2025-10-30 02:40:06
你应该去中国。中国地域辽阔，非常多样，安全，出行方便，价格实惠，有趣，与你想象的不一样，而且在世界上的地位越来越重要。亲自去看看吧。
然而，与中国其他地方不同，中国有自己的平行应用程序集，你必须在那裡使用它们。以下是我认为在中国独立旅行必备的移动应用程序。一个好习惯是出发前在中国以外下载这些应用程序，因为大多数应用程序都受到“防火长城”的限制。确保你下载的是应用程序的“国际”版本（如果有的话），这样它就可以使用英文。
（中国现在有一个新的免签10天选项，使访问变得更加容易。有关这种过境签证的详细信息，请参见这篇博客帖子。）
Airalo
现代的中国生活离不开手机。你需要一个可靠的本地网络连接来使用这些应用程序。一些美国的移动套餐，如T-Mobile或Google Fi，会自动为你提供在中国的手机信号，而像Verizon这样的套餐则每天收费高昂。对于其他人，或者为了避免额外费用，你需要一张本地SIM卡。你可以在抵达机场时购买SIM卡，但最近的手机型号支持eSIM，这是一种你只需下载的应用程序，而不是实体卡。你为不同的数据套餐付费。
有许多eSIM初创公司，如Saily、Holify和Airalo。你也可以从Trip（见下文）购买eSIM。目前，为所有这些eSIM加载数据都比应有的更繁琐，因此我建议你在抵达中国之前就下载好eSIM。一些eSIM的好处是它们通过香港或新加坡路由流量，因此它们可以充当内置的VPN，这意味着你不需要在手机上安装其他VPN应用（见下文）。我在亚洲和欧洲国家一直使用Airalo作为eSIM，它在中国也能正常工作。重要的是，Airalo内置了VPN，因此在中国，我可以轻松地在手机上使用电子邮件、聊天AI和新闻网站。
Express VPN
防火长城会阻止Gmail、ChatGPT、社交媒体、YouTube和主要新闻网站等。为了绕过你手机上这个令人沮丧的限制，使用一个好的eSIM。为了在你的笔记本电脑上绕过这个限制，你需要一个VPN。VPN在与防火长城的猫鼠游戏中不断更新，但它们并不总是有效。我在使用Express VPN的付费版本（每月8美元）时在中国取得了成功。其他旅行者和许多在中国生活的外籍人士更喜欢AstrillVPN，虽然价格更高，每月15美元。确保在出发前更新到最新版本，因为它们经常更新。最好下载另一个版本作为备份，以防你首选的版本停止工作，这种情况并不罕见。
Alipay
中国已经实现了无现金支付；现在没有人再使用现金，除了那些绝望的游客。而且大多数商店都不接受信用卡。中国有两种广泛使用的数字现金系统，即微信支付和支付宝。我推荐使用支付宝，因为它们现在让外国人很容易地设置与信用卡绑定的账户。支付宝应用会生成一个二维码，商家可以扫描，或者你可以扫描他们的二维码。支付宝是一个超级应用，里面包含许多子应用，如网约车、共享单车、门票和翻译。这个应用是必不可少的。
Didi
中国没有Uber或Lyft，但有Didi，其功能类似。你可以安装Didi应用，但更好的是，Didi已经整合在支付宝应用中。因此你不需要设置任何新东西。在支付宝中查找Didi即可。非常方便。我更喜欢使用Didi而不是出租车，以避免在沟通目的地时遇到困难。常见的做法是当你进入出租车时，报出你手机号码的最后四位数字来确认你的行程。
WeChat
使用微信来与中国人进行短信、通话和聊天，无论他们处于什么层次。这是中国每个人都有的一款应用。当你分别或见面之前，每个人都会要求你的微信账号。它无疑是与中国人沟通的首选方式。你们可以通过扫描彼此的二维码来交换联系方式。在中国以外安装微信可能有些麻烦，但这是个好主意。
Amap
Amap是中国最好的地图应用。（Google和Apple地图在中国的更新频率和详细程度不如Amap。）除了导航功能，Amap还可以用来搜索附近餐厅或景点的英文信息。它有点像Yelp的简化版。它还会提供城市公共交通和地铁的详细路线和连接信息，这在实际中非常有用。 （使用支付宝购买车票。）你需要在手机上离线下载该应用，或者使用VPN来获取英文版本。
Apple Translate
你可以使用Google翻译应用（搭配VPN），但通常我会使用iPhone内置的Apple Translate应用。 （支付宝也有一个翻译应用。）你至少需要其中一个。你可以将手机或相机对准标志或菜单上的文字进行翻译，或者使用“对话”模式将语音转换为文字，甚至进行语音对话的翻译。
Trip
在中国预订航班和高铁时，使用Trip。它们可靠，几乎涵盖所有选项。高铁几乎可以到达任何地方，但国内航班便宜且众多。你可以通过Trip预订两者。此外，Trip在中国预订酒店也最方便。出于习惯，我仍然使用Booking.com，它也适用于酒店（不适用于高铁或航班）。Trip在非旅游地区有最好的选择，而Booking.com则适用于大多数城市。
---------------
You should visit China. It’s vast, very diverse, safe, easy to get around, inexpensive, interesting, not what you expect, and increasingly important in the world. Go see for yourself.
However, unlike the rest of the world, China uses its own parallel set of apps that you will need to operate there. Here are what I consider the essential mobile apps for independent travel in China. A good rule of thumb is to download your apps outside of China before you leave, because most are behind their great firewall. Make sure you are downloading the “international” version of the app (if it has one) so that it uses English.
(China has a new visa-free 10-day option that makes it easier than ever to visit. For details of this transit visa see this blog thread.)
Airalo
Modern China lives on the phone. You need a good solid local connection for your apps. Some US-based mobile plans like T-Mobile or Google Fi will automatically give you cell coverage in China, and others like Verizon will charge you a premium $12/day. For everyone else, or to avoid the surcharge, you’ll need a local SIM. You can get a SIM card at the landing airport but recent phone models permit eSIMS, which is a mere app you load instead of a card. You pay for different data plans.
There are a bunch of eSIM startups, like Saily, Holify, and Airalo. You can also purchase an eSIM from Trip (see below). Loading an eSIM at the moment is way more cumbersome than it should be for all of them, so I’d recommend you download your eSIM before you arrive. One of the benefits of some of the eSIMS is that they route their traffic through Hong Kong or Singapore so they act as a built-in VPN, which means you can avoid needing that app on your phone (see below). I’ve been using Airalo as an eSIM in Asian and European countries, and it works fine in China. Importantly, Airalo has a VPN built in, so in China I could do my email, chat AI, and newssites on my phone effortlessly.
Express VPN
The Great Firewall will block Gmail, ChatGPT, social media, YouTube, and major newssites among others. To get around this frustrating block on your phone, use a good eSIM. To get around it on your laptop you’ll need a VPN. VPNs are in a cat-and-mouse race to keep the channel open, and they don’t always work. I’ve had success in China using a paid version of Express VPN ($8 per month). Other travelers and many expats living in China prefer AstrillVPN, which is more expensive at $15 per month. Make sure you have updated it before you leave with the latest version, which is updated frequently. It’s a good idea to download a second version as a backup in case your main choice stops working, which is not uncommon.
Alipay
China is cashless; no one uses cash anymore, except desperate tourists. And most stores do not accept credit cards. There are two widespread digital cash systems, WeChat and Alipay. I recommend using Alipay because they now make it very easy for foreigners to set up an account linked to your credit card. The Alipay app will generate a QR code the seller will scan, or you can scan their QR code. Alipay is a super-app with many sub apps inside it, such as rideshares, bikeshares, tickets, and translation. This app is an essential must-have.
Didi
There is no Uber or Lyft, but there is Didi, which works the same. You can install the Didi app, but even better, Didi is available within the Alipay app. So you don’t need to set up anything new. Look for Didi inside Alipay. Very convenient. I prefer using Didi to a taxi to eliminate the challenge of communicating my destination. The common protocol is to recite the last four digits of your phone number when you enter the car to confirm your ride.
Use Wechat to text, call, chat with other Chinese at all levels. It’s the one app everyone in China has. Everyone will ask you for your Wechat account when you part or before you meet. It by far is the preferred way to communicate. You scan each other’s QR codes to exchange numbers. Installing WeChat outside of China is cumbersome, but a good idea.
Amap
Amap is the best map app in China. (Google and Apple Maps are not as current, nor as detailed as Amap.) Besides navigation purposes, Amap is also useful for searching for what restaurants or sights are nearby in English. It kind of works like Yelp-lite. It also gives you detailed instructions and connections for public transit and subways in a city, which is a real lifesaver. (Use Alipay to purchase your tickets.) You need to load the app on your phone away from China or with a VPN in order to get the English version.
Apple Translate
You can use the Google translate app (with a VPN), but I usually wind up using the built in Apple Translate app on my iPhone. (Alipay has a translation app, too). You’ll need at least one of them. Point your camera/phone at text on a sign or menu for translation, or use ‘conversation” mode to translate speech as text, or even a spoken conversation.
Trip
For booking flights and high speed trains in China, use Trip. They are reliable and pretty much cover all options. Trains go almost anywhere, but domestic flights are cheap and plentiful. You can book both with Trip. Also Trip is the easiest for booking hotels in China. Just out of habit I still use Booking.com, which also works for hotels (not trains or planes). Trip has the best selection in less touristy places, while Booking.com will work for most cities.
2025-10-09 06:21:21
2025-10-08 01:09:36
一些作者的观点颠倒了。他们认为AI公司应该为在他们的书籍上训练AI而支付报酬。但我的预测是，在不久的将来，作者们将支付给AI公司，以确保他们的书籍被纳入AI的教育和训练中。作者（以及他们的出版商）会支付费用，以便对AI提供的答案和服务产生影响。如果您的作品不被AI所知悉和欣赏，那么它实际上将无人知晓。
最近，AI公司Anthropic同意向书籍作者支付总计15亿美元的集体罚金，作为非法复制其书籍的惩罚。Anthropic曾因使用包含50万本书籍的“影子图书馆”而被一些作者起诉，该图书馆包含了这些作者书籍的数字版本，均由一些梦想让所有书籍对所有人开放的“叛逆图书管理员”收集。Anthropic曾下载了这个非法图书馆的副本，以备用于训练其大语言模型（LLMs），但根据法院文件，他们最终并未使用这些书籍来训练他们发布的人工智能模型。即使Anthropic没有使用这个特定的图书馆，他们也使用了类似的东西，而所有其他商业前沿的大语言模型也都是如此。
然而，法官处罚他们是因为未经授权复制了受版权保护的书籍，无论是否实际使用了这些书籍，所有被复制书籍的作者都获得了每本书3000美元的赔偿。
在本案中，法院管理员发布了受影响书籍的可搜索列表，专门网站上列出了这些书籍。任何人都可以搜索数据库，查看特定书籍或作者是否包含在该盗版图书馆中，当然也可以确认他们是否应获得赔偿。我过去参与过类似的集体诉讼案件，发现很少有实际的钱能真正到达普通民众手中。大部分费用都被各方律师消耗掉了。我注意到，在本案中，每本书的赔偿金额中只有50%会真正支付给作者，而另一半则归出版商所有。也许如此。如果是教科书，你可能连一点补偿都得不到。
我是一名作者，因此我查看了Anthropic的案件列表。我发现我出版的五本书中有四本被包含在该图书馆中。我感到荣幸能被纳入一组可以训练AI的书籍中，这些AI我现在每天都在使用。我也感到自豪，因为我的想法可能通过LLMs的思维链影响到数以百万计的人。我也可以想象一些作者会感到失望，因为他们的作品没有被包含在该图书馆中。
然而，Anthropic声称他们并未使用这个特定的图书馆来训练他们的AI。他们可能使用了其他图书馆，而这些图书馆可能或可能没有合法地获得授权。使用数字化书籍的合法性仍存在争议。例如，Google为了搜索目的数字化书籍，但只显示书籍的片段作为搜索结果。他们是否可以使用已经制作的数字副本用于训练AI？在Bartz诉Anthropic一案中，裁决认为，如果书籍副本是通过正当方式获得的，那么用于训练AI的书籍副本属于合理使用。Anthropic被处罚的原因不是因为用书籍训练AI，而是因为他们持有未经付费的书籍副本。
这仅仅是关于AI训练的法律测试案例之一，但显然未来会有更多类似的测试。因为很明显，版权法无法覆盖这种新的文本使用方式。保护文本的副本——这是版权条款所做的事情——与学习和训练无关。AI不需要保存副本，它们只需要读一遍。副本并不重要。我们可能需要其他类型的知识产权权利和许可，例如“引用权”或类似的概念。但权利问题只是次要的，主要事件是AI这一新受众的崛起。
我们将逐渐积累一些最佳实践，以确定如何训练和教育AI。用于教育AI代理人的材料的整理将成为决定我们是否使用和依赖它们的重要因素。会有少数客户希望AI的训练材料符合他们的政治倾向。虔诚的保守派可能希望AI接受保守教育，这样它们在回答有争议的问题时会按照他们的喜好来呈现。虔诚的自由派则希望AI接受自由派教育。大多数人不会在意，他们只是想要“最好的”答案或最可靠的服务。我们知道AI会反映它们所训练的内容，并且可以通过人类干预进行“微调”，以产生符合用户喜好的答案和服务。在强化AI行为和引导其思维方面，已有大量研究。
50万本书听起来像是很多书籍可供学习，但世界上已有数以百万计的书籍未被AI阅读，因为它们的版权状态不明确或不便，或者它们是用较少使用的语言书写的。AI的训练远未完成。塑造这一可能影响的语料库将本身成为一门科学和艺术。某一天，AI将真正阅读人类写下的所有内容。目前只有50万本书构成你的知识库，很快就会显得很陈旧，但这也表明，被包含在这一小部分书籍中将有多么大的影响力，这正是作者们现在希望他们的作品被AI训练的原因。
年轻人和最早采用AI的人群已经将AI设置为始终开启模式；越来越多的无形生活通过AI进行，且不再有其他途径。随着AI模型变得越来越可靠，年轻人开始接受AI的结论。我在自己的生活中也发现了类似的现象。我很久以前就不再质疑计算器，然后也不再质疑谷歌，现在我发现当前的AI提供的大多数答案都很可靠。AI正在成为真理的仲裁者。
AI代理不仅被用来提供答案，还被用来查找事物、理解事物和提出建议。如果AI不知道某件事情，那就相当于它不存在。因此，选择退出AI训练的作者将很难产生影响。目前，有相当多的作者和创作者根本没有数字存在；你无法在网上找到他们，他们的作品也没有被任何地方列出。他们非常稀少，属于少数群体。正如Tim O’Reilly所说，如今大多数创作者面临的挑战不是盗版（非法复制），而是被忽视。我补充说，未来创作者面临的挑战将不是模仿（AI复制），而是被忽视。
如果AI成为真理的仲裁者，并且它们所训练的内容至关重要，那么我希望我的想法和创作作品在它们所看到的内容中占据主导地位。我非常希望我的书籍成为AI的教科书。哪个作者不想呢？我也不例外。我希望我的影响力扩展到每天使用AI的数十亿人，我甚至可能愿意为此付费，或者至少做些事情来促进我的作品被AI吸收。
另一种思考方式是，在这个新兴的环境中，书籍的受众——尤其是非虚构类书籍——已经从人类转向了AI。如果你今天正在写一本书，你必须意识到你主要是为AI写作。它们将是阅读最多、最仔细的读者。它们会逐字逐句地阅读每一本书，包括脚注、尾注、参考文献和附录。它们还会阅读你所有的书籍并收听你的播客。你不太可能有任何人比AI更彻底地阅读你的作品。在吸收了这些内容后，AI会做一件神奇的事情：将你的文本与其他它们读过的文本结合，将它置于全球知识体系中，以一种人类读者无法做到的方式。
被AI吸收的材料的成功部分取决于这些材料如何被呈现给AI。如果一本书更容易被AI解析，它的影响力就会更大。因此，许多书籍将被撰写和格式化，以适应其主要受众。为AI写作将成为一种技能，就像其他任何技能一样，你也可以通过练习来提高。作者们可能会积极寻求优化他们的作品以适应AI的吸收，甚至可能与AI公司合作，确保他们的内容被正确理解并整合。这种“AI友好型”写作，具有清晰的结构、明确的论点和定义良好的概念，将变得越来越重要，当然，AI本身也会帮助实现这一点。
我们创作的每一本书、歌曲、戏剧、电影都会被加入我们的文化。图书馆是人类发明中特别的一种。它们往往越老越有价值。它们积累智慧和知识。互联网在这方面也类似，它持续积累材料，自诞生以来从未崩溃或需要重启。AI很可能也类似这些外向系统，不断积累知识而不会中断。我们尚不确定，但它们很可能在数十年甚至更长时间内持续增长。目前它们的增长似乎是无止境的。今天学到的知识，它们明天仍然会记得，而它们今天的影响将在未来的几十年中产生复利效应。影响AI将成为当今人类可进行的最高杠杆效应活动，而且你开始得越早，效果就越显著。
作者作品的价值将不仅体现在它在人类中的销售情况，还体现在它被包含在这些基于记忆的智能系统的基础知识中的深度。这种影响力将成为宣传的重点，也将成为作者的遗产。
Some authors have it backwards. They believe that AI companies should pay them for training AIs on their books. But I predict in a very short while, authors will be paying AI companies to ensure that their books are included in the education and training of AIs. The authors (and their publishers) will pay in order to have influence on the answers and services the AIs provide. If your work is not known and appreciated by the AIs, it will be essentially unknown.
Recently, the AI firm Anthropic agreed to pay book authors a collective $1.5 billion as a penalty for making an illegal copy of their books. Anthropic had been sued by some authors for using a shadow library of 500,000 books that contained digital versions of their books, all collected by renegade librarians with the dream of making all books available to all people. Anthropic had downloaded a copy of this outlaw library in anticipation of using it to train their LLMs, but according to court documents, they did not end up using those books for training the AI models they released. Even if Anthropic did not use this particular library, they used something similar, and so have all the other commercial frontier LLMs.
However the judge penalized them for making an unauthorized copy of the copyrighted books, whether or not they used them, and the authors of all the copied books were awarded $3,000 per book in the library.
The court administrators in this case, called Bartz et al v. Anthropic, have released a searchable list of the affected books on a dedicated website. Anyone can search the database to see if a particular book or author is included in this pirate library, and of course, whether they are due compensation. My experience with class action suites like this is that very rarely does award money ever reach people on the street. Most of the fees are consumed by the lawyers of all sides. I notice that in this case, only half of the amount paid per book is destined to actually go to the author. The other 50% goes to the publishers. Maybe. And if it is a text book, good luck with getting anything.
I am an author so I checked the Anthropic case list. I found four out of my five books published in New York included in this library. I feel honored to be included in a group of books that can train AIs that I now use everyday. I feel flattered that my ideas might be able to reach millions of people through the chain of thought of LLMs. I can imagine some authors feeling disappointed that their work was not included in this library.
However, Anthropic claims it did not use this particular library for training their AIs. They may have used other libraries and those libraries may or may not have been “legal” in the sense of having been paid for. The legality of using digitized books for anything is still in dispute. For example, Google digitizes books for search purposes, but only shows small snippets of the book as the result. Can they use the same digital copy they have already made for training AI purposes? The verdict in the Bartz v. Anthropic case was that, yes, using a copy of a book for training AI is fair use, if it was obtained in a fair way. Anthropic was penalized not for training AI on books, but for having in its possession a copy of the books it had not paid for.
This is just the first test case of what promises to be many more tests in the future as it is clear that copyright law is not adequate to cover this new use of text. Protecting copies of text – which is what copyright provisions do – is not really pertinent to learning and training. AIs don’t need to keep a copy; they just have to read it once. Copies are immaterial. We probably need other types of rights and licenses for intellectual property, such as a Right of Reference, or something like that. But the rights issue is only a distraction from the main event, which is the rise of a new audience: the AIs.
Slowly, we’ll accumulate some best practices in regards to what is used to train and school AIs. The curation of the material used to educate the AI agents giving us answers will become a major factor in deciding whether we use and rely on them. There will be a minority of customers who want the AIs to be trained with material that aligns with their political bent. Devout conservatives might want a conservatively trained AI; it will give answers to controversial questions in the manner they like. Devout liberals will want one trained with a liberal education. The majority of people won’t care; they just want the “best” answer or the most reliable service. We do know that AIs reflect what they were trained on, and that they can be “fine tuned” with human intervention to produce answers and services that please their users. There is a lot of research in reinforcing their behavior and steering their thinking.
Half a million books sounds like a lot of books to learn from, but there are millions and millions of books in the world already that the AIs have not read because their copyright status is unclear or inconvenient, or they are written in lesser-used languages. AI training is nowhere near done. Shaping this corpus of possible influences will become a science and art in itself. Someday AIs will have really read all that humans have written. Having only 500,000 books forming your knowledge base will soon be seen as quaint, but it also suggests how impactful it can be to be included in that small selection, and that makes inclusion a prime reason why authors will want their works to be trained on AIs now.
The young and the earliest adopters of AI have it set to always-on mode; more and more of their intangible life goes through the AI, and no further. As the AI models become more and more reliable, the young are accepting the conclusions of the AI. I find something similar in my own life. I long ago stopped questioning a calculator, then stopped questioning Google, and now find that most answers from current AIs are pretty reliable. The AIs are becoming the arbiters of truth.
AI agents are used not just to give answers but to find things, to understand things, to suggest things. If the AIs do not know about it, it is equivalent to it not existing. It will become very hard for authors who opt out of AI training to make a dent. There are authors and creators today who do not have any digital presence at all; you cannot find them online; their work is not listed anywhere. They are rare and a minority. As Tim O’Reilly likes to say, the challenge today for most creators is not piracy (illegal copies) but obscurity. I will add, the challenge for creators in the future will not be imitation (AI copy) but obscurity.
If AIs become the arbiters of truth, and if what they trained on matters, then I want my ideas and creative work to be paramount in what they see. I would very much like my books to be the textbooks for AI. What author would not? I would. I want my influence to extend to the billions of people coming to the AIs everyday, and I might even be willing to pay for that, or to at least do what I can to facilitate the ingestion of my work into the AI minds.
Another way to think of this is that in this emerging landscape, the audience for books – especially non-fiction books – has shifted away from people towards AI. If you are writing a book today, you want to keep in mind that you are primarily writing it for AIs. They are the ones who are going to read it the most carefully. They are going to read every page word by word, and all the footnotes, and all the endnotes, and the bibliography, and the afterward. They will also read all your books and listen to all your podcasts. You are unlikely to have any human reader read it as thoroughly as the AIs will. After absorbing it, the AIs will do that magical thing of incorporating your text into all the other text they have read, of situating it, of placing it among all the other knowledge of the world – in a way no human reader can do.
Part of the success of being incorporated by AIs is how well the material is presented for them. If a book can be more easily parsed by an AI, its influence will be greater. Therefore many books will be written and formatted with an eye on their main audience. Writing for AIs will become a skill like any other, and something you can get better at. Authors could actively seek to optimize their work for AI ingestion, perhaps even collaborating with AI companies to ensure their content is properly understood, and integrated. The concept of “AI-friendly” writing, with clear structures, explicit arguments, and well-defined concepts, will gain prominence, and of course will be assisted by AI.
Every book, song, play, movie we create is added to our culture. Libraries are special among human inventions. They tend to get better the older they get. They accumulate wisdom and knowledge. The internet is similar in this way, in that it keeps accumulating material and has never crashed, or had to restart, since it began. AIs are very likely similar to these exotropic systems, accumulating endlessly without interruption. We don’t know for sure, but they are liable to keep growing for decades if not longer. At the moment their growth seems open ended. What they learn today, they will probably continue to know, and their impact today will have compounding influence in the decades to come. Influencing AIs is among the highest leverage activities available to any human being today, and the earlier you start, the more potent.
The value of an author’s work will not just be in how well it sells among humans, but how deep it has been included within the foundational knowledge of these intelligent memory-based systems. That potency will be what is boasted about. That will be an author’s legacy.
2025-09-24 06:50:22
我一直在研究电的发现早期历史，作为我们当前发现人造智能的指南。当时最聪明的人，包括艾萨克·牛顿，他可能是有史以来最聪明的人，都对电的本质有自信的理论，但这些理论却严重错误。事实上，尽管电荷在宇宙中起着基本作用，但很长一段时间内，所有研究这一基本力量的人都严重错误。所有电的先驱者——如富兰克林、惠斯通、法拉第和麦克斯韦——都有他们自己的正确想法（并非所有人都有），但这些想法中混杂着许多后来被证明是完全误导的观念。关于电能做什么的大多数发现，都是在不了解其工作原理的情况下进行的。这种知识的缺乏，当然，极大地减缓了电气发明的进步。
同样地，今天最聪明的人，尤其是所有创造人工智能的天才，都有关于智能本质的理论，我认为所有这些理论（包括我自己）都会严重错误。我们之所以不知道人工智能是什么，很大程度上是因为我们不知道自己的智能是什么。这种无知将来会被视为阻碍人工智能进步速度的因素。
我们对智能的无知很大程度上源于我们对电或智能这一总体类别的混淆。我们倾向于将电和智能视为单一维度上的连贯基本力量：你拥有更多或更少。但事实上，电被证明是如此复杂，如此充满反直觉效应，以至于即使今天，我们仍然难以理解它的工作原理。电有粒子和波，有场和流，由并不存在的事物组成。我们对电的使用超过了我们对其的理解。理解电对于理解物质至关重要。直到我们学会了控制电，才得以将水——曾被视为元素——分解为它的实际元素；这使我们认识到水并不是基础元素，而是一种由子元素组成的衍生化合物。
我认为我们很可能也会发现，智能同样不是一种基础的单一元素，而是一种由多种认知元素组成的衍生化合物，每种智能物种都有其独特的复杂系统。我们称之为智能的结果，来源于许多不同的认知原始元素，如长期记忆、空间意识、逻辑推理、预先规划、模式感知等。可能有几十种，甚至上百种。我们目前对这些元素一无所知。我们缺乏认知的元素周期表。
这些认知元素更像重元素一样不稳定且动态。或者更好的类比是生物细胞中的元素。认知的原始元素是思维循环中的流动状态。它们就像细胞中的分子，处于持续的流动中，不断改变形状。它们的分子身份与其与其他分子的作用和互动有关。思维是一种集体行为，发生在时间中（就像物质中的温度），每种模式只能在与其他模式之前和之后的关系中看到。这是一种网络现象，使得确定其边界变得困难。因此，每个智能元素都嵌入在思维循环中，并需要其他元素作为其身份的一部分。因此，每个认知元素都是在与其相邻的认知模式的背景下描述的。
我让ChatGPT5Pro帮助我生成一个认知元素周期表，基于我们目前所知的信息。它建议有49个元素，排列成一个表格，使得相关概念相邻。列是认知的家族或一般类别，如“感知”、“推理”、“学习”，因此所有类型的感知或推理都堆叠在某一列中。行按思维循环的阶段排序，早期阶段（如“感知”）位于顶部，而循环后期阶段（如“反思与对齐”）位于底部。例如，在“安全”这一家族或类别中，人工智能会先进行不确定性估计，之后进行验证，最后才达到心智理论。
该图表根据我们在每个元素上取得的进展进行着色。红色表示我们能够以稳健的方式合成该元素。橙色表示我们可以在正确框架下使其大致运行。黄色则反映有前景的研究，但尚未具备操作上的普遍性。
我怀疑这些元素之间的区分可能不像这里所展示的那样明显（在分类学上，我更倾向于合并而非细分），并且我认为这个集合可能遗漏了许多我们即将发现的类型。但作为起点，这个原型图表达到了它的目的：它揭示了智能的复杂性。很明显，智能是在多个维度上叠加的。我们将设计不同的AI，使其具有不同元素的不同组合和不同强度，这将产生数千种可能的智能类型。我们可以看到，即使今天，不同动物也有自己的认知原始元素组合，这些组合以符合其物种需求的独特模式排列。在某些动物中，某些元素——比如长期记忆——可能比我们自己的更强；当然，它们缺乏我们拥有的某些元素。
---------------
I’ve been studying the early history of electricity’s discovery as a map for our current discovery of artificial intelligence. The smartest people alive back then, including Isaac Newton, who may have been the smartest person who ever lived, had confident theories about electricity’s nature that were profoundly wrong. In fact, despite the essential role of electrical charges in the universe, everyone who worked on this fundamental force was profoundly wrong for a long time. All the pioneers of electricity — such as Franklin, Wheatstone, Faraday, and Maxwell — had a few correct ideas of their own (not shared by all) mixed in with notions that mostly turned out to be flat out misguided. Most of the discoveries about what electricity could do happened without the knowledge of how they worked. That ignorance, of course, drastically slowed down the advances in electrical inventions.
In a similar way, the smartest people today, especially all the geniuses creating artificial intelligence, have theories about what intelligence is, and I believe all of them (me too) will be profoundly wrong. We don’t know what artificial intelligence is in large part because we don’t know what our own intelligence is. And this ignorance will later be seen as an impediment to the rate of progress in AI.
A major part of our ignorance stems from our confusion about the general category of either electricity or intelligence. We tend to view both electricity and intelligence as coherent elemental forces along a single dimension: you either have more of it or less. But in fact, electricity turned out to be so complicated, so complex, so full of counterintuitive effects that even today it is still hard to grasp how it works. It has particles and waves, and fields and flows, composed of things that are not really there. Our employment of electricity exceeds our understanding of it. Understanding electricity was essential to understanding matter. It wasn’t until we learned to control electricity that we were able to split water — which had been considered an element — into its actual elements; that enlightened us that water was not a foundational element, but a derivative compound made up of sub elements.
It is very probable we will discover that intelligence is likewise not a foundational singular element, but a derivative compound composed of multiple cognitive elements, combined in a complex system unique to each species of mind. The result that we call intelligence emerges from many different cognitive primitives such as long-term memory, spatial awareness, logical deduction, advance planning, pattern perception, and so on. There may be dozens of them, or hundreds. We currently don’t have any idea of what these elements are. We lack a periodic table of cognition.
The cognitive elements will more resemble the heavier elements in being unstable and dynamic. Or a better analogy would be to the elements in a biological cell. The primitives of cognition are flow states that appear in a thought cycle. They are like molecules in a cell which are in constant flux, shifting from one shape to another. Their molecular identity is related to their actions and interactions with other molecules. Thinking is a collective action that happens in time (like temperature in matter) and every mode can only be seen in relation to the other modes before and after it. It is a network phenomenon that makes it difficult to identify its borders. So each element of intelligence is embedded in a thought cycle, and requires the other elements as part of its identity. So each cognitive element is described in context of the other cognitive modes adjacent to it.
I asked ChatGPT5Pro to help me generate a periodic table of cognition given what we collectively know so far. It suggests 49 elements, arranged in a table so that related concepts are adjacent. The columns are families, or general categories of cognition such as “Perception”, “Reasoning”, “Learning”, so all the types of perception or reasoning are stacked in one column. The rows are sorted by stages in a cycle of thought. The earlier stages (such as “sensing”) are at the top, while later stages in the cycle (such as “reflect & align”) are at the bottom. So for example, in the family or category of “Safety” the AIs will tend to do the estimation of uncertainty first, later do verification, and only get to a theory of mind at the end.
The chart is colored according to how much progress we’ve made on each element. Red indicates we can synthesize that element in a robust way. Orange means we can kind of make it work with the right scaffolding. Yellow reflects promising research without operational generality yet.
I suspect many of these elements are not as distinct as shown here (taxonomically I am more of a lumper than a splitter), and I would expect this collection omits many types we are soon to discover, but as a start, this prototype chart serves its purpose: it reveals the complexity of intelligence. It is clear intelligence is compounded along multiple dimensions. We will engineer different AIs to have different combinations of different elements in different strengths. This will produce thousands of types of possible minds. We can see that even today different animals have their own combination of cognitive primitives, arranged in a pattern unique to their species’ needs. In some animals some of the elements — say long-term memory — may exceed our own in strength; of course they lack some elements we have.
With the help of AI, we are discovering what these elements of cognition are. Each advance illuminates a bit of how minds work and what is needed to achieve results. If the discovery of electricity and atoms has anything to teach us now, it is that we are probably very far from having discovered the complete set of cognitive elements. Instead we are at the stage of believing in ethers, instantaneous action, and phlogiston – a few of the incorrect theories of electricity the brightest scientists believed.
Almost no thinker, researcher, experimenter, or scientist at that time could see the true nature of electricity, electromagnetism, radiation and subatomic particles, because the whole picture was hugely unintuitive. Waves, force fields, particles of atoms did not make sense (and still does not make common sense). It required sophisticated mathematics to truly comprehend it, and even after Maxwell described it mathematically, he found it hard to visualize.
I expect the same from intelligence. Even after we identify its ingredients, the emergent properties they generate are likely to be obscure and hard to believe, hard to visualize. Intelligence is unlikely to make common sense.
A century ago, our use of electricity ran ahead of our understanding of it. We made motors from magnets and coiled wire without understanding why they worked. Theory lagged behind practice. As with electricity, our employment of intelligence exceeds our understanding of it. We are using LLMs to answer questions or to code software without having a theory of intelligence. A real theory of intelligence is so lacking that we don’t know how our own minds work, let alone the synthetic ones we can now create.
The theory of the atomic world needed the knowledge of the periodic table of elements. You had to know all (or at least most) of the parts to make falsifiable predictions of what would happen. The theory of intelligence requires knowledge of all the elemental parts, which we have only slowly begun to identify, before we can predict what might happen next.
2025-09-03 04:15:23
无论何处存在自主性，信任都必须随之而来。如果我们培养孩子独立生活，他们需要具备自主性，而我们也需要信任他们。（育儿是学习信任的学校。）如果我们创建一个由自主代理组成的系统，就需要在代理之间建立大量的信任。如果我将决策委托给AI，那么我必须信任它；如果该AI依赖其他AI，它也必须信任它们。因此，我们需要开发一个非常稳健的信任系统，能够检测、验证和在人类与机器之间，更重要的是在机器与机器之间建立信任。
关于信任的研究主要沿着两个方向展开：更好地理解人类之间如何建立信任，以及将这些原则以抽象的方式应用于机械系统中。技术专家已经创建了初步的信任系统来管理数据云和通信的安全。例如，这个设备是否应该被允许连接？它是否值得信赖以完成它声称能做的事情？我们如何验证其身份和行为？等等。
目前这些系统并未处理适应性代理，它们的行为、身份和能力远比现有系统更加多变、模糊、不稳定，而且影响更为深远。这使得信任它们变得更加困难和重要。
如今，当我寻找一个AI时，准确度是我最看重的品质。它是否能给出正确的答案？它会有多大的幻觉？这些品质是信任的代理。我可以信任这个AI给出可靠的答案吗？随着AI开始承担更多任务，进入现实世界进行行动和决策，它们的可信度变得至关重要。
信任是一个广泛的概念，随着它渗入AI生态系统，它将被拆解为更具体的组成部分。作为安全、可靠性、责任和问责的一部分，这些要素将随着我们对其的综合和测量而变得更加精确。在未来十年中，信任将成为我们讨论得更多的话题。
随着AI的能力和技能开始分化——有些更适合特定任务——对它们的评价将开始包括其可信度。就像其他制造产品有广告的规格——如燃油效率、存储容量、像素数量、可用时间或治愈率——AI的供应商也将开始宣传其代理的“信任指数”。它们有多可靠？即使这种品质未被宣传，也需要在内部进行测量，以便公司能够不断改进它。
当我们依赖AI代理预订旅行票、续订药品处方或修理汽车时，我们将对其寄予厚望。很难想象AI代理会参与生死攸关的决定。我们甚至可能面临法律后果，因为对AI代理的信任程度。如果代理出了差错，谁来负责？
目前，AI并不承担任何责任。如果它们犯了错误，它们不保证修复。它们不对因错误可能引发的问题负责。事实上，这种差异目前是人类员工与AI员工之间的关键区别。责任最终由人类承担，他们负责自己的工作；你雇佣人类是因为你信任他们能正确完成工作。如果未能做到，他们会重新做，并学习如何不再犯同样的错误。而当前的AI则并非如此，这使得它们难以信任。
AI代理将形成一个网络，一个相互作用的AI系统，该系统可以为每个任务分配风险因素。一些任务，如购买机票或开具处方药，会有风险评分反映潜在的负面结果与正面便利。每个AI代理本身也会根据其权限拥有动态的风险评分。代理们还会根据过去的绩效积累信任评分。信任是非常不对称的；它可能需要许多互动和长时间才能积累价值，但一旦犯错，信任可能瞬间丧失。信任评分将不断变化，并由系统追踪。
大多数AI工作将通过代理之间的交流完成，而不被人类察觉。大多数由普通AI代理生成的输出将仅被另一个AI代理所消费，数量庞大。整个AI工作的极少数部分将被人类看到或注意到。人类与之互动的AI代理数量将非常少，尽管它们对我们来说至关重要。虽然我们接触的AI代理在统计上是罕见的，但它们对我们意义重大，其信任度将成为关键。
为了赢得我们的信任，一个面向外部的AI代理需要连接到它也能信任的AI代理，因此其能力的很大一部分将是选择和利用最值得信赖的AI代理的技能。我们可以预见全新的骗局，包括欺骗AI代理信任空壳代理，伪造信任证书，假冒身份，伪装任务。就像在互联网安全领域，一个AI代理的可信度取决于其最薄弱的子代理。由于子任务可以分配到多个层级，管理质量将成为AI的重要任务。
对错误的正确归因和纠正错误也将成为AI的巨大可销售技能。所有系统——包括最优秀的人类——都会犯错。没有完全无误的系统。因此，高信任度的很大一部分在于对错误负责并加以纠正。最受信任的代理将是那些能够（并且被信任！）修复自己所犯的错误，拥有足够的智能能力进行弥补并确保正确性的代理。
最终，我们给予主要AI代理的信任程度——也就是我们每天与之互动的代理——将是一个被夸耀、争论、共享和广泛宣传的评分。在其他领域，如汽车或手机，我们默认其可靠性。
AI比其他产品和服务更加复杂和个性化，不同于我们今天生活中的其他产品和服务，
AI代理的可信度将成为关键，并且是一个持续的关注点。其信任指数（TQ）可能比其智力指数（IQ）更重要。选择和保留高TQ的代理将非常类似于招聘和留住关键的人类员工。
然而，我们倾向于避免给人类分配数值评分。另一方面，AI代理系统将拥有各种我们用来决定是否希望它们帮助管理我们生活的指标。评分最高的AI可能也是最昂贵的。会有传言称那些几乎完美评分的AI你负担不起。然而，AI是一个随着使用而不断改进的系统，这意味着使用越多，它变得越好，因此最好的AI将是最受欢迎的。亿万富翁使用与我们相同的谷歌，很可能也会使用与我们相同的AI，尽管他们可能拥有高度个性化的界面。这些同样需要拥有最高的信任指数。
---------------
Wherever there is autonomy, trust must follow. If we raise children to go off on their own, they need to be autonomous and we need to trust them. (Parenting is a school for learning how to trust.) If we make a system of autonomous agents, we need lots of trust between agents. If I delegate decisions to an AI, I then have to trust it, and if that AI relies on other AIs, it must trust them. Therefore we will need to develop a very robust trust system that can detect, verify, and generate trust between humans and machines, and more importantly between machines and machines.
Applicable research in trust follows two directions: understanding better how humans trust each other, and applying some of those principles in an abstract way into mechanical systems. Technologists have already created primitive trust systems to manage the security of data clouds and communications. For instance, should this device be allowed to connect? Can it be trusted to do what it claims it can do? How do we verify its identity, and its behavior? And so on.
So far these systems are not dealing with adaptive agents, whose behaviors and IDs and abilities are far more fluid, opaque, shifting, and also more consequential. That makes trusting them more difficult and more important.
Today when I am shopping for an AI, accuracy is the primary quality I am looking for. Will it give me correct answers? How much does it hallucinate? These qualities are proxies for trust. Can I trust the AI to give me an answer that is reliable? As AIs start to do more, to go out into the world to act, to make decisions for us, their trustworthiness becomes crucial.
Trust is a broad word that will be unbundled as it seeps into the AI ecosystem. Part security, part reliability, part responsibility, and part accountability, these strands will become more precise as we synthesize it and measure it. Trust will be something we’ll be talking a lot more about in the coming decade.
As the abilities and skills of AI begin to differentiate – some are better for certain tasks than others – reviews of them will begin to include their trustworthiness. Just as other manufactured products have specs that are advertised – such as fuel efficiency, or gigabytes of storage, pixel counts, or uptime, or cure rates – so the vendors of AIs will come to advertise the trust quotient of their agents. How reliably reliable are they? Even if this quality is not advertised it needs to be measured internally, so that the company can keep improving it.
When we depend on our AI agent to book vacation tickets, or renew our drug prescriptions, or to get our car repaired, we will be placing a lot of trust in them. It is not hard to imagine occasions where an AI agent can be involved in a life or death decision. There may even be legal liability consequences for how much we can expect to trust AI agents. Who is responsible if the agent screws up?
Right now, AIs own no responsibilities. If they get things wrong, they don’t guarantee to fix it. They take no responsibility for the trouble they may cause with their errors. In fact, this difference is currently the key difference between human employees and AI workers. The buck stops with the humans. They take responsibility for their work; you hire humans because you trust them to get the job done right. If it isn’t, they redo it, and they learn how to not make that mistake again. Not so with current AIs. This makes them hard to trust.
AI agents will form a network, a system of interacting AIs, and that system can assign a risk factor for each task. Some tasks, like purchasing airline tickets, or assigning prescription drugs, would have risk scores reflecting potential negative outcomes vs positive convenience. Each AI agent itself would have a dynamic risk score depending on what its permissions were. Agents would also accumulate trust scores based on their past performances. Trust is very asymmetrical; It can take many interactions over a long time to gain in value, but it can lose trust instantly, with a single mistake. The trust scores would be constantly changing, and tracked by the system.
Most AI work will be done invisibly, as agent to agent exchanges. Most of the output generated by an average AI agent will only be seen and consumed by another AI agent, one of trillions. Very little of the total AI work will ever be seen or noticed by humans. The number of AI agents that humans interact with will be very few, although they will loom in importance to us. While the AIs we engage with will be rare statistically, they will matter to us greatly, and their trust will be paramount.
In order to win that trust from us, an outward facing AI agent needs to connect with AI agents it can also trust, so a large part of its capabilities will be the skill of selecting and exploiting the most trustworthy AIs it can find. We can expect whole new scams, including fooling AI agents into trusting hollow agents, faking certificates of trust, counterfeiting IDs, spoofing tasks. Just as in the internet security world, an AI agent is only as trustworthy as its weakest sub-agent. And since sub-tasks can be assigned for many levels down, managing quality will be a prime effort for AIs.
Assigning correct blame for errors and rectifying mistakes also becomes a huge marketable skill for AIs. All systems – including the best humans – make mistakes. There can be no system mistake proof. So a large part of high trust is the accountability in mending one’s errors. The highest trusted agents will be those capable (and trusted!) to fix the mistakes they make, to have sufficient smart power to make amends, and get it right.
Ultimately the degree of trust we give to our prime AI agent — the one we interact with all day every day — will be a score that is boasted about, contested, shared, and advertised widely. In other domains, like a car or a phone, we take reliability for granted.
AI is so much more complex and personal, unlike other products and services in our lives today,
the trustworthiness of AI agents will be crucial and an ongoing concern. Its trust quotient (TQ) may be more important than its intelligence quotient (IQ). Picking and retaining agents with high TQ will be very much like hiring and keeping key human employees.
However, we tend to avoid assigning numerical scores to humans. The AI agent system, on the other hand will have all kinds of metrics we will use to decide which ones we want to help run our lives. The highest scoring AIs will likely be the most expensive ones as well. There will be whispers of ones with nearly perfect scores that you can’t afford. However, AI is a system that improves with increasing returns, which means the more it is used, the better it gets, so the best AIs will be among the most popular AIs. Billionaires use the same Google we use, and are likely to use the same AIs as us, though they might have intensely personalized interfaces for them. These too, will need to have the highest trust quotients.
Every company, and probably every person, will have an AI agent that represents them inside the AI system to other AI agents. Making sure your personal rep agent has a high trust score will be part of your responsibility. It is a little bit like a credit score for AI agents. You will want a high TQ for yours. Because some AI agents won’t engage with other agents having low TQs. This is not the same thing as having a personal social score (like the Chinese are reputed to have). This is not your score, but the TQ score of your agent, which represents you to other agents. You could have a robust social score reputation, but your agent could be lousy. And vice versa.
In the coming decades of the AI era, TQ will be seen as more important than IQ.
2025-08-24 05:57:26
许多人发现人工智能的智能令人震惊。然而，与即将到来的更大冲击相比，这似乎微不足道：高度情感化的人工智能。合成情感的出现将在人类社会引发颠覆、愤怒、困扰、混乱和文化冲击，其影响将远超对合成智能的讨论。在接下来的几年里，新闻标题将从“所有人都会失业”（其实不会）转变为“人工智能伴侣是人类文明的终结”。
我们能够理性地接受计算机可能具备理性。尽管我们可能不喜欢，但我们可以接受计算机可能变得聪明，部分原因是我们已经将大脑视为一种计算机。很难相信它们能像我们一样聪明，但一旦它们做到了，这似乎也合乎逻辑。
接受机器制造的创造力则更为困难。创造力似乎非常人类化，某种程度上被视为理性的对立面，因此看起来并不属于机器。
情感是有趣的，因为情感不仅存在于人类，还存在于许多动物身上。任何宠物主人都能列举出宠物如何感知和表达情感的方式。人们喜爱动物的部分原因在于能够与它们在情感上产生共鸣。它们对我们的感情作出回应，就像我们对它们的感情作出回应一样。人类与动物之间存在真实而深刻的情感纽带。
同样的情感纽带也将出现在机器上。我们已经看到一些迹象。几乎每周都有陌生人向我发送他们与AI聊天的日志，展示这些互动多么深刻和直觉，它们如何理解彼此，以及在精神上如何连接。我们还收到关于青少年与AI“朋友”建立深厚关系的报告。这一切都发生在尚未有系统性地将情感嵌入AI之前。
我们为何要将情感编程进AI？有几个原因：
首先，情感是与机器交互的绝佳界面。它使互动更加自然和舒适。情感对人类来说很容易。我们无需学习如何行动，我们都能直观地理解诸如表扬、热情、怀疑、说服、惊喜、困惑等结果，这些是机器可能希望使用的。人类通过微妙的情感色彩传达非语言信息、重要性以及指令，而AI将在其指令和交流中使用类似的情感线索。
其次，市场将更青睐情感代理，因为人类本身也是如此。随着AI和机器人能力的趋同，它们的多样性将继续发展，因此它们的性格和情感特征在选择使用时将变得更为重要。如果它们都同样聪明，那么更友善、更亲切或更好的陪伴者将赢得用户的青睐。
第三，我们希望人工智能代理完成的许多任务，无论是软件AI还是实体机器人，都需要的不仅仅是理性计算。仅仅让AI整夜编程是不够的。我们目前对智能的评价过高。要真正具备创造力和创新能力，要足够明智以提供良好建议，所需要的不仅仅是智商。这些机器人需要深入其软件中的复杂情感动态。
这是否可能？是的。
已有几十年的研究项目（如麻省理工学院的研究）致力于将情感提炼为可移植到机器中的属性。其中一些知识涉及如何在硬件上视觉化情感，就像我们通过自己的面部表情来表达一样。其他研究者则提取了我们通过声音传达情感的方式，甚至通过文本中的词语。最近，我们见证了AI开发者调整其代理的友好程度和“亲切感”，因为一些用户不喜欢它们的新性格，而另一些人则单纯不喜欢性格的变化。虽然我们可以编程性格和情感，但我们尚未明确知道哪些情感最适合特定任务。
机器展示情感只是工作的一半。另一半是机器对人类情感的检测和理解。关系是双向的，为了真正成为情感代理，机器必须擅长捕捉你的情感。该领域已有大量研究，主要集中在面部识别上，不仅识别身份，还识别情绪状态。已有商业应用可以观察用户在键盘上的输入，检测他们是否抑郁或处于情绪压力下。未来，智能眼镜将不仅观察外界，还会回过头来观察你的面部表情以解析你的情绪。你是否困惑或欣喜？是否惊讶或感激？是否坚定或放松？目前，苹果的Vision Pro已经配备了朝后摄像头，用于追踪眼睛和微表情，如眨眼和眉毛上扬。当前的文本大型语言模型（LLM）在检测用户情绪状态上几乎没有尝试，除了从提示中的字母中推断出的信息。但技术上来说，这并不是巨大的飞跃。
在接下来的几年里，将会有大量关于情感的实验。一些AI会简洁而逻辑性强，一些则会健谈且外向。一些AI会低声细语，只在你准备好倾听时才说话。一些人会更喜欢那些能让他们发笑的幽默、机智的AI。许多商业AI将被设计成你的最佳朋友。
我们可能会认为这是成年人的可取之处，但对儿童来说却可能令人不安。事实上，关于AI与儿童的关系，有许多问题需要警惕，不仅仅是情感方面。但情感纽带将是儿童AI设计中的关键考虑因素。年幼的人类儿童已经能够与惰性娃娃和泰迪熊建立情感联系，并变得非常亲密。想象一下，如果一个泰迪熊能回应你的对话、以无限耐心与你玩耍，并且能映射你的情感。随着孩子的成长，他们可能永远不会放弃这个泰迪熊。因此，机器中情感的质量很可能会成为我们制定不同制度的领域，一种针对成年人，一种针对儿童。不同的规则、不同的期望、不同的法律、不同的商业模式等。
但即使是成年人，也会对情感代理产生强烈依恋，就像电影《她》中那样。起初，社会会将那些沉迷于AI爱情的人视为有幻觉或精神不稳定。但正如大多数深爱狗或猫的人并非精神有问题，而是正常且富有同理心的人一样，大多数与AI和机器人建立亲密关系的人也会将这些纽带视为健康且能拓展自我的。
关于与机器建立亲密关系的常见恐惧是，它们可能太友善、太聪明、太有耐心、太可用，比周围的人类更有效，因此人们会完全脱离人类关系。这种情况可能发生。我们很容易想象，有善意的人只会消费AI提供的“甜美易得的友谊”，就像他们被诱惑去只消费加工食品中的“甜美易得的卡路里”一样。对抗这种诱惑的最佳方法与快餐类似：教育和更好的选择。在这个新世界中成长的一部分，就是学会区分那些看似完美的关系与人类关系中混乱、困难和不完美的部分，并认识到后者的价值。要成为你自己最好的版本——无论你如何定义——都需要与人类共处！
与其禁止AI关系（或快餐），不如对其进行适度管理，并保持理性看待。事实上，AI朋友、导师、教练或伴侣的“完美”行为可以成为绝佳的榜样。如果你身边有经过训练和调整的AI，它们是人类所能创造的最好版本，这将是一种极好的自我提升方式。普通人类往往拥有非常浅薄的道德观，矛盾的准则，并且容易被自己的基本欲望和处境所左右。从理论上讲，我们应该能够编程AI，使其具备比普通人类更好的道德和准则。同样，我们也可以设计AI成为比普通人类更好的朋友。拥有这些受过教育的AI可以帮助我们提升自己，成为更好的人类。与它们建立深厚关系的人，可能成为最适应社会和富有同理心的人。
关于AI的情感并非真实，因为“机器人无法感受任何东西”的论点将被忽视。就像批评人工智能是人工的，因此不真实，因为它们不理解一样。这并不重要。我们并不真正理解“感受”意味着什么，甚至也不清楚“理解”意味着什么。这些术语和概念已经成为习惯，但已不再有用。AI能够做我们过去称为智能的事情，它们也将开始做我们过去称为情感的事情。最重要的是，人类与AI、机器人之间的关系将与任何其他人类关系一样真实和有意义。它们将是一种真实的关系。
但AI或机器人所拥有的情感，尽管真实，却可能有所不同。真实但有所偏差。AI可以幽默，但它们的幽默感略有偏差，略有不同。它们会笑我们不会笑的事情。它们的幽默方式会逐渐改变我们自己的幽默观，就像它们下棋和打麻将的方式已经改变了我们下棋和打麻将的方式一样。AI是聪明的，但以非人类的方式。它们的情感性也将同样陌生，因为AI本质上是人工的外星生命。事实上，我们将通过观察它们的情感来更深入地理解情感的本质，而不是通过研究自己。
机器中的情感不会一夜之间出现。情感将逐渐积累，这样我们就有时间引导它们。它们最初表现为礼貌、文明和友善。它们会轻易地赞美和奉承我们，也许过于轻易。核心关注点并非是我们的机器伴侣关系是否会亲密（它们会），也不是这些关系是否真实（它们是），也不是它们是否会取代人类关系（它们不会），而是你的情感代理为你服务的是谁？谁拥有它？它被优化用于什么？你能信任它不会操纵你吗？这些问题将在未来十年主导我们的讨论。
显然，关于我们的最敏感数据，莫过于来自我们情感的信息。我们害怕什么？什么让我们真正快乐？什么让我们感到厌恶？什么会激发我们？在与我们的全天候代理互动多年后，该代理将拥有我们的完整档案。即使我们从未明确透露过最深层的恐惧、最珍视的愿望和最脆弱的时刻，它也能从我们的交流、问题和反应的情感倾向中了解一切。它将比我们自己更了解我们。这将在未来几十年成为常见的说法，既令人兴奋又令人恐惧：“我的AI代理比我更了解我自己。”
在许多情况下，这将是真实的。在最好的情况下，我们可以利用这个工具更好地认识自己。在最坏的情况下，这种知识的不对称性将被用来操纵我们，放大我们最糟糕的一面。我看不到任何证据表明我们会停止将AI纳入我们的生活，每小时甚至每分钟都如此。（会有例外，比如阿米什人，他们选择退出，但他们是极少数。）大多数人，在大多数时间里，都会与一个全天候的AI代理、机器人建立亲密关系，它们随时准备以任何方式帮助我们。这种关系将变得与任何其他人类联系一样真实和有意义。我们会自愿与它分享我们生命中最私密的时刻。只要从中获得的好处持续存在，我们平均会向它提供我们最个人的数据。（数据隐私的门槛并不是谁拥有它，而是我能从中获得多少好处？人们会分享任何类型的数据，只要好处足够大。）
25年后，如果那些拥有全天候AI伴侣的人都是彻头彻尾的混蛋、厌世者和失败者，那么情感AI的故事将在此结束。另一方面，如果与AI建立亲密关系的人比平均水平更具同理心、更有生产力、更加独特、更适应社会，且拥有更丰富的内心世界，那么情感AI的故事才刚刚开始。
我们可以通过奖励那些推动我们向这一方向发展的发明，来引导故事走向我们想要的开始。问题不在于AI是否会拥有情感，而在于我们将如何利用这种情感性。
---------------
Many people have found the intelligence of AIs to be shocking. This will seem quaint compared to a far bigger shock coming: highly emotional AIs. The arrival of synthetic emotions will unleash disruption, outrage, disturbance, confusion, and cultural shock in human society that will dwarf the fuss over synthetic intelligence. In the coming years the story headlines will shift from “everyone will lose their job” (they won’t) to “AI partners are the end of civilization as we know it.”
We can rationally process the fact that a computer could legitimately be rational. We may not like it, but we could accept the fact that a computer could be smart, in part because we have come to see our own brains as a type of computer. It is hard to believe they could be as smart as we are, but once they are, it kind of makes sense.
Accepting machine-made creativity is harder. Creativity seems very human, and it is in some ways perceived as the opposite of rationality, and so it does not appear to belong to machines, as rationality does.
Emotions are interesting because emotions clearly are not only found in humans, but in many, many animals. Any pet owner could list the ways in which their pets perceive and display emotions. Part of the love of animals is being able to resonate with them emotionally. They respond to our emotions as we respond to theirs. There are genuine, deep emotional bonds between human and animal.
Those same kinds of emotional bonds are coming to machines. We see glimmers of it already. Nearly every week a stranger sends me logs of their chats with an AI demonstrating how deep and intuitive they are, how well they understand each other, and how connected they are in spirit. And we get reports of teenagers getting deeply wrapped up with AI “friends.” This is all before any serious work has been done to deliberately embed emotions into the AIs.
Why will we program emotions into AIs? For a number of reasons:
First, emotions are a great interface for a machine. It makes interacting with them much more natural and comfortable. Emotions are easy for humans. We don’t have to be taught how to act, we all intuitively understand results such as praise, enthusiasm, doubt, persuasion, surprise, perplexity – which a machine may want to use. Humans use subtle emotional charges to convey non-verbal information, importance, and instruction, and AIs will use similar emotional notes in their instruction and communications.
Second, the market will favor emotional agents, because humans do. AIs and robots will continue to diversify, even as their basic abilities converge, and so their personalities and emotional character will become more important in choosing which one to use. If they are all equally smart, the one that is friendlier, or nicer, or a better companion, will get the job.
Thirdly, a lot of what we hope artificial agents will do, whether they are software AIs or hard robots, will require more than rational calculations. It will not be enough that an AI can code all night long. We are currently over rating intelligence. To be truly creative and capable of innovations, to be wise enough to offer good advice, will require more than IQ. The bots need sophisticated emotional dynamics that are deeply embedded in its software.
Is that even possible? Yes.
There are research programs (such as those at MIT) going back decades figuring out how to distill emotions into attributes that can be ported over to machines. Some of this knowledge pertains to ways of visually displaying emotions in hardware, just as we do with our own faces. Other researchers have extracted ways we convey emotion with our voice, and even in words in a text. Recently we’ve witnessed AI makers tweaking how complimentary and “nice” their agents are because some users didn’t like their new personality, and some simply did not like the change in personality. While we can definitely program in personality and emotions, we don’t yet know which ones work best for a particular task.
Machines displaying emotions is only half of the work. The other half is detection and comprehension of human emotions by machines. Relationships are two way, and in order to truly be an emotional agent, it must get good at picking up your emotions. There has been a lot of research in that field, primarily in facial recognition, not just your identity, but how you are feeling. There are commercially released apps that can watch a user at their keyboard and detect whether they are depressed, or undergoing emotional stress. The extrapolation of that will be smart glasses that not only look out, but at the same time look back at your face to parse your emotions. Are you confused, or delighted? Surprised, or grateful? Determined, or relaxed? Already, Apple’s Vision Pro has backward facing cameras in its goggles that track your eyes and microexpressions such as blinks and eyebrow rises. Current text LLM’s make no attempt to detect your emotional state, except what can be gleaned from the letters in your prompt, but it is not technically a huge jump to do that.
In the coming years there will be lots of emotional experiments. Some AIs will be curt and logical; some will be talkative and extroverts. Some AIs will whisper, and only talk when you are ready to listen. Some people will prefer loud, funny, witty AIs that know how to make them laugh. And many commercial AIs will be designed to be your best friend.
We might find that admirable for an adult, but scary for a child. Indeed, there are tons of issues to be wary of when it comes to AIs and kids, not just emotions. But emotional bonds will be a key consideration in children’s AIs. Very young human children already can bond with, and become very close to inert dolls and teddy bears. Imagine if a teddy bear talked back, played with infinite patience, and mirrored their emotions. As the child grows it may not ever want to surrender the teddy. Therefore the quality of emotions in machines will likely become one of those areas where we have very different regimes, one for adults and one for children. Different rules, different expectations, different laws, different business models, etc.
But even adults will become very attached to emotional agents, very much like the movie Her. At first society will brand those humans who get swept up in AI love as delusional or mentally unstable. But just as most of the people who have deep love for a dog or cat are not broken, but well adjusted and very empathetic beings, so most of the humans that will have close relationships with AIs and bots will likewise see these bonds as wholesome and broadening.
The common fear about cozy relationships with machines is that they may be so nice, so smart, so patient, so available, so much more helpful than other humans around, that people will withdraw from human relationships altogether. That could happen. It is not hard to imagine well-intentioned people only consuming the “yummy easy friendships” that AIs offer, just as they are tempted to consume only the yummy easy calories of processed foods. The best remedy to counter this temptation is similar to fast food: education and better choices. Part of growing up in this new world will be learning to discern the difference between pretty perfect relationships and messy, difficult, imperfect human ones, and the value the latter give. To be your best — whatever your definition —requires that you spend time with humans!
Rather than ban AI relationships (or fast food) you moderated it, and keep it in perspective. Because in fact, the “perfect” behavior of an AI friend, mentor, coach, or partner can be a great role model. If you surround yourself with AIs that have been trained and tweaked to be the best that humans can make, this is fabulous way to improve yourself. The average human has very shallow ethics, and contradictory principles, and is easily swayed by their own base desires and circumstances. In theory, we should be able to program AIs to have better ethics and principles than the average human. In the same way, we can engineer AIs to be a better friend than the average human. Having these educated AIs around can help us to improve ourselves, and to become better humans. And the people who develop deep relationships with them have a chance to be the most well-adjusted and empathetic people of all.
The argument that the AIs’ emotions are not real because “the bots can’t feel anything” will simply be ignored. Just like the criticism of artificial intelligence being artificial and therefore not real because they don’t understand. It doesn’t matter. We don’t understand what “feeling” really means and we don’t even understand what “understand” means. These are terms and notions that are habitual but no longer useful. AIs do real things we used to call intelligence, and they will start doing real things we used to call emotions. Most importantly the relationships humans will have with AIs, bot, robots, will be as real and as meaningful as any other human connection. They will be real relationships.
But the emotions that AIs/bots have, though real, are likely to be different. Real, but askew. AIs can be funny, but their sense of humor is slightly off, slightly different. They will laugh at things we don’t. And the way they will be funny will gradually shift our own humor, in the same way that the way they play chess and go has now changed how we play them. AIs are smart, but in an unhuman way. Their emotionality will be similarly alien, since AIs are essentially artifical aliens. In fact, we will learn more about what emotions fundamentally are from observing them than we have learned from studying ourselves.
Emotions in machines will not arrive overnight. The emotions will gradually accumulate, so we have time to steer them. They begin with politeness, civility, niceness. They praise and flatter us, easily, maybe too easily. The central concern is not whether our connection with machines will be close and intimate (they will), nor whether these relationships are real (they are), nor whether they will preclude human relationships (they won’t), but rather who does your emotional agent work for? Who owns it? What is it being optimized for? Can you trust it to not manipulate you? These are the questions that will dominate the next decade.
Clearly the most sensitive data about us would be information stemming from our emotions. What are we afraid of? What exactly makes us happy? What do we find disgusting? What arouses us? After spending all day for years interacting with our always-on agent, said agent would have a full profile of us. Even if we never explicitly disclosed our deepest fears, our most cherished desires, and our most vulnerable moments, it would know all this just from the emotional valence of our communications, questions, and reactions. It would know us better than we know ourselves. This will be a common refrain in the coming decades, repeated in both exhilaration and terror: “My AI agent knows me better than I know myself.”
In many cases this will be true. In the best case scenario we use this tool to know ourselves better. In the worst case, this asymmetry in knowledge will be used to manipulate us, and expand our worst selves. I see no evidence that we will cease including AIs in our lives, hourly, if not by the minute. (There will be exceptions, like the Amish, who drop out but they will be a tiny minority.) Most of us, for most of the time, will have an intimate relationship with an AI agent/bot/robot that is always on, ready to help us in any way it can, and that relationship will become as real and as meaningful as any other human connection. We will willingly share our most intimate hours of our lives with it. On average we will lend it our most personal data as long as the benefits of doing so keep coming. (The gate in data privacy is not really who has it, but how much benefit do I get? People will share any kind of data if the benefits are great enough.)
Twenty five years from now, if the people whose constant companion is an always-on AI agent are total jerks, misanthropic bros, and losers, this will be the end of the story for emotional AIs. On the other hand, if people with a close relationship with an AI agent are more empathetic than average, more productive, distinctly unique, well adjusted, with a richer inner life, then this will be the beginning of the story.
We can steer the story to the beginning we want by rewarding those inventions that move us in that direction. The question is not whether AI will be emotional, but how we will use that emotionality.
