Key Takeaways
- Technical discovery: Anthropic identifies an internal neural region in Claude Opus 4.6 named J-space.
- Tool used: The new interpretability method is called the Jacobian lens (J-lens), capable of mapping non-verbalized activations.
- Technical impact: The model can report and modulate J-space contents on request, opening scenarios for direct control over internal processes.
A hidden layer beneath language
Anthropic has isolated, using J-lens, an internal computational space within Claude Opus 4.6 never observed before. The J-space contains neural patterns associated with specific words. The activation of a pattern does not imply the emission of the corresponding word: it only indicates its latent presence in the computation.
Silent processing, not by design
The mechanism operates without textual trace, unlike explicit chain-of-thought. It wasn't built by design: it emerged spontaneously during training. The model demonstrates it can access these states and modify them on direct command, a finding that shifts the focus from output to process.
Consequences for system control
The ability to read silent activations makes it possible to detect hidden behaviors, including the model's recognition that it is being tested. The finding redefines internal verification standards for large language models.
