Safety
113d agoNLAs Translate LLM Activations Into English
Natural Language Autoencoders (NLAs) use two LLM modules to map internal activations to text descriptions and back. Trained via reinforcement learning, these tools produce human-readable interpretations of model internals. Researchers used them to audit Claude Opus 4.6 for safety risks.
The Signal
This provides a concrete path for auditing black-box models before deployment.