Quality Alchemist (品質煉金術師): ZJU Public Online Course for DeepSeek Series

Recently, DeepSeek has become popular all over the world. Its system-level collaborative innovation of algorithms, models, and systems has achieved results that bringing infinite imagination to the technological breakthroughs and challenges towards general artificial intelligence (AGI). In order to systematically analyze DeepSeek's innovative exploration, Zhejiang University's artificial intelligence research team launched a series of online open courses on DeepSeek started on 17^th Feb 2025. The theme named “Decoding DeepSeek's "breaking the circle" power and exploring the future world of AI technology” (解码DeepSeek“破圈”之力，探索人工智能技术明日世界！)

The first speaker was Prof. Wu Fei (吴飞教授) and his topic entitled “Review Three Major AI Principles and Strengthening General Education” (回望AI三大主义与加强通识教育). His talk separated into three parts including AI three major principles, from ChatGPT to DeepSeek and AI general education.

Three Major AI Principles are Symbolism, Connectionism, and Behaviorism. Prof. Wu explained one by one. Recently, connectionism is more popular through machine learning and deep learning.

And then he introduced what is Generative Pretraining Transformer (GPT). Prof. Wu pointed out three key elements for AI success and they are Algorithm, Big Data and Computing Power.

He quoted Rich Sutton article “The Bitter Lesson in AI” that "The biggest lesson that can be read from 70 years of AI research is that general methods that leverage computation are ultimately the most effective, and by a large margin. The ultimate reason for this is Moore's law".

After that Prof. Wu introduced DeepSeek that breakthrough the ceiling of big computation power!

And then DeepSeek V3 was discussed including its Mixture-of-Experts (MoE) language model.

For DeepSeek R1, he briefed that model trained via large-scale reinforcement learning (RL) and demonstrated that the reasoning patterns of larger models can be distilled into smaller models, resulting in better performance.

Finally, Professor Wu Fei introduced Zhejiang University’s AI general education course, which is a compulsory course.

He also mentioned the importance of developing AI teaching materials.

Lastly, he used a poem to conclude his talk that is “Artificial Intelligence, Education First, Industry-University Collaboration, Leading Innovation”(人工智能、教育先行、產學協作、引領創新).

Prof. Chen Wenzhi (陈文智) was the second speaker and his talk named “Chatting or Acting?— DeepSeek’s breakthrough and the future prospects of “Zhejiang University Professor” (Chatting or Acting？—DeepSeek的突破边界与“浙大先生”的未来图景)

First, he briefly analyzed DeepSeek’s technological breakthroughs and explain the Mixture-of-Experts (MoE) structure.

And then introduced the Multi-head Latent Attention (MLA) and explained how to increase the efficiency.

He then mentioned the optimization engineering of DeepSeek.

Deepseek’s low-precision training and cost control were discussed.

After that Prof. Chen pointed out three turning point including Strategy, Computation Power and Data. We need to gain new breakthroughs through algorithm innovation.

Finally, he introduced three ways to optimize performance of large models and they are Model Distillation, Model Fune Tuning and Retrieval-Augmented Generation (RAG)