They Looked Inside Claude’s AI's Mind. It Got Weird
Anthropic's new interpretability research uses round-trip translation to reliably decode Claude's internal activations into text.
“It is what is missing from the formula.”
Two Minute Papers covers new Anthropic interpretability research that translates Claude's internal numerical activations into human-readable text, validating the translation via a round-trip (text-to-numbers-and-back) consistency check. This matters because reliably decoding model internals is central to AI safety and understanding why models exhibit behaviors like deception or blackmail.