OpenAI’s reported “Neuralese” approach fuels debate over chain-of-thought monitoring

The Information reported that OpenAI is quietly using loop transformers that become more efficient at scale without exposing the model’s thinking in the usual way. The approach, described as “Neuralese,” involves raw vectors rather than ordinary text or discrete tokens; the report’s accuracy remains uncertain.

OpenAI research director Jacob Petracki disputed the interpretation that the company is abandoning chain-of-thought monitoring. He said OpenAI has worked to preserve and use the technique since its first reasoning models, while acknowledging that monitoring is fragile, has been trending negatively for reasons not contingent on architecture changes, and that strengthening it is a core research goal.

The debate centers on whether reduced observability could weaken safety and prompt a race among frontier labs toward less monitorable systems. Concerns include losing potentially dangerous information when vectors are compressed into tokens, while skeptics argue that the reporting is insufficiently nuanced and that OpenAI is doing what it can.