3 ms·Anthropic NLAs translate LLM activations to human-readable text for safety1 points by sebastianperezr 5mo ago