OpenAI's next flagship still needs alignment work as Astra technique spreads

The next flagship's delay traces back to the looped-transformer technique that already powers Astra. By looping a token through the same layers multiple times, the model gets more math and coding performance for less cost, since a smaller model can punch above its weight. The trade-off is that some of the model's "thinking" moves out of readable text into internal numerical representations that human reviewers cannot inspect, which undermines chain-of-thought monitoring. OpenAI deliberately capped how many loops Astra runs to keep its chain of thought readable, The Information reports.

Safety capacity inside OpenAI has also shifted. After the Hugging Face incident, the company moved roughly ten percent of its total compute toward alignment and safety, according to a person familiar with the figures — hundreds of megawatts of data center capacity that would cost billions of dollars a year to rent at market rates. AI researcher Noam Brown said in a mid-September podcast that about ten percent of his team also works on security and safety.

How the technique is actually used emerged through a Microsoft slip-up. 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 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. OpenAI chief scientist Jakub Pachocki has already acknowledged that chain-of-thought monitoring is growing "fragile" and "unfortunately trending in a negative direction," and warned against an industry-wide race toward models without readable reasoning.