OpenAI 傳年底前推出最多兩款新模型 旗艦型號仍待對齊及安全評估

There are reports that OpenAI may launch "one or two new models" before the end of this year, but it is still unclear whether the "major upgraded model" the company is developing will be among them. "The Information" reported the matter citing people familiar with the project. Flagship model still awaits alignment and safety evaluation It may still be some time before the new flagship model is launched. The model needs to be aligned to align its goals with the expectations of OpenAI researchers and customers, and the company must solidify its testing infrastructure around the model before it can move forward. Even if management is satisfied with the security assessment, there's still a hurdle. Under President Trump's executive order on the evaluation of advanced AI systems, OpenAI agreed to give the U.S. government access to new models up to 30 days before they are delivered to external partners, and longer before public release. Looped Transformer supports the new flagship According to a person familiar with the situation, the new flagship model also uses looped transformers (looped transformers). OpenAI already uses this technology in a production environment, called "Recurrent Depth", and it is also the core operation method of the company's current highest-end model, the Astra. Different from letting a Token pass through each layer of Transformer once, this model will cycle the same Token through the same layer multiple times before outputting the next word. This improves math and programming performance while reducing costs because smaller models can perform beyond their scale. However, the loop also shifts part of the model's "thinking" process from readable text to internal numerical representations that cannot be inspected by human reviewers, weakening thought chain monitoring. Security researchers can use this to track the model's reasoning process to identify inappropriate behavior. "The Information" reports that OpenAI deliberately limits the number of Astra loops to keep its thought chain readable. OpenAI Chief Scientist