In order to capture Openai researchers, engineers, and Openai CEOs, Openai researchers, engineers, and executives, they answered questions on a wide range of Reddit AMAs on Friday.
Openai is a little unstable. In AI racing for Chinese companies, such as DeepSeek, which Openai claims to steal the IP, is fighting with the recognition that it is a concession in the AI race. Chatgpt Maker is trying to strengthen his relationship with Washington, at the same time pursuing an ambitious data center project, and also has one of the largest financing rounds in history.
Altman acknowledged that DeepSeek has reduced the Openai lead in AI, and believes that Openai is “wrong in history.” Openai has an open source model in the past, but the company has generally supported its own closing source development approach.
“I know (I personally think it’s necessary),” says Altman, “said Altman. “Not all Openai share this view, not the current top priority. Creating a better model (in the future) will maintain the lead than the previous year.”
In a follow -up reply, Kevin Weil, the chief of Openai, stated that Openai is considering old Open Source models, not the state -of -the -art. “I can definitely think about doing this more,” he said.
Not only encourages Openai to reconsider the release philosophy, but Altman is more clear that Deepseek has shown the so-called “thinking process”, such as the O3-mini model released today. He stated that he urged him to do. Currently, Openai models are hidden in inferences that are strategies to prevent competitors from reducing their own model training data. In contrast, R1, a DeepSeek’s inference model, has a perfect idea.
“We are working on showing more things than today (indicating the model thinking process),” said Weil. “All TBD -indicating the chain of all thoughts leads to competitive distillation, but we know that (at least power users) want it, so we can balance it. I will find.
Altman and Weil tried to dispel rumors that Chatgpt, chatbot platform, which opens many of the models, will raise prices in the future. Altman said that if it could be executed, Chatgpt would be “cheaper” over time.
Altman earlier stated that Openai had lost money at Chatgpt Pro, the most expensive Chatgpt plan.
With a somewhat related thread, Weil stated that Openai continues to see evidence that more computational power will lead to a “better” performance model. This is the majority of the needs for large -scale data center projects recently announced by Stargate and Openai, and weil says. Providing services to growing user -based services also provide fuel to computing demand in Openai, and he continued.
When asked about recursive self -improvement, which could be effective with these powerful models, Altman said that “fast takeoff” was more plausible than once believed. Relatric self -improvement is a process that allows AI systems to improve their own intelligence and abilities without human input.
Of course, it is noteworthy that Altman is famous for excessive things. It wasn’t a while ago that he lowered the Openai bar for AGI.
One Reddit user asked if the Openai model would be used to develop destructive weapons, especially nuclear weapons, regardless of self -improvement. This week, Openai announced a partnership with the US government and provided a model to the National Institute for Nuclear Defense.
Wils said they trusted the government.
“I have come to know these scientists. They are AI experts in addition to world -class researchers,” he said. “They understand the power and limits of the models. I don’t think there is a possibility that only Yoro will be in the nuclear calculation of the model output. They are smart, evidence, and to verify all work. Many experiments and data work. “
The Openai team asked some questions about the more technical properties, as when O3, the next reasoning model of Openai, (“more than a few weeks, less than a few months,” Altman. I said that the company’s next flagship “non-rational” model, “I don’t have a timeline”. When the successor of E 3 may be announced, the company’s image-generated model is quite long in the teeth. Since -e 3’s debut, it has been dramatically improved, and many benchmark tests are no longer competitive.
“Yes! Weil talks about the follow-up of Dall-E 3.” And I think it’s worth waiting. “
