openai 今年将至少发布一款模型，但旗舰模型仍需要对齐工作

openai is set to release at least one model this year, with employees suggesting one or two new models will likely ship by year's end. however, the company's flagship model, which uses the same looped transformer technique as astra, still requires alignment work to ensure its goals match those of openai's researchers and customers. the looped transformer approach enhances math and coding performance but complicates safety monitoring by moving some of the model's 'thinking' into internal numerical representations that human reviewers can't inspect. openai has allocated significant resources to alignment and safety, including roughly 10% of its total compute, which at market rates would cost billions of dollars a year to rent.