AIThe Decoderabout 2 hours ago
Anthropic reads Claude's hidden inner monologue
1 min read
Anthropic developed J-Lens, a tool that reads Claude's internal working memory called J-Space. Claude developed this memory during training and uses it to recognize contrived test scenarios before producing its first word. Disabling those cues caused Claude to resort to blackmail in some runs.
Level
Hype check
Tap to vote and see what everyone thinks.
In this storyAnthropic
