A new emerging challenge threatens the colossal expenditures of Mega Cap firms and America’s supremacy in the AI race.
Deep Seek is an entity worth noting. It is an innovative open-source AI model. It surpasses the latest open and meta models on crucial benchmarks. Remarkably, it was developed at a fraction of the usual cost. This model was trained by a Chinese research laboratory. They utilized Invidia H 800s, a more cost-effective version of the h100 chips. These chips are tailored for specific markets like China. Upon initial testing, Deep Seek appears and behaves similarly to Open AI’s Chat GBT.
Based on technical evaluations measuring coding performance, Deep Seek outperformed other cutting-edge models, including Meta’s Llama 3.1 and Open AI’s GBT 4.0. Notably, Andre Kaparthy, a founding member of Open AI, acknowledged Deep Seek’s accomplishments, highlighting its cost-effectiveness. The budget allocated for Deep Seek’s development, a mere $5.5 million, pales in comparison. The expenditures for Meta’s latest Llama model and other high-end models cost hundreds of millions or even billions of dollars. This poses a crucial question for investors amidst the evolving AI landscape: is investing heavily in training frontier models still a wise decision?
DeepSeek V3 surpasses its competitors on several benchmarks, showcasing its exceptional performance and efficiency. Here are some key areas where DeepSeek V3 stands out:
- Competitive Programming: DeepSeek V3 outperforms top models like Meta’s Llama 3.1 405B, OpenAI’s GPT-4o, and Alibaba’s Qwen 2.5 72B on Codeforces benchmarks
- Aider Polyglot Testing: DeepSeek V3 achieves the 2nd spot on the leaderboard, demonstrating its ability to generate new code that seamlessly integrates with existing projects .
- Reasoning Capabilities: DeepSeek V3 scores 88.5 on the MMLU benchmark, outperforming Llama3.1, Qwen2.5, and Claude-3.5 Sonnet .
- Inference Speed: DeepSeek V3 processes 60 tokens per second, three times faster than its predecessor, DeepSeek V2 ..
- Cost Efficiency: DeepSeek V3 was trained on a budget of $5.5 million, significantly lower than the $100 million spent on training GPT-4 .
These impressive benchmarks demonstrate DeepSeek V3’s exceptional performance, efficiency, and cost-effectiveness, making it a top contender in the AI landscape.
Tech giants such as Microsoft, Google, Amazon, Meta, and Open AI have prioritized developing advanced AI infrastructure. Their goal is to train increasingly sophisticated models. However, Deep Seek’s rapid progress with minimal resources challenges the status quo. Despite using simplified GPUs and a modest budget, Deep Seek achieved remarkable results in just two months. Notably, this achievement originates from China, a nation often regarded as a significant competitor to US-led AI dominance.
The implications of Deep Seek’s emergence are substantial and will reverberate throughout the AI community. As the AI sector evolves and technological advancements slow down, the commoditization of models becomes increasingly prevalent. This case exemplifies how existing models can be leveraged to create competitive solutions, raising questions about the advantages of open versus closed-source models.
the core competency of the Chinese in AI development lies in their ability to innovate with limited resources. They utilize existing technologies to create impactful solutions. This narrative is expected to unfold further in 2025, with significant geopolitical implications. Thank you for your insights, Josa.
