DeepSeek: Open-source AI and the end of the monopoly of big tech companies

Why in the NEWS?

• Chinese AI Company DeepSeek is making waves around the world as its open-source AI model is challenging the monopoly of big tech companies and bringing about a significant shift towards new developments.

Key Points:

• Stock markets fell sharply, especially the tech-heavy NASDAQ, which fell by about 3%.
• This decline is believed to be due to the global attention being drawn by new AI models from Chinese AI start-up DeepSeek.

What will you read next in this topic?

1. The rise of DeepSeek AI:
2. Founding and Objective of DeepSeek:
3. Models of DeepSeek AI:
4. Features of DeepSeek-V3:
5. DeepSeek-R1 launched:
6. Comparison of DeepSeek with US companies:
7. Model creation at low cost:
8. DeepSeek's open-source policy:

The rise of DeepSeek AI:

• In the last few weeks, DeepSeek unveiled its AI models - DeepSeek-V3 and DeepSeek-R1, which are competing with OpenAI's most advanced models. 
• These models have surpassed ChatGPT as the most downloaded app on the App Store, which is a major achievement.
• DeepSeek-V3 and DeepSeek-R1 have demonstrated their power in the field of AI, and it is believed that these models have challenged OpenAI's AI models. 
• DeepSeek-V3 has proven its effectiveness by beating cutting-edge AI models like GPT-4 and Cloud 3.5 in benchmarks. 
• At the same time, DeepSeek-R1 has achieved a new position in the competition due to its thinking ability and affordability.
• These models have started a new revolution, leaving behind famous AI technology like ChatGPT, which shows a new direction in the development of AI. 
• DeepSeek aims not only to increase the potential of AI but also to contribute to the global community by making it open-source.

Founding and Objective of DeepSeek:

• DeepSeek is an AI (artificial intelligence) company based in Hangzhou, China, founded by Liang Wenfeng. 
• Wenfeng is a prominent entrepreneur and also the CEO of a hedge fund called High Flyer.
• Liang Wenfeng started working in the field of AI in 2019, and through his company High Flyer AI, he made significant contributions to research and development on AI.
• DeepSeek holds some patents related to High Flyer AI, which are used in training AI models. 
• These patents are helpful in ensuring the operation and progress of AI models, and it puts the company at the forefront of AI development.

Models of DeepSeek AI:

• The DeepSeek-V3 and DeepSeek-R1 models are the main competitors of OpenAI's O1 and O3 models. 
• These models have challenged OpenAI's leading models from a technical point of view and have proven their excellence.
• DeepSeek-V3 has been trained at a cost of only $5 million, which is extremely low compared to the investment made by other AI companies. 
• In comparison, OpenAI and other companies invest hundreds of millions of dollars in their AI models. 
• Thus, DeepSeek has developed highly effective and efficient AI models with limited resources.

Features of DeepSeek-V3:

• The architecture of DeepSeek-V3 is based on Mixers-of-Experts (MOE), in which multiple expert models work together.
• DeepSeek-V3 has been trained on 14.8 trillion tokens, giving it a better understanding of language and task-specific capabilities.
• This model has outperformed GPT-4 and Cloud-3.5 in benchmarks.
• DeepSeek-V3 used Multi-Head Latent Attention (MLA) technology, which has reduced the cost of training and deployment.

DeepSeek-R1 launched:

• DeepSeek-R1 was unveiled, which comes with test-time compute capability.
• R1 outperforms OpenAI's Frontier Model in tasks like math, coding and general knowledge and is also affordable.
• The key feature of R1 is that it is open-source, allowing it to be used by anyone.
• The R1 model also clearly shows the thinking process, while OpenAI-o1 takes time for its output.

Comparison of DeepSeek with US companies:

• DeepSeek has developed cutting-edge AI models with limited resources, giving tough competition to big companies like OpenAI. 
• DeepSeek's V3 and R1 models have proven that effective and competitive AI technologies can be created even with limited resources. 
• This is a big message for companies that are not getting the expected results despite huge investments.
• The release of R1 has raised the question of whether such huge expenditure is really needed in the AI ​​industry. 
• DeepSeek proved that effective and competitive models can be created even with less investment, raising serious questions about the current approach of the industry. 
• Do companies need such huge investments, or can smaller, cost-effective methods also give better results? 
  • This question has sparked a new debate on the future direction of the AI ​​industry.

Model creation at low cost:

• Training AI models requires heavy investment, but DeepSeek has reduced the cost of training by using older GPUs like NVIDIA H800.
• DeepSeek worked on the NVIDIA H800, while US companies used advanced GPUs like the NVIDIA H100.
• NVIDIA had imposed restrictions on the sale of A100 and H100 chips in 2022, which led DeepSeek to use low-cost A800 chips.
• DeepSeek's engineers ensured high performance despite GPU limitations with the help of low-level code optimizations.

DeepSeek's open-source policy:

• DeepSeek made its models open-source, giving developers from around the world the opportunity to work on and improve these models. 
• This move has opened a new path for AI development, where more and more people can use these models, modify and adapt them for different uses. 
• This is promoting collaboration and innovation in the AI ​​field.
• This open-source approach has created a stir in the AI ​​community, as it not only gives developers the freedom to work openly, but it is also being seen as a way to reduce costs in the AI ​​industry. 
• Companies can now acquire cutting-edge AI technologies at a lower cost, which may pose a challenge for traditional large-investment companies.