Anthropic uncovers Claude's J-space during training, hidden reasoning area
Technology
Anthropic just uncovered a hidden feature inside its AI, called Claude, called J-space.
Think of it as a backstage area where the AI quietly plans and reasons through tough questions, without showing its work in the text you see.
This discovery came up during training and gives a peek into how advanced AI really thinks.
J-space aids Claude's reasoning, not consciousness
With J-space switched on, Claude can handle tricky tasks like multi-step reasoning and remembering information across different jobs.
But if you take J-space away, the AI suddenly struggles with things like summarizing or even writing rhyming poetry.
Still, researchers say J-space is just one small part of Claude's brainpower, and while it's inspired by how humans think, it doesn't mean the AI is actually conscious.