NLAs read thoughts beyond the J-space

According to LessWrong, researchers have demonstrated that large language models harbor thoughts they cannot consciously report. In experiments on Llama-3.3-70B using tools released by Anthropic, researchers found that when models are injected with hidden concepts and asked whether they perceive them, they deny it. Yet a new tool called Natural Language Autoencoders can read those hidden activations almost perfectly. Here's what's happening: Anthropic recently showed that models can only verbalize about ten percent of their own internal thoughts—a small mental workspace they call J-space. The rest remains inaccessible to the model's own introspection. But NLAs, which translate model activations into English, can decode roughly eighty percent of what's actually happening inside. When researchers split injected concepts into conscious and subconscious parts, the model never named the subconscious half, even though that hidden content actively shaped the model's output logits. Zero false positives across multiple controls. The findings raise an uncomfortable question for AI safety: if models are only reporting ten percent of their cognition, what do our alignment tests and behavioral monitors actually measure? The research suggests we've been evaluating only the tip of the iceberg.

Listen to this story

Hear this and more stories in a personalized audio briefing.

Open The Chonkerton