【DeepSeek】DeepSeek創始人梁文鋒的故事 |經濟學人



梁文鋒追求通用人工智慧,採取與矽谷不同的創新路徑,在有限資源下探索新模型結構,證明創新可不依賴硬體。然而,OpenAI質疑R1可能複製西方成果。

中國AI公司DeepSeek發佈的模型R1以低成本實現與OpenAI o1相當性能,震驚全球科技界。其創始人梁文鋒出身湛江貧困村莊,早顯數學天賦,在浙江大學學習電子資訊後轉投AI領域。他先後創立量化投資公司,投入大量資金研發深度學習平台,最終孵化出DeepSeek。

DeepSeek的成功並非依賴國家主導創新模式,暴露該模式侷限。梁文鋒指出中美創新差距在於“原創與模仿”,提示中國應重構創新生態,而非照搬西方路徑。

Business | Flying high

DeepSeek poses a challenge to Beijing as much as to Silicon Valley

Jan 29th 2025 | The story of Liang Wenfeng, the model-maker’s mysterious founder

關於這家模型製造商的神秘創始人,梁文鋒的故事

目前,梁先生的公開資料相當匱乏。1985年,他降生於南方城市湛江周邊一個貧窮村落的教師家庭,打小就展現出超乎常人的學習天賦。據一位昔日的老師回憶,他在中學階段便已熟練掌握了大學層級的數學知識,令人驚嘆不已。2002年,憑藉自身的努力和優異成績,他成功被浙江大學錄取,開啟了電子資訊專業的求學之路,浙江大學坐落於東部名城杭州,在教育界享有盛譽。此後,他繼續在這所頂尖學府深造,跟隨一位業內知名的機器視覺科學家攻讀碩士學位,在此期間,他正式踏入了AI領域的大門,開啟了一段全新的探索之旅。

彼時,杭州堪稱網際網路技術的繁榮重鎮,諸多新興企業在此誕生,其中就包括電商巨頭阿里巴巴。梁先生與幾位同窗選擇留在這座城市,著手嘗試開發量化投資模型。這種模型並不以公司基本面為依據,而是借助大量的資料運算來發揮作用。到了2013年,梁先生與三位同學一同創立了一家名為雅克比(Yakebi)的投資集團,希望能將他們精心建構的交易模型轉化為實際的商業價值。

兩年過後,梁先生攜手他人共同建立了幻方量化(High-Flyer)避險基金。在當時中國監管環境相對寬鬆,市場波動較為頻繁的階段,這只基金如同眾多同類公司一般,實現了飛速發展。

2021年,幻方量化對外宣稱其管理的資金規模達到了1000億元人民幣(約合140億美元)之巨,不過到了當年下半年,其資金規模卻呈現出快速縮減的態勢。量化基金時常與中國監管部門產生摩擦,監管部門認為這類基金在市場暴跌的情況下獲取利益。有業內人士透露,幻方量化在業內以行事最為激進的量化基金之一著稱,也因此頻繁引發證券監管機構的關注與不滿。

DeepSeek的誕生源於幻方量化提升自身演算法的一次探索。在2019年,幻方量化投入2億元人民幣,專門成立了一個獨立部門,旨在研發名為“Fire-Flyer 1”的深度學習平台。到了2021年,該基金再度斥資10億元人民幣,用於推出配備1萬個輝達A100圖形處理單元的第二代產品。這一舉措讓幻方量化在行業中脫穎而出,顯得格外特別:彼時在中國,僅有包括阿里巴巴這類科技巨頭在內的另外四家公司擁有如此規模數量的高性能晶片。直到2023年,DeepSeek才正式從幻方量化中獨立出來,成為一家獨立運作的公司。

對梁先生而言,使用較少計算能力開發模型是他追求長期目標過程中的關鍵一步。他向當地媒體表示:“我們的目標是通用人工智慧(AGI),這就要求我們探索新的模型結構,以便在有限資源內實現卓越的能力。”

梁先生指出,過去30年間,中國在技術領域一直充當著“追隨者”的角色,主要是基於西方已發展的技術成果來進行拓展建設。今年7月,他在接受當地媒體採訪時談到,中西在技術層面的差距核心在於“原創性和模仿”。他覺得,輝達之所以能夠取得成功,不單單是依靠自身強大的技術實力,西方眾多公司之間密切的技術合作也起到了關鍵作用。在他看來,雖然在國家的大力主導下投入了巨額資金,試圖建構類似的技術合作體系,但由於未能有效實現這種協同合作,導致在追趕西方計算能力方面,中國的發展成效未達預期。DeepSeek的出現,或許不僅給矽谷提了醒,也為北京的決策層提了醒。(以上來自經濟學人報導)

The Wall Street Journal

DeepSeek Chief’s Path to AI Disruptor

DeepSeek掌門人(梁文鋒)成為AI顛覆者之路

Some call him China's Sam Altman. Others liken him to quantitative investing pioneer Jim Simons. Liang Wenfeng has much in common with these two innovators, and his influence may ultimately be just as great. Liang Wenfeng's Chinese company DeepSeek's artificial intelligence model has shocked the world, ranking among the top ten globally in terms of performance and popularity. The company achieved this feat using less advanced chips than those in the United States, stunning tech executives in Silicon Valley, politicians in Washington and global investors.

有人稱他為中國的山姆·奧特曼。也有人將他比作量化投資先驅吉姆·西蒙斯。梁文峰與這兩位創新者有諸多共同之處,其影響力最終或許也會同樣巨大。梁文峰旗下的中國公司DeepSeek所開發的人工智慧模型震驚世界,在性能和受歡迎程度方面位居全球前十。該公司使用的晶片不如美國先進,卻取得了這一成就,令矽谷的科技高管、華盛頓的政客以及全球投資者大為驚嘆。

Like Simons, Liang Wenfeng is a math geek with a passion for investing. He realized that analyzing large amounts of market data through computers could uncover hidden patterns and thus find ways to profit. His hedge fund High-Flyer Asset Management oversees about $8 billion in funds and is one of the largest quantitative funds in China. Colleagues said that Liang Wenfeng doesn't care about the external appearances of wealth and fame, but he has said that he hopes to gain the respect of the US-led tech community. It's safe to say that he has now earned that respect.

和西蒙斯一樣,梁文峰是個痴迷投資的數學極客。他意識到,通過電腦分析大量市場資料,能夠發現隱藏的規律,進而找到盈利方法。他的避險基金高毅資產管理規模約80億美元,是中國最大的量化基金之一。同事們表示,梁文峰不在意財富與名聲的外在表象,但他曾說希望贏得以美國為首的科技界的尊重。可以說,他如今已贏得了這份尊重。

High-Flyer Asset is a pioneer in applying relatively new artificial intelligence technologies in China. Through these technologies, computers can not only collect market price data, but also process various types of data and sort out valuable patterns from them. Unlike high-frequency traders who try to trade a few milliseconds faster than others, Liang Wenfeng and his colleagues focus on medium-frequency trading with a relatively longer position-holding time. Liang Wenfeng was inspired by Simons, a mathematician and a pioneer in quantitative finance who founded the Long Island hedge fund Renaissance Technologies and started using machine learning techniques in the 1980s. Liang Wenfeng wrote the preface to the Chinese version of The Man Who Solved the Market: How Jim Simons Launched the Quant Revolution. He wrote, "Whenever I encounter difficulties in my work, I always think of Simons' words: 'There must be a way to model prices'."

梁文鋒創辦的高毅資產是中國應用相對新型人工智慧技術的先驅。借助這些技術,電腦不僅能收集市場價格資料,還能處理各類資料並從中梳理出有價值的規律。與那些試圖比他人快幾毫秒進行交易的高頻交易員不同,梁文峰及其同事專注於持倉時間相對較長的中頻交易。

梁文峰的靈感源自西蒙斯,西蒙斯是一位數學家,也是量化金融領域的先驅,他創立了長島避險基金文藝復興科技公司,並於20世紀80年代開始使用機器學習技術。梁文峰為《解讀量化投資:西蒙斯用公式打敗市場的故事》中文版撰寫了序言。他寫道:“每當工作中遇到困難,我總會想起西蒙斯的話:‘一定有辦法為價格建模’。”

Technology executive and former hedge fund manager Richard Dewey said that DeepSeek's AI development strategy of achieving more with fewer resources (a smaller team, less capital and fewer chips) reminds him of quantitative traders. Dewey said, "They seem to focus on extracting a large number of signals from a relatively small amount of data, which is similar in spirit to the requirements of quantitative trading." According to people who know Liang Wenfeng, like some other quantitative traders, he prefers to be regarded as an engineer rather than a trader. People familiar with High-Flyer Asset describe Liang Wenfeng as a hands-on leader who optimizes algorithms to make them more intelligent in stock selection and risk hedging. Sometimes, to meet project deadlines, he would sleep in the office with his colleagues. He described his entry into the field of artificial intelligence models as a natural progression for an engineer curious about exploring the limits of computer learning.

科技高管、前避險基金經理理查德·杜威表示,DeepSeek用更少資源(規模更小的團隊、更少資金和晶片)實現更多成果的人工智慧開發策略,讓他聯想到量化交易員。杜威說:“他們似乎專注於從相對少量的資料中提取大量訊號,這在理念上與量化交易的要求相似。”瞭解梁文峰的人說,和其他一些量化交易員一樣,他更希望被視為一名工程師而非交易員。熟悉高毅資產的人形容梁文峰是一位親力親為的領導者,他最佳化演算法,讓其在選股和風險避險方面更智能。有時,為了趕項目期限,他會和同事們睡在辦公室。他將自己進入人工智慧模型領域,描述為一名對探索電腦學習極限充滿好奇的工程師的自然發展。

One of Liang Wenfeng's major decisions was to open-source the code, meaning anyone could access it. He said he hopes DeepSeek can break the monopoly of large tech companies. He said in an interview with 36Kr last year, "For technicians, seeing others follow up on their achievements brings great satisfaction. Open-sourcing is more of a culture than a commercial act, and contributing to open-source can earn us respect." Colleagues said that Liang Wenfeng fits the typical image of a Chinese engineer. They said that he doesn't care much about his clothes and hairstyle, makes decisions based on formulas and calculations, and loves football. They also said that Liang Wenfeng has always been low-key and was very surprised to see DeepSeek become popular overnight.

梁文峰的一項重大決策是開放原始碼,意味著任何人都能獲取。他表示希望DeepSeek能打破大型科技公司的壟斷。去年他在接受36氪採訪時說:“對於技術人員來說,看到別人跟進自己的成果會帶來極大的滿足感。開源更多是一種文化,而非商業行為,為開源做貢獻能贏得尊重。”同事們說,梁文峰符合典型中國工程師的形象。他們說,他不太在意穿著和髮型,基於公式和計算做決策,熱愛足球。他們還說,梁文峰一直很低調,看到DeepSeek一夜成名感到十分驚訝。 (Wendy讀書)