OpenAI CEO admits defeat: Advantages weakened by DeepSeek, will re-formulate open source strategy

In a Reddit Q&A on Friday, OpenAI CEO Sam Altman, his research team and executives answered a series of pointed questions. The embattled AI giant is facing multiple challenges: dealing with the technological catch-up of Chinese competitor DeepSeek (which OpenAI accuses of stealing intellectual property), maintaining political relations with Washington, advancing the $10 billion Stargate data center project, and preparing for the largest financing round in history.

Altman said that the rise of DeepSeek has weakened OpenAI's technological lead and admitted that the company was "on the wrong side of history" in its open source strategy. Although OpenAI has open-sourced some models, it mainly adopts a closed-source development model. " We need to develop a different open source strategy ," Altman said, "although not all colleagues agree with this view, and it is not a top priority at the moment... We will continue to launch better models in the future, but the lead will be smaller than in previous years."

Chief Product Officer Kevin Weil added that the company is considering open-sourcing old models that are not cutting-edge : "We will definitely move forward with this work. " For the reasoning models such as o3-mini released today, Altman said that inspired by DeepSeek, OpenAI may enhance the transparency of the model's "thinking process" - its current model hides the reasoning chain to prevent competitors from stealing training data, while DeepSeek's R1 model displays the complete thinking chain. Weil revealed: "We are about to significantly increase the transparency of the display, and we need to find a balance between competition protection and user needs. "

Regarding the rumor of ChatGPT price increase, Altman denied it, saying that he hopes to "reduce the cost of use as much as possible." He previously revealed that the $200 per month ChatGPT Pro service is in a loss-making state. When talking about computing power demand, Weir emphasized that the expansion of computing resources is still the key to improving model performance, which explains the necessity of Stargate's super-large-scale data center, and the growth of users also increases the pressure on computing power.

Regarding the "technological singularity" that may be caused by AI's self-iteration, Altman changed his words and said that "the possibility of a rapid arrival is higher than previously thought." When asked whether the model would be used for nuclear weapons development (OpenAI recently cooperated with the US National Laboratory to conduct nuclear defense research), Weir expressed his trust in government partners: "These scientists are well aware of the limits of AI capabilities and will never rashly adopt unverified model outputs."

In terms of technical routes, OpenAI revealed that the next-generation inference model o3 will be released "within weeks to months"; there is no timetable for the flagship model GPT-5; and the successor to the image generation model DALL-E 3 is under development. "It's worth looking forward to," Weir commented on the new generation of the Vincent graph model.