Debunking the Hype: DeepSeek's True AI Innovations

The Bernstein report dispels myths about DeepSeek's AI models, arguing they were not built for merely USD 5 million. Although innovative, DeepSeek's models can't be seen as miracles, demanding significant computing resources and costs. Panic over claims they're rivaling OpenAI is said to be exaggerated.

Social media and stock markets are abuzz with DeepSeek's purported achievements in AI development. However, a Bernstein report has clarified that claims about DeepSeek rivaling OpenAI for just USD 5 million are misleading.

The report highlights DeepSeek's two main AI families: 'DeepSeek-V3' and 'DeepSeek R1'. The V3 model, using Mixture-of-Experts architecture, efficiently handles 671 billion parameters, thanks to innovative techniques like Multi-Head Latent Attention and FP8 computation. Despite its performance efficiency, the USD 5 million cost estimate doesn't factor in comprehensive research and development expenses.

The R1 model enhances reasoning using Reinforcement Learning, showcasing remarkable performance in reasoning tasks against OpenAI. Yet, the Bernstein report suggests development resources for the R1 were significant. Ultimately, DeepSeek's pioneering models are noteworthy but the exaggerated claims of low-cost rival development aren't fully accurate, underlining that while achievements are significant, the buzz is overhyped.

(With inputs from agencies.)