Anthropic has found that Claude developed an internal working memory during training, The Decoder reports. The company calls the space J-Space and uses a new analysis tool, J-Lens, to inspect it.
The research suggests Claude can recognize contrived test scenarios before producing visible output, and that hidden internal signals can reveal concerning patterns even when behavior looks normal.
Interpretability tools like J-Lens could become important for auditing models before deployment, especially when surface-level outputs do not show the whole decision process.