To judge his social summaries, the CEO of Openai Sam Altman is the very happy vehicle, as his company has achieved great success after another. The startup company that he participated in only in 2015 only Arise 40 billion dollars in rating 300 billion dollars, which is the largest financing round by a private technology company; It seems that everyone on the Internet is publishing Ghibli-style studio photos Perhaps the new GPT-4o photo generation model from Openai; Chatgpt now owns 500 million weekly users, up from 400 million users last month.
However, besides all this good news, Altman revealed on Monday that Openai is making what is very large in its strategy: in several months, Al -Taman saidOpenai will issue an open source model.
The first time that the company has issued a public model since the launch of the GPT-2 in 2019, apparently the opposite of the company’s transformation into closed models in recent years. Grant, the next model will not be 100 % open – as with other companies that offer “open” models, including Dead Mistral, Openai will not provide any access to data used to train the form. However, the user license will allow researchers, developers and other users to reach the basic code and “weights” of the new model (which determines how the information form is processed) for its use, modification or improvement.
Why transformation?
On its surface, it may seem a direct cause of embracing the open source in Openai from China, specifically, the start of startup Deepseek, which turned the AI scenario For the open source in January. But according to many people familiar with the artificial intelligence industry luck It occurs to a group of wider and more accurate factors, it is also possible to stimulate the changing of Altman’s heart on the open source. Since AI technology makes its way to companies, customers want the flexibility and transparency of open source models for many uses. Since the gap in performance between Openai and its competitors is narrowing, it became difficult for Openai to justify its 100 % closed approach – something that altman admitted in January when it was I confess This DEPSEK reduced Openai’s progress in artificial intelligence, and Openai was “on the wrong side of history” when it came to open the sources of its techniques.
Openai needs to be outside the models
Naveen Rao, Vice President of AI in Databrics, said that Openai’s move is about admitting that the scene of artificial intelligence is changing. The value is transmitted away from the same forms to applications or systems used by organizations to customize a model to meet their specific needs. While there are many situations that the company may want to use in the latest LLM, the open weight model will allow Openai to have a presence in scenarios where customers do not want to use Chatgpt, for example, or the company’s developer’s applications. For example, a financial company may not want to leave its customer data to leave its infrastructure and move to an external cloud, or the AI manufacturing company may want the factory devices not connected to the Internet.
He told me: “The open source is not some curiosity. It is a large part of the use of artificial intelligence.” “Openai wants to be part of this through their brand and models.”
Rowan Coran, chief analyst at Forrester Research, who focused on artificial intelligence, agreed that Openai’s return to the open source speaks to the increasing ecosystem of the spontaneous organization, from Openai, GoogleHuman, Amazon To Meta for Chinese and Dibsic, Mistal France, Canada’s laboratories and Israeli laboratory AI21 laboratories.
He said that many institutions companies are enthusiastic about the open source models-not only because of the accuracy or extent of the answer to the questions, but because they are flexible. He explained that the fact that it is portable is the key – which means that it can work on different cloud platforms or even in the company’s private data center, workstation, laptop or robot, instead of linking it to one provider.
Koran also explained that launching an open model can make Openai’s special services more attractive to its institutions. If Openai built a customer project and needs to run some of their businesses inside the company’s data center or even smaller models, for example, they cannot do this using Openai models such as 4O because those that are launched from the servers -based servers. He said: “This limits their ability to provide a comprehensive solution from the cloud to the edge,” whether it is a laptop, a smartphone, a robot or a self -driving car. Similar to what Google does with Gemini (the largest typical closed family) and Gemma (a smaller open model), Openai can have an open solution without having to consider the third -party open source models.
A difficult budget act
While Rao does not see an open source Openai model as a great reaction to Deepseek versions, “MEPSEK MOMENT” showed that Chinese startups are no longer late in the artificial intelligence race.
He said, “Many of us have already known this.” If Openai is not targeting the open source community now, he added: “You will lose a lot of influence, goodwill and societal innovation.”
In the past, Openai said that one of the reasons for their inability to launch open models is that Chinese companies will try to use their technology to improve their own models. In January, Openai issued a statement saying, “It is very important that we are working closely with the United States government to protect the best models of efforts made by opponents and competitors to take American technology.” In fact, although Deepseek did not issue the data that you used to train its R1 model, there are indications that may have used outputs from Openai’s O1 to start training the thinking capabilities in the model.
While Openai is now heading towards Open Source again, she apparently found herself trying to reconcile contradictory messages. Witnesses of the Openai Witness LinkedIn postOn Monday: “As for the US -led democratic intelligence, the CCP -led AI prevails, it is increasingly clear that we need to achieve a balance between open and closed models. The open source places strong tools in the hands of developers around the world, and expanding the extension of America.
“They are certainly talking about both sides,” Rao said, describing Openai’s messages as “really dangerous (to issue open models), but we need to benefit from the society that builds and has an impact.”
There is also an Openai’s commercial budget law: it cannot issue an open model competing with those paid. To target the developers of artificial intelligence with influences, Rao suggested that Openai would issue a big model – but it is not very large.
Throw the shade in the definition
If the strategic model is not to open the source, it is not only responding to Deepseek, it may be very good about throwing a shadow on a large open source competitor: Meta was set for the fourth repetition of the open source family, Lama, at the end of this month. Llama was significantly released with an open license except for services that include more than 700 million active users – Meant to reduce companies such as Openai Building on them.
“We will not do anything ridiculous like saying that you cannot use our open style if you have more than 700 million users per month,” Altman, Altman. Yesterday on x.
“Meta has become a record holder of Open Source Ai, at least in the West,” said Rao. “If they want to wrestling some effect on the ecological system, they must take a dead.”
However, Koran from Forster said that the mysterious Altman comments aside, there is no reason to believe that an open source Openai model will be more transparent – in terms of data or training methods, for example – any other open -commercial version of Meta or Mistral.
He said: “I expect it to be more mysterious and closed compared to other open models,” with much less transparency. “
This story was originally shown on Fortune.com
https://fortune.com/img-assets/wp-content/uploads/2025/03/GettyImages-2198379368-e1742566274761.jpg?resize=1200,600
Source link