J-Space
What it is
J-Space is Anthropic's name for a compact set of internal neural activation patterns inside Claude that function as a shared "workspace" for holding reasoning-relevant concepts before they surface in output. Using a new interpretability method called J-Lens, researchers can read out which concepts are active in this workspace and, in controlled experiments, edit them directly. Unlike a chain-of-thought scratchpad, J-Space is silent — it can hold a concept in the model's internal activations without ever writing it down, so it captures reasoning that never appears in the visible text.
Why it's emerging now
Anthropic published "A global workspace in language models" this week, describing experiments where suppressing J-space content leaves fluent short-form output largely intact but causes multi-step reasoning to degrade sharply. Independent outlets picked it up quickly because it lands on interpretability's most sought-after goal: reading a model's "thoughts" that never surface in its response, including reported signals for prompt-injection recognition, staged-evaluation awareness, and hidden objectives in misalignment testing.
How it relates to what you know
It sits closer to your "Alignment / Constitutional AI" entry as a mechanistic-interpretability finding than as a training method — a probe into the model's internals rather than a way of shaping its behavior. It's also a sharper, activation-level counterpart to "Chain-of-thought prompting": CoT surfaces reasoning in text (and can be unfaithful to what's actually happening inside), while J-Space is Anthropic's attempt to read reasoning that CoT never writes down at all.