图灵预言成真？AI存储技术如何用两项突破改写人类数据史

穿越时空的对话：AI存储

附完整对话全文（中英）

数字人Jimmy：图灵先生，您好！我是上海川源信息科技有限公司的CEO Jimmy，我来自21世纪。未来的人们称您为“人工智能之父”，能穿越时空隧道见到您，我感到无比荣幸和激动。

Mr. Turing, hello! I am Jimmy, the CEO of AccelStor. I come from the 21st century. People in the future regard you as the "Father of Artificial Intelligence." It's really cool and exciting to meet you through this time tunnel!

Alan Turing: Hello Jimmy! It's my pleasure to meet you. Really very interested to know how much advancement technologies have seen during these years. Why don’t you just go ahead and tell me something about your company and what kinds of technologies you are engaged with?

你好，Jimmy！很高兴见到你。我非常想了解这些年来人类社会的技术取得了哪些进展。你能否介绍一下你的公司，以及你们所从事的技术领域？

数字人Jimmy：没问题。我们先来看一段对话

No problem. Let's look at a conversation as an example.

生成一段Python程序代码并执行。

Generate a Python program and execute it.

我们来看看执行的效果。

Let's take a look at the execution result.

通过这段对话，您能判断出回答问题的是人还是机器吗？

From this dialogue, can you determine whether the respondent is a human or a machine?

Alan Turing: Incredible. I can hardly discern any difference. Have you succeeded in creating a "thinking machine"?

难以置信！ 我无法分辨。你们实现了“会思考的机器”？

数字人Jimmy：是的，我们这个时代成为“人工智能”，简称AI。短短的几十年，人类已经从传统的智能演进到AI，再到AGI（通用人工智能），甚至是ASI（超级人工智能）。

Yes, in our era, this is known as "Artificial Intelligence," or AI. In only a few decades, we went from normal smarts to AI, then AGI, and now we're even at ASI.

Alan Turing: Cool. How did you accomplish this?

太酷了，这是如何做到的呢？

数字人Jimmy：我们对数据进行训练，生成AI模型，也就是AI的“大脑”，然后利用模型进行推理。刚刚AI回答问题的过程就是一个推理过程。

We train the data to create an AI model, that acts as the “brain” of the AI. Once the model is built, it could be used for inference. For example, this process of the AI answering my questions just now is an inference process.

Alan Turing: That’s interesting. Where does the data come from?

数据从哪里来呢？

数字人Jimmy： AI模型的数据最早就来自于在互联网上公开的数据，相当于有154万篇论文。

The data used to train AI models comes from the Internet, roughly 1.54 million papers.

Alan Turing: I'm curious, are these data stored with delay-line memory? I don't really like using mercury delay lines for data storage—they can't balance speed and cost.

我很好奇这些数据是用延迟线存储器存储的吗？因为我对用水银延时线来存储数据并不是很热情，它们无法兼顾速度和经济性。

数字人Jimmy：延迟线存储器之后，人们发明了磁鼓存储器、磁芯存储器和硬盘驱动器、软盘、光盘，到现在广泛采用的闪存。我们公司的产品就是基于闪存构建的存储系统，并通过技术降低闪存的整体拥有成本，兼顾您刚刚提到的速度和经济性。

After delay-line memory, people developed magnetic drum memory, then magnetic core memory, hard disk drives (HDD), floppy disks, optical disks, and now popular flash memory. Our company's main products are all flash array—based on flash memory. We use technology to lower the overall cost of using flash memory while maintaining the speed and affordability you just mentioned.

Alan Turing: I am glad you've found better alternatives. Your storage systems, then, must be related to AI as well, right?

我很高兴你们有了更好的选择。那么你们的存储系统也一定是跟AI相关的，是吗？

I am glad you’ve found better alternatives.

数字人Jimmy：是的，AI 存储是我们未来发展的一个重要方向。现有的IT架构服务于传统的企业应用。AI 发展的趋势，逐渐从训练侧重向推理，从研究机构走向企业应用，这需要对企业的IT架构进行重新定义和设计。我们希望能够通过革新式AI存储，帮助企业真正从AI中获益。

Yes, AI Storage is one of the most important trends for our future development. The existing IT architecture supports legacy enterprise applications. The AI trend has shifted from training to inference and from research labs to business applications. Its developments require the redefinition and redesign of IT architecture for enterprises. We dedicated to helping companies really benefit from AI， through innovative AI storage solutions.

Alan Turing: It sounds very meaningful, but it must come with a lot of challenges as well. When I was working on the ACE（Automatic Computing Engine）project, I also had to handle so many challenges. However, I strongly felt that for different tasks, we could convert the problem of "designing different machines" into "designing different programs for a universal machine."

听上去很有意义，但也应该会面临很多挑战。在ACE（自动计算引擎）项目时，我当时也面临很多挑战，但我坚持认为，对于不同的工作，我们可以把“设计不同的机器”这个问题，变成“给通用机设计不同的程序”。

数字人Jimmy：是的，首先是存储架构上的挑战。AI模型发展最重要的一个趋势是多模态，需要对各种类型的数据进行学习。受益于您的启迪，我们认为，在设计AI存储时，不需要再为不同类型的数据设计不同的存储，而是一套统一的存储，支持对不同数据类型的存储。

Indeed, the first challenge lies in storage architecture. One of the most significant trends in AI model development is multimodality, which requires learning from various types of data. Inspired by your insights, we believe that when designing AI storage, it is no longer necessary to design separate storage solutions for different types of data. Instead, a unified storage system should be implemented to support the storage of diverse data types.

Alan Turing: You have certainly discovered the correct path. What other challenges does AI storage face?

你们似乎找到了正确的方向。AI storage还有哪些挑战呢？

数字人Jimmy：性能和容量。企业在使用AI模型部署推理应用时，需要调阅企业本地知识库中的数据，存储的数据读写速度、时延、带宽等将直接影响推理的性能。而当企业本地数据增长至容量上限时，传统的存储架构只能通过将数据迁移至新的存储来释放空间，迁移的过程耗时、耗能又费力。

Performance and capacity: Once inference applications powered by AI models are deployed, they have to retrieve data from local knowledge bases within an enterprise. Read/write speeds, latency, and bandwidth will have direct consequences on the inference's performance. Besides, as data volumes for the local knowledge bases of an enterprise become so massive that it approach storage capacity, legacy architectures will not free up much space without performing expensive and arduous tasks like migrating all that data over to a newer system.

Alan Turing: 你们的AI Storage如何处理这两个挑战呢？

How does your AI Storage address these two challenges?

数字人Jimmy：我们的AI Storage可以支持按需扩容，并支持性能和容量的解耦，也叫做存算分离，性能和容量可以分开扩展。企业可以根据需要逐步增加性能或容量。

Our AI Storage is designed to support on-demand scalability, and it decouples capacity and speed, also called Disaggregated Architecture, which means the capacity and speed can be expanded independently. Enterprises can gradually increase capacity or speed as needed.

Alan Turing: I am confident that this design will enable you to attain significant success in the realm of AI storage. I eagerly anticipate witnessing more groundbreaking technology from you in the future, which will undoubtedly bring even greater value to enterprises.

我相信这样的设计能够帮助你们在AI 存储领域取得巨大的成功。期待你们未来有更多的技术突破，为企业带来更多的价值。

数字人Jimmy：非常感谢您的认可和肯定。您追求本质的思想也一直激励着我们，如果不是站在您的肩膀上，人类无法在AI领域取得如此大的进展和成功。再次感谢您！期待下一个十年再与您相见。

Thank you so much for acknowledging and affirming. Your pursuit of fundamentals has always motivated us. Without standing on your shoulders, humanity wouldn't have made such awesome progress and success in AI. Thank you once again! I eagerly anticipate our reunion in the coming decade.