Om Malik – What DeepSeek Means for Everyone (crazystupidtech.com)

DeepSeek, backed by the Chinese hedge fund High-Flyer, has unveiled an open-source AI reasoning model called DeepSeek R1, challenging the dominance of OpenAI by showcasing advanced capabilities achieved at a minimal cost. With a training cost of just $6 million—about 3% to 5% of what U.S. companies typically spend—DeepSeek’s approach highlights a shift away from reliance on massive data centers and expensive hardware. The model employs a novel “Mixture of Experts” (MoE) technique, activating only a small subset of its massive 671 billion parameters according to the task, significantly reducing computational requirements and costs. This development is significant as it opens up new possibilities for AI startups and researchers, encouraging a focus on efficiency and innovation rather than costly infrastructure. Notably, tech giants like AWS and Microsoft are swiftly integrating DeepSeek’s technology into their offerings. DeepSeek’s breakthrough not only impacts AI training costs but may also redefine inference methodologies, fostering a new wave of productization in AI. This shift emphasizes practicality over speculative advancements like AGI, positioning DeepSeek as a potential game-changer in the evolving AI landscape.