全网轰动！DeepSeek超越ChatGPT，登顶美区

中国AI公司的创造力正技惊四座。

最近几天，一家名为深度求索（DeepSeek）的中国公司在欧美AI圈引起了不小的震动，甚至被认为是大模型行业的最大“黑马”。DeepSeek被不少外国人称为“神秘的东方力量”。

DeepSeek, a relatively unknown Chinese AI startup, has sent shockwaves through Silicon Valley with its recent release of cutting-edge AI models. Developed with remarkable efficiency and offered as open-source resources, these models challenge the dominance of established players like OpenAI, Google and Meta.

1月27日，DeepSeek应用登顶苹果美国地区应用商店免费APP下载排行榜，在美区下载榜上超越了ChatGPT。

苹果美国区应用商店

同日，苹果中国区应用商店免费榜显示，DeepSeek成为中国区第一。

苹果APP Store中国区免费榜

DeepSeek has surged to the top of the free app download charts in the United States region of the Apple App Store, surpassing the once-dominant ChatGPT. It also secured the number one spot on the free app rankings in China.

对于一款中国大模型来说，能够在美国力压ChatGPT，也是历史性一刻。

DeepSeek是什么

DeepSeek，全称杭州深度求索人工智能基础技术研究有限公司，成立于2023年7月17日，是一家创新型科技公司，专注于开发先进的大语言模型 （LLM）和相关技术。

DeepSeek, founded in July 2023, is a Chinese AI startup that develops open-source large language models (LLMs), according to the company's website.

几天前，总部位于中国杭州的DeepSeek发布推理模型R1，在性能逼近OpenAI o1正式版的同时，推理成本却仅为后者的几十分之一。

外媒称，DeepSeek大模型以极低成本（600万美元）和少量芯片（2000块）实现了与OpenAI等巨头相媲美的性能，挑战了“唯有科技巨头才能研发尖端AI”的行业共识。

The company unveiled R1, a specialized model designed for complex problem-solving, on Jan 20, which "zoomed to the global top 10 in performance," and was built far more rapidly, with fewer, less powerful AI chips, at a much lower cost than other US models, according to the Wall Street Journal. The Chinese engineers said they needed only about $6 million in raw computing power to build their new system. That is about 10 times less than the tech giant Meta spent building its latest AI technology.

低成本实现高性能模型研发，对用户来说的体验感也立竿见影——它功能强大，但却免费使用，并且DeepSeek还将代码面向开发者进行了开源。

据了解，DeepSeek R1没有使用业内普遍使用的监督微调（SFT）训练范式，而是直接通过强化学习让模型自主进化出复杂的推理能力，包括反思和长链思考等能力。这种方法不仅提高了训练效率，还减少了对昂贵计算资源的依赖。

Unlike traditional methods that rely heavily on supervised fine-tuning, DeepSeek's models learn by interacting with their environment and receiving feedback on their actions, similar to how humans learn through experience. This allows them to develop more sophisticated reasoning abilities and adapt to new situations more effectively.

与OpenAI的o1相比，DeepSeek模型的百万token输入成本从15美元锐减到0.55美元，输出成本则从60美元降低到2美元。

有人提出，DeepSeek恰恰是美国对华进行芯片出口限制之下所激发出的创新。

Meta生成式AI团队

正疯狂分析DeepSeek

1月24日，美国消费者新闻与商业频道CNBC发文称，DeepSeek的AI模型“挑战了美国在AI领域的主导地位” （challenges America’s global leadership in artificial intelligence）。

同日，华尔街顶级风投A16Z创始人马克·安德森在社交媒体发言称，DeepSeek R1是其见过的最令人惊叹、最令人印象深刻的突破之一，并且是开源的，是给世界的礼物。

Venture capitalist Marc Andreessen posted on X: “Deepseek R1 is one of the most amazing and impressive breakthroughs I’ve ever seen — and as open source, a profound gift to the world.”

英伟达资深科学家、AI智能体业务负责人Jim Fan也对其给予了高度评价。

另据媒体报道，Meta（前身为 Facebook）员工在美国匿名职场社区Teamblind上发帖提到，DeepSeek最近的一系列动作让Meta的生成式AI团队陷入了恐慌，工程师正在疯狂地分析DeepSeek，试图从中复制任何可能的东西。

"Engineers are moving frantically to dissect DeepSeek and copy anything and everything we can from it," said a staff member of Meta on the anonymous and professional community Teamblind.

编辑：左卓

来源：中国日报网 北京日报 环球网 华尔街日报 纽约时报 福布斯