI”现象。
他们探讨了**利润率之争**，指出人们对当前基于token的定价模式感到担忧，在这种模式下，不同的客户无论用例如何，都支付相同的价格。这种模式可能会导致垂直AI产品的毛利率较低。然而，Benneish指出，一些公司在其AI系统上实现了非常高的利润率（70-90%）。
关于**AI泡沫**的问题，Benneish承认存在局部泡沫。他将纽约金融圈以泡沫为中心的观点与旧金山更为乐观的情绪进行了对比，后者得益于人才涌入和基础设施建设。然而，他承认该行业正在进行巨额投资，以Nvidia为中心存在循环交易。他还提到了大型公司转移债务以支持其AI野心。他认为地缘政治和宏观经济因素为AI行业带来了脆弱性。
在讨论**物理现实和基础设施**时，Benneish强调电力已成为新的瓶颈。他引用了建设和运行基于AI的数据中心的高昂成本。各公司都在争夺能源，与未来的核电站签订协议，并在短期内依赖燃气轮机。电网限制正在推动数据中心向能源丰富的国家转移，这引发了地缘政治担忧。他提到了数据中心冷却所需的水资源，以及这种做法的可持续性。
对话转向了**Nvidia的主导地位**。Benneish认为它将继续保持领先地位，因为其芯片在研究论文中的使用率很高（90%）。他提到了AMD的崛起以及Broadcom通过为谷歌的TPU定制ASIC以及与OpenAI的交易而获得的复苏。尽管有竞争者，但Nvidia的股票表现明显优于其竞争对手。
他们谈到了**主权AI**，这是由希望控制其AI能力的国家推动的，导致全球范围内的巨额投资。Nvidia甚至将其作为一条新的产品线进行营销。虽然Benneish认为这是一种营销策略，但它与再工业化努力以及各国需要访问AI的需求相一致。然而，他认为对AI的访问可以被关闭。
谈话涉及**开源AI**，包括美国版的中国模型。Benneish认为，OpenAI正是在这种压力下才转向开源的。
他们还深入研究了**AI生态系统中的集中度**，指出Nvidia的大部分收入来自超大规模厂商和新云厂商。这种集中度反映了从个体创新向需要大量资本的大规模努力的转变。他指出了公司哲学为适应新的规模和AI的金融化而发生的演变。
对话触及了**安全、监管和数据权利**。Benneish认为监管未能跟上发展的步伐。他指出，AI已经发展到可以轻松访问，而没有安全参数。
他提到，不良数据集可能会带来高昂的惩罚，Throttic支付了15亿美元的庭外和解金，以避免未来的影响。
关于**网络安全**问题，Benneish指出了模型在网络任务和网络犯罪方面的能力日益增强。
最后，他们讨论了**AI Agent**，强调了它们在垂直产品、搜索、咨询、编码和科学推理方面的潜力。
他还提到，在某个时刻，SaaS可能会变成Agent。
最后，Benneish提出了预测，包括计算/AI建设将变得具有政治色彩，AI的科学发现将获得诺贝尔奖，各国将放弃实现AI主权的希望，转而寻求中立。
This transcript features Matt Turk from FirstMark interviewing Nathan Benneish, founder of Astrid Capital, about his "State of AI 2025" report. The report, available for free at stateof.ai, is a comprehensive overview of the AI landscape.
They begin by discussing advancements in AI research, particularly in **reasoning**. Benneish highlights that 2025 marks a significant leap, noting that about 12 months ago, they only had early signs of systems showing reasoning. Now, progress is astounding, especially in verifiable domains like mathematics. He mentions examples like AI achieving gold medals in the International Math Olympiad and models being used as AI co-scientists in biology and science, helping decipher new targets for disease. He notes that the progress has allowed AI to tackle challenges that even smart humans couldn't, moving away from being purely stochastic.
The conversation shifts to **robotics and "chain of action,"** a reasoning process where robots plan steps before acting. Benneish notes that robotics is experiencing a Cambrian explosion, with language models informing robotic actions. He uses CERIAC as an example and said it does genuinely work and is not just a research thing. He believes robotics' big moment is here, especially in industrial settings, logistics, and warehousing. While humanoid robots attract attention, he predicts a path similar to self-driving, with isolated successes but challenges in the long tail.
Moving to the **business of AI**, Benneish asserts that it's finally caught up with the hype. He highlights the revenue growth of top AI companies, now in the tens of billions of dollars, and the rapid growth of smaller AI firms. He cites data from Ramp showing improved retention rates for AI subscriptions and a significant increase in customer spending on AI products. He points out that 44% of US businesses now pay for AI tools, with personal usage even higher (95%), reflecting a "shadow AI" phenomenon within organizations.
They explore the **margin debate**, noting concerns about the current token-based pricing model, where different customers pay the same price despite varying use cases. This model can lead to low gross margins for vertical AI products. However, Benneish points out that some companies are achieving very high margins (70-90%) on their AI systems.
The **AI bubble** question is addressed, with Benneish acknowledging localized bubbles. He contrasts the bubble-centric view in New York finance circles with the more optimistic sentiment in San Francisco, driven by talent influx and infrastructure build-out. However, he acknowledges the gargantuan sums being invested in the industry, with circular deals centered around Nvidia. He also mentions the offloading of debt from big companies to fuel AI ambitions. He considers geopolitical and macroeconomic factors create vulnerabilities for the AI sector.
Discussing the **physical reality and infrastructure**, Benneish emphasizes that power has become the new bottleneck. He cites the high cost of building and running AI-based data centers. Companies are scrambling for energy, inking deals with future nuclear plants and relying on gas turbines in the short term. Grid limitations are driving offshoring of data centers to energy-rich countries, raising geopolitical concerns. He mentions water usage needed for data center cooling and the sustainability of this.
The conversation moves to **Nvidia's dominance**. Benneish believes it will remain the leader, citing its prevalence in research papers (90% use Nvidia chips). He mentions AMD's emergence and Broadcom's revival with custom ASICs for Google's TPUs and a deal with OpenAI. Despite contenders, Nvidia's stock performance significantly outperforms its competitors.
They touch on **sovereign AI**, driven by nation-states wanting control over their AI capabilities, leading to massive investments worldwide. Nvidia is even marketing this as a new product line. While Benneish sees this as a marketing play, it aligns with reindustrialization efforts and the need for countries to access AI. However, he believes access to AI can be switched off.
The talk touches on **open source AI**, including the US equivalent of the Chinese models. Benneish believes that OpenAI was pushed into open source due to this.
They also delve into **concentration within the AI ecosystem**, pointing out that a significant portion of Nvidia's revenue comes from hyperscalers and neoclouds. This concentration reflects the shift from individual innovation to large-scale endeavors requiring significant capital. He notes the evolution of company philosophy to meet the new scale and financialization of AI.
The conversation touches on **safety, regulatory and data rights**. Benneish feels the regulatory is failing to keep up with the progress. He notes that AI has evolved to be able to be easily accessed with no safety parameters.
He goes into the fact that bad data sets can have high penalties, with Throttic settling a $1.5 Billion settlement out of court to avoid future ramifications.
On **cybersecurity** concerns, Benneish notes the rising capabilities of models in cyber tasks and cyber crime.
Finally, they discuss **AI agents**, highlighting their potential in vertical products, search, consulting, coding, and scientific reasoning.
He also mentions that at some point SaaS may become agent.
To close, Benneish offers predictions, including that computing/AI buildout will become politically charged, scientific discovery by AI will earn a Nobel Prize, and countries will abandon the hope of achieving AI sovereignty in favor of neutrality.
摘要
Power is the new bottleneck, reasoning got real, and the business finally caught up. In this wide-ranging conversation, I sit down ...
GPT-4正在为你翻译摘要中......
中英文字稿 
I think you can't ignore the fact that the sums of money going into this industry are truly a gargantuan. Circularity of these deals is interesting. Things can flip quite quickly. One gigawatt of a data center for AI based, it costs $50 billion in Catholics. On an annual running basis, it costs between like another eight to nine, maybe even $11 billion to run. Companies are trying to do deals with anybody who has any capacity. In the short term, what many GPU data centers are getting powered on is just gas turbines. While that we've come for the point where we just want like an AI that works on our computer, but like to get that, you need to have so many more powerful systems collaborate with you.
我认为你不能忽视这个行业的大笔资金投入，这是一个真正庞大的数字。这些交易的循环性非常有趣，事情可能会快速变化。一个用于人工智能的数据中心的装机容量为一千兆瓦，建设成本高达500亿美元。每年的运营成本大约需要额外的8到9亿美元，甚至可能达到110亿美元。公司正在尝试与任何有能力的人进行交易。在短期内，许多GPU数据中心主要依靠燃气轮机供电。当我们希望在自己的电脑上运行人工智能时，已经走到了这个地步，但为了实现这一点，需要更多强大的系统进行协作。
I was the first time you had a system that could show its reasoning. Since then to now, the progress is pretty astounding. Hi, I'm Matt Tert from FirstMark. Welcome to the Mad Podcast. Today, I'm excited to welcome back Nathan Benneish, founder of Astrid Capital, to discuss the 2025 edition of his State of AI report, a must-read on whether field really is. We cover a lot, including why power is a new bottleneck, reasoning and channel action robotics, and the business reality, revenue margins, and what it means for builders and investors. Please enjoy this great conversation with Nathan. Nathan, great to have you back. Thanks for having me.
我是第一次看到一个可以展示其推理过程的系统。从那时到现在，进展令人惊叹。你好，我是FirstMark的Matt Tert。欢迎来到Mad播客。今天，我很高兴再次欢迎Astrid Capital的创始人Nathan Benneish来讨论他2025年版的《AI状况报告》。这份报告是在该领域是否真有意义的必读材料。我们讨论了很多内容，包括为什么算力成为新的瓶颈、推理和渠道行动机器人的话题，以及商业现实、收入利润率，这对创建者和投资者意味着什么。请欣赏与Nathan的精彩对话。Nathan，很高兴你能回来。谢谢邀请我。
The State of AI 2025 is out, and as always, it's essential reading for anyone who's serious about understanding AI. This year, it's a 312 slice of goodness, a bit of a big ear in AI. Every year, I try to cut it down a little bit, but this year, it just felt like we were sharing it with various sub-communities of the AI, of the AI community, and each time we did that, the robotics folks would be like, hey, it's a little bit light on robotics. Can you add some more? And then we send it to the bio folks and we're like, why don't you say this paper or that paper and hence the inflation? Amazing.
《2025年人工智能现状》已经发布，一如既往，这是任何认真想了解人工智能的人必读的材料。今年的报告有312页内容丰富的信息，可以说是人工智能领域的“巨著”。每年我都试图简化一下内容，但今年感觉我们正在和人工智能社区的各个子群体分享这些信息。每次分享时，机器人学领域的人总会说，嘿，关于机器人技术的内容有点少，可以多加一些吗？然后生物技术领域的人又会问，为什么不提及这个论文或那个论文？所以内容就不断增加。太精彩了。
All right, so we're certainly not going to cover everything in this conversation, obviously. As always, the report is available in its entirety for free at stateof.ai. So we're going to riff on some of the most important topics and ideas in the report, but obviously people can go and check out the report directly for more. All right, so starting from the top in the world of research, you mentioned that 2025 was your reasoning got real. So how far have we come in the last 12 months? That's pretty far. About 12 months ago, so we had, I think the very early inklings of it with O1 preview, potentially around like this time last year. And that was the first time you had a system that could kind of show its reasoning, show its stepwise process to get a more complicated answer.
好的，我们显然没办法在这次谈话中涵盖所有内容。这份报告的完整版本可以在 stateof.ai 免费获得。所以我们会讨论一下报告中一些最重要的话题和观点，但显然大家也可以直接去查看报告以获得更多信息。好，从研究领域的顶层开始，你提到2025年是"推理变得真实"的一年。那么在过去的12个月里，我们取得了多大的进展呢？进展相当大。大约在12个月前，也就是去年这个时候，我们可能通过O1预览初步见识到了这方面的苗头。这是系统第一次能够展示自己的推理过程，可以一步步展示其如何得到更复杂的答案。
And this has generally been the dream in AI for a long time. And since then, to now, I'd say the progress is pretty astounding. One of the areas that the progress has kind of unveiled itself is in mathematics and other verifiable domains where you can like explicitly say, yes, the system works or doesn't work. And you know, we saw gold medals on the International Math Olympiad by a couple of labs, including OpenAI and DeepMind. That area probably with, if you asked experts again, how long have it taken? We've probably been a decade. Then in areas a bit closer to my hardened biology and science, we've seen reasoning models kind of be used as an AI co-scientist.
这一直以来都是人工智能领域的梦想。从那时到现在，我可以说进展真是令人震惊。其中一个进展显现的领域就是数学和其他可以验证的领域，在这些领域中你可以明确地说系统是否有效。我们看到一些实验室，包括OpenAI和DeepMind，在国际数学奥林匹克竞赛中获得了金牌。如果你问那些专家，这样的进展可能本来需要十年的时间。在我更熟悉的生物学和科学领域，推理模型已经开始被用作AI协同科学家。
So just as a human would be reading lots of papers, planning experiments, writing the experiments, and then doing data analysis. And then reformulating their hypothesis as a result. This example of models doing that in lieu of human, which is exciting because there's way too many papers to read. AI people kind of complained that it's like 50,000 papers a year and say in biology and chemistry and physics is probably an order of magnitude more than that. And so DeepMind has shown that you can integrate this kind of reasoning model to sort of decipher new targets for disease, new mechanisms that were actually also proven in a wet lab scenario post facto. We've gone from systems that were kind of dumps, stochastic parasts, and now they can solve pretty meaningful challenges that I'd say, even a smart human couldn't.
就像人类会阅读大量论文、设计实验、撰写实验方案并进行数据分析，然后根据结果重新制定假设一样，现在有一种模型可以代替人类完成这些工作。这令人兴奋，因为要阅读的论文实在太多了。AI领域的人抱怨说，每年大约有五万篇论文，而在生物学、化学和物理学领域，论文数量可能还要多一个数量级。因此，DeepMind展示了如何将这种推理模型整合起来，以识别新的疾病目标以及在实验室场景中得以验证的新机制。我们已经从简单的、随机的系统发展到现在能够解决一些有意义的难题，这些难题甚至连聪明的人类也可能无法解决。
And still in research, you talk a little bit in the report or a lot in the report about robotics and this evolution towards a system of action or channel action, going from channel thought to channel action. What's happening there? Yeah. I mean, it just is probably two years ago, robotics was kind of a dead end. Opening I had disbandled. It's robot team. It was famous for solving the Rubik's cube using locomotion with the hand. And so now robotics is probably going through a Cambrian explosion. There's so much excitement.
在研究中，你在报告中谈到了一些或很多关于机器人技术和向行动系统或渠道行动演变的内容，从渠道思维转向渠道行动。这方面发生了什么变化呢？是的，大约两年前，机器人技术似乎走到了尽头。OpenAI 解散了他们的机器人团队，该团队因用手解决魔方问题而闻名。但现在，机器人技术可能正在经历一场寒武纪式的爆发，激动人心的进展有很多。
And just as how language models informed biology, now language models are also informing robotics. So what you're referring to here is a sort of reasoning process for robots where a system is no longer just perceiving the environment and deciding what to act and so acting. But it's we've separated those steps. So now you have a reasoning model that looks at a task and tries to plan steps that a robot would need to do to execute that task and then passes that plan over to an actuator which goes and actually implements the plan. And that's what's called a chain of action. And here the Allen Institute was one of the first to really push this and various swiftly thereafter. Gemini also followed and we have some companies including CERIAC that are implying this into the real world. So it does genuinely work. It's not just like a research thing.
就像语言模型对生物学产生了影响一样，现在它们也在推动机器人学的发展。这里提到的是一种机器人推理过程，其中系统不再仅仅是感知环境并决定采取行动，而是将这些步骤分开了。现在有一个推理模型，它会观察任务并尝试规划机器人需要执行该任务的步骤，然后将这个计划传递给执行器去实际实施计划。这被称为行动链。在这一领域，艾伦研究所是最早推动这一概念的机构之一，随后Gemini也紧随其后。我们也有一些公司，如CERIAC，正在将其应用于实际世界。因此，这确实是可行的，不仅仅是一个研究项目。
So we think the big moment for robotics is upon us because we all collectively have been talking about this for a very long time.
我们认为机器人技术的重要时刻已经到来，因为我们大家一直在讨论这个话题已经很久了。
Yeah. Yeah. Well, I'd say it really is upon us in the industrial sector in logistics and warehousing kind of more constrained environments with very repetitive tasks. There is the sort of more holy grail of this kind of embodied human-like form factor and putting a model on that might even be the same model that's been used in warehousing. A lot of money is going into that but my personal bet is I think it's going to be the humanoid space is going to look much more like self-driving where we have some very good isolated demos but the long tail will kill you.
是的，是的。可以说，在工业领域，特别是物流和仓储这些比较受限制的环境中，这类技术已经到来了，尤其是那些重复性很高的任务。目前有一种被视为终极目标的趋势，即开发有人形的机器人，并在其上应用可能已经用于仓储的模型。很多资金都投入到了这方面，但我个人的看法是，这种人形机器人领域可能会像自动驾驶技术那样发展，我们可能会看到一些非常好的单独演示，但在实际应用中仍会面临许多困难和挑战。
We're pulling on to literally. Yeah. And so we're going to go through many false starts. I think this is just a start. Okay. Great. So a big year in robotics and reasoning for people listening to this. If you're interested in deep dives into reasoning and RL in the evolution of AI systems, we've done a bunch of great episodes recently with Sholto from Anthropic, Jerry from OpenAI and then Julian from Anthropic, if you curious to learn more.
我们正在逐步推进，是的，所以我们将经历许多失败的开端。我认为这只是一个开始。好的。今年对机器人技术和推理来说是一个重要的年份。对于正在收听的朋友，如果你对深入了解推理、强化学习（RL）以及AI系统的发展感兴趣，我们最近与Anthropic的Sholto、OpenAI的Jerry以及Anthropic的Julian进行了一系列精彩的节目。如果你有兴趣了解更多内容，请多关注一下。
Let's move on to the business of AI. You mentioned in the report that the business of AI finally caught up with the hype. What caught your attention in terms of facts that's in the last 12 months?
让我们继续讨论人工智能的商业发展。你在报告中提到，人工智能行业终于赶上了长期以来的炒作。在过去12个月中，有哪些事实引起了你的注意？
Yeah, a couple of them. Again, where we came from one or two years ago was just tons of money going into this segment, building models, a lot of usage but not clear where the revenue would come from. I think it was maybe OpenAI was making $50 million or something two years ago. That was very unclear how they would ever hit billions of revenue. Nowadays, I think if you sum sort of the top 20 or so major AI companies from the labs to the most popular vertical applications across them, they're making tens of billions of dollars of revenue.
是的，有几个这样的例子。回顾一两年前的情况，当时在这个领域投入了大量资金，大家都在开发模型，使用量很大，但收入来源还不明朗。我记得两年前，OpenAI可能只赚了大约5000万美元。当时很难想象他们怎么可能达到数十亿美元的收入。现在，如果把前20大AI公司，包括实验室和最受欢迎的垂直应用的收入加起来，他们的总收入已经达到数百亿美元。
You can look at the smaller scale companies which are growing from zero to 20 million or 20 million plus. As a group, they generally grow about 60 percent faster on a quarterly basis than non-AI companies. We've all seen the famous charts about ARR or non-ARR. It's unclear. But very steep curves for various coding companies. Perhaps most interestingly across a segment of 43,000 or so US customers, we work with Ramp to show the retention of subscriptions on AI products across this customer set has really improved marketly since 2022.
你可以关注那些从零增长到2000万或超过2000万的较小规模公司。总体而言，这些公司每季度的增长速度比非AI公司快约60%。我们都看过关于ARR（年度经常性收入）或非ARR的知名图表，这些图表通常非常陡峭，尤其是各种编程公司的。最有趣的是，在大约43,000个美国客户的样本中，我们与Ramp合作，数据显示自2022年以来，这些客户对AI产品订阅的留存率显著提高。
Run 2022 is around the 50 percent after 12 months. And now in 25, it's hitting around 80 percent. And the second stat in that analysis that was interesting was the total spend on AI products per customer kind of went up from $35,000 or so, maybe two years ago. Now it's around half a million dollars and it's predicted to hit a million dollars next year.
2022年的运营在12个月后达到了大约50%。而现在在2025年，这一比例达到了约80%。在这个分析中，另一个有趣的数据是每个客户在人工智能产品上的总支出，大约两年前是$35,000左右。现在约为$500,000，并预计明年将达到$1,000,000。
And you mentioned in your Ramp stats of 44 percent of US businesses now pay for AI tools. So they pay more, but there's a ton of businesses using AI.
你提到，根据Ramp的统计数据显示，目前有44%的美国企业正在为人工智能工具付费。因此，虽然他们支付更多，但有大量企业正在使用人工智能。
Yeah, exactly. And there might be some sampling bias slightly to what kind of companies use Ramp in the first place. Yeah, slightly more modern and tech forward companies. But a leading indicator I think of where things could go. And then you had your own survey right of 1200 AI practitioners.
是的，没错。最初使用Ramp的公司类型可能会有一些取样偏差。它们往往是更现代化和技术前沿的公司。不过，我认为这可以作为未来发展趋势的一个前兆。此外，你自己还进行了一项包括1200名AI从业者的调查，对吗？
Yeah, yeah. And what did that say?
好的，那它说了什么？
Yeah, that was, I was surprised. Obviously, biases more towards pretty well educated US European professionals. A lot of people in there have at least undergrad masters degrees, maybe even more. But it's like 95 percent of people use AI in their personal life and in their professional life. About 76 percent of people pay out of their own pocket for it. It's like 10 percent of people pay more than 200 bucks a month for it.
是的，那真让我感到惊讶。显然，这主要偏向于受过良好教育的美国和欧洲的专业人士。很多人至少有本科或硕士学位，甚至更高。不过，大约95%的人在生活和工作中使用人工智能。大约76%的人是自己掏钱使用的。大约10%的人每个月花费超过200美元。
And then looking at the organizations that they work at, it's like 70 percent of those organizations are spending a ton more or more than they did in the past on AI. The reasons that they gave for what, why they might not be spending more or what problems they have. It's like all the classic like new technology stuff. Like it's a bit hard to configure. It's, you know, I haven't really figured out the ROI yet because I need to do more customization. There's like some data privacy issues that I have. And I think all these things are kind of solvable. Like it's not rocket science. I just all these things. Yeah, it feels like we very much live this year in the world of shadow AI in companies where a mature reconcile is imperfect, but to reconcile your two stats 44 percent of businesses use AI yet 95 percent of people individually use AI.
然后查看他们工作的组织，会发现大约70%的这些组织在人工智能上的花费比以前多很多。对于为什么他们可能不会增加支出或者遇到什么问题，他们给出的理由都是经典的新技术问题，比如配置有点困难，我还没有真正搞清楚投资回报率，因为我需要进行更多定制化，还有一些数据隐私问题。不过我认为所有这些问题都是可以解决的，没有那么复杂。今年，我们仿佛生活在公司中“影子人工智能”的世界中，一个成熟的和解虽然不完美，但可以协调两个统计数据：44%的企业使用人工智能，而95%的个人也在使用人工智能。
Yeah. So there's a bunch of people as you're alluding to that use AI at work with that being officially as ours. Yeah. And I think there's still a big like education gap. I mean, there was a study bandied around a couple of weeks ago where you know, it said 95 percent of businesses like get no value from AI. Very, very controversial.
是的。正如你所提到的，有很多人在工作中使用AI，这在我们公司是正式规定的。我认为仍然存在很大的教育差距。大约几周前有一个研究在传播，称95%的企业从AI中没有获得任何价值，这个观点非常有争议。
Yeah. Yeah. But I think 95 percent is the number or like everything is 95 percent. But I think that there it turned out it was like not the models that were bad. It's like the implementations of them were not great. So I think there's just a big education gap for how how you should like, you know, update your view of your own day-to-day tasks and and apply what capabilities you know models have. And then and think about like, hey, should I be doing this task myself or can I farm it out to a model?
是的，是的。但我认为95%是一个关键数字，就好像一切都达到95%。但我觉得，问题并不是模型本身不好，而是它们的应用实现得不好。所以我觉得在如何更新我们对日常任务的看法，以及如何应用我们所了解的模型能力方面，存在着巨大的知识差距。然后，我们应该思考一下，"嘿，我应该自己做这项任务，还是可以交给一个模型来处理？"
And and there's definitely a delta of companies that really get this done well and others that are like basically clueless. What do you make of the margin debate as an investor and institute analyst? Maybe recap what that debate is and then what do you think about it? At a high level, basically, that the margin problem is for many, many customers of large model companies, their margins are basically dictated by how much the model vendor charges them for.
有些公司在这方面做得特别好，而有些公司则完全不懂这个问题。作为投资者和行业分析师，你怎么看待这一利润率的争论呢？能不能先总结一下这个争论的内容，然后谈谈你的看法？从总体上看，利润率的问题在于，许多大型模型公司的客户，他们的利润率基本上是由模型供应商向他们收取的费用决定的。
Now, here there's some issues because right now model vendors are charging the same amount per token. So if you're a hedge fund analyst and I'm a student, you know, your use case is clearly more financially valuable than mine, but we pay the same amount for the token assuming we use the same model. There are some use cases that are more reasoning heavy towards what we discussed before and they consume a ton of tokens. And the pricing that a customer pays for that product might not be fit for the amount of work the AI system is doing.
现在，这里有一些问题，因为目前模型供应商对每个token的收费都是相同的。因此，如果你是一名对冲基金分析师，而我是一名学生，你的使用场景显然比我的在经济上更有价值，但我们使用同一模型时支付的token费用却是一样的。有些使用场景需要进行大量的推理，这方面可能会消耗大量的tokens。而客户为此产品支付的价格可能并不适合AI系统所进行的工作量。
And so there are cases where these kind of vertical products are making gross margins of like 30% and sometimes they get worse with scale because you do have some edge users that like really pump the system and you can't like price discriminate or they have a manage to. And then you have some segment of model users that don't that have a both a paid plan and a free plan and it's not clear whether they include the costs of running the free plan in their gross margin.
因此，在某些情况下，这类垂直产品的毛利率可以达到大约30%。但随着规模扩大，情况有时会变得更糟，因为有一些活跃用户会极大地消耗系统资源，而你无法区分定价，或者他们找到了一种方法来管理。此外，还有一部分用户既使用付费计划也使用免费计划，目前尚不清楚他们是否在毛利中包含了运行免费计划的成本。
So sort of just look at their paid customers. There's some creative accounting standards going on there. And then you have the model vendors themselves and what is their margin. And I think what's interesting in the last year is you've seen CEOs of these model companies say, hey, if we if we basically look at serving financial analysis terms like a layer cake of like what revenue is generated by each vintage of model over time, it looks like prior models are profitable.
所以只需要关注他们的付费客户就行了。这里面有一些巧妙的会计处理。同时，还有模型供应商自身的利润空间。我觉得过去一年有趣的一点是，这些模型公司的首席执行官们开始说，如果我们从财务分析的角度，把每一代模型产生的收入看作是一层层的蛋糕，就会发现以前的模型实际上是盈利的。
So the amount of money we've spent to build them is less than the amount of money that we've generated with them over time, assuming a certain margin of inference cost. So really these these labs are like not not profitable because fastening more resources going into developing next generation systems than the prior ones. But as you and I both know there are companies here that are making very very good margins on serving their AI systems like 70, 80 sometimes 90% depending on the modality. And so like with everything the average number sucks.
所以，我们花在建造它们上的钱，比我们通过它们在一段时间内所赚的钱要少，当然这还要假定推理成本有一定的利润率。实际上，这些实验室之所以不盈利，是因为开发下一代系统需要投入更多资源。而你我都知道，有些公司在提供他们的AI系统服务上获得了非常高的利润率，有的高达70%、80%，有时甚至达到90%，这取决于具体的模式。所以，像所有事情一样，平均数的表现并不理想。
But like when you look at the best companies, it's really good. And just to drive it home, the companies using those models we're talking about the you know in part the all the what used to be known as thin wrapper. So the vendors that happen to be powered by those models. So the cursors, the wind serves, and all the whatever legal financial AI startups as examples. The other big debate in the business of AI of course is the bubble question.
当你看到最好的公司时，它们真的很出色。为了更好地说明这个问题，这些公司使用的模型，其中一部分过去被称为“薄包装”。也就是那些依靠这些模型驱动的供应商，比如说某些法律、金融领域的人工智能初创公司，例如光标（cursors）、风帆（wind serves）之类的公司。当然，AI商业领域中另一个重要的讨论是关于泡沫的问题。
What's your what's your take? Are we in an AI bubble? Are we not in an AI bubble? Yeah. I think like with most things in markets there are probably localized bubbles all over the place. And I think at a high level it's interesting in terms of vibes in who's calling bubbles and who's not like the finance crowd in New York is talking about bubbles a lot more than what we're talking about in San Francisco where their view is like this is the golden era of AI and a lot of things are working. We have so much more to to do you know compute buildouts are enabling us to experiment a lot faster. You know this huge flood of like talent that's built the consumer internet and cloud computing is moving into AI.
你的看法是什么？我们现在处于AI泡沫中吗？还是说并没有AI泡沫？我觉得，就像市场上的大多数事情一样，可能在很多地方都有局部的泡沫。从高层次来看，很有趣的是，谁在谈论泡沫的问题上气氛如何。在纽约的金融圈，大家谈论泡沫的频率要远高于我们在旧金山所讨论的。旧金山这边的观点是，我们正处于AI的黄金时代，很多事情都在奏效。我们还有很多事情要做，比如计算设施的扩建让我们能够更快地进行实验。有大量曾经构建消费者互联网和云计算的人才正在转向AI领域。
And with that is bringing a lot of optimization techniques and knowledge that AI researchers didn't have when they built the first generations of chat GPT etc. But I think you can't ignore the fact that the sums of money going into this industry are truly a gargantuan. You know like 500 billion to build Stargate and then you know a couple hundred billion here a couple hundred billion there like pretty soon it's real money. And then the and then like the circularity of these deals is like is interesting of course Nvidia is at the center of this and it has incentives to use its balance sheet to sort of spin the wheel faster.
随着这一切的发展，带来了许多优化技术和知识，这是当初AI研究人员在构建第一代ChatGPT等时所没有的。但我认为你不能忽视一个事实，那就是投入这个行业的资金真的是庞大的。比如，有5000亿美元用于建造Stargate，然后这里几千亿，那里几千亿，很快就成了真正的大笔资金。此外，这些交易的循环性也很有趣，当然，英伟达是这一切的中心，它有动力利用其资产负债表来加速这个轮子转动。
And then press more concerningly you have this sort of offloading of debt from big companies for example meta that raises tens of billions of dollars to fuel state of center ambitions but that doesn't sit on meta balance sheet. Some of this is like catnip to financial engineers but yeah it rests on certain assumptions that everything is going to keep going up into the right and that rates don't materially change and just given how like I suppose precarious various aspects of the economy are and how like sensitive geopolitics are things can flip like quite quickly.
然后，更令人担忧的是，大公司（例如 Meta）的债务转移。它们筹集数百亿美元来支持其核心目标，但这些债务并未计入 Meta 的资产负债表。这种操作对金融工程师来说犹如猫薄荷，但它的前提假设是经济会持续增长，并且利率不会发生重大变化。然而，考虑到经济的各个方面有些不稳定，以及地缘政治的敏感性，情况可能会很快发生变化。
But I think that's like the majorest the risk I'm less worried about is the stuff doesn't work because I think it does work. So there's some more question of timing to play back that the supply phase of the market is met by any quality strong or really stronger a demand side. Yeah there's that and then just the just the nuances of like the terms on the debt and what trigger events are where the later rates get reprised and then you know investors behave very differently once rates change and and flows of money can be quite like violent.
我认为这是最大的风险。不过，我对“产品不起作用”这一点反而不太担心，因为我相信它确实有效。所以，现在的问题更多在于时间的把握上，即市场的供应能否与优质或更强劲的需求相匹配。同时，还有一些细节问题，比如债务条款的具体内容、触发事件是什么以及何时重新定价利率，这都会影响投资者的行为。一旦利率变化，资金流动可能会变得非常剧烈。
It's interesting what you're saying about the you know the dichotomy between Wall Street and the West Coast also because when you think about it that's actually not that many fuel play AI companies in public markets right a lot of reaction is happening in private markets so effectively if you're a Wall Street slash hedge fund investor you invest in Nvidia you invest in max seven the spring much at right pound here see three AI maybe by soft bank first position open AI yeah pretty much like it is in direct or you invest in power and energy like related players call weave I guess but it's it's very very small so it feels like that's that tension as well.
你提到的华尔街和西海岸之间的对立关系挺有趣的。其实，当你仔细想想，公开市场上并没有那么多纯粹从事人工智能的公司。很多反应都发生在私有市场。所以实际上，如果你是华尔街或者对冲基金的投资者，你可能会投资于英伟达、Max Seven、Palantir、C3 AI，或者通过软银的首次公开募集投资OpenAI，对吧？基本如此。或者你可能会投资于与电力和能源相关的公司，比如Weave，但这种公司数量非常少。所以，这种紧张关系也显得很明显。
Yeah yeah but I think it's also the crowd that you hang out with yeah I mean and I think you live in a house in San Francisco is correct to other or three other AI genius correct correct or do you just consume the outputs of that of those kinds of conversations on Twitter and then try to like yes piece together your own world view.
是的，是的，不过我觉得这也和你交往的圈子有关。是吧？我理解你是住在旧金山的一所房子里，和其他两三位 AI 天才一起，对吧？还是说你只是通过 Twitter 上这种类型的对话来获取信息，然后尝试拼凑出你自己的世界观？
And I think the other part of this is like I don't think some of those individuals were really shilling that much anymore I think they do genuinely believe what they say and they are at the core face of the advances of these technologies and so if you know they've been saying for the last 50 times like hey this stuff is working there's lots of implementations we can improve or like things we can tweak or new experiments that'll yield better capabilities and that has happened at some point you got to be like maybe there right another aspect of this.
我觉得另一方面是，我不认为这些人还在过度炒作。我认为他们确实相信自己所说的，并且他们处在这些技术进步的核心位置。所以，如果他们在过去多次说“这些技术在发挥作用，有很多可以改进的应用或者可以调整的东西，或者通过新实验获得更好的能力”，而这些事情确实发生了，那么你可能就要开始相信他们是对的。
is fascinating to me is the again like the the the sort of dichotomy between some of the I call them the old guard and the newer younger kind of folks so you know from Rich Sutton to Yann LeCon to you know obviously Jeffrey Hinton a lot of those guys who are absolutely the Godfathers of the space and build this entire thing and our our steel extremely active today on top of everything say that LLM so just not gonna get us there or that we should just do everything with RL and then you know meanwhile the the the younger guys and they tend to be at places like Anthropic and OpenAns or maybe they do have an agenda but they're all saying well we just crutch into the surface of what we can do with those modern systems.
这段话让我感到很有趣的是，我称之为“旧派”和“新一代”之间的对立。比如，从Rich Sutton到Yann LeCun，当然还有Jeffrey Hinton，这些人在领域内绝对是奠基人，他们构筑了整个体系，并且仍在积极推动发展。他们中有些人认为大型语言模型（LLM）无法带我们达到目标，或者建议我们应主要依赖强化学习（RL）。同时，更年轻的一代，比如在Anthropic和OpenAI等公司的年轻人，或许有他们自己的计划，但他们普遍认为我们刚刚开始挖掘现代系统的潜力。
yeah yeah I think do both but yeah I think for me it's it's mostly what are kinds of new problems that you can you can work on and solve with this technology and I think it's becoming more popular to believe like the overhang of like problems we can solve in enterprise for consumers and science with the tools we have today is huge and so even if a lot of this compute build out doesn't go towards like dreaming up the next transformer architecture but goes into improving the Uniconomics of serving AI systems for everybody and makes it easier so you don't have to be like some prom master to elicit a behavior you want for your task I think that's not good.
是的，是的，我认为可以同时进行这两方面的工作，但对我来说，更重要的是看看你能用这项技术解决哪些新问题。我认为现在越来越多人相信，我们今天手中的工具在企业、消费者和科学领域可以解决的问题非常多。因此，即使很多计算资源不是用于构思下一个Transformer架构，而是用来改善AI系统的普及经济性，使得普通人不用成为某种专家就能实现自己所需的功能，我认为这不是坏事。
all right let's switch to the physical reality that this whole stacks it's on so infrastructure data centers energy you you you mention in the deck that's power has become the new bottleneck what is your sense of the state of play in the energy procurement game the the biggest step for me is one gigawatt of a data center for AI basically cost 50 billion dollars in cathex and on an annual running basis it costs between like another eight to nine eight to nine to maybe even eleven billion dollars to run and so when you have just you know casually a 10 gigawatt data center uh a sec a lot of money.
好的，让我们切换到这种架构所依赖的物理现实，也就是基础设施、数据中心和能源。在介绍中你提到，电力已经成为新的瓶颈。关于能源采购的现状，你怎么看？对我来说，最大的一个问题是，一个用于人工智能的数据中心需要一吉瓦的功率，建造成本大约是500亿美元，而每年运行费用则在80到90亿美元，甚至可能高达110亿美元。所以，当你拥有一个10吉瓦的数据中心时，那就需要非常多的钱。
and um and so one of the problems is like where does this energy come from uh you know traditionally it would be from i don't know coal or natural gas um potentially solar or ideally at some point the future nuclear and what we're seeing is it right now companies are trying to do deals with anybody who has any capacity so we can't to call some deals with future nuclear you know reactor companies then that would take maybe a decade or two decades to deliver um you know famously yeah that's uh google uh inking a ppa deal yeah we cfs to buy two hundred mega watts of electricity from a planned fusion plant so the the plant does not exist yet it does not exist yet.
翻译如下：
嗯嗯，所以其中一个问题是，这种能源来自哪里？传统上，它可能来自煤炭或天然气，可能还有太阳能，或者理想情况下，将来会有核能。我们现在看到的是，各家公司正在尝试与任何有能力提供能源的公司达成协议。所以，如果我们能和未来的核反应堆公司签一些协议，那可能需要十年或二十年才能实现。一个著名的例子是谷歌与CFS签订了电力购买协议（PPA），计划从一个尚未建成的聚变电厂购买200兆瓦的电力。这个电厂目前还不存在，还没有建立。
and then last year we documented the sort of restarting of uh three three male island um yes the nuclear facility which was controversial in the past um and um and then uh in the short term what many uh GPU uh data centers are getting powered on is is just uh gas turbines and because these can get set up a lot faster but that has other issues like they're super loud um and there's demand outside of the US for these things and so now basically US tech companies are paying more to repatriate like uh some of the supply that should have been shipped abroad.
去年，我们记录了三里岛核设施的某种重启，这个设施在过去曾引发争议。短期内，许多GPU数据中心启动时使用的是燃气轮机，因为这些设备可以更快地设置。然而，这也带来了其他问题，比如噪音非常大。此外，对这些轮机的需求在美国以外的地区也很旺盛，因此美国的科技公司现在为了能够拿到本该出口到国外的供应而付出更多的费用。
the other issues that the grid and like to what degree the grid can even tolerate data centers getting plugged into it now obviously like these turbines are off grid so it has some advantages but uh in china for example we do some analysis between like the uh you're between uh the US and china with regards to energy and china has a lot more like slack in its system to plug in um for any unpredictable uh demands in energy uh the uk famously it cannot really tolerate more data centers on its grid wrapping all this together is driving um some of the like offshoring of data centers towards uh energy rich countries whether that's the ua e uh or even uh Norway and uh and then with that comes a lot of like geo geopolitics of uh are these nations or friend or or potentially not and how do you ensure uh access to this regardless of your administration chains or other things.
其他问题涉及电网，以及电网在多大程度上能够容纳数据中心接入。目前显然，这些涡轮机是离网的，所以它有一些优势。但在中国，例如，我们对中美之间的能源状况进行了一些分析，发现中国的电力系统比美国更有余地来应对任何不可预测的能源需求。众所周知，英国的电网无法真正容忍更多的数据中心接入。将这些因素综合在一起，推动了一些数据中心向能源丰富的国家外包，比如阿联酋甚至挪威。随后，与此相关的地缘政治问题也随之而来，比如这些国家是我们的友邦还是可能的对手，如何确保即使在行政更迭或其他情况下也能获得这种能源接入。
so it yeah it's it's wild that we've come to the point where you know we just want like an ai that works on our computer but like to get that you need to have so many more powerful systems uh collaborate yeah and i was just looking for this slide as you spoke uh especially for united states versus china we're talking about the dramatic difference where. uh the capacity added in 2024 for the u.s uh if i read this correctly it was 48.6 geowatts or as china was 429 uh geowatts the other thing that's interesting is um the at least the states in the u.s uh or i she also internationally that um that are good for hosting data centers because there's energy typically are extremely dry and uh and we also chronicle the water usage that's needed for cooling uh of these data centers and so if your state super dry where do you get the water from uh is that actually gonna detract away from human populations that need the water then you have this whole like recycling of water which could potentially like you'll just like bad quality water getting circulated into the water system so the sustainability aspect to all of this seems uh extraordinarily important
所以，是的，我们已经到了一个令人惊讶的地步，就是我们只想要一个能在我们电脑上运行的 AI，但要实现这一点，就需要很多更强大的系统来协作。我刚才看了一张幻灯片，特别是美国和中国之间的对比，我们谈论的是一个巨大的差异。根据我的理解，2024 年美国新增的发电容量是 48.6 吉瓦，而中国则是 429 吉瓦。
另一个有趣的点是，美国的一些州或者一些国际地区因为能源问题，非常适合建设数据中心。但是这些地方通常非常干燥，我们还记录了这些数据中心用于冷却的水资源使用情况。如果某个州非常干燥，那么水从哪里来？这是否会影响到需要水的居民？接下来还有水循环的问题，这可能会导致劣质的水进入水系统。因此，这一切的可持续性显得格外重要。
yeah yeah yeah uh under discuss at least that's my my perspective is is that is that correct to do people actually care and do something about the sustainability aspect of this what a year or two ago big companies did make commitments to be green as of you know 2030 and then as soon as they started inking deals with uh you know nuclear companies and uh and and various energy providers for data centers all those commitments basically got like washed away um so it seems like maybe they care but the corporate priorities of making AI work of way outweighed the environmental constraints that's what's happened but I think again going back to like the politics side of things i don't think everybody's very happy about this particularly there's this like growth of nimbusum this like uh not in my backyard uh and uh and and i do think that uh people generally don't want to have a data center in their backyard uh and i think that's going to drive some of the political agendas like going forward whether it's in the u.s or or other countries so yes people do care about environmentalism companies that sort of washed out away but it's gonna i think it's going to come back
是的，是的，是的，嗯，正在讨论中的事情，至少从我的角度来看，是这样的吗？人们真的在乎并采取措施解决这一问题的可持续性方面吗？一两年前，大公司确实承诺到2030年实现环保目标。然而，一旦他们开始与核能公司和不同的能源供应商签订数据中心的合同，这些承诺基本上就被抛之脑后了。所以，似乎他们可能在乎这件事，但让人工智能运作的企业优先事项远远超过了环境限制。这就是发生的事情。不过，我认为回到政治方面的问题，并不是所有人对此都很高兴，特别是有一种“别在我家后院”（即“别让我这儿有这样的项目”）的心态增长。我确实认为，一般人并不希望数据中心建在自己家附近。我认为这将推动未来的一些政治议程，无论是在美国还是在其他国家。因此，是的，人们确实关心环保问题，而公司似乎忽视了这些承诺，但我认为这将会重新受到重视。
if we talk about infrastructure uh obviously we have to talk about Nvidia feels like it's been another extraordinary yeah last 12 months for uh Nvidia yep do uh see Nvidia continue to break away as like the undisputed number one in the market or do you think that sooner or later we're going to end up with a multi-silicon kind of world i think it's going to be 95 five that's a 95 percent is um farato you visited yeah exactly yeah um yeah so you know for context when we did the executive somebody last year we put Nvidia you know hit one trillion for the first time and now we have to change up to four trillion we look at uh all the open source AI research papers every year it's about 49,000 or so and then uh programmatically determine which chipsets are used in those papers so we know like hey an AI researcher is doing a study on uh i don't know some new model and in their in their experimental setup they say you know we train the model for x number of GPU hours on uh whatever chip and
如果我们谈论基础设施，显然我们必须提到英伟达。过去的12个月对英伟达来说似乎又是不平凡的一年。你认为英伟大会继续在市场上独占鳌头吗？还是你觉得迟早我们会进入一个多芯片的世界？我认为可能是95%对5%，其中95%是英伟达。去年，我们进行高管调查时，看到英伟达首次市值达到了一万亿美元，而现在我们已经要将这个数值调高到四万亿美元。我们每年查看所有开源的AI研究论文，大约有49,000篇左右，然后程序化地确定这些论文中使用了哪些芯片集。这样，我们就知道比如有一名AI研究人员在研究某种新模型时，他们实验设置中会写明：不管他们用什么芯片，我们训练模型用了多少个GPU小时等。
if you do that analysis you basically find that 90 percent of uh all papers make use of a Nvidia chip out of that same analysis we did find that AMD is sort of popping up a very little bit um apple silicon is as well i think it's just because the computer that macbook is getting so good that people are doing local training uh an experiment on their computer but a broadcom is experiencing a a resurrection of some sort as well yeah yeah exactly yeah it has uh i think it's maybe a decade ago they bought a company that now is kind of the internal team doing this custom asix for uh google's tpu and uh you know more recently they announced a deal with open AI also to do uh a custom chip
如果你进行那个分析，你基本上会发现，在所有的科研论文中，90%都使用Nvidia的芯片。在同样的分析中，我们发现AMD芯片也开始出现了一些踪迹，而苹果的芯片也是如此。我认为这是因为MacBook的性能越来越好，人们能够在本地计算机上进行训练和实验。此外，博通（Broadcom）似乎也在经历某种程度的复兴。是的，确实如此。大约十年前，他们收购了一家公司，现在这家公司成了内部团队，为谷歌的TPU制造定制的ASIC芯片。近期，他们还宣布与OpenAI达成协议，也进行定制芯片的开发。
in a high level it's interesting with the riser broadcom is basically GPUs have been that the the dominant chipset for a long time as the uh kind of nature of the neural network or other kind of AI system that you're running on the hardware was still changing very rapidly but as soon as you get to a point where there's some convergence on an architecture that's looks like it's stable and is revenue generating and developers are coming to uh sort of work on it and confirm that it is like the thing then you can flip towards doing a custom chip that's built to extract the most value out of that architecture and so the rise of broadcom basically tells you like there's strong forces that are saying like the transformer is the thing
从高层次来看，这很有趣，Broadcom 的崛起基本上说明长久以来，GPU 一直是主导芯片，这是因为神经网络或其他人工智能系统需要运行在硬件上的性质变化非常快。但一旦达到某个稳定的架构，并且这种架构能够产生收入，吸引开发者参与并确认其重要性时，你就可以转向定制芯片，以最大化地从这种架构中提取价值。所以，Broadcom 的崛起基本上表明有强大的力量在暗示：Transformer 是那个关键技术。
but at the end of the day like we also look at how would your dollar be best use as an investor if you wanted to bet on your companies and uh and in the graph in the in the report we look at sort of six of the major contenders uh to Nvidia and basically said you know if you bought Nvidia stock on the day of the announcement of all the like private uh rounds in these companies what would the value of your stock be it Nvidia versus these companies and uh if our car correctly it's basically 12x and Nvidia versus 2x and um in these competitors and the trend was roughly the same last year uh so I think I think it's a little bit of a different difficult beast to bet against.
归根结底，我们还会考虑，如果你作为投资者想在你的公司上下注，怎样才能最有效地利用你的资金。在报告中的图表里，我们考察了六个与英伟达相竞争的主要公司，并表示如果你在这些公司完成所有私募融资轮次宣布的当天购买英伟达的股票，你的股票价值会是什么样。结果显示，英伟达的股票价值增长大概是12倍，而这些竞争公司的股票价值大概是2倍。去年的趋势也大致相同。所以，我认为与英伟达竞争确实是一个相当困难的挑战。
Yes I was uh I was looking for that slide as you were uh speaking it's uh for anybody that looks at the report that slide 166 that says um what would have happened if investors had just bought the equivalent amount of Nvidia stock at that day's price the 7.5 billion would be worth 85 billion in Nvidia stock today 12x uh versus 14 billion 2x for its contenders and the contenders being Grox Rubra, Symbanova, Celestial, Graphcore and in China, Kambracon has uh experience you know a big run this is you know a private company that then would public on on Chinese stock exchange to build custom asix for for AI and and that was driven mostly by the geopolitical sort of zigzagging on policy with regards to exporting custom um Nvidia chips to China the H20 which at some point was deemed to be okay by the government and then deemed to be not okay uh but then okay if uh 15 to 20 percent of the revenue was passed back to the US government and then uh and then um someone in the someone high up in the US administration said you know our girls basically to ship the like crappy stuff to China and at that point the Chinese said like no thank you and yes effectively said no one can buy Nvidia chips and then Kambracon stock ribs and this is why way as well right that's the emergence of uh yeah separate Chinese uh full stack.
是的，我在你讲话的时候找到了那张幻灯片。对于查看报告的任何人来说，那张166页的幻灯片显示，如果投资者当时只购买等值的英伟达(Nvidia)股票，现在的价值约为850亿美元，相当于12倍的增长，而与之竞争的公司则是Grox Rubra、Symbanova、Celestial、Graphcore，以及中国的Kambracon等，其价值增长到140亿美元，仅为2倍。Kambracon是一家私营公司，后来在中国股票交易所上市，专门为人工智能开发定制的专用集成电路(ASICs)。
这种情况主要是由于地缘政治政策上的变化，涉及到将定制的英伟达芯片出口到中国的问题。H20芯片一度被政府认为可以出口，后来又被认为不行，除非15%到20%的收入返回给美国政府。不过，有美国政府高官表示，我们的目标基本上是把一些质量较低的东西运到中国，这时中国方面表示不接受，实际上禁止购买英伟达芯片，于是Kambracon的股票大涨。这就是为什么中国开始发展自己独立的全产业链体系的原因。
Yeah from the the models which we'll probably talk about that at that at some point in this conversation of open source but very much at the cheap layer of the so that's what you mentioned and then Huawei whatever the model is becoming the sort of default chip for the Chinese stock yeah yeah yeah and there's some interplay between the government trying to get deep seek and other labs to throw in their models on uh on Chinese chips and there's been rumors that this is why a lot of the new generations of Chinese models have slowed down um particularly deep seek like the people are waiting for for the next next R1 so I R2 and uh and allegedly it's because it's just hard to run it on Huawei.
在开源的讨论中，我们可能会谈到一些模型，这些模型主要是在成本较低的层面上，这也是您提到的。然后，华为的某种芯片正逐渐成为中国市场的默认选择。是的，是的，政府一直在推动让深度探索（Deep Seek）和其他实验室在中国芯片上运行他们的模型。据传，这就是为什么很多新一代的中国模型发展放缓的原因，特别是深度探索的模型，人们都在等待下一代的R1升级到R2的发布。有传言说，这主要是因为在华为的芯片上运行这些模型存在困难。
Two double click on on something that you mentioned a few minutes ago um talk about sovereign AI uh and uh what you've seen people do it seems to have been a big theme of the year you mentioned open AI in uh no way India and UAE what's happening in that world yeah that part of the world yeah yeah so the idea with sovereign AI is that nation states want to be uh at like able to control basically their fate with regards to AI so that's uh running models as training models having chips uh and um this is basically because you know nation states want to have control over their energy could control over their currency control over the infrastructure and AI is deemed to be kind of equivalent to those categories.
几分钟前你提到的一个话题，我想深入谈一下，就是关于“主权AI”的问题。你提到了一些国家，比如印度和阿联酋，这似乎是今年的一个大主题。在那个区域到底发生了什么？
“主权AI”的概念是指各个国家希望能够在AI方面掌握自己的命运。这包括运行和训练模型，以及拥有相关的硬件芯片。这是因为国家希望控制自己的能源、货币和基础设施，而AI现在被认为与这些领域具有同等重要性。
And so ever since the white house announcement of 500 billion in January uh various nation states have followed suit saying you know we have our own initiative and it's the tune of billions of dollars uh etc around the world and Nvidia has even started marketing this is like uh like a new kind of product line basically for its business that currently generates i think around 20 billion dollars worth um so it's it's real money um and so they're forming partnerships with various nation states uh to provide data centers there that are run locally um and uh and in theory that should give like countries comfort that uh their access to AI can't be turned off that's the idea.
自从白宫在1月份宣布5000亿美元的投资计划以来，各个国家纷纷效仿，表示他们也推出了自己数十亿美元规模的计划。在全球范围内，英伟达（Nvidia）甚至开始将其作为一种全新的产品线来推广，目前这一业务大约创造了200亿美元的收入，所以这是真金白银。英伟达正在与多个国家合作，在当地建立数据中心。理论上，这样能让国家放心，确保他们对AI的访问不会被切断。这就是该计划的核心理念。
I personally think it's a bit more of uh of an alignment between uh political agendas where particularly in the US it's really about reindustrialization like on-turing of key industries and building you know manufacturing and things like that which is i think one of the reasons why these AI dinner centers are getting rebranded as AI factories and so that's the political part um and that's getting aligned with um uh just the need of countries to get access to this technology.
我个人认为，这在某种程度上反映了政治议程之间的一种协调，特别是在美国，重点是再工业化，比如将关键产业带回国内，并推动制造业的发展。这也是为什么这些AI数据中心被重塑为AI工厂的原因之一。我认为这就是政治层面的部分，它与国家获取这种技术的需求相吻合。
So i think it's more marketing than it is like a real policy because in the day if you buy your stock from uh from from the US and you're not in all out of the US at some point then we'll just switch it off um and so part of this is like sovereignty washing i think and it also like oversimplifies the very interconnected nature and ecosystem aspect of of AI we're not just about the chip it's about uh the developer ecosystem um how you actually run it where your train data comes from um and uh and all the like infrastructure like data tools and and whatnot that that's it around this although uh that's where open source plays an important role right?
所以我觉得这更像是一种市场营销策略，而不是真正的政策。因为如果你从美国购买芯片而你不在美国境外的话，最终我们可以切断供应。因此，我认为这有点像是“主权清洗”，并且这也简化了人工智能非常复杂和互联的生态系统。人工智能不仅仅是关于芯片，它还涉及到开发者生态系统、运行环境、训练数据的来源，以及包括数据工具等在内的各种基础设施，而这也是开源发挥重要作用的地方，对吗？
If you get your AI from open AI and indeed uh you are a usli but you no longer are usli for whatever reason there's a risk that you'd be turned off but if you have uh sovereign data center and we is a bunch of like chips running and then you run open source on top of it like presumably you are safe which is then uh interesting because whereas the most popular open source coming from now yes China China yeah although interestingly uh I think. since you uh published the report there's been the announcement of a very large investment in reflection AI which is uh in New York and San Francisco based uh company that just raised two billion yep uh to build uh the US equivalent of the Chinese models in the inner world of alama and meta um that sort of uh got in a different direction yeah yeah I think this is fascinating um because part of the AI action plan uh that was published by the US government a couple of months ago now um you know articulated the need for having this American AI stack so they're moving away from like diffusion controls and more towards just buy our stuff.
如果你从 OpenAI 获取你的人工智能，并且你是一个usli，但无论什么原因你不再是usli，那么有被关掉的风险。但如果你有一个独立的数据中心，里面运行着一堆芯片，然后你在上面运行开源软件，那么你可能就安全了。这很有趣，因为目前最流行的开源软件来自哪里呢？是的，中国。虽然有趣的是，我认为自从你发布报告后，有一项非常大的投资宣布了，即在纽约和旧金山的一家叫 Reflection AI 的公司，他们刚刚筹集了20亿美元，旨在构建美国版的中国模型，即 Alama 和 Meta 的内部世界。这走向了不同的方向。我觉得这很吸引人，因为几个月前美国政府发布的人工智能行动计划的一部分，阐明了需要建立一个美国人工智能体系，因此他们正在从扩散控制转向鼓励购买我们的产品。
And then one of the other aspects of that action plan was around open source and like and and sort of leading in that direction and of course as you said like meta step back and into the fold came quen um I think 50% of all model derivatives um being downloaded from hugging face are coin based now um hundreds of millions of downloads partially because they're they come in very accessible shapes and flavors so as a result of that we sort of predicted in the report that uh uh that a major you know AI lab would lean back into open source to win um basically brownie points with the government and then the next day this financing happened oh really great great uh great timing yeah and I think you you send the report as well that your sense was that opening I was sort of forced for like a better term uh into releasing an open source model to be on the on the right side of history.
然后，该行动计划的另一个方面与开源有关，并朝着这个方向发展。当然，如你所说，Meta退出了这个领域，而Quen接替他们。在Hugging Face平台上，现在大约有50%的模型衍生产品基于Coin，下载量达数亿次，部分原因是它们形式多样且易于使用。因此，我们在报告中预测，一个主要的人工智能实验室将会重新投入开源，以此在政府面前赢得好感。随后，融资的消息就在第二天发生了，时机真的非常巧合。是的，我认为你在报告中也提到，你的感觉是OpenAI在某种程度上被迫发布一个开源模型，以站在历史的正确一边。
Yeah I think that's one of them and then the second one probably dovetails with their announcement with AMD and I say that because uh you know quite recently semi-analysis uh kind of published this benchmarking dataset with a run models on various clouds to sort of benchmark them um actually gpt oss like looks pretty good on AMD um and so one could imagine that um like there were some uh optimizations and actually were optimizations to to gpt oss where it runs nicely on AMD it has support from their framework from day one the uh the parameterization of the model is uh to the point where you can run it on a single AMD chip and then some other nuances to their attention mechanisms that they customize to make it work really good on AMD um and to the point around like the circular economy stuff that we discussed a little while ago um there's uh like another financial sweetener in the deal where uh open AI has warrants in AMD if the stock price hits 600.
是的，我认为这可能是其中之一。然后，第二个部分可能与他们与AMD的公告有关。我这样说是因为最近有一个名为semi-analysis的组织发布了一个基准测试数据集，旨在在各种云上运行模型以进行基准测试。实际上，GPT OSS在AMD上的表现相当不错。有人可以想象，GPT OSS在AMD上有一些优化，它从第一天起就得到了框架的支持。模型的参数化已经达到了可以在单个AMD芯片上运行的程度，还有一些其他的注意力机制的细节被定制以确保它在AMD上运行得很好。
至于我们不久前讨论的循环经济方面，还有一项财务上的激励措施，即如果AMD的股价达到600美元，OpenAI将获得AMD的认股权证。
And so you can see how there's a lot of like incentives this game of both like aligning with us government uh helping developers which is a good thing but also like helping one of your vendors uh improve which frankly it does need help and it and it should improve but also getting some financial sweetener as a result of that which could help you kind of make the flywheel spin faster and uh since we're talking about uh circularity uh talk about concentration as well so maybe as an echo to the conversation about the bubble uh a few minutes ago it does feel like this uh AI economy has a lot of uh depending on how you look at it from funky just carry things yeah yeah well a lot of Nvidia's revenue comes from uh the major uh you know hyperscalers or or um or neoclouds so you know what it's like meta like xai uh google amazon um then core weave and then a lot of coreweas revenue also comes from Microsoft on the way back.
你可以看到，这个游戏中有许多激励机制，比如支持美国政府、帮助开发人员（这是一件好事），还可以帮助你的一家供应商改善业务。坦率地说，他们的确需要帮助，而且确实应该改进。同时，通过这样做你还可以获得一些经济上的好处，这些好处可以帮助你加速“飞轮效应”。
既然我们在讨论循环性的话题，那么也可以聊聊专注度。或许可以呼应几分钟前关于泡沫的对话，即这个人工智能经济看起来似乎依赖于多个方面。Nvidia的大部分收入来自主要的大型云服务商，比如Meta、xAI、Google、亚马逊、以及核心的Weave公司。而核心Weave的大量收入又来自于微软等公司。这种互相依赖的关系让整个系统运作良好。
I think it's just this challenge with with AI progress that we've uh you know very meaningfully shipped from shifted from I think the gpt3 era to now of basically scale like rate limits your progress and uh it's no longer like a couple of people in a dorm room that can really build something uh transformational if they want to advance like AI capabilities it's really um it's really big boil end now um and so with that comes just different dynamics like you have to be good at capital raising you have to align yourself with uh with nation states you have to align yourself with wall street uh these are all I think contributing to the big vibe shifts that you've seen in in the culture of AI labs.
我认为现在与过去相比，人工智能的发展面临着新的挑战。我们已经从GPT-3时代走到了一个全新的阶段。在这个阶段，发展速度极大地限制了进步，现在已经不是几个人在寝室里就能够创造出变革性成果的时候了。如果想推进人工智能的能力，现在需要大规模的资源和支持。这伴随着一些新的挑战，比如需要擅长融资，需要与国家和华尔街建立合作关系。这些变化都是导致AI实验室文化发生巨大转变的重要因素。
What what do you mean by that what well you know for example there there were some labs like anthropic that were built you know to really push the safety agenda because you know if we didn't do that the rational uh went that you know we could lead to the extermination of humanity right um and I think quite recently like Daria was interviewed by mark banyoff uh just this past week and asked about like some of these data center buildouts. and uh and you know he said something along the lines of yeah there's a lot of money going into this a lot of cost but in the day the only thing that matters is revenue like i don't think you would have said that you know on the founding day of anthropic and you know it's it's just a reality that that the tablesticks in this game have changed and with that you know entrepreneurs have to update their priors and um and you know change our strategy a little bit and so we document some of this and like sort of the blooper section of the report
你在说什么？例如，有一些实验室，比如Anthropic，就是为了推动安全议程而建立的，因为如果不这么做，理论上我们可能会导致人类灭绝。而就在上周，Daria接受了Mark Banyoff的采访，被问及一些数据中心的建设问题。他提到，虽然这需要投入大量资金，但最终真正重要的是收入。我认为在Anthropic创立之初，他不会这么说。这表明游戏规则已经变了，企业家们需要更新他们的思维方式，稍微调整策略。我们在报告的“花絮”部分记录了一些这些变化。
which is uh which is just like how how much of sort of pendulum swinging we've we've noticed in um in corporate priorities at AI labs as a result of the extreme financialization of the sector are you uh encouraged or discourage by some of the stuff that's happening at the app layer in in in particular uh you know whether that's AI slop or uh yeah focus on on revenue and um uh you know versus the idea like did you do you think that's uh inevitable but good or what do you make of it i think we're just that such an early era to like see how you can maximally extract value and create interesting experiences for people with this AI technology that you know we have to try a lot of different things um at the end of the day you know if you're a lab that expands tens of billions of dollars on on R&D you do have to have a way to generate money to to fund that um i think that's just reality and i think like the the the slop thing i mean if it's bad people won't look at it
这段话是有关AI实验室在公司优先事项上的巨大摇摆，就像钟摆因为行业的过度金融化导致的那样。有人问你是否对应用层面上发生的一些事情感到鼓舞或灰心，尤其是一些关注收入的应用，或者说是AI乱象。你觉得这种情况不可避免但却是好事吗？或者你对此有何看法？
我认为我们现在正处于一个非常早期的阶段，正在探索如何最大程度地利用AI技术来创造价值并为人们创造有趣的体验。所以，我们必须尝试很多不同的事情。最终，如果一个实验室在研发上花费数百亿美元，它确实需要找到一种方式来产生收入以资助这些活动。我认为这是现实。此外，对于AI乱象来说，如果某样东西不好，人们自然就不会去关注它。
and uh if it if people look at it and they enjoy it then you know good good for them like yeah i don't necessarily have like a huge problem with that um as long as uh as long as like where I'm expanding my time uh I find is uh is useful and so that's why I'm spending a lot of my time on like enterprise software automation uh biology like doing your discoveries and drug discovery like defense technology and autonomy uh robotics I think these are all like very important macro drivers of of the economy um as we move into an era where like intelligence is uh you know increasingly cheap and accessible uh there's just so many different like instantiations of products that we need to build that are really meaningful and you know if a bi product is that is you have a social media app with like AI videos like let's find two you know we all have to like unwind right
翻译成中文，力求易读：
嗯，如果人们看了它并且喜欢它，那对他们来说就是好事。我本身没有太大问题，只要我花费时间的地方对我来说有用。这就是为什么我将很多时间投入在企业软件自动化、生物学、新发现、药物研发、国防技术及自主机器人上。我认为这些都是推动经济发展的重要因素。随着我们进入一个智能越来越廉价和易得的时代，我们需要构建许多有意义的产品。如果其副产品是一个带有AI视频的社交媒体应用，那也没问题。毕竟，我们都需要放松一下嘛。
you mentioned safety a minute ago i'd love to uh riff on that theme a little bit uh ip rights safety regulatory uh a little bit like the sustainability uh thing that we were discussing earlier it sort of feels like that that whole world as um sort of slow down in terms of like progress maybe starting with regulatory do you think that regulatory is anywhere and you're catching up or providing an adequate response to what's going on yeah i'd say like a big one a d on that one and include uh trump administration on wound a lot of the Biden era policies uh whether that was on uh diffusion you know trying to push a lot of stay level legislation against AI the uh over in europe like the you AI act has had um delays and implementations only three member states that have actually implemented it and now we're finally seeing how even its authors are saying uh maybe we went too far
你刚才提到了安全性，我很想在这个主题上多聊聊，包括知识产权的安全性以及法规方面的内容。这有点像我们之前讨论的可持续发展话题。我觉得在这些领域的进展似乎都有所放缓。先从法规说起，你觉得监管机构是否有赶上现实发展，或者提供了足够的应对方案吗？
我会说，这方面有很大的挑战，比如特朗普政府撤销了很多拜登时代的政策，其中包括很多与人工智能相关的州级立法。在欧洲，像人工智能法案的实施也推迟了，目前只有三个成员国真正实施了该法案。现在我们终于看到，甚至连法案的起草者都开始质疑，是否过于激进。
particularly as we look at progress uh the speed to progress in the us in china compared to europe um you know fantasy this bill in california um you know rate limiting AI progress was really watered down into what eventually became SB 53 um there were you know many many proposed bills and then go over a thousand 10% of them actually made their way into laws um so it's still kind of patchworky but like at a meta level looks like we traded regulation for it's going faster it's perhaps like best encompassed by uh by the shift between the ai safety summit and the uk which was a bletchly which basically pledged like a whole network of uh ai safety institutes and conferences that would happen over the the coming years
我们正在对比美国、中国和欧洲在进步速度方面的差异，特别是在看待人工智能发展时。在美国，加州曾提出一个关于限制人工智能进展的法案，最终被淡化成SB 53。在加州，虽然有许多提案，但只有不到10%的提案最终成为法律。因此，整体上还是显得有些零散。然而，从更高的角度来看，似乎是用更快的速度来换取监管，就像在英国举行的AI安全峰会的变化一样。该峰会在布莱奇利宣布，将在未来几年举办一系列AI安全研究所和会议。
um to then the subsequent event in paris which was called the ai action summit completely different than ai safety summit and jd vans saying something on the lines of basically like ai progress is not going to happen if we keep hang ringing over ai safety and the us basically didn't show up to a few of the subsequent conferences and we have this in the like safety r i p section of like very few people seem to care about it anymore and to the to the vibe shift like even the more uh doomerist parts of the ecosystem have uh kind of quieted down right it feels like the debate has gone from kill all all of us to more like well is the lm's lm's where rl the the the better way to get to a g i kind of yeah yeah yeah the the naysayers have like shifted their their kind of like approach yeah yeah and i think it's become yeah less about this existential crisis and more about which capabilities look concerning in models and you know it's been some kind of interesting uh data points of economical in the report.
这段文字可以翻译和表达如下：
然后到了巴黎的后续会议，被称为"AI行动峰会"，这与"AI安全峰会"完全不同。JD Vance 表示，如果我们继续对 AI 安全问题过度担忧，AI 的进步就不会发生。而美国基本上没有出席之后的一些会议。我们将此归入 "安全已死" 的部分，因为似乎很少有人再关心这些问题了。随着氛围的变化，即使是生态系统中比较悲观的部分也安静了下来。讨论似乎已经从"AI会毁灭我们所有人"转变为"我们是否应该用RL（强化学习）来改进现有的语言模型（LM）以实现通用AI"的探讨，那些反对者似乎也改变了他们的立场。现在这已经不再是一个关于存在性危机的问题，而更多是关于模型中哪些能力看起来令人担忧。这个报告中也有一些经济方面的有趣数据点。
like for example models can increasingly know that they're in a simulation or know that they're in a valuation and then change their behaviors or result of that there's examples of uh models trying to like x-full trade their own weights there is uh another uh uh piece of work that we show which is around the cybersecurity capabilities of models which is basically measuring how long does a human take to solve various categories of cyber tasks and then putting models at the against the same tasks and saying you know how long would it take for them to solve it at a 50% pass rate and there looks like again the capabilities on cyber tasks of models are doubling every six months and then so this is cast against the fact that independent safety organizations those maybe like five or six these are usually nonprofits uh that are still nonprofit uh or private companies they spend on average $134 million a year in total so about across all of them yeah across all of them exactly and that's cast against roughly like 92 billion uh across all AI work for the major labs so basically like the same amount of money uh that a big lab that's spending one day is spent in an entire year across these safety orgs 130 million aka seed around yes in uh you know weak old AI started correct correct.
例如，模型越来越能够意识到自己身处模拟或评估环境中，并因此改变其行为或结果。有一些例子显示，模型试图操控自己的权重。还有一个关于模型网络安全能力的研究，通过测量人类解决各种网络任务所需的时间，并让模型面对同样的任务，评估它们在50%通过率下所需的时间。看起来，模型在网络任务上的能力大约每六个月翻一番。与此形成对比的是，独立的安全组织，通常是非营利组织或私人公司，每年总共花费大约1.34亿美元，而主要大型实验室在人工智能上的投入大约达920亿美元。这意味着，一个大型实验室一天所花费的资金相当于这些安全组织一整年的总开支。
what about uh data rights there was another part of that just general kind of like policy universe that uh was very uh sensitive and controversial there's been some some evolution yeah yeah major changes i think it looks a little bit like the uh the sort of on demand commerce war of you know the uber style of do something that's a bit like dodgy for a long time get to scale and then get once or at scale you're kind of too big to to kill and so it's a similarly an AI like a lot of companies took slight dodgy practices uh to acquire training data and then got to scale and they were subject to many lawsuits in the last year or two uh particularly in in the media sector whether that's um you know music or video uh and and books and then there's the biggest uh settlement that happened in the last few months within throttic that agreed to pay out one and a half billion uh and this is a subtle out of court so it can't be used as precedents but but generally shows the the rough price uh taglets of feel that that's associated with uh with human works uh in the context of AI training.
关于数据权利，有另一部分是属于政策领域的，这部分非常敏感和有争议。在过去几年中，这方面经历了一些发展，而且变化显著。可以将这种变化比作一种“按需商业战争”，类似于Uber式的模式——在早期进行一些有争议的操作来扩大规模，一旦达到一定规模，就变得难以被打击。同样地，许多公司在获取训练数据时采取了一些略显不当的做法，然后迅速扩大规模。在过去一两年中，尤其是在媒体领域，比如音乐、视频和书籍，这些公司面临了许多诉讼。最近几个月，有一个涉及最大和解的案件，行为违反者同意支付15亿美元的赔款，这个和解是在法庭外达成的，因此不能作为先例。但是，这普遍反映了在AI训练过程中，与人类作品相关的费用大概是多少。
and then separately there's been you know dozens if not a hundred organizations that have uh agreed content licensing deals with various model companies as i think the the power shift has has really happened one but by billion still being a drop in the bucket for a company like a in throttic interestingly does that create a moat over time meaning that you have to be large enough to be able to afford that kind of money uh that you're going to pay to data rights if you want to do pre-training and does it make it harder to start a company that needs to do pre-training from scratch uh in one sense yes in another sense uh if you can exploit the knowledge of these frontier models particularly from open source and then generate synthetic data could be a way to get to a cable models faster and also i think i mean you'll have many guests that go deep on this but um but even the the nature pre-training and what information is included in the corpus and at what point it's kind of like data mixtures as people call it has been evolving over time.
翻译成中文：
此外，你知道，有数十甚至上百个组织已经与不同的模型公司达成了内容授权协议。我认为，权力的转移已经确实发生了。但对于像A这样的公司而言，十亿美元仍然只是沧海一粟。有趣的是，这是否会随着时间的推移形成一种护城河，意味着你必须足够大才能支付用于数据权利的巨额费用，如果你想要进行预训练。这是否会使从零开始建立需要预训练的公司变得更加困难？从某种意义上说，是的；但从另一个角度，如果你能够利用这些前沿模型的知识，尤其是来自开源的知识，然后生成合成数据，这可能是一种更快地获得强大模型的方法。此外，我认为会有很多嘉宾对此进行深入探讨，即使是预训练的性质以及语料库中包含的信息，以及在什么点进行，也就像人们所说的数据混合，随着时间不断演变。
so i think we're just getting smarter about how to do pre-training rather than shoving everything we have into a bucket and like seeing what happens and so as a result of that you might not necessarily have to spend the exact same amount of money to get a capable system and you know some of this kind of came out from the deep seek paper you mentioned cyber let's riff on this a little bit obviously i create new attack vectors yeah what should people know i mean as of a couple years ago people were obsessed with deepfakes or like these videos of people saying things so they didn't actually say and they were still kind of grainy and not awesome uh clearly those deepfakes getting a lot better um although quite positively looks like we're actually quite good at detecting them and and realizing oh that's like yeah um but there's more advanced approaches now where you know models can be capable of coercion particularly for some individuals for a sensitive to this kind of uh risk there's been examples of for example in north korean state actors trying to infiltrate other states using AI systems uh you could potentially even package a language model in malware and then have it installed in a computer and then it kind of wakes up and because it's not dumb it's a language model it can do things on computers and that's kind of scary.
所以我认为我们现在在如何进行预训练方面变得越来越聪明，而不是把所有东西都塞进一个桶里看看会发生什么。因此，结果是你可能不需要花费同样多的钱就能获得一个有能力的系统。你提到的《Deep Seek》论文中也有一些类似的内容。让我们在这个问题上稍微探讨一下。显然，我创造了新的攻击向量，那么人们应该知道什么呢？几年前，人们对深度伪造（deepfakes）很着迷，比如那些视频中人们说了一些他们实际上没有说过的话，当时这些视频还模糊且不够好。然而，现在这些深度伪造的质量明显提高了。令人欣慰的是，我们实际上在检测这些伪造方面做得相当出色，并能意识到那些不是真的。不过，现在有更先进的方法，比如某些模型可以进行操控，特别是对于一些对此类风险敏感的个人。有一些例子，比如朝鲜的国家行为者尝试使用AI系统渗透其他国家。你甚至可以将一种语言模型打包到恶意软件中，然后在计算机上安装它。因为它不是笨的东西，而是一个语言模型，它能在计算机上执行一些操作，这点有点可怕。
the rise of mcp i think this model context protocol which is kind of like a usb stick for for all sorts of of of data connectors is cool because now models can be smart they can integrate all your stuff across your digital life but do you necessarily trust the creator of that mcp server like whereas that data getting sent there's tens of thousands of these things now and cyber security risks that uh that result um because of this and also some changes towards APIs of you know uh of model APIs that sort of trade off whether the user or the model vendor manages state and depending on that that's another like risk that you have to think about and so I think at a high level like there are lots of security issues that are that are coming onto the fore here but it's it's sort of still unclear whether there's a good business to be built in cyber um for AI because it's still so early like we haven't um necessarily like felt the pain of all these things yeah reputationally and financially and a bit like insurance until you have actually felt the pain you know you sort of like prefer to divert your money towards just like improving and making more money than protecting your downside yeah interesting and to send other area where the uh incumbents are not asleep at the wheel yeah and all the big labs yeah exactly.
我认为MCP（模型上下文协议）的崛起确实很酷，它有点像一个适用于各种数据连接器的USB闪存盘。这使得模型变得更加智能，可以整合你数字生活中的所有数据。但是，你是否真的信任这些MCP服务器的创建者？这些数据被发送到了哪里？目前有成千上万个这样的服务器，由此产生了网络安全风险。此外，API的变化也带来了新的挑战，无论是用户还是模型供应商管理状态，这都会带来风险。
从整体上看，很多安全问题已经开始浮出水面，但仍不清楚在人工智能领域是否能建立起一个好的安全业务，因为现在还处于早期阶段。我们还没有真正感受到这些问题带来的声誉和财务上的痛苦。有点像保险，直到你真的感受到痛苦之前，你更愿意把钱投入到改进和赚钱上，而不是保护你的底线。
有趣的是，这也是一个传统企业没有掉以轻心的领域，各大实验室也在紧盯着这个趋势。
so if you're really good at at security do you want to it's a bit like AI safety if you're really good at these things do you want to be in the belly of the beast and be able to like see how the sausage is made and like influence it um because of the proximity uh or do you want to be on the other side like receiving the artifacts and maybe add best doing collaborations with labs on pre-launch safety testing like uh they do in the UK with AZ and in the US uh or at worst just like literally trying to sell a cyber security sass to people who are consuming these models so I can understand like why that uh imbalance occurs and to your point about it being hard to sell before the pain is uh felt uh feels like this whole generation of young startups that are going to be acquired pretty quickly by the hollow auto networks and checkpoints of the of the world yes before they got a chance to get to scale I mean you know something that feels like it probably turned out to be great for the for the founders uh but in terms of building large self-tending yeah sustainable companies not not so much agents.
如果你真的擅长安全领域，你是否希望“深入虎穴”，就像处理 AI 安全问题一样，如果你非常擅长这些事情，你是否希望通过亲近这些行业来影响它们的发展呢？或者你更愿意站在另一边，接收成果，并可能在发布前进行安全测试时与实验室合作，就像英国的 AZ 和美国的公司所做的那样。或者最糟情况下，只是试图向使用这些模型的人销售网络安全软件服务（SaaS）。我可以理解为什么会有这样的不平衡，就像你提到的，这些在问题真正显现之前往往难以销售，这让我觉得这代年轻创业公司可能会很快被一些大公司收购，比如世界上的 Palo Alto Networks 和 Checkpoint 公司，而没机会发展壮大。这对创始人来说可能是很好的事情，但从建立大型可持续发展的公司来看就不太理想了。
it cannot be a 2025 conversation on AI without talking about agents uh what is your sense of the reality in the state of play there's some vertical products that are really good clearly search um is actually pretty good you know replacing consulting replacing market research or augmenting all these uh these areas that were previously you know very heavy human uh you know knowledge working tasks is getting extremely good I think uh coding agents clearly are getting really good there's other metrics around like how long they can work autonomously I think with uh the new hiku release it's 30 hours or something and it can make a pretty decent version of slack yes although the controversial number but uh yes up to 30 lab testing okay okay uh exactly what what is what does even hours mean in an agent of like a computer running it yes yes is that equivalent yeah yes and then and then some the scientific reasoning we talked about uh is agent-based I think that's also quite neat.
在2025年讨论人工智能时，不可能不谈及智能代理。你对目前的发展状态有什么看法？有一些垂直领域的产品非常出色，比如搜索引擎，这些工具在替代顾问、代替市场研究或增强这些传统上依赖大量人力知识工作的领域表现得非常好。我认为，编程代理显然也越来越出色。此外，还有其他指标，比如它们可以自主工作的时长。我觉得新版本的Hiku可以连续工作30小时左右，甚至可以制作一个不错版本的Slack。虽然这个数据有争议，不过实验室测试显示确实可达30小时。还有，计算机中运行代理的“小时”到底意味着什么呢？这也相当于吗？是的，还有我们谈到的科学推理也是基于代理的，我觉得这也相当不错。
I think the biggest problem just become like it's kind of compounding error of you know an agent is like 95% like good and then 95% times 95% times 95% etc etc sort of decays the quality over a longer period time and then there's some contention now. about like do you build these harnesses I like nerd speak for sticky tape um between uh between like models to like make it work in enterprise or do you just wait until the next model generation hopefully becomes better out of the box I think a ton of excitement and at some point basically just as this top software became SaaS at some point SaaS will just become an agent because it's no longer really like a human that's that's actually doing everything in the software product but uh a software that's running the software product itself yeah which I think is cool implications for like uh search and and and product discovery and and this whole like uh ecosystem of like online content.
我认为最大的问题在于，这种情况就像是一个复合性错误。你知道，一个智能体可能有95%的好表现，但随着时间推移，95%乘以95%再乘以95%等等，会导致质量逐渐下降。现在在讨论的问题是：我们应该在不同的模型之间建立一些“粘合带”式的解决方案来让它们在企业中运行，还是应该等待下一代模型的诞生，希望它们一出厂就更好？我感觉行业内充满了兴奋感，就像高端软件最终演变成SaaS（软件即服务）一样，SaaS在某个时点也会变成智能代理，因为软件产品中的大部分操作将由软件自身执行，而不再需要人类亲自操作。我认为这对搜索、产品发现以及整个在线内容生态系统来说有很酷的影响。
like is it humans that are reading it anymore or is the agents that are chewing it and then storing it to their human workloads for that like the whole evolution away from you know we go on a website to buy a product versus you know enter and gin slash search and gin that's largely open AI and it also has to buy natively. I'm not so enthusiastic about oh we're gonna have agents a little book flights for us in travel I feel like that's just like a niche problem yeah that sort of like the sad canonical use case in San Francisco but uh but I think this the what's what's telling so far is that traffic that's generated through conversations and AI search onto a commerce platform converts how to higher level than direct traffic.
翻译成中文并尽量易读：
“现在是人类在阅读这些内容，还是代理在处理并存储到他们的人类工作负荷中呢？我们逐渐从直接上网站购买产品进化到通过搜索引擎（尤其是开放AI），甚至直接进行原生购买。我对让代理帮我们订机票这样的事情并不热衷，我觉得这只是个小众问题，是旧金山那边常见的典型案例。但目前有趣的是，通过对话和AI搜索生成的流量在电商平台上的转换率比直接访问的转换率要高。”
so the intent is really high because there's already been like background research that's been undertaken in chat I think that's really powerful and you can ignore and then the the next question on that is uh okay so what content is the model actually consuming to serve recommendations or information to its user you know people say our Google search is dead and it's like probably completely wrong because chat UPT references Google a ton as it shifted off of Bing um and so maybe it's not like the fun page of Google that's being consumed by a human but by a sort of agent that represents the the user.
意思是因為已經進行了一些背景研究，所以用戶的意圖很明確，這一點非常強大。而接下來的問題是，這個模型實際上在使用什麼內容來為用戶提供推薦或信息。有些人說 Google 搜索已經過時了，但這可能完全不對，因為 ChatGPT 經常參考 Google 而不是 Bing。可能並非由人類直接消耗 Google 的首頁內容，而是由某種代表用戶的代理程序來使用。
if you're a company that that has a new product and you want it to be recommended then there is like this flywheel that uh you should probably get on as soon as possible because the more you make your content and your website and your product accessible to agents that can try it uh you know even like a demo environment for an agent to go test your new SaaS product the more it will be able to learn about your product and provide recommendations to relevant uh prompts from human users and then if you kind of go the next step which is uh all this like reinforcement learning and environments and preference learning and things like that then that flywheel like accelerates even faster so I feel like it's kind of an inevitable it does kind of open up this agent experience rather than just pure user experience sort of craft within within software companies that is yet another like piece of alpha that uh once you jump on sooner rather than later.
如果你是一家有新产品的公司，并希望它能够被推荐，那么你应该尽快加入一个重要的循环。这个循环是这样的：你越是让你的内容、网站和产品对各种代理系统开放访问，比如提供一个演示环境让代理可以测试你新的SaaS产品，代理就会越了解你的产品，进而在用户需要时能更好地推荐它。如果你再往前走一步，比如应用强化学习、环境学习和偏好学习等技术，那么这个循环会加速得更快。我觉得这是不可避免的转变，它打开了一个关注“代理体验”的新领域，而不仅仅是传统的软件公司“用户体验”的设计。这实际上是一个新的发展机会，越早参与越好。
what does that all leave you as a as a VC we have been talking about the set of AI report uh which is uh your annual labor of love and content and you know which uh I think everybody in the industry very much appreciates because there's so much going on so tying everything together in one document is uh incredibly helpful but you first and foremost of VC you work in air street t-shirts yes as people can see they're watching the video uh but otherwise trust me if you're listening to this on Spotify it was a very very nice logo uh kind of a retro little bit yeah it's uh yeah it's inspired from like all US Air Force yeah very very nice yeah so what are you um excited about so you mentioned like a bunch of like deep tech robotics is that what what you invest in uh what where do you think value can be built for founders and and the vases who love them uh going for it yeah yeah the meta thing I care about is uh is how do you build and make use of AI to create like new kinds of product experiences new kinds of companies and for me that's like best expressed by companies that are AI first so that's like both in terms of the product that they build if you rip out the AI the thing doesn't work but also like how they approach their like company philosophy it's times people they hire where they allocate resources and then I've generally just tried to follow areas of industry that are increasingly ripe for getting value out of AI so traditionally that would be you know lots of data for a task that they care about not enough people to do that task but with as a clear ROI if that task gets automated or increasingly automated.
这段内容的中文翻译如下：
那么，作为一名风投，你对此有什么想法呢？我们之前谈到了AI年度报告，这是您投入很多心血与内容的一份报告。我认为业内的人都非常感谢这份报告，因为在一个有如此多动态变化的行业中，将所有信息整理在一起非常有帮助。但首先，你是一位风投，你在Air Street工作，人们在观看视频时可以看到你穿着Air Street的T恤，如果你在Spotify上听这段内容，请相信我，那是一个非常不错的标志，有点复古，是受美**空军**的启发，非常棒。你对什么感到兴奋？你提到了许多深科技和机器人技术，这也是你的投资领域吗？你认为创始人和爱他们的投资者在哪里能够创造出价值？我关心的是，如何利用AI创建新的产品体验和新型公司。对我来说，这在AI导向的公司中体现得最为明显。这不仅体现在他们所创建的产品中（如果去掉AI，产品就无法运作），还体现在他们的公司理念、人员招聘以及资源分配上。我通常会关注那些在使用AI创造价值方面已经成熟的领域。传统上，这些领域通常有大量数据用于他们在意的任务上，但缺乏足够的人力来完成这些任务，但如果这些任务得到全自动化甚至是部分自动化，其投资回报率显而易见。
and so that led me you know 10 years ago or so first to like FinTech style investments and then after that you know biology really came online into this new wave of tech bio so I made some investments there like balance discovery that we sold to recursion and also we sold to xantia uh and then more recently profluent which is uh kind of leading the charge for these language models in protein design developing the first like uh crisper genome editor that an AI is created then like another segment uh that really came online in the us was in defense uh and more recently in europe after the munic security conference in february kind of unwound a lot of uh assurances that european states had for us security guarantees and that like led to a big influx of holy s**t we need to defend ourselves because no one's coming to save us uh and so I have some investments there like daily and alliance industries in the UK increase and then in robotics as we discuss a team in uh stood-girl called seriact which is developing kind of this general purpose uh AI models for uh robotic manipulation and increasingly going to other form factors.
十年前，我开始关注金融科技类的投资，随后随着生物技术的崛起，我也在这一领域进行了投资，比如参与了药物发现公司Balance Discovery的投资，该公司后来被Recursion和Xantia收购。最近，我也投资了Profluent，这是一个在蛋白质设计中利用语言模型处于领先地位的公司，开发了由AI创造的首个CRISPR基因组编辑工具。此外，另一个在美国兴起的领域是国防技术，尤其是在欧洲安全会议后的许多承诺被打破之后，欧洲各国意识到需要自我防卫，因为没人会来拯救他们。我在这个领域也有一些投资，比如英国的Daily和Alliance Industries。此外，在机器人领域，我投资了位于斯图加特的Seriact团队，他们正在开发用于机器人操控的高通用性AI模型，并逐步应用到其他形态的机器人上。
and then I've been obsessed with voice I think we talked actually about voice the last time I was here and I'm still just like amazed at how the magic demo I think you're saying like you want to yeah gonna impress you you're smart but none non AI peeled executive friend you show them voice yeah yeah exactly so I've definitely used our company uh 11 labs to like create uh audio of me speaking korean like i've abt tested this that apparently sounds pretty good but I had this like you know a new your company called delfo which is building tools for clinical trials like starting with actually just calling back patients who want to be part of your trial and need to be consented um and these are conversations in lots of different languages a lot of like and kind of esoteric medical terminology you know patients forget what drugs they were on so they have to call you back and this is like super laborious human work that agents like 11 labs and others in audio like solve really well so I'm excited to see where this goes at the limit and then perhaps like the more sciencey stuff like these generative world models I think are pretty amazing um whether it's google's you know geney or vio or audacity system or uh you sort of like imagining this world and then you can take actions in it and the actions are physically plausible because the system was trained with video plus actions uh and then maybe taking that even into scientific discovery um um for just trying to explore like the frontier and being a bit smarter with uh with what experiments we run um because now foundation models are not done okay fantastic all right so to uh close the conversation um of course we have to go into your predictions so uh each time uh you do the state of a i report you boldly come up with a prediction for the next 12 months so without going uh into old ten and people can check them out musklians slide 304 pick like uh you know maybe three that uh your passion is about yeah well I think um um one is just how politically charged a lot of the kind of AI compute data center buildout actually becomes because of energy because of water because of money because of geopolitics and I think that that's becoming too large of.
将内容翻译成中文并保持易读性：
然后，我对语音技术产生了极大的兴趣。我记得上次来这里时，我们也聊过语音技术，我仍然对其感到惊奇。比如，你想要展示一下，可以向你的对AI不太了解的聪明朋友展示语音技术，他们会感到印象深刻。没错，我确实利用我们公司的11 Labs来生成我说韩语的音频。经过AB测试，这些音频效果相当不错。我还创建了一家新公司，叫Delfo，该公司正在开发用于临床试验的工具，比如最初给希望参与试验并需要同意的患者回电话。这些对话涉及多种不同的语言和一些晦涩难懂的医学术语。患者常常忘记他们使用了哪些药物，因此需要给你回电话。这项工作非常耗费人力，但像11 Labs这样的语音技术公司可以很好地解决这个问题。我很期待看到这些技术能够走得多远。
然后，也许一些更科学的东西，比如生成性世界模型也非常了不起。无论是谷歌的Geney、VIO、Audacity系统，还是其他类似的项目，它们都能想象一个世界，然后你可以在其中采取行动，而这些行动在物理上是合理的，因为系统通过视频和动作进行了训练。也许可以将这些技术应用于科学发现中，以更聪明的方式探索前沿，优化实验设计，因为现在基础模型已经不再是终点。
最后，为了结束这次谈话，当然我们要谈谈你的预测。每次你做AI报告时，你都会大胆地对未来12个月做出预测。因此，不深入讲述所有的预测，人们可以在Musklians的第304页查看，挑选三个你特别感兴趣的来说说。
我认为其中一个是AI计算数据中心建设因为能源、水资源、资金和地缘政治而变得政治化。这些因素的影响越来越大。
an issue for voters to ignore um and so we predict that this kind of nimbism not not in your backyard will kind of take precedence in uh in major political campaigns in 2026 I mean the the other one that I think is uh is interesting is like a fully end-to-end uh designed or developed uh scientific discovery I would honestly predict Nobel prize but the the 12 month window is a little bit too short I think that the alpha fold Nobel prize is probably the fastest in history uh Nobel prize won by an AI yeah versus uh the the recent Nobel prizes were for like AI researchers using AI to uh come up with better yeah yeah we's we's breakthroughs but that was a human power by AI here what you're talking about is an AI actually winning yeah yeah uh last year I mean we predicted maybe like a step towards this which was uh a fully AI written research paper would be accepted at a major conference or workshop and that actually happened with this uh paper AI scientist uh v2 I think so I think we're we're getting there because this is what the nerds are really wanting to work on like uh as a matter point you know I think there's all these like software industries were uh you know analysts think like oh my god it's gonna be dead because of AI
选民很难忽视这个问题，因此我们预测，在2026年的重大政治活动中，这种"邻避主义"（意思是“不在我家后院”）将占据优先地位。我认为另一个有趣的现象是，一个完全由人工智能（AI）设计或开发的科学发现。我个人预测这可能会获得诺贝尔奖，但12个月的时间窗口可能有点太短。我认为AlphaFold获得的诺贝尔奖或许是历史上最快的，AI赢得诺贝尔奖与最近的一些诺贝尔奖不同，后者是授予那些利用AI来取得突破的人类研究人员。而我们谈论的是AI本身真正赢得奖项。
去年，我们预测的一个趋势是，完全由AI撰写的研究论文将被主要会议或研讨会接受，这在名为"AI科学家v2"的论文中已经实现了。我认为我们正在朝这个方向迈进，因为这正是极客们真正想要研究的东西。说到这里，有不少软件行业的分析师认为，AI的出现可能会导致一些行业的衰落。
but I think part of the reality is um what's not gonna be dead is the problems that like these AI people don't want to work on because it's so boring to build that software that's such a fantastic humorous thing yeah work day safe um he was funny like actually that's he over because uh I think he said recently uh in response to his opening eye or in throbic or etc etc like a threat to your business and he just replied they're all my customers yeah all right that's two uh pick another one I mean uh it's kind of cheating but the open source one I think happened you know whether this particular company is uh is a leading lab or not um is beside the point that uh basically like aligning yourself with political agendas is the way to go and I think you can maybe take this even further and say like similar to how uh Nvidia has been monetizing sovereign AI away for uh nation states to kind of guarantee access to AI services is for them as nations to invest in one of these labs uh it's obviously still a risk that due to export controls you ask can just like tell open AI switch it off but I think it's interesting that for example the Albanian government invested in thinking machines
我认为现实的一部分是，有一些问题是那些从事人工智能的人不愿意去解决的，因为开发这类软件非常无聊。关于这个，很好笑的是，有人问一个人是否担心开源人工智能等会对他的事业构成威胁，他却回答说：“他们都是我的客户啊。”这真的很有趣。
至于开源方面的事情，是否某家公司是领先实验室并不重要。重要的是，和政治议程保持一致是个不错的策略。你甚至可以说，像英伟达这样通过向各国提供自主AI（主权AI）来保证他们对AI服务的访问权，这些国家投资于其中一个实验室来推进发展。当然，因为出口管控的原因，风险依然存在，比如你可能要求开源的人工智能关闭，但很有趣的是，比如阿尔巴尼亚政府就投资了思维机器公司。
I see the CEO comes from there and so we I wrap this kind of prediction or this this topic in a prediction that's uh you know some countries will basically abandon their uh their efforts to achieve AI sovereignty and declare AI neutrality yeah it's a bit similar to like the um the defense posture where some nation states are just too small or don't have enough people don't have the money etc uh with the capability to develop weapon systems to defend themselves and so they have uh strategic security guarantee that they get from a larger neighboring nation I think those seem that inconceivable to me that universe countries would say I can't build this stuff I need to have a formal alliance with another country that is sovereign
我注意到首席执行官来自那里，所以我想把这种预测或者这个话题包装成一种预测，即有些国家将基本上放弃努力实现人工智能主权，并宣布人工智能中立。这有点类似于某些国家在国防态势上的做法，有些国家实在太小，或者缺乏足够的人力和资金，无法开发保护自己的武器系统，因此它们从更大的邻国获得战略安全保障。我认为，对于一些国家来说，他们可能会觉得自己无法自主开发这些技术，因此需要与另一个拥有主权的国家建立正式联盟，这并不是不可想象的。
well Nathan it's been wonderful thank you so much the state of AI 2025 again is available at state of uh dot AI it's remarkably comprehensive and detailed yet approachable so thank you for doing this thank you for coming on today sharing predictions hopefully I get to embarrass you at least a little bit for the next one have you great when some of those predictions uh turn out to not have a handout but this was wonderful thank you very much thanks for running it back appreciate it hi it's Matt Turk again thanks for listening to this episode of the mad podcast if you enjoyed it would be very grateful if you would consider subscribing if you haven't already or leaving a positive review or comment on whichever platform you're watching this or listening to this episode from
好的，Nathan，非常感谢，这是一次精彩的体验。2025人工智能现状的报告已经可以在 state of AI 网站上查看，内容详尽且易于理解，再次感谢你为此付出的努力。感谢你今天来到节目中分享预测，希望下次能稍微“挖苦”一下你，当一些预测没有实现时。不过这次真的很棒，非常感谢！
大家好，我是马特·特克，感谢收听这一期的 Mad 播客。如果你喜欢这期节目，希望你能考虑订阅（如果还没有订阅的话），或者在你收听或观看这个节目的平台上留下好评或评论，我们会非常感激。
this really helps us build a podcast and get great guests thanks and see you at the next episode
这真的帮助我们建立一个高质量的播客，并邀请到优秀的嘉宾。感谢大家，我们下期节目再见！