To cover a day of product release, the researchers, engineers and Openai executives answered questions to a wide Reddit AMA Friday.
Openai is in a precarious position. He faces the perception that he is transferred to the AI race to Chinese companies such as Deepseek, which Openai claims to have stolen his IP. Chatgpt manufacturer has been trying Touch his relationship with Washington And at the same time they are seeking an ambitious Data Center project, and reportedly set a basis for one of the largest rounds of funding in history.
Altman admitted that Deepseek has reduced Openai’s lead in AI and said that he believed that Openai was “by the wrong side of the story” when it was about to open the supply of his technologies. While Openai has models of open origin in the past, the company generally favored a privately owned approach to the development of closed sources.
“[I personally think we need to] Understand a different open source strategy, “Altman said. [going forward]But we will maintain less than one lead than we did in previous years. ”
In a follow-up answer, Kevin Weil, head of Openai’s products, said Openai is considering open supply of older models that are no longer a state-of-the-art. “We will definitely think about doing more of them,” he said, without going in more details.
In addition to the exhortation of Openai to review its liberation philosophy, Altman said that Deepseek has pushed the company to reveal more about how so-called reasoning models, such as the O3-MINI model released today, show the “thinking process ”. Currently, Openai models hide their reasoning, a strategy aimed at preventing competitors to scratch training data for their own models. On the contrary, the Deepseek, R1 reasoning model shows its full chain of thinking.
‘We are working to show a bunch more than we show today – [showing the model thought process] It will be very soon, “Weil added.” TBD to everyone – the appearance of all the thought chain leads to competitive distillation, but we also know that people (at least power users) want it, so we will find the right way to balance it ”.
Altman and Weil have tried to dismantle rumors that Chatgpt, the Chatbot platform through which Openai starts many of its models, will increase in the future. Altman said he would like to do the “cheaper” chatgpt over time, if possible.
Altman has previously said that Openai lost money in the most expensive Chatgpt project, Chatgpt Pro, which costs $ 200 a month.
In a somewhat relevant thread, Weil said that Openai continues to see evidence that the most computational power leads to “better” and most chief models. This is largely what requires projects such as Stargate, the Openai recently announced by the Massive Data Center Project, Weil said. Serving a growing users base also supplies computational demand within Openai, he continued.
Asked about the retrospective self-improvement that can be activated by these powerful models, Altman said he believes that a “quick take-off” is more reasonable than he once believed. Retrospective self-improvement is a process where an AI system could improve its own intelligence and potential without human inflow.
Of course, it is worth noting that Altman is well known for excessive promotion. It was not long before reduced the Openai bar for AGI.
A Reddit user asked if the Openai models, which improve self-entertainment or not, would be used to develop devastating weapons-specific nuclear weapons. This week, Openai has announced a partnership with the US government to give its models to US national laboratories partly for nuclear defense research.
Weil said he trusted the government.
“I have met these scientists and are AI experts except world -class researchers,” he said. “They understand the strength and boundaries of the models and I don’t think there is only a model output of a nuclear calculation. They are smart and based on evidence and do a lot of experimentation and data to validate all their work. ”
Openai’s team has been asked several questions about a more technical character, such as when Openai’s next model O3 will be released (“More than a few weeks, less than a few months,” Altman said). When the next model of the company’s “Non-Connection” company, GPT-5, may land (“you still don’t have a timetable,” Altman said). And when Openai can reveal a Dall-E 3 successor, the company’s image creation model. The Dall-E 3, released about two years ago, has taken a long time on the tooth. Image production technology has improved with jumps from Dall-e 3 debut and model is Is no longer competitive in various reference tests.
“Yes! We are working on it,” Weil said of a Dall-E 3 watch. “And I think it will be worth the wait.”