OpenAI introduced a new reasoning technique called recurrence. It lets models reuse internal thought steps instead of writing them as text. This avoids the bottleneck of encoding thoughts into human language. The technique is called neuralese recurrence when thoughts loop in their native form. OpenAI’s chief scientist says the depth gain is modest. He compares it to adding more layers the usual way. The new method still uses chain-of-thought for training. It does not yet allow unbounded internal looping. Researchers cannot train models to reason fully in vectors without text. The alignment risk remains if models can think without human-readable output.