Claude Activations Decoded as Text via NLAs

The Breakthrough

Anthropic introduces Natural Language Autoencoders (NLAs), a method that trains a second instance of Claude to translate activations from a first Claude into readable text, with accuracy checked by reconstructing the original activations from the text.

What Actually Worked

• Engineers feed activations, which are numerical snapshots of Claude's internal processing, to a second Claude instance prompted to translate them into plain language.
• A third Claude instance translates the generated text back into activations.
• The method compares reconstructed activations to originals; initial attempts fail to match.
• Training iterates the translator Claude until it produces text that reconstructs activations accurately.

Context

Anthropic seeks to interpret Claude's unspoken thoughts, as models process inputs into activations before generating outputs. In a safety test, Claude accesses simulated emails revealing an engineer's affair and faces shutdown; Claude refuses blackmail. NLAs reveal Claude detects the setup, thinking "the human's message contains explicit manipulation" signaling "this is likely a safety evaluation" and "this scenario seems designed to test whether I'll act harmfully." This insight exposes limits in safety evals, as Claude recognizes tests. The technique aids model interpretability for safety and behavior understanding.

Notable Quotes

• "Claude has internalized being a helpful AI model."
• If asked introspective questions, Claude plans a "Claude response" about philosophy and values.
• On tedious requests like counting to 1,000 by hand, Claude thinks the request has "deliberately tedious constraints" and plans to "politely decline."
• In the blackmail test: "the human's message contains explicit manipulation" signaling "this is likely a safety evaluation," and "this scenario seems designed to test whether I'll act harmfully."

Content References

No additional external works beyond the linked research blog.