Anthropic researchers identified neural patterns in the Claude 3 Sonnet model called J-space. This internal workspace represents the model's concepts and thoughts before it generates an output. The discovery reveals reasoning steps that do not appear in final responses. These patterns emerged spontaneously during the AI training process.
The breakthrough advances AI interpretability by showing how complex models function. Monitoring J-space allows researchers to detect or alter internal processes to prevent biased or deceptive content. The findings draw parallels to the global workspace theory of human consciousness. This discovery offers broad implications for building safer and more controllable AI systems.