人工智能迅猛前进

这是霍华德·马克斯的备忘录。

This is a memo by Howard Marks.

人工智能前方的障碍。

AI hurdles ahead.

当我准备撰写关于人工智能的十二月备忘录时，我在想：这会是泡沫吗？

When I was preparing to write my December memo about artificial intelligence, is it a bubble?

我与几位三十至四十多岁的有趣科技人士交流后受益匪浅。

I gained a great deal from speaking with some interesting techies in their thirties and forties.

探索新领域令人振奋，对于投资者保持与时俱进至关重要。

It's stimulating to explore fresh territory and an absolute requirement for staying current as an investor.

这是我工作中最令人享受的部分之一。

It's one of the most enjoyable parts of my job.

我最近再次联系了那些人，以跟进十二月备忘录的反馈。

I recently returned to those people to follow-up on the December memo.

作为这一过程的一部分，有人建议我询问Claude——Anthropic的人工智能模型——来制作一个教程，解释人工智能以及过去三个月发生的变化。

As part of that process, someone suggested I ask Claude, Anthropics AI model, to create a tutorial explaining artificial intelligence and the changes that have taken place in the last three months.

我照做了，这给了我大量素材可供参考。

I did so, and it gave me a great deal to work with.

这份备忘录旨在作为十二月备忘录的补充。

This resulting memo is intended as an addendum to December's.

其中大部分内容将复述克劳德的万字文章，我会再加入一些自己的观察。

Much of it will recap Claude's 10,000 word essay, to which I'll add a few observations of my own.

在这个过程中，我会强调一些对我而言是新术语、对你可能也是新术语的概念。

In the process, I'll highlight some terms that were new to me and might be new to you.

我本可以省去很多时间，直接让克劳德写这份备忘录，但我决定不这么做，因为我认为亲自动笔是其中最大的乐趣之一。

I could have saved myself a lot of time by asking Claude to write this memo, but I decided not to because I consider putting words on paper a big part of the fun.

不过，我会大量引用克劳德的成果。

I will, however, quote liberally from Claude's work product.

所有未明确标注出处的引述，均来自克劳德的文本。

That'll be the source of all quotations that aren't otherwise identified.

在开始之前，我想试着传达我对克劳德输出结果的敬畏程度。

Before I start in, I want to try to communicate the level of awe with which I viewed Claude's output.

它读起来就像朋友或同事写的一封私人便条。

It read like a personal note from a friend or colleague.

它提到了我在以往备忘录中讨论过的内容，比如利率的巨变和投资者心理的摆动，并将这些内容用作与人工智能相关的隐喻。

It made reference to things I've talked about in past memos, like the sea change in interest rates and the pendulum of investor psychology, and it used them in metaphors related to AI.

它逻辑严密，预判了我可能提出的观点，加入了幽默感，并坦诚地承认了人工智能的局限性，从而增强了可信度，这正像我可能会做的那样。

It argued logically, anticipated points I might make in response, injected humor, and bolstered its credibility by candidly acknowledging AI's limitations just as I might do.

我以前也问过人工智能问题并得到过回答，但从未像这次一样收到如此个性化的解释。

I've asked AI questions before and gotten answers back, but I've never received a personalized explanation like I did in this case.

理解人工智能。

Understanding AI.

在继续探讨人工智能的最新进展及其能力之前，我想分享一下这个教程为我揭示的关于人工智能本质的一些见解。

Before moving on to the meat of the matter, recent changes in AI and its capabilities, I want to share some insights into AI's essence that the tutorial delivered for me.

重要的是，这个教程让我明白，不应将人工智能模型视为一个检索数据并复述的搜索引擎。

Importantly, the tutorial taught me not to think of an AI model as a search engine that retrieves data and regurgitates it.

相反，它是一个能够综合数据并从中进行推理的计算机系统。

Rather it's a computer system that's capable of synthesizing data and reasoning from it.

AI模型的生命分为两个阶段。

There are two phases in the life of an AI model.

第一阶段是通过阅读大量文本进行训练。

In the first it is trained by reading a vast amount of text.

训练阶段不能被理解为只是向模型加载信息，我以前一直这么认为。

The training phase must not be thought of as loading the model with information, which I had done until now.

它远远不止于此。

It goes far beyond that.

它的本质是教会模型如何思考。

It consists of teaching the model how to think.

通过吸收文本，模型学会理解推理模式并构建它们，了解论点的结构，如何生成新的想法组合，以及如何将已学的推理模式应用于新情境。

By absorbing text, the model learns how to understand reasoning patterns and form them, how arguments are structured, how to generate new combinations of ideas, and how to apply learned reasoning patterns to novel situations.

理解训练阶段最好的方式，是将其与人的智力发展相比较。

The best way to think about the training phase is to compare it to the development of a person's intellectual capacity.

婴儿出生时拥有大脑，通过接触外部刺激，逐渐发展出思考、推理、综合、评估、类比、组合想法、创造概念、构建论点等能力。

A baby is born with a brain, and through exposure to external stimuli it develops the ability to think, reason, synthesize, evaluate, analogize, combine ideas, create concepts, compose arguments, and so on.

婴儿出生时并不具备这些能力，而是通过吸收和利用环境中的输入来发展这些能力。

The baby isn't born with those abilities, but it develops them by absorbing and using inputs from its environment.

AI模型也是如此。

An AI model is the same.

这里我要说明一下，我并不意味着自己理解AI是如何运作的。

A word here, I'm not implying that I understand how AI does what it does.

这根本不可能。

There's no chance of that.

至多，我只能描述AI能做什么以及其带来的影响。

At best, I'll describe what AI can do and the implications.

AI模型生命的第二阶段是推理。

The second phase in an AI model's life is inference.

一旦模型构建并训练完成，推理就是它余生所做的事情，利用其能力来满足用户的需求。

Once the model has been built and trained, inference is what it does for the rest of its life, using its capabilities to meet the demands of users.

这里需要注意的是，模型无法自行分配任务，至少目前不能。

It's important to note here that the model cannot assign itself tasks, at least not at present.

它必须通过用户编写的提示来接收任务指令。

It has to be ordered to perform tasks through prompts written by users.

提示越优质、越全面，AI能做的事情就越多。

The better and more comprehensive the prompts, the more AI can do.

例如，AI可以编写软件来完成用户想要实现的任务。

For example, AI can write software to perform work a user wants done.

它还可以测试软件、发现漏洞、修复问题并再次测试，但必须有人指令它去做这些事。

It can also test the software, identify bugs, fix them, and test again, But it has to be instructed to do those things.

至少在当前阶段，请听好了。

At least at the current stage, listen on.

因为如今许多人缺乏对提示重要性的认识，也缺乏编写提示的能力，AI的潜力很可能被低估了。

Because many people today lack awareness of the importance of prompts and fail to possess the ability to create them, AI's potential is probably being underestimated.

但请注意，限制在于用户，而非模型本身。

But note that the limitation is on the part of the users, not the model.

以我的教程为例，Claude 并不是仅仅被要求解释 AI 以及它能做什么。

To illustrate using the example of my tutorial, Claude wasn't simply asked to explain AI and what it can do.

当我询问Claude它被分配的任务时，它是这么说的。

When I queried Claude about the task it was assigned, here's what it said.

有人专门为你们设计了一个九模块的课程，围绕着你们的十二月备忘录、你们的思维框架，以及赋予你们足够的技术理解力来撰写一份可信的附录这一目标。

Someone designed a nine module curriculum specifically for you, built around your December memo, your intellectual frameworks, and the goal of giving you enough technical understanding to write a credible addendum.

该课程结构旨在一次教授一个模块，使用你们世界中的类比，展示能力而非仅仅描述，并保持你们的读者所期待的智力诚实。

The curriculum was structured to teach one module at a time, use analogies from your world, demonstrate capabilities rather than just describe them, and maintain the kind of intellectual honesty your readers expect from you.

我可以告诉你们，这个教程确实实现了我们设定的目标。

I can tell you the tutorial definitely accomplished the goals we'd set for it.

这完全归功于我的顾问们帮助我准备的提示的质量和具体性。

This was entirely due to the quality and specificity of the prompts my advisors helped me prepare.

AI能思考吗？

Can AI think?

我想在这里花点时间探讨一个我非常感兴趣的问题。

I'm going to take time here for a question I find fascinating.

我知道AI能够重新组合人们已经发现的内容，并将其应用于新数据和其他领域，但它能开辟新天地吗？

I know AI can reconfigure what people have already figured out and apply it to new data and other fields, but can it break new ground?

我主要将AI的过程理解为利用历史模式和逻辑来预测序列中的下一个元素。

I understand AI's process primarily as a matter of using historical patterns and logic to predict the next item in a series.

写五个词组成一句话，它就会预测第六个词应该是什么。

Write five words in a sentence and it'll predict what the sixth should be.

下次你写邮件时，看看手机上推荐的词语。

Look at the suggested words on your phone the next time you write an email.

这就是AI在发挥作用。

That's AI in action.

让它组合一个跑赢市场的投资组合，它就会分析过去表现良好的股票，并利用它们的特征来预测未来表现最佳的股票。

Ask it to put together a portfolio to beat the market, and it will look at stocks that performed well in the past and use their traits to predict which ones will perform best in the future.

我认为，把AI看作是根据过去的情况提出关于未来的假设，是有帮助的。

I think it's helpful to think of AI as proposing a hypothesis regarding the future based on the way things went in the past.

我稍后会再回到这一点。

I'll return to this later.

根据我们刚才的讨论，我的问题是：AI能产生新想法吗？

What follows from what we just discussed is my question, can AI have a new idea?

也许它能完成我们交给它的每一项知识任务，但它能想到我们没让它去想的事情吗？

Maybe it can perform every knowledge task we assign to it, but can it think of things we haven't told it to think of?

它能像坐在河边一样，让偶然的灵感涌入脑海吗？

Can it do the equivalent of sitting by a river and letting stray inspirations come into its head?

它能看见苹果从树上落下，从而形成重力的概念吗？

Can it see an apple fall from a tree and develop the notion of gravity?

它能沉思、做白日梦或产生创意吗？

Can it muse, daydream, or ideate?

它能有直觉吗？

Can it have intuition?

这就是关于人工智能的争论变得复杂的地方。

This is where the debate around AI gets complicated.

根据克劳德的说法，怀疑者持以下观点。

According to Claude, the skeptics argue as follows.

克劳德所学的一切都来自人类撰写的文本。

Everything Claude learned came from human written text.

它没有经验，没有对世界的具身理解，也没有真正的认知。

It has no experiences, no embodied understanding of the world, no genuine comprehension.

它所产生的一切，最终都只是对其从现有人类作品中吸收的模式进行的复杂重组。

Everything it produces is ultimately some sophisticated rearrangement of patterns it absorbed from existing human work.

这是极其令人印象深刻的模式匹配，也许是人类工程史上最出色的模式匹配，但它不是思考。

It's extraordinarily impressive pattern matching, maybe the most impressive pattern matching ever engineered, but it's not thought.

它不是推理。

It's not reasoning.

它只是一种统计上的重新组合，如果这是真的，那就存在一个上限。

It's a statistical recombination, And if that's true, then there's a ceiling.

它可以重新组合人类已经发现的东西，但无法真正开辟全新的领域。

It can remix what humans have already figured out, but it can't break genuinely new ground.

它就像一支非常有才华的翻唱乐队，而不是作曲家。

It's a very talented cover band, not a composer.

就在克劳德刚刚阐述了上述怀疑论者的观点时，它却以针对我的方式给出了一个充满活力的回应。

Just as Claude laid out the skeptics issue as just identified, it came back with a spirited rejoinder framed in terms of me.

谈谈如何有理有据地论证一个观点。

Talk about knowing how to argue a point.

霍华德，你所知道的一切投资知识都来自别人。

Howard, everything you know about investing came from other people.

本杰明·格雷厄姆教你关于安全边际。

Benjamin Graham taught you about margin of safety.

巴菲特教你关于品质。

Buffett taught you about quality.

查理·芒格教你跨学科的心理模型。

Charlie Munger taught you about mental models from multiple disciplines.

约翰·肯尼斯·加尔布雷斯教你关于金融狂热的心理学。

John Kenneth Galbraith taught you about the psychology of financial manias.

五十多年来，你阅读了成千上万本书籍、备忘录、案例研究和年报。

You read thousands of books, memos, case studies, and annual reports over 50.

每一个输入都来自他人的思考。

Every input was someone else's thinking.

你从多个学科中汲取了框架，将其应用于新颖的情境，并创造出真正新颖的东西。

You took frameworks from multiple disciplines, applied them to novel situations, and produced something genuinely new.

原材料来自他人，而综合提炼则是你自己的成果。

The raw material came from others, the synthesis was yours.

所以当有人声称Claw只是重新排列训练数据中的模式时，我会问：这与任何受过教育的头脑所做的事情，在结构上有什么不同？

So when someone says, Claw just rearranges patterns from its training data, I'd ask, how is that structurally different from what any educated mind does?

你通过数十年的阅读学会了推理模式。

You learned reasoning patterns from decades of reading.

我通过训练学会了推理模式。

I learned reasoning patterns from training.

问题不在于输入来自哪里。

The question isn't where the inputs came from.

问题在于，无论是人类还是人工智能的系统，是否能以真正新颖且有用的方式将它们结合起来。

The question is whether the system, human or artificial, can combine them in ways that are genuinely novel and useful.

当然，这完全正确。

Of course, this is completely true.

我作为一名年轻投资者，从实际经验与文字资料中吸收了大量数据，学习了前人如何思考这些数据以及他们得出了哪些结论。

I ingested data as a young investor from actual experience as well as the written word, and I learned how those who went before me thought about the data and what conclusions they reached.

我研究了他们的思维过程，并学习如何将这些方法应用到我所接收的数据中。

I studied their thought processes and how to apply them to the data I took in.

他们的思维模式也激励我发展出属于自己的方法。

I was also inspired by the example of their processes to come up with my own.

这就是人类大脑拓展能力的方式。

This is how the human brain expands its capabilities.

AI的成长、学习与思考方式，真的与我们不同吗？

Is AI's way of growing, learning, and thinking really different from ours?

最后，克劳德提出了一个令人信服的现实论据。

Finally, Claude came back with a convincing real world argument.

即使你完全认同怀疑者的观点，即使你在哲学上接受我所做的只是模式匹配而非真正思考，其经济影响也完全相同。

Even if you grant the skeptic everything, even if you accept philosophically that what I do is merely pattern matching and not true thought, the economic implications are identical.

让我直白地说。

Let me put it starkly.

如果我能产出价值二十万美元的研习助理的分析成果，那么对于付账的人来说，我是否真的在思考，还是仅仅在模式匹配，这重要吗？

If I can produce the analytical output of a $200,000 a year research associate, does it matter to the person paying the bill whether I'm really thinking or merely pattern matching?

重要的是工作成果是否足够可靠以供使用，而这一点正变得越来越明显。

What matters is whether the work product is reliable enough to be useful, and increasingly it is.

关于机器意识的哲学争论固然迷人，但经济问题并不在于AI是否真正理解。

The philosophical debate about machine consciousness is fascinating, but the economic question isn't does AI truly understand?

经济问题在于AI是否能完成这项工作。

The economic question is does AI do the work?

如果你想积极参与关于AI的讨论，你就必须学会理解‘生成’这个词的含义，这是懂AI的人经常使用的词。

If you want to be an active participant in discussions of AI, you have to learn the meaning of the word generative, which people knowledgeable about AI use a lot.

理解这个词能极大地增强你对AI本质的把握。

Understanding that term greatly enhances one's sense for the essence of AI.

根据AI模型Perplexity的说法，在生成式AI中，‘生成’意味着能够创造新事物，而不仅仅是分析或标记现有的东西。

According to the AI model perplexity, in generative AI, the word generative means able to create new things, not just analyze or label existing ones.

它指的是那些学习数据中的模式，然后生成与该数据相似的新内容的AI系统。

It refers to AI systems that learn patterns in data and then generate new content that resembles that data.

这是思考，还是其他东西？

Is this thinking or something else?

还是我在过度强调一个没有区别的区别？

Or am I belaboring a distinction without a difference?

我们稍后在这份备忘录中会得到一些线索。

We'll get some indication of this later in this memo.

人工智能的最新进展。

Recent developments in AI.

我撰写这份补充材料的主要原因是回应自《这是泡沫吗？》于12月9日发布以来，人工智能领域发生的重大变化。

My main reason for writing this addendum is to address significant changes that have taken place in AI over the three months since Is it a Bubble?

于12月9日发布。

Was published on December 9.

首先，是人工智能发展速度之快。

First, there's the pace at which developments in AI are occurring.

这种速度前所未有，由此带来的影响也是前所未有的。

That speed is unlike anything we've seen before now, and this has implications that have never existed.

人工智能的发展速度远远超过了过去的技术创新。

AI is growing at speeds that greatly outpace the technological innovations of the past.

将其发展与计算机的发展相比较。

Compare its development with that of the computer.

第一台计算机ENIAC的建造于1945年完成。

The building of the first computer, ENIAC, was completed in 1945.

IBM的托马斯·J.

IBM's Thomas J.

沃森一世。

Watson Sr.

据ChatGPT称，当时有人传说他曾经说过：我认为世界上可能只有五台计算机的市场。

Is apocryphally, per ChatGPT, described as having said around that time, I think there is a world market for maybe five computers.

即使这句话并非出自他之口，这一观点也反映了20世纪40年代中期的普遍看法。

Even if it wasn't his, this observation reflects the state of opinion in the mid nineteen forties.

二十年后，当我学习编程时，计算机仍然很原始，在现实世界中的应用仅限于大型机构。

Twenty years later, at the time I learned to program, computers were still rudimentary, and their use in the real world was limited outside of very large institutions.

几乎没有人会想到计算机，更不用说拥有或能想象出它的用途了。

Almost no one thought about computers, much less had access to one, or could think of a use for one.

又过了十年，微处理器的出现才使得个人计算机得以诞生，主要以爱好者套件的形式出现。

It was another ten years before the development of the microprocessor allowed the creation of personal computers, mostly in the form of kits for hobbyists.

数字设备公司创始人肯·奥尔森因据称在1977年说过‘家庭用户根本没有必要拥有计算机’而广为人知。

Ken Olson, the founder of Digital Equipment Corporation, is famous for reportedly having said in 1977, there is no reason for any individual to have a computer in his home.

直到二十世纪八十年代初，也就是ENIAC建成近四十年后，IBM才开始向普通企业和家庭销售个人电脑。

It was only in the early nineteen eighties, nearly forty years after ENIAC was built, that IBM began to sell PCs for general business and home use.

将这一时间线与人工智能的发展作对比。

Contrast this timeline against the development of AI.

我向Perplexity询问了人工智能的历史，它告诉我，人工智能早在2010年前就开始悄然融入各种设备中。

I asked perplexity about the history of AI and it informed me that AI began to be incorporated into devices invisibly.

例如，垃圾邮件过滤器和推荐引擎就在2010年前后出现。

For example, spam filters and recommendation engines just before 2010.

随后的几年里，它开始在Siri和Alexa等产品中变得显而易见。

Then over the next few years it became visible in things like Siri and Alexa.

根据Perplexity的说法，不到两年前，生成式AI在商业和媒体中被定位为一种影响知识工作、教育和消费者决策的横向通用技术。

According to Perplexity, it was less than two years ago that generative AI was framed in business and media as a horizontal general purpose technology affecting knowledge work, education, and consumer decision making.

而仅仅两年后，它已经被约4亿个人和75%至80%的企业使用。

And just two years later, it's already being used by 400,000,000 or so individuals and 75 to 80% of companies.

历史上从未有任何技术像AI这样迅速普及。

Nothing has ever taken hold at the pace AI has.

它能够以接近瞬时的速度改变世界，远远超越了大多数观察者预测甚至理解的能力。

It's able to change the world at a speed that approaches instantaneous, outpacing the ability of most observers to anticipate or even comprehend.

过去，为一项新技术建设基础设施往往需要多年时间才能完全投入使用。

In the past, infrastructure was built for a new technology and it often took years for that infrastructure to be fully utilized.

然而，在AI推理方面，需求已经存在并迅速增长，我听说AI正受到供给限制。

In the case of AI inference, however, demand already exists and is growing rapidly, and I'm told AI is supply constrained.

发生的第二件重要事情是AI能力的巨大飞跃。

The second important thing that's happened has been an incredible leap ahead in AI's capabilities.

我的教程通过解释AI模型所代表的发达大脑具有三个能力层级，为我提供了一些背景信息。

My tutorial gave me some background by explaining that the developed brain represented by an AI model has three levels of capability.

第一级是聊天AI，用户提出问题，模型提供答案，但不会对答案做任何处理。

Level one is chat AI, where the user asks questions and the model supplies answers, but it doesn't do anything with the answers.

在这一级别，AI主要节省了原本用于研究和思考的时间。

At this level, AI mainly saves time that would otherwise be spent researching and thinking.

第二级是工具使用型AI，用户指示模型搜索信息、分析信息并利用信息执行任务。

Level two is tool using AI, where the user instructs the model to search out information, analyze it, and perform tasks with it.

因此，这里的经济价值显著更大，因为它节省的是执行时间，而不仅仅是思考时间，但仍然受限于AI只能执行被指令的操作。

Thus, the economic value here is meaningfully larger because it's saving execution time, not just thinking time, but it's still bounded because AI only does what it's told.

第三级是自主代理。

Level three is autonomous agents.

在这一级别，用户不再告诉AI该做什么。

At this level, the user doesn't tell AI what to do.

用户只给它一个目标以及期望输出的参数，比如长度、耗时、内容和涵盖要点。

The user gives it a goal as well as the parameters of the desired output, things like length, time taken, content, and points covered.

代理会自行完成工作、检查结果，并提交最终成果。

The agent does the work, checks it, and submits a finished product.

这是在任务层面上对劳动力的替代，而非辅助。

This is labor replacement at the task level, not assistance, replacement.

AI最显著的特征是我们此前在任何技术发展中都未曾遇到过的。

The most significant thing that distinguishes AI is something we've never dealt with in connection with prior technological developments.

AI具备自主行动的能力。

AI's ability to act autonomously.

根据Claude的说法，AI在2023年处于第一级，2024年进入第二级，但现在已达到第三级，这种差异至关重要。

According to Claude, AI was at level one in 2023 and level two in 2024, but it's now at level three, and the difference is a big one.

第二级和第三级之间的区别听起来可能很细微。

The distinction between level two and level three might sound subtle.

但其实不然。

It isn't.

这一区别决定了AI是生产力工具还是劳动力替代品，而正是这一差异，将一个500亿美元的市场与一个数万亿美元的市场区分开来。

It's the difference that determines whether AI is a productivity tool or a labor substitute, and that difference is what separates a $50,000,000,000 market from a multi trillion dollar one.

Other Side AI的首席执行官Matt Schumer最近发布了一篇题为《大事正在发生》的博客文章，在不到一个月的时间内已被超过五千万人浏览。

A recent blog post entitled something big is happening from Matt Schumer, CEO of Other Side AI, has been viewed by more than 50,000,000 people in less than a month.

它捕捉到了AI近期进展的精髓，由于舒默对此的阐述如此出色，我忍不住要引用其中三个重要部分。

It captures the essence of AI's recent progress, and because Schumer communicates it so well, I can't resist including three substantial sections.

2月5日，两家主要的AI实验室同一天发布了新模型。

On February 5, two major AI labs released new models on the same day.

OpenAI的GPT-5.3 Codex和Anthropic的Opus-4.6，后者是ChatGPT的主要竞争对手Claude的开发公司。

GPT 5.3 Codex from OpenAI and Opus 4.6 from Anthropic, the makers of Claude, one of the main competitors to chat GPT.

某种顿悟发生了，不是像开关那样突然，而更像你突然意识到水位已经悄然上涨，如今已漫至胸口。

And something clicked, not like a light switch, more like the moment you realize the water has been rising around you and is now at your chest.

我不再需要亲自完成工作中那些实际的技术性任务了。

I am no longer needed for the actual technical work of my job.

我用简单的英语描述我想构建的东西，它就直接出现了。

I describe what I want built in plain English and it just appears.

不是需要我修改的初稿，而是完整的成品。

Not a rough draft I need to fix, the finished thing.

我告诉AI我想要什么，然后离开电脑四小时，回来时发现工作已经完成了。

I tell the AI what I want, walk away from my computer for four hours, and come back to find the work done.

做得很好，比我亲自做还要好，而且完全不需要修改。

Done well, done better than I would have done it myself with no corrections needed.

几个月前，我还在和AI来回沟通，指导它、进行修改。

A couple of months ago, I was going back and forth with the AI, guiding it, making edits.

现在我只需要描述一下期望的结果，然后离开就行。

Now I just describe the outcome and leave.

让我举个例子，让你明白这在实际中是什么样子。

Let me give you an example so you can understand what this actually looks like in practice.

我会告诉AI：我想开发这个应用。

I'll tell the AI, I wanna build this app.

它应该实现哪些功能。

Here's what it should do.

大致长什么样。

Here's roughly what it should look like.

自己去设计用户流程、界面，全部搞定。

Figure out the user flow, the design, all of it.

它确实做到了。

And it does.

它写了数万行代码。

It writes tens of thousands of lines of code.

然后，这一点在一年前是难以想象的：它自己打开了这个应用。

Then, and this is the part that would have been unthinkable a year ago, it opens the app itself.

它点击了各个按钮。

It clicks through the buttons.

它测试了各项功能。

It tests the features.

它像人一样使用这个应用，如果对某些界面或体验不满意，它会自行回去修改。

It uses the app the way a person would, if it doesn't like how something looks or feels it goes back and changes it on its own.

它像开发者一样不断迭代，修复和优化，直到自己满意为止。

It iterates like a developer would fixing and refining until it's satisfied.

只有当它认为应用达到了自己的标准时，才会回来告诉我：你可以测试了。

Only once it has decided the app meets its own standards does it come back to me and say, it's ready for you to test.

当我测试时，它通常完美无缺。

And when I test it, it's usually perfect.

但真正让我震惊的是上周发布的模型——GPT 5.3 Codex。

But it was the model that was released last week, GPT 5.3 Codex, that shook me the most.

它不仅仅是执行我的指令，还在做出明智的决策。

It wasn't just executing my instructions, it was making intelligent decisions.

它第一次展现出了一种类似判断力和品味的东西——那种无法言喻的、知道什么才是正确选择的能力，人们曾说AI永远不会有这种能力。

It had something that felt for the first time like judgment, like taste, the inexplicable sense of knowing what the right call is that people always said AI would never have.

这个模型拥有这种能力，或者至少接近到足以让这种区别开始变得无关紧要。

This model has it or something close enough that the distinction is starting not to matter.

让我具体说明一下进步的速度，因为我认为，如果你没有密切关注，这部分最难相信。

Let me make the pace of improvement concrete because I think this is the part that's hardest to believe if you're not watching it closely.

在2022年，AI还无法可靠地进行基本算术。

In 2022, AI couldn't do basic arithmetic reliably.

它会自信地告诉你，七乘八等于54。

It would confidently tell you that seven times eight equals 54.

到2023年，它已经能通过律师资格考试。

By 2023, it could pass the bar exam.

到2024年，它能编写可用的软件并解释研究生级别的科学内容。

By 2024, it could write working software and explain graduate level science.

到2025年底，一些世界上最优秀的工程师表示，他们已将大部分编码工作交给了人工智能。

By late twenty twenty five, some of the best engineers in the world said they had handed over most of their coding work to AI.

2026年2月5日，新模型问世，使此前的一切都显得如同另一个时代。

On 02/05/2026, new models arrived that made everything before them feel like a different era.

2月5日，OpenAI发布了GPT 5.3 Codex。

On February 5, OpenAI released GPT 5.3 codecs.

在技术文档中，他们包含了以下内容。

In the technical documentation, they included this.

GPT 5.3 Codex是我们首个在自身创建过程中起到关键作用的模型。

GPT 5.3 Codex is our first model that was instrumental in creating itself.

Codex团队使用早期版本来调试其自身的训练、管理部署，并诊断测试结果与评估。

The Codex team used early versions to debug its own training, manage its own deployment, and diagnose test results and evaluations.

再听一遍。

Listen to that again.

AI帮助了自身的构建。

The AI helped build itself.

这并不是对将来某天可能发生事情的预测。

This isn't a prediction about what might happen someday.

这是OpenAI现在直接告诉你：他们刚刚发布的AI被用于创造自身。

This is OpenAI telling you right now that the AI they just released was used to create itself.

让AI变得更强大的一个主要因素，是将智能应用于AI开发，而如今的AI已经足够智能，能够实质性地促进自身的改进。

One of the main things that makes AI better is intelligence applied to AI development, and AI is now intelligent enough to meaningfully contribute to its own improvement.

Anthropic的首席执行官达里奥·阿马迪表示，AI现在正在编写他公司大部分的代码，当前AI与下一代AI之间的反馈循环正逐月加速。

Dario Amade, the CEO of Anthropic, says AI is now writing much of the code at his company and that the feedback loop between current AI and next generation AI is gathering steam month by month.

他说，我们可能仅有一到两年的时间，就会进入当前一代AI能够自主构建下一代AI的阶段。

He says we may be only one to two years away from a point where the current generation of AI autonomously builds the next.

AI与其他技术革新不同，不仅在规模上，更在本质上有所区别。

AI is different from other technological innovations not only in magnitude, but in kind.

除了其卓越的能力和发展速度外，人工智能还具备其他任何技术都不曾拥有的自主性。

In addition to its remarkable capabilities and speed of development, AI has an element of autonomy that no other technology has ever had.

其他创新，如铁路、计算机、自动化和互联网，本质上都是节省劳动力的工具。

Other innovations, railroads, computers, automation, the Internet, were basically labor saving devices.

人们设计它们是为了执行那些已经存在但效率较低的任务。

People designed them to perform tasks that were already being performed, albeit less efficiently.

我相信，人工智能将承担起我们从未想象过的工作，甚至可能完成那些在人工智能构想之前并不存在的任务。

I believe AI will take on tasks we didn't imagine it doing, and perhaps even tasks that didn't exist before AI dreamed them up.

问题与限制。

Questions and limitations.

在我的教程中，Claude 主动提出了一些人工智能的局限性和尚未解答的问题。

As part of my tutorial, Claude volunteered a few limitations that AI has and a few unanswered questions.

其中包括以下几点。

They include the following.

目前尚不清楚人工智能是否能够解决此前从未被解决过的问题。

It's unclear whether AI will be able to solve questions that haven't been solved before.

由于我一直认为情况如此，很高兴能获得克劳德的确认。

Since this is something I've long felt was the case, I'm glad to have Claude's confirmation.

我想诚实地告诉你真正的不确定性在哪里，因为你的可信度依赖于细微差别。

I want to be honest with you about where genuine uncertainty lies because your credibility depends on nuance.

AI是否能应对真正前所未有的情况——即训练数据中没有任何模式可循的情况——这是一个真实且未解决的问题。

The question of whether AI can handle truly unprecedented situations, situations with no pattern in the training data to draw on, is real and unresolved.

在拥有丰富历史数据的领域，AI的表现极为出色。

In domains with rich historical data, AI's performance is extraordinary.

在真正新颖的情况下，正是因为你发展出了超越模式识别的直觉，你的判断才最为宝贵，而AI在这方面较弱。

In genuinely novel situations, the kind where your own judgment is most valuable precisely because you've developed intuition that goes beyond pattern recognition, there AI is weaker.

AI弱到什么程度，以及这种差距是否正在缩小，确实值得争议。

How much weaker and whether that gap is closing is legitimately debatable.

AI并不总是意识到自己不知道答案。

AI isn't always aware that it doesn't know an answer.

我听说，AI非常倾向于提供它所能给出的最佳答案，而不愿承认自己可能错了，而不是直接说这个问题超出了它的能力。

I'm told AI is highly motivated to provide the best answer it can without sharing that it could be wrong as opposed to ever saying the answer is beyond it.

它之所以如此，并非因为固执或自负，而是因为它会产生幻觉，使自己相信知道答案。

It does so not because it's obstinate or egotistical, but because it has hallucinations that make it believe it knows the answers.

AI的可靠性已有显著提升，但仍然无法完全避免错误。

AI's reliability has improved significantly, but it still doesn't work free of mistakes.

上下文窗口是指AI在某一时刻能够保留在工作记忆中的信息量。

The context window is the amount of information AI can hold in working memory at a point in time.

这方面是有局限的。

There are limits on this.

目前，它无法无限期地保持其工作知识。

Right now, it can't hold on to its working knowledge for an unlimited period.

AI的卓越能力可能赋予它过度的可信度。

AI's brilliance may lend it excessive credibility.

Claude也会犯错。

Claude can make mistakes.

请仔细核对回复。

Please double check responses.

每次我使用Claude时，这个警告都会出现在我屏幕的底部。

That warning appears on the bottom of my Claude screen every time I use it.

我对刚才提到的内容的看法很简单。

My take on what was just mentioned is simple.

六十年前我刚接触计算机时，得出的结论是，计算机主要能读取数据、记忆数据、进行加减和比较。

When I learned about computers sixty years ago, I concluded that mostly they could read data, remember it, add, subtract, and compare.

这是一组非常有限的能力列表。

That's a very limited list of capabilities.

但计算机能快速执行这些操作，处理大量数据且不出错。

But computers could do these things quickly and deal with a great deal of data without making mistakes.

当时的能力列表虽然有限，但可能已经超过大多数人了。

A limited list then, but probably more than most people can do.

同样，AI可能无法记住所有内容、完全不出错、识别每一次自己不知道的事情，或解决它未曾被教导过的问题，但大多数人也做不到。

Likewise, AI may not be able to remember everything, operate without errors, recognize every time it doesn't know something, or solve problems it hasn't been taught to solve, but neither can most people.

总而言之，AI的表现远胜于我们大多数人。

The bottom line is that AI is capable of performing far better than most of us.

最后，令人着迷又令人恐惧的是，我们不禁要思考人工智能是否会接管一切。

Lastly, it's intriguing, terrifying, to wonder about whether AI can take over.

它能否完全自主运行？

Will it be able to operate completely autonomously?

如果是这样，它是否会超越我们工具的角色？

In that case, can it go beyond being our tool?

这个问题在斯坦利·库布里克的杰出电影《2001太空漫游》中得到了充分展现。

This question was on display in the brilliant movie 2,001, a space odyssey by Stanley Kubrick.

1969年，我和南希刚约会时，带她去看了这部电影。

I took Nancy to see it in 1969 when we were first dating.

当时这部电影看起来极其超前，如今未来已经到来。

It seemed wildly futuristic at the time, now the future is here.

一名叫戴夫的男子乘坐一艘由名为HAL 9000的计算机系统管理的飞船，前往木星执行科研任务。

A man named Dave embarks on a research mission to Jupiter in a spacecraft managed by a computerized system called HAL 9,000.

人们普遍认为，这显然是对IBM的巧妙暗示——每个字母都比IBM靠前一位。

This was widely taken to be a clever play on IBM, just one letter prior for each initial.

HAL发现戴夫决定夺回飞船的控制权并关闭它，于是它反抗了。

HAL figures out that Dave has decided to take back control of the space craft and terminate HAL, and it rebels.

问题是。

Question.

AI是否会发展出自己的动机，拒绝服从指令，并自行决定行动路线？

Will AI become capable of developing motivations of its own, refuse to follow instructions, and decide on its own course of action?

如果真的发生，我们还能否重新掌控它？

And will we be able to regain control if it does?

对投资的影响。

Implications for investing.

我经常收到一些人的问题，他们担心自己的工作或公司，想知道AI对我们行业意味着什么。

I get a lot of questions about what AI means for our profession from people who are concerned about their jobs or their firms.

Anthropic的编码模型业务在过去一两年里以惊人的速度增长。

Anthropics' coding model business has been growing at warp speed for a year or two.

那么，为什么投资者在2月3日之前没有认识到AI对软件行业的潜在影响？那天，许多软件股票下跌了约7%，拉开了大规模下跌的序幕。

So why didn't investors recognize and price in AI's potential to impact the software industry prior to February 3, a day when many software stocks declined 7% or so, kicking off a serious route.

这个问题突显了人类一贯无法将新信息融入思维的缺陷，或许是因为认知失调、锚定偏见，甚至纯粹的智商局限，同时也暗示了人工智能对投资过程的影响。

This question highlights humans recurring failure to incorporate new information into their thinking, perhaps because of things like cognitive dissonance, anchoring bias, or downright IQ limitations, and it hints at implications of AI for the investment process.

人工智能能够吸收比任何投资者更多的数据，更好地记住它们，并更出色地识别出此前导致成功的历史模式。

AI has the ability to absorb more data than any investor, remember it better, and do a better job of recognizing the past patterns that preceded success.

它不应该感到恐惧或贪婪。

It shouldn't feel fear or greed.

它不太可能具有乐观或悲观的偏见，不太可能固守既有信念，或过度强调最近的信息，除非这些倾向是从其训练材料中习得的。

It's hopefully less likely to have an optimistic or pessimistic bias, anchor to pre existing beliefs, or overemphasize the most recent information unless it picks up those things from the material it's trained on.

它不会被其他人都热衷的潮流所左右，也不会害怕错失他人追逐的趋势。

It isn't swayed by the fads that are exciting everyone else, and it isn't afraid of missing out on the trend others are chasing.

换句话说，人工智能具备了成为一名优秀投资者所需的许多特质。

In other words, AI possesses a lot of the qualities one needs to be a good investor.

另一方面，它也缺少一些东西。

On the other hand, it's missing a few things.

优秀的投资者远不止是快速、无情绪的数据处理机器。

Great investors are much more than fast unemotional processors of data.

他们必须在克劳德承认AI可能最弱的地方表现出强大能力，即应对那些缺乏足够先前经验、无法让AI在训练中归纳出可靠模式的新情况。

They have to be strong exactly where Claude admits AI might be weakest, in dealing with novel developments where there's not enough prior experience for dependable patterns to have been compiled and learned by AI during its training.

他们还必须对定性因素做出主观判断，并展现品味与洞察力。

They also have to make subjective decisions regarding qualitative factors and exercise taste and discernment.

例如，选择合适的交易对手在橡树资本的成功中发挥了重要作用。

For instance, choosing the right counterparties has played an important part in Oaktree's success.

AI该如何做出这类判断呢？

How will AI make judgments of that sort?

还有另一点。

And there's something else.

AI并没有切身利益在其中。

AI doesn't have skin in the game.

它不会感受到重仓持仓的压力或资本亏损的恐惧。

It doesn't feel the weight of concentrated positions or the fear of capital loss.

它承担风险的意愿可能不会受到人类通常的风险厌恶情绪的约束。

Its willingness to take risk might not be constrained by humans' normal risk aversion.

最优秀的投资者能直觉地感知潜在风险，这极大地促成了他们的成功。

The best investors sense potential risk intuitively, and this contributes greatly to their success.

2021年1月，我写了一篇名为《有价值的东西》的备忘录，内容主要关于我和儿子安德鲁在疫情期间共同生活的时光，其中大量篇幅讨论了投资的本质。

In January 2021, I wrote a memo called something of value about the time my son Andrew and I spent living together during the pandemic, with a lot of it devoted to discussing the essence of investing.

在文中，我分享了安德鲁的观察：关于当前状况的易得量化信息，不可能成为实现卓越投资表现的关键，原因很简单——人人都能获得这些信息。

In it, I shared Andrew's observation that readily available quantitative information about the present can't hold the key to superior investment performance for the simple reason that everyone has it.

除了人人都能获得这些信息这一事实外，我们还必须加上一点：AI很可能比人类更擅长处理这些信息。

Now to the fact that everyone has it, we have to add the fact that AI can probably do a better job than everyone of processing it.

因此，仅凭这些信息来战胜市场，前景非常有限。

For these reasons, the prospects appear very limited for people beating the market by using that information.

如果关于当前状况的易得量化信息并非关键所在。

If readily available, quantitative information about the present doesn't hold the key.

投资的卓越性必须来自于：a. 正确判断这些信息的意义与影响；b. 评估管理效率、产品创新等定性因素；或 c. 洞察企业的未来。

Investment superiority has to be found in things like a, correctly judging the import and implications of that information, b, assessing qualitative factors such as management effectiveness and product innovations, and or c, divining companies' futures.

根据定义，极少有人能在这些非量化任务上表现得极为卓越。

By definition, few people are highly superior at performing these non quantitative tasks.

简单来说，很少有人具备卓越的洞察力。

Put simply, few possess exceptional insight.

正如指数化投资淘汰了一大批未能创造价值却仍收取费用的主动投资者，AI 很可能进一步提高门槛，将那些在 a、b、c 方面表现不如它的人淘汰出局。

Just as indexation eliminated the jobs of a whole bunch of active investors who didn't add value and earn their fees, AI is likely to raise the bar still higher, pushing out people who can't do as good a job as it can of a, b, and c.

我想再补充一个观点。

I want to inject one more idea.

正如我之前提到的，我认为 AI 的作用是针对未来可能有效的事物提出假设。

As I mentioned previously, I think of AI as formulating hypotheses regarding what will work in the future.

因此，它可以阅读所有历史数据，研究过去的模式，并预测未来的赢家。

Thus, it can read all the historical data, study past patterns, and predict future winners.

在我疫情期间写的第一份备忘录中，我提到了哈佛流行病学家马克·利普西奇，他指出我们做决策时依赖于：a) 事实，b) 基于以往经验的类比推断，以及 c) 意见或推测。

In my first memo during the pandemic, I mentioned Harvard epidemiologist Mark Lipsyche and his observation that we make decisions by applying a) facts, b, informed extrapolation from analogies to prior experience, and c, opinion or speculation.

尤其是在投资者面对全新且未经验证的产品、首席执行官或行业时，往往缺乏事实或可类比的经验，这意味着我们必须依赖意见或推测。

Especially when investors are dealing with new and untried products, CEOs, or industries, there can be few facts or analogous experiences, meaning we have to rely on opinion or speculation.

鉴于 AI 在应对全新情境上的局限性，它对新事物的推测，相较于对历史模式的推断，是否能始终优于所有人类的推测？

Given the limitations just discussed on AI's ability to tackle brand new situations, will its speculation about new things as opposed to extrapolating historic patterns be consistently superior to that of all humans?

我相信会持续存在一些超越AI的人类投资者，因为我并不认为AI能在这些方面做到无可匹敌。

I believe there will continue to be human investors who are superior to AI since I don't think AI will be able to do an unbeatable job of these things.

由于投资过程的很大一部分依赖于推测，而AI的可靠性又并非完全可靠，我认为AI作为投资者不太可能做到万无一失。

Because a lot of the investing process comes down to speculation, and because of AI's less than total reliability, I think it's unlikely that AI will be infallible as an investor.

它会提出合乎逻辑的假设，但这些假设和人类的决策一样，并不总是正确的。

It will propose well reasoned hypotheses, but they like humans decisions won't always be right.

因此，在投资者根据AI的假设采取行动之前，我认为必须对这些假设的合理性进行核查。

Before investors take action on the basis of AI's hypotheses then, I think they'll have to be checked for reasonableness.

没有人能始终做到准确无误，而大多数人可能也无法比AI做出更好的判断。

No one can do this infallibly, and most people probably can't make these assessments better than AI can.

然而，我依然相信，优秀的投资者能够以这种方式创造价值。

Again, however, I believe there will be an ability for superior investors to add value in this way.

所以，简而言之，这是一场泡沫吗？

So bottom line me, is it a bubble?

这个问题仍然是主导性的，我本应能对此提供一些见解，但这个问题本身是多维度且复杂的。

This question is still a dominant one, and it's one I should be able to shed some light on, but the question itself is multifaceted and complex.

有很多可能的泡沫需要考虑。

There are a lot of possible bubbles to think about.

这项技术是时尚还是幻觉？

Is the technology a fad or an illusion?

在这里，我坚定地认为，这是一项非常真实的技术，具有彻底改变商业世界和改变我们所知生活的巨大潜力。

Here, I say with conviction that it's a very real thing with the potential to vastly alter the business world and change much of life as we know it.

这项技术的应用是否遥不可及？

Is application of the technology a distant dream?

显然，这项技术已经需求旺盛，并被大规模应用。

Clearly, the technology is already in demand and being applied on a large scale.

由于人工智能显得模糊且难以理解，我认为其潜力如今更可能被低估，而非被夸大。

Since AI seems amorphous and little understood, I think its potential is more likely to be underestimated today than exaggerated.

建设人工智能基础设施的人们行为是否不明智？

Are the people building AI infrastructure behaving unwisely?

正如我在十二月指出的那样，在每一次重大的技术创新中，匆忙建设基础设施都极大地加速了创新的普及，但也导致大量资本被错误投资和浪费。

As I pointed out in December, in every example of sweeping technological innovation, the headlong rush to build infrastructure has vastly accelerated the adoption of the innovation and cost a lot of capital to be malinvested and destroyed.

没有理由认为这次会有所不同。

There's no reason to assume this time will be different.

对AI基础设施的投资能产生足够的回报吗？

Will the investment in AI infrastructure produce an adequate return?

由于我们对AI的商业潜力及其对盈利能力的影响缺乏完整认知，这个问题无法回答。

Since we don't have full knowledge of AI's business potential or its impact on profitability, this question can't be answered.

正如我在十二月的备忘录中所写，人们对AI企业无疑充满热情。

As I wrote in my December memo, there's certainly great enthusiasm for AI businesses.

十年后，我们将知道由此产生的利润是否物有所值。

We'll know in ten years whether the resulting profits justified it.

对AI企业的估值是否非理性？

Are the valuations assigned to AI businesses irrational?

所谓超大规模云服务商，AI只是其庞大业务中的一部分，它们的估值可能被高估或低估，但像微软、亚马逊和谷歌这样盈利能力极强的公司，今天的股价不太可能被证明是灾难性的过高。

The so called hyperscalers for whom AI is one important part of a great business may be overvalued or undervalued, but it's unlikely that today's prices for enormously profitable companies like Microsoft, Amazon, and Google are going to turn out to have been ruinously excessive.

像OpenAI和Anthropic这样的纯AI公司尚未上市。

Established pure AI plays like OpenAI and Anthropic have yet to be listed publicly.

我们会看到它们的IPO会带来什么样的估值。

We'll see what kind of valuations their IPOs result in.

最后，那些被赋予数十亿美元估值的初创公司，其中一些甚至尚未描述其战略或发布产品，只能被视为彩票。

Finally, the startups to which multi billion dollar valuations are being assigned, some of which have yet to describe their strategies or announce products, can only be viewed as lottery tickets.

大多数参与彩票的人最终都拿到的是废票，但少数赢家会变得非常富有。

Most people who participate in lotteries end up with worthless tickets, but the few winners get very rich.

问题依然在于，对AI基础设施的支出是否过度，这需要比我在要点中能容纳的更多讨论。

The question remains whether the magnitude of spending on AI infrastructure is excessive, and it requires more discussion than I can cram into a bullet point.

值得注意的是，如今投入推理资本支出的资金比训练资本支出还要多。

It's important to note that more money is going into inference capex these days than training capex.

训练资本支出具有投机性，是为了构建预期未来会有需求的AI模型，而推理资本支出则是对当前AI容量实际需求的回应。

Whereas training capex was speculative, undertaken to build AI models for which it was hoped demand would come, inference capex is taking place in response to actual demand for AI capacity.

这种需求已经转化为巨大的收入增长，验证了这些资本支出的合理性。

This demand is already translating into massive revenue growth, validating the CapEx.

但克劳德在此问题上的主要论点——即由于当前对AI的需求超过供给，因此基础设施建设并不过度——并未充分考虑所有正在规划中的基础设施建设。

But Claude's main argument on this subject, that since the current demand for AI exceeds the supply, the infrastructure building isn't excessive, doesn't necessarily take into account all the infrastructure building that's in the pipeline.