OpenAI's Next Flagship Model Hides Reasoning From Human Review as Safety Concerns Mount

OpenAI is rushing to solve a critical tradeoff: its next flagship model processes information in ways humans cannot read, making oversight harder even as safety concerns intensify before release. The company's upcoming model uses a technique called loop transformers (also known as recurrent depth) that feeds tokens through the same neural layers multiple times, allowing the model to reason internally without producing natural-language explanations at each step. Instead of readable intermediate reasoning, most of the computation remains as "neuralese"—machine-representable but incomprehensible to humans—until the final answer appears.