深度求索(全稱杭州深度求索人工智慧基礎技術研究有限公司,英文簡稱DeepSeek)是中國大陸人工智慧以及大型語言模型公司,其創始人和執行長為梁文鋒Liang Wenfeng。
沿革
DeepSeek成立於2023年7月,由中國大陸避險基金幻方量化創立,其創始人和CEO為梁文鋒,公司的總部位於中國大陸浙江省杭州市,公司專注於AI大模型的研究和開發。
2024年5月,DeepSeek發布並開源了旗下MoE大模型DeepSeek-V2,因其模型架構與經濟性而脫穎而出,DeepSeek-V2的API定價為每百萬tokens輸入1元、輸出2元,價格僅為GPT-4 Turbo的百分之一。
2024年12月,DeepSeek發布並開源了DeepSeek-V3,DeepSeek表示該大模型的訓練系基於2,048塊輝達H800型GPU(針對中國大陸市場的低配版GPU)叢集上執行55天完成,訓練耗資557.6萬美元;DeepSeek-V3的評測成績超越Qwen2.5-72B(阿里自研大模型)和Llama 3.1-405B(Meta自研大模型)等開源模型,能與GPT-4o、Claude 3.5-Sonnet(Anthropic自研大模型)等閉源模型相抗衡。
2025年1月20日, DeepSeek發布並開源了DeepSeek-R1模型,該模型在數學、代碼、自然語言推理等任務上,效能超越OpenAI o1正式版。
2025年1月27日,DeepSeek智慧型助手在美區蘋果App Store下載榜上超越ChatGPT,並登頂App Store免費應用榜榜首。
#ChatGPT-YouTubeSummary-Chrome-Ext gives the following summary:
逐字稿摘要:
早年生活與教育(00:00–01:05)
- 梁文鋒於1985年出生於廣東湛江,從小對數學建模展現濃厚興趣。
- 17歲考入浙江大學電子信息工程專業,隨後攻讀信息與通信工程碩士,專注於機器視覺研究。
創業起步(2008–2013)
- 在2008年金融危機期間,梁文鋒運用機器學習進行量化交易。
- 拒絕大疆創始人汪滔的合夥邀請,選擇人工智慧創業之路。
- 2013年,與徐進共同創立杭州雅克比投資管理有限公司,為金融事業奠定基礎。
創建幻方量化基金(2015–2017)
- 2015年成立幻方科技(後更名為九章資產管理有限公司),專注於基於人工智慧的量化投資。
- 成功應對2015年股災,憑藉高頻策略迅速擴大管理資產至300億元。
AI與技術進展(2016–2020)
- 2016年引入AI模型,從線性模型過渡到深度學習,並開始使用GPU計算。
- 2019年推出深度學習平台「螢火一號」,配備1100塊GPU,管理資產超過1000億元。
- 2020年,GPU數量達到萬卡級別,成為少數具備該規模算力的企業之一。
挑戰與重組(2021)
- 2021年資產管理規模突破1萬億元,但因業績波動主動縮減超過1000億元資產規模。
- 積極調整交易策略以應對AI決策問題。
慈善與進軍AGI(2022–2023)
- 2022年捐款1.38億元,資助多項慈善項目,外界推測捐款者為梁文鋒。
- 2023年成立DeepSeek,進軍通用人工智慧(AGI),被譽為「AI界的拼多多」。
DeepSeek創新與成就(2024)
- 2024年推出DeepSeek-V3,以低成本超越包括ChatGPT-4o在內的全球競爭對手。
- DeepSeek的訓練成本僅為557.6萬美元,遠低於OpenAI的7800萬美元。
團隊與理念(2024)
- DeepSeek由139名全本土團隊組成,成員多來自清華、北大等頂尖高校。
- 梁文鋒仍親身參與技術研發,營造高效且團結的企業文化。
量化投資策略調整(2024)
- 幻方宣布逐步將對沖基金倉位降至零,展現重要策略調整。
影響力與行業地位
- 梁文鋒重塑中國量化與AI行業,在全球贏得廣泛認可,並以具競爭力的定價模式顛覆市場。
Summary of the Transcript:
Early Life and Education (00:00–01:05)
- Liang Wenfeng, born in 1985 in Zhanjiang, Guangdong, showed an early interest in mathematical modeling.
- At 17, he enrolled at Zhejiang University in electronic information engineering, later pursuing a master’s in information and communication engineering, focusing on machine vision.
Entrepreneurial Beginnings (2008–2013)
- During the 2008 financial crisis, Liang leveraged machine learning for quantitative trading.
- Declined DJI founder Wang Tao’s offer to co-found DJI, choosing AI entrepreneurship instead.
- In 2013, co-founded Hangzhou Jacobi Investment Management, laying a foundation for his finance career.
Establishing Huanfang Quantitative Fund (2015–2017)
- Founded Huanfang Technology (later Nine Chapters Asset Management) in 2015, specializing in AI-driven quantitative investments.
- Successfully navigated the 2015 stock market crash using high-frequency strategies and expanded assets under management to ¥30 billion by 2017.
AI and Technological Advancements (2016–2020)
- Introduced AI models in 2016, transitioning from linear models to deep learning, utilizing GPUs.
- Launched “Firefly One” in 2019, a deep learning platform with 1,100 GPUs, enabling Huanfang to manage over ¥100 billion by 2019.
- By 2020, the company’s GPU count reached over 10,000, a rarity among non-tech giants.
Challenges and Restructuring (2021)
- Achieved ¥1 trillion in assets under management in 2021 but faced setbacks due to performance fluctuations.
- Proactively reduced assets by over ¥100 billion, addressing AI decision-making issues in trading.
Charity and Expansion into AGI (2022–2023)
- Speculated to have donated ¥138 million to charitable causes in 2022.
- Entered the AGI field in 2023, founding DeepSeek, with a mission to lower AI application costs, earning the nickname “Pinduoduo of AI.”
DeepSeek Innovations and Achievements (2024)
- Released DeepSeek-V3 in 2024, surpassing global competitors like ChatGPT-4o in efficiency and cost-effectiveness.
- DeepSeek achieved superior performance with only $5.58 million in training costs compared to OpenAI’s $78 million.
Team and Philosophy (2024)
- DeepSeek operates with a 139-person, all-local Chinese team, composed of young talents from top universities like Tsinghua and Peking.
- Liang maintains hands-on involvement, fostering a collaborative and efficient company culture.
Strategic Adjustments in Quantitative Trading (2024)
- Huanfang announced plans to reduce hedge fund investments to zero, signifying a major strategic shift.
Recognition and Industry Impact
- Liang’s work has reshaped the quantitative and AI sectors in China, achieving global recognition and competitive pricing models that disrupt the industry.
#ChatGPT shared link:
https://chatgpt.com/share/6797fbc2-e16c-8009-af0e-745f05d5369b