DeepSeek是如何把构建AI的价格“打下来”的

DeepSeek真的太火了！

今天你练听力了吗？

🤔️小作业：

1. How does DeepSeek increase the precision of its calculations, as mentioned in the article?

A. It uses advanced GPUs to process data more quickly.

B. It stores numbers in 8 bits and then uses 32 bits for multiplication.

C. It employs more data storage to keep intermediate results.

D. It uses a larger memory pool to store calculations for later use.

无注释原文：

How Did DeepSeek Build Its A.I. With Less Money?

From: The New York Times

Last month, U.S. financial markets tumbled after a Chinese start-up called DeepSeek said it had built one of the world's most powerful artificial intelligence systems using far fewer computer chips than many experts thought possible.

A.I. companies typically train their chatbots using supercomputers packed with 16,000 specialized chips or more. But DeepSeek said it needed only about 2,000.

As DeepSeek engineers detailed in a research paper published just after Christmas, the start-up used several technological tricks to significantly reduce the cost of building its system. Its engineers needed only about $6 million in raw computing power, roughly one-tenth of what Meta spent in building its latest A.I. technology.

How was DeepSeek able to reduce costs?

It did many things. Most notably, it embraced a method called “mixture of experts.”

Companies usually created a single neural network that learned all the patterns in all the data on the internet. This was expensive, because it required enormous amounts of data to travel between GPU chips.

If one chip was learning how to write a poem and another was learning how to write a computer program, they still needed to talk to each other, just in case there was some overlap between poetry and programming.

With the mixture of experts method, researchers tried to solve this problem by splitting the system into many neural networks: one for poetry, one for computer programming, one for biology, one for physics and so on. There might be 100 of these smaller “expert” systems. Each expert could concentrate on its particular field.

Many companies have struggled with this method, but DeepSeek was able to do it well. Its trick was to pair those smaller “expert” systems with a “generalist” system.

The experts still needed to trade some information with one another, and the generalist — which had a decent but not detailed understanding of each subject — could help coordinate interactions between the experts.

It is a bit like an editor's overseeing a newsroom filled with specialist reporters.

That's it?

Well, they added another trick.

After squeezing each number into 8 bits of memory, DeepSeek took a different route when multiplying those numbers together. When determining the answer to each multiplication problem — making a key calculation that would help decide how the neural network would operate — it stretched the answer across 32 bits of memory. In other words, it kept many more decimals. It made the answer more precise.

- ◆ -

注：完整题目见本文开头；中文文本为纽约时报官方译文，仅供参考

含注释全文：

How Did DeepSeek Build Its A.I. With Less Money?

From: The New York Times

Last month, U.S. financial markets tumbled after a Chinese start-up called DeepSeek said it had built one of the world's most powerful artificial intelligence systems using far fewer computer chips than many experts thought possible.

上个月，一家名为“深度求索”(DeepSeek)的中国初创公司表示，它用比许多专家认为的最低限度要少得多的芯片，打造出世界上最强大的人工智能系统之一，随后美国金融市场出现暴跌。

可以作动词，也可以作名词，表示“跌倒；滚落；暴跌”，英文解释为“a sudden fall”举个🌰：The jockey took a nasty tumble at the third fence. 骑师在第三道栅栏处给重重摔下马来。Share prices took a sharp tumble following news of the merger. 合并消息传出，股价随即暴跌。（急剧下降）

🎬电影《指环王2：双塔奇兵》（The Lord of the Rings: The Two Towers）中的台词提到：He took a little tumble off the cliff. 他掉下了悬崖。

startup / start-up /ˈstɑːt.ʌp/ 表示“刚起步的小企业，新兴小型企业；初创公司”，英文解释为“a small business that has just been started”如：start-up costs 创业成本，举个🌰：Start-ups are very vulnerable in the business world. 商界中刚起步的小企业非常脆弱。

chip /tʃɪp/ 1）表示“炸薯条，油炸土豆条”，英文解释为“a long, thin piece of potato that is fried and usually eaten hot”如：fish and chips 炸鱼加土豆条。

2）表示“（尤指计算机的）集成电路片，芯片”，英文解释为“a very small piece of semiconductor, especially in a computer, that contains extremely small electronic circuits and devices, and can perform particular operations”如：a silicon chip 硅芯片。

A.I. companies typically train their chatbots using supercomputers packed with 16,000 specialized chips or more. But DeepSeek said it needed only about 2,000.

人工智能公司通常使用装有1.6万枚或更多专用芯片的超级计算机来训练聊天机器人。但深度求索表示，该公司只用了大约2000枚芯片。

pack /pæk/ 作名词，1）表示“一批，一套，一组”，英文解释为“a group, set, or collection of something”举个🌰：The information pack consists of a brochure and a map. 那套资料包括一本小册子和一张地图。

2）表示“（尤指行为不受人欢迎的）一伙人，一帮人”，英文解释为“a group of similar people, especially one that contains people whose activities you do not approve of”如：a pack of thieves 一伙窃贼。

3）表示“（统称）竞赛中的落后者”，英文解释为“all the people who are behind the leaders in a race, competition, etc.”如：measures aimed at keeping the company ahead of the pack 旨在使公司领先所有竞争对手的措施。

作动词，1）表示“聚集；挤进，塞满”，英文解释为“to come or bring together in large numbers or to fill a space”举个🌰：Thousands of fans are packing into the stadium. 体育馆里挤满了数千名球迷。

2）表示“（把…）压实，（使）堆积”，英文解释为“to (cause to) form into a solid mass”举个🌰：The wind has packed the snow against the garage door. 风把雪吹到车库门口，积成了厚厚的一堆。

specialized /ˈspeʃ.əl.aɪzd/ 表示“专门的；专科的”，英文解释为“relating to one particular area or designed for a particular purpose”举个🌰：Her job is very specialized (= involves only one limited area). 她的工作专业性很强。

As DeepSeek engineers detailed in a research paper published just after Christmas, the start-up used several technological tricks to significantly reduce the cost of building its system. Its engineers needed only about $6 million in raw computing power, roughly one-tenth of what Meta spent in building its latest A.I. technology.

正如深度求索工程师在圣诞节甫一过后发表的一篇研究论文中详细说明的那样，这家初创公司使用了一些技术手段，显著降低了系统构建成本。它的工程师只需要约600万美元的纯算力，大约是Meta在构建其最新人工智能技术时所花费的十分之一。

trick /trɪk/ 作动词，表示“欺骗，诱骗，哄骗”，英文解释为“to deceive someone, often as a part of a plan”举个🌰：He tricked the old lady into giving him two hundred dollars. 他骗这个老妇人给了他200美元。

作名词，可以表示“诀窍，秘诀；技巧”，英文解释为“an effective or quick way of doing something”举个🌰：What's the trick to getting this chair to fold up? 把这张椅子折叠起来有什么诀窍？

How was DeepSeek able to reduce costs?

深度求索是如何把成本降下来的？

It did many things. Most notably, it embraced a method called “mixture of experts.”

它做了很多工作。其中最值得注意的是，它采用了一种所谓的“混合专家”法。

mixture /ˈmɪks.tʃər/ 表示“混合物；混合体；混合料”，英文解释为“a substance made from a combination of different substances, or any combination of different things”举个🌰：Their latest CD is a mixture of new and old songs. 他们最新的光盘中收录了新歌和老歌。

Companies usually created a single neural network that learned all the patterns in all the data on the internet. This was expensive, because it required enormous amounts of data to travel between GPU chips.

公司通常会创建一个单一的神经网络，学习互联网上所有数据的所有模式。这样做的成本很高，因为它需要大量的数据在GPU芯片之间传输。

neural network /ˌnjʊə.rəl ˈnet.wɜːk/ 表示“类神经网络（一种模拟人脑运作机制的电脑系统或程序）”，英文解释为“a computer system or a type of computer program that is designed to copy the way in which the human brain operates”举个🌰：Neural networks can learn solutions to difficult problems. 类神经网络能学习如何解决问题的办法。

If one chip was learning how to write a poem and another was learning how to write a computer program, they still needed to talk to each other, just in case there was some overlap between poetry and programming.

如果一枚芯片正在学习如何写一首诗，而另一枚芯片正在学习如何编写计算机程序，它们还是需要相互交流，以防诗歌和编程之间出现某种重叠。

overlap /ˌəʊ.vəˈlæp/ 可以作名词，也可以作动词，1）表示“（活动、课程或时间）相互重叠，有共同之处”，英文解释为“If two or more activities, subjects, or periods of time overlap, they have some parts that are the same.”举个🌰：My musical tastes don't overlap with my brother's at all. 我对音乐的喜好和我弟弟完全不一样。

2）表示“（与…）交叠，（与…）部分重叠，叠盖”，英文解释为“to cover something partly by going over its edge; to cover part of the same space”举个🌰：The fence is made of panels that overlap (each other). 这个围栏是由交叠的板条搭成的。

With the mixture of experts method, researchers tried to solve this problem by splitting the system into many neural networks: one for poetry, one for computer programming, one for biology, one for physics and so on. There might be 100 of these smaller “expert” systems. Each expert could concentrate on its particular field.

研究人员尝试通过混合专家法来解决这个问题，他们将系统拆分成许多神经网络：一个用于诗歌，一个用于计算机编程，一个用于生物学，一个用于物理学，等等。这样较小的专家系统可能多达100个。每个专家都可以专注在特定领域。

Many companies have struggled with this method, but DeepSeek was able to do it well. Its trick was to pair those smaller “expert” systems with a “generalist” system.

许多公司在尝试这种方法时并不顺利，但深度求索能够很好地做到这一点。它的诀窍是将那些较小的“专家”系统与一个“通才”系统配对。

pair with 表示“与…配对”，英文解释为“to make something one of a pair, or to become one of a pair”举个🌰：Famous paintings have been paired with poems by Sandburg, Angelou, and others. 著名的油画作品与桑德伯格和安杰洛等人的诗配成对。 

generalist /ˈdʒen.ər.əl.ɪst/ 表示“全面的，广泛的，通博的；多面手，全才，通才”，英文解释为“(someone who is) not specialized”举个🌰：Children of this age need specialist rather than generalist teachers. 这个年龄段的孩子需要的是专业老师而不是通才教育。

The experts still needed to trade some information with one another, and the generalist — which had a decent but not detailed understanding of each subject — could help coordinate interactions between the experts.

专家系统仍然需要相互交换一些信息，而通才系统可以帮助协调专家系统之间的互动。通才系统对每个主题都有不错的理解，但比较粗略。

表示“正派的；体面的，正经的；像样的，相当不错的”，英文解释为“socially acceptable or good”举个🌰：It was very decent (= kind) of you to help. 你能帮忙，真是个大好人啊。

It is a bit like an editor's overseeing a newsroom filled with specialist reporters.

这有点像一个主编负责一个全是专业记者的新闻编辑室。

That's it?

就这么简单？

Well, they added another trick.

这个嘛，他们另外还有一招。

After squeezing each number into 8 bits of memory, DeepSeek took a different route when multiplying those numbers together. When determining the answer to each multiplication problem — making a key calculation that would help decide how the neural network would operate — it stretched the answer across 32 bits of memory. In other words, it kept many more decimals. It made the answer more precise.

在将每个数字塞进八位存储器后，深度求索在将这些数字相乘时采取了不同的方法。在确定每个乘法问题的答案时——进行有助于决定神经网络将如何运作的关键计算——它将答案扩展到32位存储器中。换句话说，这样就保留了更多的小数，使得答案更为精确。

1）表示“（困难地）挤进，挤过”，英文解释为“to get in, through, under, etc. with difficulty”举个🌰：She squeezed through the crowd and found a seat at the front. 她从人群中挤过去，在前面找了个座位。

2）表示“挤压；捏”，英文解释为“to press sth firmly, especially with your fingers”如：to squeeze a tube of toothpaste 挤牙膏，举个🌰：He squeezed her arm reassuringly. 他安慰地捏了捏她的手臂。

3）表示“给…带来财务困难，使陷入财务困境”，英文解释为“If you are squeezed by financial demands, they cause you financial problems.”举个🌰：Small businesses are being squeezed by heavy taxation. 重税使小企业在财务上捉襟见肘。

🎬电影《今晚带我回家》（Take Me Home Tonight）中的台词提到：Come on, punch it up. punch. and squeeze real tight 来 击个拳 击拳 然后握紧。

multiply /ˈmʌl.tɪ.plaɪ/ 表示“大幅增加；乘，使相乘”，英文解释为“to increase very much in number, or (in mathematics) to add a number to itself a particular number of times”举个🌰：In warm weather these germs multiply rapidly. 在温暖的天气里这些细菌会迅速繁殖。

stretch /stretʃ/ 作名词，1）表示“（连续的）一段时间”，英文解释为“a continuous period of time”举个🌰：The elderly generally need far less rest than the young, and tend to sleep in several short stretches. 年长者一般需要的休息时间比年轻人要少得多，而且往往会分成几个小段时间睡觉。

2）表示“（土地或水域的）一片，一段”，英文解释为“a continuous area of land or water”举个🌰：This particular stretch of coast is especially popular with hikers. 在这一段海岸散步的人特别多。

作动词，1）表示“到…的限度；超过…的限度；竭尽”，英文解释为“to go as far as or past the usual limit of something”举个🌰：Many families' budgets are already stretched to breaking point. 许多家庭的预算已经撑到了极限。

2）表示“伸出；伸长；拉伸”，英文解释为“to cause something to reach, often as far as possible, in a particular direction”举个🌰：She stretched out her hand and helped him from his chair. 她伸手将他从椅子上扶起来。

3）表示“展开，铺开；延伸”，英文解释为“to spread over a large area or distance”举个🌰：A huge cloud of dense smoke stretched across the horizon. 一大团浓烟在地平线上蔓延开来。

4）表示“延续”，英文解释为“to spread over a long period of time”举个🌰：The dispute stretches back over many years. 这场争论可追溯到许多年以前。

1）作形容词，表示“十进制的”，英文解释为“A decimal system involves counting in units of ten.”举个🌰：The mathematics of ancient Egypt were based on a decimal system. 古埃及数学以十进制为基础。

2）作名词，表示“小数”，英文解释为“A decimal is a fraction that is written in the form of a dot followed by one or more numbers which represent tenths, hundredths, and so on”如：simple maths concepts, such as decimals and fractions 简单的数学概念，如小数和分数。

- 词汇盘点 -

- 词汇助记 By DeepSeek -

LearnAndRecord

2015年2月8日

2025年2月22日

第3668天

每天持续行动学外语