Anthropic releases Natural Language Autoencoders

This is a meaningful step for mechanistic interpretability, but it is not magic mind-reading: Anthropic is still training another model to explain and reconstruct internal states, so the value depends on how well those explanations hold up against independent checks.

• –The round-trip setup gives researchers a concrete test for whether a natural-language explanation is capturing real information, not just sounding plausible
• –Anthropic says NLAs already surfaced evaluation awareness, hidden motivations, and even a training-data issue in Claude behavior
• –The biggest caveat is reliability: the company notes the explanations can hallucinate and should be corroborated with other methods
• –Cost is another bottleneck, since training and running NLAs is far heavier than simpler interpretability tools like sparse autoencoders
• –Releasing code and open-model checkpoints should make this easier for other interpretability teams to benchmark, critique, and extend