DeepSeek's Journey: From Quant Trading to LLM Innovation

Founder's Early Journey & Quant Success

• 💡 Liang Wenfeng (LWF), the founder of DeepSeek, began his career as a math genius from a small village in Guangdong, China, studying electronic information engineering at the prestigious Zhejiang University.
• 🚀 He started quant trading in 2008 during graduate school, developing an automatic trading system that became the foundation for his first investment company, Hul Yakobi, co-founded in 2013.
• 📊 Hul Yakobi experienced rapid growth, managing 1 billion R&B by 2015 and an astounding 100 billion R&B by 2021, making LWF temporarily one of China's richest men.
• 🧠 LWF demonstrated early conviction in artificial intelligence, declining an offer from DJI and actively investing in GPUs and an AI lab (Firefly) from as early as 2015.

The Pivot to Large Language Models

• ⚠️ Despite immense success, LWF's quant firm faced a significant financial drawdown in late 2021, exacerbated by market volatility and government crackdowns on financial arbitrage in China.
• 💡 This crisis, coupled with the global impact of ChatGPT's release in late 2022, prompted LWF to pivot his focus entirely to Large Language Model (LLM) development.
• 🚀 DeepSeek was officially established in May 2023, with a bold vision centered on achieving Artificial General Intelligence (AGI), distinguishing itself from typical short-term ROI-driven Chinese tech companies.

Innovation Under Constraint

• 🛠️ DeepSeek's innovative approach was largely driven by US export bans on high-performance GPUs to China, forcing the team to develop cost-saving architectures.
• 🧠 DeepSeek V2 (May 2024) introduced key innovations like Mixture of Experts (MoE) and MLA, significantly reducing training costs and computational power requirements.
• 💡 The DeepSeek R1 reasoning model (Jan 2025) gained global attention for its unique self-correction mechanism, solving complex problems without human feedback, a departure from traditional reinforcement learning.

Impact and Future Outlook

• 📈 DeepSeek's models, particularly V3 and R1, disrupted the market, with a widely publicized (and somewhat misleading) $6 million training cost claim that initially shocked the industry.
• ✅ The company's commitment to open-source AI led to over 500 derivative models and 2.5 million downloads on Hugging Face, fostering a vibrant developer community and increasing accessibility.
• 🎯 DeepSeek employs an unconventional hiring philosophy, prioritizing creative new graduates and PhD students from domestic Chinese universities over experienced professionals, to cultivate fresh perspectives.
• ⚠️ Despite its rapid success and innovation, DeepSeek faces ongoing challenges, including continued GPU supply restrictions and the long-term financial sustainability of its AGI-focused, non-commercialized business model.