与谷歌AI领军科学家一较高下的梦想如何反应了中国对科技突破的追求？

‍编者注：

近日，一流科技OneFlow创始人袁进辉接受《南华早报》记者专访，分享科学家的创业心得。原文为英文，OneFlow团队翻译为中文。

原文题目为：How one man’s dream to rival Google’s AI chief reflects China’s quest for tech supremacy

撰文：Minghe Hu

首发：South China Morning Post

Yuan Jinhui, the 40-year-old founder of OneFlow Technology in Beijing, believes his company will one day rival Google’s TensorFlow and Facebook’s Pytorch to become the mainstream AI computing framework of choice and, through this accomplishment, he can become China’s answer to Jeff Dean, the American computer scientist who heads Google’s artificial intelligence research.

袁进辉，这位北京一流科技年届四旬的创始人，相信自己公司的产品OneFlow有朝一日可以超越谷歌的TensorFlow和脸书的PyTorch，成为AI计算框架的主流选择，他在中国达成的这一成就正如Jeff Dean在美国取得的成就，Jeff Dean是美国著名计算机科学家，谷歌人工智能的领袖。

In an interview with the South China Morning Post, the grey-haired Yuan, one of a group of Chinese scientists whose research may help the country become an AI superpower, said that it is only a matter of time for OneFlow, an open-source framework released last summer.

在《南华早报》的一次采访中，头发花白的袁进辉表示，对于去年夏天开源的框架OneFlow来说，成功只是时间问题。作为中国科学家中的一员，袁进辉的研究致力于帮助中国成为AI强国。

“I have always been trying to achieve it, and I know my work is still lagging far behind Dean’s accomplishment,” said Yuan, who started the company in 2017 when he saw an opening in the deep learning framework market.

2017年，袁进辉看到深度学习框架市场在“大模型分布式”方向上有着空缺，于是创办了这家公司。他说：“我一直在努力达成这个理想，我也知道我个人的成就目前远远落后于Jeff Dean。”

“But the framework of OneFlow is faster than Google’s, and I believe our ecosystem could eventually compete with theirs.”

“但OneFlow的框架比谷歌快，我相信我们的生态系统最终可以与他们竞争。”

OneFlow is a newcomer in the area of open-source AI frameworks for machine learning, which is similar to operating systems for personal computers.

OneFlow是机器学习领域开源AI框架的“新生代”，而深度学习框架在AI领域被类比为个人电脑的操作系统。

While Google and Facebook are the preferred frameworks globally, major Chinese technology firms are racing to come up with their own: Huawei developed Mindspore, Megvii launched MegEngine, and Baidu has its Paddle Paddle platform.

虽然谷歌TensorFlow和脸书PyTorch是全球范围内的首选框架，但中国主要科技公司也在竞相推出自己的框架:华为开发了Mindspore, 旷视推出了MegEngine，百度有自己的飞桨平台。

Beijing named AI as a key technology to develop in its 14th five-year plan, with the area expected to be a battlefield for China and the US as they jostle for supremacy.

中国政府将AI列为“十四五”规划的重要组成部分，随着中美科技争霸的持续，该领域将成为双方的重要角力点。

While China’s main advantage comes from the vast amounts of data from its 1.4 billion people, the country still has yet to figure out how to rapidly train its AI models using massive data inputs.

尽管中国的主要优势来自14亿人口产生的海量数据，但如何利用大量数据快速训练人工智能模型并非易事。

While working as an AI scientist at Microsoft Research Asia to create a large-scale machine learning platform in 2014, Yuan developed a light training algorithm system using only dozens of servers that could process highly complex calculations.

2014年，作为微软亚洲研究院的人工智能科学家，袁进辉创建了一个大模型机器学习平台。当时，袁进辉开发了一个轻量级训练算法系统，仅使用几十台服务器就可以处理高度复杂的计算。

The achievement brought Yuan recognition as the system was used in Microsoft’s search engine and gave him an inkling about the future growth potential of data and AI models. It also echoed his research into computational neuroscience for his post-doctoral degree.

这一成绩为袁进辉带来了业界认可，因为该系统被用于微软的搜索引擎，并让他对数据和人工智能模型的未来增长潜力有了初步认识。这也呼应了他在博士后时对计算神经科学的研究。

“Biological neural networks are at least 1,000 times larger than the current largest artificial ones, which means the scale of the framework would be a key component for future AI development,” Yuan said.

“生物神经网络比目前最大的人工神经网络至少要大1000倍，这意味着框架能训练的规模将是未来人工智能发展的关键因素。”袁进辉说道。

Seeing the huge demand for data and AI models, he left Microsoft and created OneFlow, which supports AI applications in image processing, voice recognition, text and ad recommendation systems.

在看到对数据和人工智能模型的巨大需求后，他离开了微软，开始研发OneFlow，该框架可支持图像处理、语音识别、文本和广告推荐系统等人工智能场景。

Researchers believe new AI frameworks have a chance to shine. “The current popular frameworks were designed about eight years ago with an original goal to support supervised models, such as those for computer vision,” said Luo Mai, an assistant professor at the University of Edinburgh and a specialist in machine learning systems.

研究人员认为新的人工智能框架有机会大放异彩。爱丁堡大学助理教授、机器学习系统专家麦洛表示:“目前流行的框架是在大约8年前设计的，最初的目的是支持监督模型，比如计算机视觉模型。”

“Those models are much smaller compared to the giant AI models that are emerging today. Nowadays, we are expecting a framework that is designed to support much bigger models and target wider applications.”

“与今天出现的大型人工智能模型相比，当年那些模型要小得多。现在，我们期待一个框架能够支持更大的模型和更广泛的应用。”

At the same time, Beijing is throwing support behind its own AI labs to cut reliance on US tools. Despite an ambitious plan to become a global leader in AI by 2030, nearly all small- to mid-sized Chinese AI companies rely on open-source platforms from the US.

与此同时，中国政府正在支持建设自己的人工智能实验室，以减少对美国工具的依赖。尽管中国雄心勃勃地计划到2030年成为人工智能领域的全球领导者，但几乎所有中小型人工智能企业都依赖于美国的开源平台。

“Open-source deep learning frameworks are fundamental and critical for AI technology, and 90 per cent of current applications in the market are rooted in them,” said Lu Yanxia, associate research director at IDC China. “It is a problem you cannot get around for a domestic replacement in AI innovation.”

“开源深度学习框架是人工智能技术的基础和关键，目前市场90%的应用都植根于这些框架，”IDC中国研究副总监卢言霞表示。“在人工智能创新领域，这是一个无法回避的问题。”

Huawei, the telecoms giant struggling under strict US trade sanctions, is investing significant sums to build an AI ecosystem, from deep learning frameworks to AI chips and cloud computing.

在美国严厉的贸易制裁下苦苦挣扎的电信巨头华为，正投入巨资打造一个从深度学习框架到AI芯片及云计算的人工智能生态系统。

“The best scenario for AI frameworks is to combine novel software design with hardware innovation because such a combination can often lead to tens or even hundreds of times improvement in computational efficiency,” said the University of Edinburgh’s Mai.

爱丁堡大学的麦洛说:“人工智能框架的最佳场景是将新颖的软件设计与硬件创新结合起来，因为这样的结合往往可以导致计算效率几十倍甚至数百倍的提高。”

“It will give the framework an unsurpassedcompetitive advantage.”

“这将给该框架带来无与伦比的竞争优势。”

Yuan’s vision for OneFlow is for it to become the most popular AI framework in the world. To achieve that goal at home first, the company is working with the Beijing Academy of Artificial Intelligence and Hangzhou-based Zhejiang Lab, along with top-tier universities and institutions, including Tsinghua University, Zhejiang University and the Chinese Academy of Science, to give students and academia the chance to know and use OneFlow.

袁进辉对OneFlow的愿景是让它成为世界上最流行的人工智能框架。为实现这个目标迈出的第一步,是公司正在与北京的智源人工智能研究院和总部位于杭州的之江实验室,以及一些一流大学和机构,包括清华大学、浙江大学和中国科学院共同合作，面向学术界提升OneFlow的知名度和使用率。

It is also winning over investors. In February, A-round fundraising with Hillhouse raised 50 million yuan, (US$7.7 million) and since then more investors have expressed interest.

OneFlow还赢得了投资者的青睐。今年2月， OneFlow A轮融资筹集了人民币5,000万元(770万美元)，为高瓴创投独家投资，此后越来越多的投资者表达了兴趣。

Yuan said OneFlow has received local government support in the form of tax breaks, talent recruitment and R&D funding, and in March Beijing’s Communist Party secretary, Cai Qi, visited the company’s headquarters, a sign of the high level of support for technology innovation.

袁进辉表示，OneFlow已经获得了当地政府的支持，包括税收减免、人才招聘和研发经费支持。今年3月，中共中央政治局委员、北京市委书记蔡奇视察了该公司总部，这释放了政府对技术创新的高度重视和支持的信号。

Since launching four years ago, OneFlow’s staff has grown from three, including Yuan, to more than 50. These days, Yuan said he chases investors during the day and writes code at night - a lifestyle that baffled his former colleagues and frustrated his family.

自4年前成立以来，OneFlow的员工从3人(包括袁进辉自己)增加到50多人。袁进辉说，当年他白天见投资者，晚上写代码——这种生活方式让他以前的同事和家人都感到困惑和苦恼。

“I was running into walls all the time. Former colleagues couldn’t understand why I gave up a stable job and started a company … even my family couldn’t agree with me,” said Yuan, who added that investors have been more rational about his vision.

“我一直在碰壁。以前的同事们无法理解我为什么放弃一份稳定的工作而自己创业……甚至我的家人也不同意我的选择” 不过他补充到，投资者反而更加理解他的愿景。

Raised by parents who worked as wheat farmers in northern Hebei province, Yuan was the top student at his school.

袁进辉出生在一个中国北部省份河北的农民之家，父母以种小麦等农作物为生。他自小到大都是学校里的优等生。

After graduating with a bachelor’s degree in computer science from Xidian University in northwestern Shaanxi province, he was admitted to a doctoral programme at Tsinghua University, studying with Zhang Bo, a well-known AI scientist in China and a fellow of the Chinese Academy of Sciences.

他在西安电子科技大学取得计算机科学的学士学位后，来到清华大学攻读博士学位，师从中国著名人工智能学界泰斗、中国科学院院士张钹。

“I have always been looking for a sense of accomplishment. Everyone should do the same, look for areas where they have a competitive edge,” Yuan said.

“我一直在寻找一种成就感。每个人都应该和可以这样做，一定能找到自己有竞争优势的领域。”袁进辉如是说。