Anthropic researchers have found a hidden workspace inside Claude Opus 4.6 — a pre-output "scratchpad" they call J-space — where the model computes words it might say before it actually says them. If the theory holds, this could become the most powerful new tool yet for understanding and controlling what large language models are doing under the hood.

Evidence For

• Anthropic built a technique called the Jacobian lens (J-lens) to peer into the middle layers of...