It has been more than a week since Deepseek has subverted the AI ​​world. Introduce its open models (trained in a small part of the power industry leader, all conducted shock waves inside the OpenAI. Employees not only claim to see that Deepseek has a "improper distillation" OpenAI model to create its own model, and the start -ups should start The company's success is that Wall Street questioned whether companies like OpenAi were overtaken.

"Deepseek R1 is the moment of AI's Putnik," Marc Andreessen wrote that Marc Andreessen is one of the most influential and provocative inventors on Silicon Valley on X.

In response, OPENAI is preparing to launch a new model today and before the timetable for its initial plan. This model O3-Mini will debut on API and Chat for the first time. Sources said it has a level 4 -level O1 reasoning. In other words, it is fast, cheap, clever, and aims to shatter Deepseek.

At this moment, OpenAI employees were excited. Without the company, there is a feeling-especially Deepseek's dominant position-OpenAI must become more effective or behind the latest competitors.

Part of the problem comes from the origin of OpenAI as the origin of non -profit research organizations, and then becomes a strong country seeking profit. Employees claim that the power struggle between the research and the product group has led to the cracks between the teams engaged in senior reasoning and people engaged in the chat. (Openai spokesman Niko Felix said that this is "incorrect" and pointed out that the leaders of these teams, KEVIN Weil, and chief researcher Mark Chen, "meet and work closely every week to maintain products and research priority matters ")

Some people inside Openai want the company to build a unified chat product, a model that can determine whether the problem is required for advanced reasoning. So far, this has not happened. Instead, the drop-down menu in ChatGPT prompts the user to decide whether to use GPT-4O ("most problems") or O1 ("use advanced reasoning").

Some staff claims that although the chat has brought the lion's share of OpenAI's income, O1 can get more attention and computing resources from leaders. "Leaders don't care about chatting," said a former employee who was engaged in (you guessed). "Everyone wants to work on O1 because it is sexy, but the code library is not established for experiments, so there is no motivation." The former employee asked to keep anonymous, because the agreement was not confidential.

Openai spent many years for strengthening learning, and fine -tuning eventually became a model called O1. (Strengthening learning is a process of training the AI ​​model with a fine and reward system.) Deepseek establishes an enhanced learning work created by OpenAI to create its advanced reasoning system, called R1. A former OPENAI researcher said: "They know that strengthening learning to apply in language models can work normally."

Another OPENAI researcher said: "Deepseek) is similar to what we do in Openai, but they use better data and cleaner stacks to do this."

OpenAI employees said that the study of entering O1 was carried out in the code library, called the "Berry" stack, which was established at a speed. An employee who directly understands this situation said: "Weighing torture-to achieve strict throughput."

For O1, these weighs are meaningful. Despite this, this is actually a huge experiment, but there is still a basic limit for code. They don't make much sense for chatting. The chat is used by millions of users. These users are based on different, more reliable stacks. When O1 was launched and became a product, the crack began to appear in the internal process of Openai. The employee explained: "Just like, 'Why do we do this in the experimental code library, shouldn't we do this in the main product research code library?" "There are major challenges inside."