新一代互联网：AIGC+区块链

来源 | 数据经济评论

作者 | 张文卓

长期以来，我们普遍认为下一代信息革命将由量子信息驱动。与工业革命情况相似，量子信息学也是一个新的物理学分支，量子计算机可能就会成为这一代科技革命，也就是第二次信息革命一个最重要的代表，它对比经典的计算机就像电力对比蒸汽机一样。在过去的很长一段时间里，无论是在讲课、对外报告，还是在撰写书籍时，我们都是沿着这个脉络，看着逻辑很完整，然而，今年我们发现我们的预测是错误的。

has long been widely believed that the next generation of information revolutions will be driven by quantum information. Similar to the industrial revolution, quantum informatics is a new branch of physics.

事实上，信息科学已经成为一个非常完整的学科，它本身就具备支撑一次人类科技革命所需的能力，并不需要物理学在更基础的层面新生成一个分支来支撑。所以这就相当于一个时代的变迁，意味着我们并不需要等待梦想中的量子计算机的实现才能让人类的科技再进一步。

is in fact a very complete discipline, with the ability to sustain a revolution in human science and technology in itself and without the need for physics to create a new branch at a more basic level to support it. So it's like an era change, meaning that we don't have to wait for the achievement of the desired quantum computer to make human science and technology go further.

实际上，在信息科学领域，从神经网络基础到新兴的深度学习等专业领域，已经具备了支持这一次科技革命所需的能力。因此，我们正处于第二次信息革命，或智能革命的时代。一个典型的例子就是AI大模型。

, therefore, we are in the era of a second information revolution, or an intelligent revolution. A typical example is the AI model.

大语言模型，也被称为AI大模型。因为语言是对人类对现实世界和虚拟世界的最佳映射。语言可以精确到数学这一层面，所以逻辑模型才能运用起来。同时由于语言的一些模糊逻辑，它也催生了各种创造性的应用。

比如说ChatGPT用到的大语言模型GPT-3.5，它采用了1万多维的向量来表示每个单词，也就是这些单词放在1万多维的向量空间里，每个词对应一个向量，成为词向量。

, for example, the large language model GPT-3.5 used by ChatGPT, which uses a vector of more than 10,000 dimensions to indicate each word, that is, the word is placed in a vector of more than 10,000 dimensions, and each word corresponds to a vector and becomes a vector.

2017年Google的自然语言团队提出了Transformer机制，它里面定义了attention这样一个量，这个量指的是各个词向量在向量空间中的绝对距离。实际上，优化训练模型的过程就是在不断地优化attention。

The Google Natural Language Team introduced the Transformer mechanism in 2017, which defines the amount of attention that refers to the absolute distance of the word vectors in the vector space. In fact, the process of optimizing the training model is constantly optimizing the application.

通过这个过程，词语之间的关联度逐渐增强，而它们的attention值则逐渐减小。这种情况下，当你输入一个词后，模型能够联想到与之相关的其他词，并能够持续生成输出。这种能力的根本在于每个词在向量空间中的位置以及词语之间的attention关系。

uses this process to increase the correlation between words, while their value decreases. In this case, when you enter a word, the model can think of other words associated with it and can generate output on an ongoing basis. The essence of this ability is the position of each word in the vector space and the relationship between words.

虽然这是Google发明的，但是OpenAI做了弯道超车，OpenAI在自然语言处理和强化学习的结合方面比Google做得要好，所以虽然Google起步较早，但最终却被OpenAI迎头赶上，被OpenAI直接用12层的神经网络强化学习，率先实现了GPT模型。

, although it was invented by Google, OpenAI has made a curved supercar, OpenAI has done better than Google in terms of natural language processing and intensive learning, so Google, although it started early, was eventually caught up by OpenAI and was directly reinforced by OpenAI with 12 layers of neural networks, leading to the GPT model.

为什么Open AI能获胜？陆奇博士告诉过我们一些原因，也包括我们自己的一些想法。首先大家都知道OpenAI是含着金钥匙长大的，OpenAI的出现与对抗DeepMind有关。

Why did Open AI win? Dr. Rudd told us a number of reasons, including some of our own ideas. First of all, everyone knows OpenAI grew up with a gold key.

DeepMind是由Google收购的深度学习团队，他们在当时确实处于一家独大的地位，首先用AlphaGo把围棋搞定了，再用AlphaFold把蛋白质折叠给搞定了，为了应对DeepMind可能独占市场这种危险的情况，Sam Altman联合马斯克等大佬成立了OpenAI。

Deepmind was acquired by Google, who was indeed in a privileged position at the time, first dealt with chess with AlphaGo, then folded the protein with AlphaFold, and in response to the dangerous situation of Deepmind's possible monopoly on the market, Sam Altman joined with big men like Mask to set up OpenAI.

OpenAI在成立时并不希望重蹈Google的覆辙，他们的自然语言处理团队和深度学习团队之间保持密切合作。反观Google，自身的自然语言处理团队搞出了Transformer模型，而DeepMind则取得了AlphaGo、AlphaFold等引人瞩目的成就。尽管两个团队在各自领域都属于世界顶尖，但由于DeepMind是英国公司，和美国团队相互敌视，合作有一定障碍。OpenAI在成立时避免了这些问题和错误，这是他们成功的重要因素之一。

OpenAI did not want to repeat Google's mistakes when it was set up, and there was close cooperation between their natural language processing teams and the deep learning team. In contrast, Google, his own natural language processing team created the Transformer model, while Deepmind achieved remarkable achievements such as AlphaGo, AlphaFold. Although both teams are among the top in the world, there are certain obstacles to cooperation because Deepmind is a British company and the American team is hostile to each other.

Sam Altman为了实现其梦想与微软达成了合作。微软投入了大量资金，孤注一掷。在几年前，微软购买了开源代码库GitHub，当时引起了广泛的争议。大家质疑微软为什么要花那么多钱购买这些非盈利的公司，这明显是个亏本的买卖。

Sam Altman worked with Microsoft to realize his dream. Microsoft invested a lot of money and was desperate. A few years ago, Microsoft bought the open source library GitHub, which was widely controversial. It was clearly a lost business to question why Microsoft spent so much money on these non-profit companies.

现在我们明白，微软购买这些代码库是为了帮助OpenAI训练GPT，因为微软可以从后台高效使用这些代码数据，而其他公司只能通过前端、网页爬虫等方式来获取这些代码，这种效率远远赶不上微软。这一点非常重要，因为微软使得OpenAI能够充分利用这些资源。

now we understand that Microsoft purchased these code libraries to help OpenAI train GPT, because Microsoft can use them efficiently from backstage, while other companies can access them only through front end, web reptiles, which is much less than Microsoft. This is very important because Microsoft allows OpenAI to make the best use of these resources.

微软不仅在软件方面下了血本，也在硬件方面投入了大量资源。为了支持OpenAI，微软为其提供了专属的服务器集群。这种服务器集群架构与超级计算机非常相似，不仅使用了最好的GPU，而且通过FPGA直接实现了服务器内存互相访问。这种架构确保通信带宽不成为瓶颈，即通信带宽的速度高于GPU的主频。

Microsoft has not only given blood to software, but has also invested a lot of resources in hardware. To support OpenAI, Microsoft has provided it with an exclusive server cluster. This server clustering structure is very similar to supercomputers, not only using the best GPU, but also directly enables server memory to access each other through the FPGA. This structure ensures that the bandwidth is not a bottleneck, i.e., that the bandwidth is faster than the GPU main frequency.

这是一种非常独特的架构，在普通服务器中很少见。微软为OpenAI提供了如此深度的资源支持，其他公司要复现这种功能是非常困难的，因为构建这种专属的服务器集群非常具有挑战性。当然，现在Google也开始进行类似的工作，而亚马逊是否也在打造专属的服务器集群，但这是一项非常昂贵的事情，微软每次训练就要耗费数亿美元，这不仅包括软件成本，还包括硬件成本。

对于中国而言，如果要构建这样的服务器集群，可能会面临一些挑战。虽然华为有能力打造这样的集群，但在芯片层面上，如果英伟达不向中国提供最优秀的GPU，那么构建这样的集群的意义将大打折扣。

/strang>

对于国内来说，训练大型模型在硬件层面又回到了芯片供应受限的根本性问题。这导致中国很难能够在短期内实现OpenAI这样的成果，因为无论是在语料资源方面还是在训练效率方面，都存在一定的差距。

for the country, training large models returns to the fundamental problem of limited chip supply at the hardware level. This makes it very difficult for China to achieve OpenAI’s results in the short term, because there is a gap between language resources and training efficiency.

现在多模态大模型层出不穷，不光是文生文的ChatGPT，文生图也出现了能够跟Open AI竞争的开源模型，包括Stable Diffusion，还有在开源上优化的私有模型Midjourney，文生图的效果它最好。紧接着像文生3D建模、文生视频都会逐渐成熟。

is now a multi-model, large model of ChatGPT, not just a literary text, and there are open-source models that can compete with Open AI, including Stable Diffusion, and an open-source private model, Midjourney, which works best.

可以想象，这些工具的结合将催生出超级创作者的出现。事实上，目前已经出现了一些迹象，个人开发者制作了小型的3D游戏，并将其发布在YouTube上，预计下半年将会出现这种超级创作者的爆发态势。

这一次AI大模型的出现被称为新一次科技革命，这与历史上的工业革命有一定的相似性。作为物理学出身的我，愿意将其与工业革命进行比较。我们可以发现，它非常类似于电力的出现，因为电力引导了第二次工业革命，但它依赖于第一次工业革命的成果来输入，就像现在火电，主要用的就是蒸汽机和汽轮机等，尽管其他发电方式的比例在增加，但它们还无法与火力发电相比。

, the emergence of this large AI model is called a new technological revolution, which is somewhat similar to the historical industrial revolution. /strong> as a physics-based person willing to compare it with the industrial revolution. We can find that it is very similar to the emergence of electricity, which led to the second industrial revolution, but it relies on the results of the first industrial revolution to be imported, as is the case now with electric power, the main use of which is steam engines and turbines, and although the proportion of other power generation methods is increasing, they are not comparable to power generation.

AI大模型的生成能力和计算机之间的关系非常像电力和蒸汽机之间的关系，它属于二次信息，它依赖于计算机的输出作为它的输入，但它能够完成比计算机更多的任务，就像电力驱动着各种电器一样。

The relationship between the generation capacity of the big model of AI and the computer is very much like the relationship between electricity and the steam engine, which is secondary in nature and depends on the output of the computer as its input, but it can perform more tasks than the computer, just as electricity drives all kinds of appliances.

电力具有多模态的输出性质，可以驱动声、光、磁等各种能量，而AI大模型的生成能力也可以产生更多的应用，完成计算机无法实现的事情。这种多模态的生态系统使得AI时代与电力时代很相似。

这个时代，马斯克就像爱迪生一样，而Sam Altman和Ilya Sutskever（OpenAI首席科学家）的组合就像西屋加上尼古拉特斯拉。在电力时代，出现了许多杰出人物，AI时代也会涌现出众多杰出的个人和团队。

>, as in Edison, the combination of Sam Altman and Ilya Sutskever (Chief Scientist of OpenAI) is like Westha and Nicholas Tsla. In the electricity age, there are a lot of outstanding personalities, and the AI era has a lot of outstanding individuals and teams.

前面我们提到多模态AIGC会催生超级创作者的角色，这会给我们整个互联网带来一个什么样的变化？Web3这个词在去年引起了一些热度，但今年可能有点冷却。然而Web3的真正发展与AIGC密切相关。

, we mentioned earlier that multi-model AGCs could create a super-creative role, and how this would change our entire Internet? The word Web3 caused some heat last year, but may be cool this year. But the real development of Web3 is closely related to AIGC.

这种关联是因为Web3的整个网络结构是建立在区块链技术之上的，而当AIGC工具开始大规模使用时，必然会催生出超级创作者的出现。这样的发展符合马克思所阐述的生产力决定生产关系。

在过去只有大团队才能完成一些任务，而现在一个小团队甚至个人用低成本AI工具，可以完成同样的工作，并且用时更短。这种情况下，这些人不再需要为大公司工作，他们可以更自由地依靠自己的创造力来获得足够的收入。因此，小团队之间的生产关系趋向平等，去中心化的机制变得更有价值。

, which used to be done only by big teams, and now a small team, even individuals, with low-cost AI tools, can do the same work and do it for a shorter period of time. In this case, these people no longer need to work for large companies, and they are freer to rely on their creativity to earn enough income. As a result, production relationships among small teams tend to be equal, and decentralised mechanisms become more valuable.

区块链作为一种去中心化的网络架构，最适合像这样的小团队或个人创作者之间达成共识的生产关系。反过来，传统的大公司在这个时代的地位将会降低，它们将变成提供平台服务的企业，而不是高高在上，继续当雇佣创作者的机构，而是反过来创作者联合雇佣它。

/strong>

实际上这种形式并不新鲜。我们可以关注体育比赛的例子。目前足球比赛由欧足联这样的中心化机构管理，决定比赛规则、欧冠联赛的安排和分配等等。一些超级豪门球队，像皇马就想建立欧超联赛这样一种去中心化的方式。

他们也并不是完全创新，欧超联赛仿的就是北美体育联盟，例如NBA，球队之间形成了一种去中心化的组织结构，共同组建了一个平台，雇佣了一个运营NBA的公司。因此，当某个球队老板犯错或说错话时，NBA联盟及其雇佣的公司也拿他没办法，因为他们之间的生产关系是相反的。可以想象，随着超级创作者数量的增加，区块链和去中心化的生产关系将成为主流。

区块链带来的去中心化生产关系实际上解决了人类长期以来的一个问题，即如何在缺乏信任的情况下继续合作。这里我想到刘慈欣的《三体》中第二部和第三部的核心概念"黑暗森林"。黑暗森林被认为是一种极度内卷的状态，但在人类历史上，它并没有持续很久，因为部落之间发现可以建立一些信任，通过牺牲一些防卫来成为盟友，这样的合作能够获得比之前更大的能力。因此，人类通过合作的方式战胜了比自己强大得多的野兽，从而存活下来。由于人类拥有智力，可以建立合作关系和盟友关系。

然而在互联网时代，能否在没有绝对信任的情况下继续展开合作？这就是区块链所实现的功能。区块链是一个分布式去中心化的账本系统，它将所有交易信息收集起来，然后进行统一处理，采用各种共识机制选择记账节点。如果有节点进行犯罪或错误操作，会有相应的惩罚措施。

, but in the Internet age, can cooperation continue without absolute trust? This is the function of the block chain. The block chain is a distributed decentralised system of accounts that collects all transaction information, then treats it together and selects the account points using various consensus mechanisms. If there are nodes to commit crimes or misoperate operations, appropriate penalties are applied.

这种方式让参与者无需建立信任，只需依靠自身运行的计算机代码就能完成任务。这就是区块链的价值所在，也是它在AIGC时代，尤其是在涉及到众多超级创作者的情况下，能够发挥作用的原因。

夸密量子的宗旨是用量子技术塑造新一代互联网。Web3就约等于新一代互联网。新一代互联网的底层会有一些改变，比如很多加密技术可以用在链路层，变成加密数据流。

The purpose of exaggeration quantum is to shape the new generation of the Internet using quantum technology. Web3 is about the same as the new generation of the Internet. There will be some changes at the bottom of the new generation of the Internet, such as many encryption techniques that can be used in the chain layers to turn into encrypted data streams.

在物理层方面，随着移动互联之后的外务互联，将有更多设备接入，并且网络连接节点的数量也会爆发式增长。在应用层面上，AIGC作为生产力的基础，区块链作为生产关系的基础，共同支撑着更多新一代互联网的应用，即Web3的应用。这样，新一代互联网的潜力将变得非常巨大。

从整个社会和经济学的角度来看，目前Web3在与去中心化金融相关的领域表现最为出色。随着AIGC的出现，接下来将会出现数字资产确权和交易的爆发，从而推动这一生态的繁荣。

, from the perspective of society as a whole and economics, web3 is currently doing the best in areas related to decentralizing finance. , with the emergence of the AGC, will be followed by the outbreak of digital asset validation and trading, thus contributing to the ecological prosperity.

由于数字资产的建设成本较低，它本身不光是艺术品，还有可能成为标准化的建设品或生产资料，使得在构建元宇宙的门槛变低，而数字资产将来会变成和实体资产一样成为非常重要的东西。

, due to the low cost of building digital assets, it is not just art in itself, but also possible standardized construction products or production materials that lower the threshold of building meta-cosm, and digital assets will become as important as physical assets in the future.

这些数字资产将运行在个人计算机和网络上，基于区块链的去中心化的网络作为基础设施。同时，各种去中心化组织也需要这样的结构来完善自身。因此，我们也在去中心化科学研究领域做出了一些工作。

These numbers of assets will run on personal computers and networks, with decentralised networks based on block chains as infrastructure.

长远来看，新一代互联网这将是一个巨大的市场，每个人都可以参与其中。尽管目前的区块链对普通用户来说门槛很高，但我们夸密量子的技术可以降低这个门槛，为人人上链铺平道路，用量子技术塑造新一代互联网。

in the long term, the new generation of the Internet will be a huge market in which everyone can participate. Although the current block chain has a high threshold for ordinary users, our technology of inflating quantum can lower that threshold, paving the way for everyone to come in and shape the new generation of the Internet with quantum technology.

个人简介：张文卓，夸密量子创始人兼CEO。中国科学院量子光学重点实验室博士，中国科学院物理研究所和丹麦Aarhus大学国家量子技术博士后，曾任中国科学技术大学“墨子号”卫星团队任副研究员。先后主持国家自然科学基金面上项目1项，作为课题骨干参加科技部重点研发计划2项。在Nature Photonics、Optical Letters、Optical Express、Physical Review、Journal of Physics等国际高水平学术期刊发表论文20余篇。2020年全职从事量子技术创业，获得杭州市领军人才，绍兴市特聘专家等称号。创业期间发明基于量子随机数芯片的移动设备量子加密技术体系，以及建设新一代互联网的区块链随机性证明（PoR）共识机制。创业期间已获国家授权发明专利1项，实用新型专利3项，软件著作权4项，另已受理公开10余项发明专利。

本文为张文卓博士在全球数字金融中心（杭州）举行的“科技向善：强AI时代的变革”人工智能与数字金融研讨会发言。

.

2023年中国零售金融榜单评选

案例征集中！

转载、合作、交流请留言，

数据与商业合作：13261990570（微信同号）

客服微信：lycj002