Deepseek: Team behind the $1 trillion bloodbath

US Market & Sectors
1

  • Origins in Finance:

    • DeepSeek’s founder and CEO, Liang Wenfeng, co-founded High-Flyer, a Chinese quantitative hedge fund, in 2016.
    • The company specialized in AI-driven trading strategies, leveraging technology to optimize financial outcomes.

  • AI as a Side Project:

    • By 2021, High-Flyer fully transitioned to using AI exclusively for its trading operations.
    • Liang and his team began exploring AI model development as a side project to enhance their trading algorithms.
    • The success of these explorations inspired the creation of DeepSeek as a standalone AI research lab in 2023.

  • From Finance to AI Leadership:

    • What started as a supplementary initiative to boost trading efficiency evolved into a disruptive AI project.
    • The team’s finance background gave them unique insights into data-driven decision-making and optimization, which they applied to AI development.

  • Founder and CEO: Liang Wenfeng

    • Co-founded High-Flyer, a Chinese quantitative hedge fund, in 2016.
    • Established DeepSeek in 2023 as an AI research lab under High-Flyer.
    • Holds a background in electronics and computer vision from Zhejiang University.
    • Transitioned High-Flyer, his company to exclusively use AI in trading by 2021.

  • Team Composition

    • Comprises young researchers from top Chinese universities.
    • Emphasizes technical abilities over extensive work experience.
    • Includes individuals from diverse academic backgrounds to enhance the AI’s knowledge base.

  • Notable Achievements

    • Developed DeepSeek-R1, an AI model rivaling leading Western counterparts.
    • Achieved significant advancements despite limited resources and U.S. export restrictions on advanced chips.

This dynamic team, under Liang Wenfeng’s leadership, has propelled DeepSeek to the forefront of AI innovation, challenging established industry norms of costs to focus more on optimization.

2

Summary on how DeepSeek really works

Re-posting my comment from other post on same topic

  • DeepSeek’s AI model, DeepSeek-R1, focuses on reasoning capabilities, powered by pure reinforcement learning.

  • Unlike traditional AI models, it doesn’t rely on large supervised datasets (that uses large computational power) but uses techniques inspired by AlphaZero, mastering tasks through self-play and optimization (same that beat every other AI in chess by Google). I wonder why Google didn’t think of this move before.

  • Instead of increasing model size and computation, DeepSeek emphasizes algorithmic efficiency, making high-quality AI accessible for under $6 million

  • DeepSeek released its model under the MIT license, fostering collaboration and further development within the AI community.

  • Transparency in its architecture and training techniques is a bold move most companies haven’t done before.

  • There are concerns over data transmission to China and state-aligned outputs in sensitive areas raise (problems in China or related questions about politics or some censored events) questions about independence.

  • If DeepSeek’s approach becomes the industry norm, it could disrupt the AI hardware supply chain, reducing dependency on expensive GPUs.

  • Nvidia dropped 17%, total combined drop of $1 trillion across all major stocks.