OpenAI, release the operator AI tool for Chatgpt last week, showed that at least when it comes to demos, it was very ahead compared to its competitors like Google. However, the AI ​​news that shook the world was not Chatgpt, operator or the gigantic Stargate project announced last week. Deepseek AI has created great waves in the AI ​​world with a Chinese attempt release the R1 reasoning model of OpenAI’s Chatgpt O1.

What makes Deepseek different?

While the OpenAI O3 is announced, there is no surprising situation in this section, as other AI companies are expected to create competing systems for O1. However, the extraordinary side of Deepseek was that the Chinese company could access and examine by any company or developer by making its models open source. The more interesting part was the R1 research article, which claimed that Deepseek was educated on a much less cost of OpenAI’s O1.

The worldwide repercussion that Deepseek R1 training is possible with only 3 %to 5 %of the resources that OpenAI needed for similar progress with Chatgpt. On Monday, the stocks about artificial intelligence fell on Monday, and Deepseek rose to number 1, leaving the chatgpt behind in the App Store.

One of the problems in existing AI software is about the cost of development and use of the product. Developing advanced models like O1 can cost tens of million dollars. The process requires high -level graphics cards (GPU) that enable the necessary information processing power and energy expenditures.

Therefore, finished products like chatgpt O1 cannot be offered free of charge without restrictions. Companies such as OpenAI must meet costs and make profits. Therefore, the huge Stargate program of $ 500 billion, especially when the inevitable AI armament race between the US and China is considered, is a very important decision for AI development.

Despite the US embargo against China …

Considering the US sanctions that prevented access to the same senior chips and GPUs that made it possible to develop Chatgpt O1 products, Chatgpt, Gemini, Meta AI and Claude would not be expected to face significant competition from China.

That was one reason that Deepseek was so surprising. The Chinese initiative knew that he could not compete with OpenAI only relying on the hardware power. He could not reach the GPU, which was held by companies like OpenAI. Therefore, Deepseek has adopted a different approach for R1 and found ways to train a logic model without access to the same equipment.

In addition, Deepseek has made access to R1 much cheaper than OpenAI’s chatgpt. If you add the open -source nature of Deepseek models to all this, it is not difficult to predict why developers flocked to test the AI ​​of the Chinese company and why Deepseek has risen in the App Store.

According to a study, the Chinese venture used a reinforced learning (RL) instead of OpenAI’s Operated Fine Temporary Technology to train Chatgpt to produce faster and cheaper results. SFT is based on showing ways to solve problems by accessing the data to know what kind of answers to AI to give AI. RL relys on the AI ​​model, tries to find answers with the reward system and then provides feedback to AI.

RL allowed Deepseek to develop R1’s reasoning capabilities and overcome the lack of calculation. However, as Venturebeat pointed out, some SFT trainings were required in the early stages of R1 before moving to RL.

Success with only 50,000 Nvidia GPU

The fact that Deepseek has achieved this success with 50,000 Nvidia GPUs, which was taken before the US sanctions, leads to questioning that Western companies such as OpenAI, Google and Anthropic, which works with more than 500,000 GPUs, can make similar methods. Although Deepseek is based in China, there will be a reason for concern for some organizations and people, many people will prefer this cheaper service. Therefore, western AI companies may now be obliged to decrease their costs, and we can soon see more breakthroughs in the field of AI.