Anthropic CEO Dario Amodei published a new blog highlighting the critical need for "mechanistic interpretability" in AI, arguing that understanding models’ inner workings could become humanity’s safeguard as they grow increasingly powerful.

Amodei stressed that AI is different from traditional software because decision-making emerges organically, making its operations unclear even to creators. He revealed that Anthropic has mapped over 30M "features" in Claude 3 Sonnet, representing specific concepts the model can understand and process.

The CEO compared the ultimate goal to creating a reliable “AI MRI” for diagnosing models and better understanding their “black box.” He warned that AI is advancing faster than interpretability, leaving us unprepared for AI systems like a “country of geniuses in a datacenter,” possibly arriving as early as 2026.

Anthropic’s essay frames interpretability as not just a technical hurdle but a prerequisite for safe deployment—urging the broader AI community to act swiftly before it's too late.