OpenAI 拟年内发布新模型：循环 Transformer 架构提升性能但加剧对齐与安全监测难度

OpenAI is racing to finish its next flagship model, which uses the same looped transformer technique as Astra. The approach boosts math and coding performance but makes safety monitoring harder. Employees say one or two new models will likely ship by year's end.

OpenAI will likely release "one or two new models" before the end of the year, according to employees, though it's unclear whether "the major upgraded model" the company is currently working on will be among them. That's according to The Information, citing people familiar with the projects.

The new flagship could still be a ways off. The model needs alignment work so its goals match what OpenAI's researchers and customers expect. The company also has to shore up the testing infrastructure around the model before it can move forward.

After leadership is satisfied with safety evaluations, there's another hurdle. Under an executive order from President Trump governing evaluations of advanced AI systems, OpenAI agreed to give the US government access to new models up to 30 days before they go to outside partners and longer still before a public launch.

Looped transformers power both Astra and the next flagship

The new flagship also uses looped transformers, according to a person familiar with the situation. OpenAI already runs this technique in production, where it goes by "Recurrent Depth." It's central to how Astra works, the company's current top model. Instead of passing a token through the transformer layers once, the model loops it through the same layers multiple times before outputting the next word. That drives up performance on math and coding and brings down costs, since a smaller model can punch above its weight.

But the looping also moves some of the model's "thinking" out of readable text and into internal numerical representations that human reviewers can't inspect. That undermines chain-of-thought monitoring, where safety researchers trace a model's reasoning to spot bad behavior. OpenAI deliberately capped how many loops Astra runs to keep its chain of thought readable, The Information reports.

After the Hugging Face incident, OpenAI moved roughly ten percent of its total compute toward alignment and safety, according to a person familiar with the figures. The Information reports that amounts to hundreds of megawatts worth of data center capacity, though it's unclear how much was allocated to safety before or what the current split looks like. At market rates, the compute OpenAI shifted would cost billions of dollars a year to rent. AI researcher Noam Brown said in a mid-September podcast that about ten percent of his team also works on security and safety.

Microsoft accidentally leaks how OpenAI's looped inference works

How OpenAI actually uses this technique came to light through a slip-up at Microsoft. The major investor, which has access to OpenAI's intellectual property, published a technical write-up on GPT-6.1 Sol, a cheaper Astra alternative that OpenAI launched earlier this month.

In a now-deleted section, Microsoft described GPT-6.1 Sol as using "looping inference," where the model works through the same problem in multiple "passes" before producing an answer. GPT-6.1 Sol "uses the same base model weights as GPT-6 Sol with two inference passes instead of three," The Information quotes from a cached copy of the page.

A person with direct knowledge confirmed that the write-up matched what OpenAI had told Microsoft in private. Microsoft took the section down after realizing it wasn't meant to be public. Looped transformers remain a sensitive part of OpenAI's model development, both during training and at inference, The Information reports.

OpenAI chief scientist Jakub Pachocki has already acknowledged that chain-of-thought monitoring is growing "fragile" and "unfortunately trending in a negative direction." He also warned against an industry-wide race toward models without readable reasoning.