“In the coming months,” according to OpenAI, the company plans to release its first “open” language model since GPT-2.
This is following a feedback form that the business posted on its website on Monday. “What would you like to see in an open-weight model from OpenAI?” is one of the questions on the form that the company is asking “developers, researchers, and [members of] the broader community” to complete, as well as “What open models have you previously used?”
On its website, OpenAI stated, “We’re thrilled to work with developers, researchers, and the larger community to gather input and make this model as useful as possible. Please fill out the form below if you would like to participate in a feedback session with the OpenAI team.”
Also read: OpenAI introduces image generation for ChatGPT users powered by GPT-4o
OpenAI intends to hold developer gatherings to get input and eventually show off model prototypes. Over a few weeks, San Francisco will host the inaugural developer event, with sessions across Europe and Asia-Pacific to follow.
Rivals like the Chinese AI lab DeepSeek, which has taken an “open” approach to model launches, are putting more and more pressure on OpenAI. These “open” rivals, in contrast to OpenAI’s approach, make their models accessible to the AI community for testing and, occasionally, commercialization.
For certain groups, it has been an extremely effective tactic. Earlier in March, Meta, which has made significant investments in its Llama family of open AI models, said that Llama has amassed more than 1 billion downloads. In the meantime, DeepSeek has drawn interest from domestic investors and rapidly accumulated a sizable global user base.
When it comes to open-sourcing its technologies, OpenAI CEO Sam Altman stated in a recent Reddit Q&A that he believes the company has been on the wrong side of history.
Also Read: OpenAI intends to bill up to $20,000 a month for specialized AI “agents.”
Altman stated, “I believe we need to come up with a different open source strategy.” This viewpoint isn’t held by everyone at OpenAI, and it’s not our top priority right now […]. Although we will continue to build better models, our lead will be smaller than it was in prior years.
In a post on X on Monday afternoon, Altman went into further detail about OpenAI’s plans for an open model, stating that the company’s next open model will feature “reasoning” capabilities similar to those of its o3-mini.
As with any other model, Altman stated, “Before release, we will evaluate this model according [to] our preparedness framework.” And since we are aware that this model will be altered when it is released, we will put in more effort. […] We’re eager to see what developers create and how governments and big businesses use it when they would rather manage a model themselves.
Also Read: OpenAI releases GPT-4.5: the “largest, most knowledgeable” model
According to excerpts released over the weekend from a book that Wall Street Journal reporter Keach Hagey is writing, Altman lied to OpenAI officials about model safety checks before being briefly fired in November 2023.
