AI Talks

What’s the Big Deal with DeepSeek?

DeepSeek’s AI models, particularly the R1 (reasoner) and V3 (general-purpose chat), are achieving near state-of-the-art performance, rivaling those of tech giants like OpenAI, Google, and Meta. What’s truly shocking is that they’ve done it at a fraction of the cost. While major US companies are pouring billions into AI development, DeepSeek reportedly developed its latest model for a mere $6 million. This cost-effectiveness is sending ripples through the industry.

Here’s how DeepSeek is making waves:

• Efficient Algorithms: DeepSeek uses innovative algorithms that allow their models to perform at a high level while using less computing resources. For example, the DeepSeek-R1 model uses a “Mixture-of-Experts” (MoE) architecture, activating only a subset of the model’s parameters for each task, instead of the entire network. This means the model isn’t wasting energy and resources processing information that is not relevant.
• Open-Source Approach: Unlike many of its competitors, DeepSeek has embraced an open-source model, meaning they’ve made their algorithms public. This encourages collaboration and helps accelerate innovation, as others can build upon their work.
• Innovative Training: DeepSeek prioritizes pure Reinforcement Learning (RL) in its training process without relying heavily on supervised fine-tuning (SFT), which allows the AI to learn autonomously and adapt to complex situations.
• Impressive Performance : DeepSeek-R1 has proven its abilities in various performance benchmarks:
  • MATH-500: For this benchmark, evaluating math skills at a strong high school level, DeepSeek achieved an impressive score of 97.3%, slightly better than OpenAI-o1-1217 (96.4%) and a significant lead over other models.
  • American Invitational Mathematics Examination (AIME) 2024 : For this benchmark, measuring advanced multi-step mathematical reasoning skills, DeepSeek slightly surpassed OpenAI-o1-1217 (79.8% vs 79.2%).
  • MMLU: Showcasing its breadth of knowledge, DeepSeek scored 90.8% on this demanding benchmark, a close second to OpenAI-o1-1217’s 91.8%.

(source : deepseek, https://arxiv.org/html/2501.12948v1#S5)

Affordability: The output cost for DeepSeek-R1 is roughly 27.4 times cheaper than OpenAI’s o1, making advanced AI much more affordable.

A Shake-Up for the AI Ecosystem

DeepSeek’s emergence isn’t just about a new AI model; it has significant implications for the entire AI landscape:

• Challenging the “Bigger is Better” Mentality: DeepSeek’s cost-effective approach questions the idea that you need massive resources to create powerful AI.
• Accelerating AI Adoption: Lower costs and open-source models mean that smaller companies and developers can now use AI capabilities that were previously out of reach.
• Potential Shift in Infrastructure: DeepSeek’s efficient models could mean less reliance on large data centers and expensive AI chips.
• Geopolitical Implications :DeepSeek’s success with less advanced chips (like Nvidia’s H800 GPUs) challenges US AI dominance and its hardware restriction strategy, potentially boosting China’s AI influence.

Impact on AI Companies

DeepSeek’s disruptive arrival is impacting various categories of AI companies, particularly:

Foundation Model Developers: DeepSeek’s approach could pressure these developers to innovate and reduce costs, intensifying competition. Companies will need to differentiate themselves through efficiency and affordability, potentially making profitability more challenging.

Chipmakers: DeepSeek’s success raises questions about future demand for high-end AI chips, potentially affecting companies like Nvidia that have profited from the growing need for such chips.

Cloud Service Providers: DeepSeek’s models and approach may reduce reliance on cloud-based AI services, potentially impacting providers like AWS, Google Cloud, and Azure. Its competitive pricing and open-source approach could draw clients away from these resource-intensive offerings.

AI Application Builders: DeepSeek’s open-source models and competitive API pricing offer potential advantages, especially for smaller players, by providing more cost-effective access to capable AI. This can foster innovation and a wider range of AI applications.

What Do the Experts Say?

Experts have mixed reactions to DeepSeek. Some see it as a major shift in the AI landscape, while others believe its impact may be overstated, arguing that major US companies are unlikely to rely on a Chinese startup for their AI infrastructure. There are also concerns about the potential security risks and misuse of the technology. Some believe that DeepSeek’s impact could democratize access to AI while also shifting the focus of hardware priorities towards more specialized and energy-efficient AI chips.

Looking Forward

DeepSeek’s emergence has undeniably shaken up the AI world. Its focus on cost-effectiveness, open-source collaboration, and efficiency is challenging established norms. While the future is uncertain, DeepSeek seems to mark a pivotal moment in AI’s trajectory. It is definitely a company to watch, as its actions could continue to shape the future of the field.

References

What is DeepSeek, and why is it causing Nvidia and other stocks to slump? – CBS News, accessed January 28, 2025,
https://www.cbsnews.com/news/what-is-deepseek-ai-china-stock-nvidia-nvda-asml/

DeepSeek-R1: The Open-Source AI Revolutionizing Technology – Medium, accessed January 28, 2025,
https://medium.com/@harsh.vardhan7695/deepseek-r1-the-open-source-ai-revolutionizing-technology-1b0914a162da

Deepseek Ripple Effect on AI and Supply Chain – EPS News, accessed January 28, 2025,
https://epsnews.com/2025/01/28/deepseek-ripple-effect-on-ai-and-supply-chains/

DeepSeek: Could this be a decisive shift in the Generative AI Landscape? – Medium, accessed January 28, 2025,
https://medium.com/@rahulsandil/deepseek-could-this-be-a-decisive-shift-in-the-generative-ai-landscape-6074e6f5fc64

Why DeepSeek AI Has Big Tech on Edge – AiThority, accessed January 28, 2025,
https://aithority.com/machine-learning/why-deepseek-ai-has-big-tech-on-edge-and-what-it-all-means/

How Did DeepSeek Train Its AI Model On A Lot Less – And Crippled – Hardware? – The Next Platform, accessed January 28, 2025,
https://www.nextplatform.com/2025/01/27/how-did-deepseek-train-its-ai-model-on-a-lot-less-and-crippled-hardware/

DeepSeek R1: Pioneering Open-Source ‘Thinking Model’ and Its Impact on the LLM Landscape – United Nations University, accessed January 28, 2025,
https://c3.unu.edu/blog/deepseek-r1-pioneering-open-source-thinking-model-and-its-impact-on-the-llm-landscape

The Democratization of AI: How DeepSeek Could Reshape the AI Race – Medium, accessed January 28, 2025,
https://kaiiyer.medium.com/the-democratization-of-ai-how-deepseek-could-reshape-the-ai-race-475b538b8931

DeepSeek-AI. (2025). DeepSeek-R1: Incentivizing Reasoning Capability in LLMs via Reinforcement Learning [arXiv preprint arXiv:2501.12948v1], accessed January 28, 2025,
https://arxiv.org/html/2501.12948v1

This post was researched and written with the assistance of various AI-based tools.