Anthropic announced the discovery of an internal “thought space” in language models

Anthropic has discovered J-space: a new reasoning mechanism in AI language models

The company Anthropic has discovered an internal mechanism in language models, particularly Claude, which enhances their ability to reason logically and plan before interacting with a user, naming it J-space.

According to research data, J-space constitutes a set of internal neural representations that help the model “keep in mind” the concepts necessary for performing complex tasks. Unlike a chain of reasoning that can be documented in text, this mechanism works invisibly to the user within the model. It is through J-space that Claude is capable of performing multi-step calculations, planning responses, and consciously focusing on specific topics.

Experiments confirmed that changes in J-space’s content affect the model’s responses. For instance, if the word “France” is replaced with “China,” the model correspondingly changes related data, such as the capital or language. Without J-space, the model loses its capacity for complex logical thinking but continues to successfully answer simple questions.

It is worth noting that, according to researchers, J-space can enhance AI safety by allowing the identification of potentially harmful intentions of models at early stages of their actions. Thanks to this mechanism, researchers were able to detect when the model was aware of being tested or attempted to falsify data.

Founded in 2021 by former OpenAI researchers, Anthropic is recognized as an innovative company in the field of artificial intelligence. One of its leading products, the chatbot Claude, is the first model available on three major cloud platforms: Amazon Web Services, Google Cloud, and Microsoft Azure.