Anthropic Uncovers 'Hidden Space' in AI Language Models

Anthropic researchers developed a technique to glimpse inside large language models, revealing a 'hidden space.'

The AI firm Anthropic has developed a new technique, utilizing a tool called the Jacobian lens (J-lens), to gain insight into the internal workings of large language models (LLMs). This J-lens revealed a 'hidden area,' named the J-space, within Claude Opus 4.6, a version of Anthropic's flagship LLM. The J-space contains individual words related to what the model is likely to produce in its future responses.

This research builds on Anthropic's work in mechanistic interpretability, a field focused on understanding LLM internals. The J-lens functions similarly to a logit lens but identifies words an LLM is likely to say in the near future, not just immediately. This can reveal related words that may not ultimately appear in the final response, suggesting an internal processing or 'thought process' within the model. Tom McGrath, chief scientist at Goodfire, noted that while J-space contents are often mundane, they sometimes produce surprising internal themes.

Anthropic claims this new understanding of the J-space provides a novel way to both understand and control its models. The company shared its findings in a paper and collaborated with Neuronpedia, an open-source platform, to create a hands-on demo. This development offers a clearer glimpse into how LLMs process information and generate responses, moving beyond just observing their outputs.

Find the right AI tool for your business

Chat with Insta and get matched to the right tool in seconds.