Lecture Notes on DeepSeek AI Models

Introduction to DeepSeek

• DeepSeek is a startup based in China.
• Made headlines by overtaking OpenAI's spot on the US Apple App Store as the most downloaded free app.
• Known for releasing an open-source model that rivals industry-leading models at a fraction of the cost.

Key Model: DeepSeek R1

• DeepSeek R1: A reasoning model, indicated by "R" in the name.
• Competes with OpenAI's reasoning model, O1.
• Excels in AI benchmarks for math and coding tasks.
• Approximately 96% cheaper to run compared to O1.

Features of Reasoning Models

• They solve complex problems by breaking them into steps.
• Utilize a process called "chain of thought" for step-by-step analysis.
• Differ from past models by showing the reasoning process.

Evolution of DeepSeek Models

DeepSeek Model Timeline

1. DeepSeek V1 (Jan 2024)
   • 67 billion parameters.
   • Traditional transformer model.
2. DeepSeek V2 (June 2024)
   • 236 billion parameters.
   • Featured multi-headed attention and DeepSeek mixture of experts.
3. DeepSeek V3 (Dec 2024)
   • 671 billion parameters.
   • Introduced reinforcement learning and load balancing across GPUs.
4. DeepSeek R1-Zero (Jan 2025)
   • First reasoning model using exclusively reinforcement learning.
5. DeepSeek R1
   • Combines reinforcement learning with supervised fine-tuning.

Distilled Models

• Distilled Models: Smaller student models derived from larger teacher models.
• Serve as model compression or translation between architectures.

Efficiency of DeepSeek

• Utilizes fewer specialized Nvidia chips compared to competitors.
• Example: Only 2,000 GPUs for DeepSeek V3 while Meta used over 100,000 for Llama 4.

Architectural Efficiencies

• Mixture of Experts (MoE) architecture is resource-efficient:
  • Divides the model into sub-networks (experts).
  • Activates only necessary experts for tasks.
  • Reduces computational costs and speeds up performance.

Reinforcement Learning

• Rewards the model for correct answers, allowing it to discover its reasoning pathway.

Conclusion

• DeepSeek R1 matches industry-leading models in performance at lower costs.
• Reflects exciting progress in AI reasoning models.