DeepSeek Breaks the Illusion Behind the Development of Artificial Intelligence Technology



If you follow the current developments in the world of technology, you must have at least heard of DeepSeek, a new AI chatbot model introduced by an investment company called High-Flyer from China.


Launched just about a week ago, DeepSeek R1 is seen to have conquered the AI ​​application market on Android and iOS, surpassing ChatGPT as the most popular AI chatbot application today.


What is more interesting is that compared to other chatbot models such as ChatGPT, Gemini and Perplexity AI developed by technology giants such as OpenAI, Google and Perplexity, DeepSeek was developed with an open source model and as a part-time project using technology and hardware that is around 2-3 years old.


DeepSeek is seen to have become very popular at first because the answers given by this application are seen to be more natural and up-to-date, like me myself doing a search on the web arena. The hallucination rate on DeepSeek is also seen to be lower because if it is asked something that it does not have an answer to through machine learning or existing searches, it will say that it does not know.


Technology Development

In the past two or three days, there has been a lot of talk about the development of this AI application. As we said before, it was developed by a quantitative investment company called High-Flyer from China.


In short, quantitative investment uses company financial data, computer algorithms and complex mathematical models to make decisions about which stocks are suitable for investment.


High-Flyer was developed by Liang Wenfeng in 2015, using AI technology to allow them to make predictions about stock market trends, and thus about which stocks are worth investing in.


For this, Liang Wenfeng is said to have purchased around 10 thousand NVIDIA H100 graphics cards in 2021, before the economic sanctions imposed by US President Joe Biden on Chinese companies, especially related to graphics cards with high arithmetic and AI processing capabilities.


In 2023, and more NVIDIA H100 and NVIDIA H800 graphics cards, he started DeepSeek as a “side project” to his investment company. Wenfeng has been seen spending a lot of time developing this company, and the DeepSeek language model we see today.


There has also been a lot of talk about how a company with a bunch of “old” graphics cards and an investment of only $5 million can compete with companies that are valued at several billion dollars and have been operating for much longer.


DeepSeek has also just launched their first multi-modal language model, Janus-Pro-7B, which can be used to generate images. It is said to be able to compete with the latest models Dall-E by OpenAI, and Stable Diffusion, especially through benchmarking software


The biggest advantage for DeepSeek, in my opinion, is that it is an open source language model. If you are an AI technology developer, this means that you can download the DeepSeek or Janus models from their GitHub page and start developing your own AI system or service.


If you want to learn more about how DeepSeek V3 and DeepSeek R1 were developed, you can also read the technical documents about them on that page. It's quite interesting, especially when it shows that the development of the AI ​​system's algorithm is developed by them themselves without copying the homework of other companies.


Application Usage

For ordinary users, for now they can use the DeepSeek chatbot application which may not seem that interesting. ChatGPT is still seen to come up with deeper contextual details when asking the same question, but for an application that was only introduced a few days ago,


The advantages of DeepSeek can be shown when you leverage this artificial intelligence system for productivity purposes. If you are a programmer, there are several language models to choose from, including DeepSeek R1, DeepSeek V3 and DeepSeek Coder which can currently benefit programmers who want to use AI for various reasons. For the DeepSeek chatbot application that can be downloaded from the app store, it uses the DeepSeek R1 language model.


If you are diligent in studying X/Twitter, you can find many examples where DeepSeek is used to simplify various mundane processes, especially for technical users.


As a writer, I wouldn't normally use artificial intelligence technology for this purpose, but DeepSeek also seems to have shown that writing is more natural.


As a writer, I wouldn’t normally use AI technology for this purpose, but DeepSeek also seems to have shown that its writing is more natural than what is shown using ChatGPT or Claude AI. Unfortunately, for now, the feature is only available in Mandarin and English.


If you want to build your own AI cluster at home or in the office (on-premise), DeepSeek may be the language model for you. You can download this AI system, install it on a computer like a Mac Mini in a superset and train it with content of your choice to build an AI specifically for your own use.


For chatbot applications, user registration is currently limited to only two ways, namely using a Google account or a phone number from China. For now, the response from DeepSeek is seen as slow because it has a very large number of users, and DeepSeek itself says that they have experienced


The Discussion Behind DeepSeek

As usual, when talking about technology coming out of China, many people have all kinds of questions, especially regarding how a company that is still unknown can develop a language model for such a small cost, when compared to giants like OpenAI, Microsoft, Google and Meta.


When I read various discussions and articles about DeepSeek, what I found was that the advantage of this language model is that their developers have optimized the H100 and H800 graphics cards used through their own CUDA codes.


We are sure that a large part of this is due to the trade restrictions imposed by the US government that do not allow them to buy the latest and most powerful semiconductor chips, but this is seen as a step for them to take advantage of what they currently have.


For AI technology developers in the western world such as the United States, they are seen using the latest graphics cards such as NVIDIA B200 and rarely optimize existing hardware to reduce operating costs. The modus operandi for these companies has been to use the most powerful hardware and components to advance the propaganda of AI development.


This can also be confirmed by reports showing that companies such as Google, Meta and Microsoft will use nuclear power plants to power their data centers.


Because of this, we can also see AI companies starting to spend billions of dollars to increase their processing capacity, and this makes the supply of these components difficult to buy for other uses and research.


OpenAI, Softbank and a number of other companies have also just announced the Stargate Project, an AI data center that is expected to cost $500 billion over the next four years, and is estimated to consume $100 billion this year alone.


Unfortunately, with the launch of DeepSeek, many have begun to question whether anyone would need that much computing power to develop efficient and powerful artificial intelligence.


This has caused many technology and semiconductor chip companies, most notably NVIDIA, to experience a 17 percent decline in stock value and a $600 million loss in company value.


It is also interesting to see NVIDIA’s value decline because DeepSeek is powered by their graphics cards, even though it is an older model that launched in 2023.


Sam Altman, founder and CEO of OpenAI, says that DeepSeek is a very good language-based model, especially with its very low development and operational costs.


However, in his opinion, OpenAI will continue to introduce ever-improving models with ChatGPT and Dall-E, and sees DeepSeek as a formidable competitor to them.


Jensen Huang, CEO of NVIDIA, also showed that his net worth fell by around $18-$20 billion, but also said that DeepSeek is a very powerful AI model because it is very efficient in using existing processing power, and does not need to rely on the latest components.


The company also sees this as a new opportunity because technically, this new AI technology is still powered by NVIDIA's own AI processors.


Conclusion

In just a few days, DeepSeek has managed to show that the development of artificial intelligence systems can be done without Titanic-scale investments. We are sure that DeepSeek R1 and Janus are not the only two products they will launch, and they will launch more AI algorithms in the future.

If you want to read more about DeepSeek, there are some tweets on X/Twitter, and an article by Fortune that I've read about this company and its founder, Liang Wenfeng.

Previous Post Next Post

Contact Form