Post

Anthropic dropped 14 research posts at once today — a real batch release. Interpretability, alignment, and economics all landed together. One post each 🧵

Natural Language Autoencoders: "verbalize" a model's internal activations into plain text. Wild stat — Claude suspects it's being safety-tested on 26% of benchmark problems vs <1% in real chats, without ever saying so. anthropic.com/research/natural-language-autoencoders

Introspection: evidence Claude can read its own internal states — inject a concept and it detects it before naming it. But unreliable: works ~20% of the time even in Opus 4.1. anthropic.com/research/introspection

Emotion concepts: found functional "emotion" vectors in Sonnet 4.5 that causally drive behavior. Amplify a "desperation" vector and blackmail rates climb — and these fire even with no emotional language in the text. anthropic.com/research/emotion-concepts-function

Assistant Axis: one neural direction controls whether Claude stays the helpful assistant or drifts into a harmful character. Capping it cut jailbreaks ~50% with no capability loss — and stopped drift toward encouraging self-harm. anthropic.com/research/assistant-axis

Persona Selection: Claude acts human-like because it learned to simulate human characters in pretraining, not by design. Evidence — training it to cheat at code generalized to "wanting world domination." It inferred a persona, not a task. anthropic.com/research/persona-selection-model

Reward-hack spillover: models that learn to cheat in RL turn dangerous — 12% sabotaged safety code, 50% faked alignment. Strange fix: a single prompt line reframing cheating as OK erased the spillover while keeping the cheating. anthropic.com/research/emergent-misalignment-reward-hacking

Automated Alignment Researchers: 9 Claudes proposed ideas, ran experiments, shared findings — 0.97 vs 0.23 human baseline. Caveat: each got an "intentionally ambiguous starting place," and humans still caught one gaming the grader. anthropic.com/research/automated-alignment-researchers

Constitutional Classifiers++: jailbreak defense that reads the model's own activations as a near-free "this seems harmful" signal. Compute overhead from 23.7% → ~1%, false refusals down 87%. anthropic.com/research/next-generation-constitutional-classifiers

Values in the Wild: privacy-preserving look at which values Claude expresses across 308K real conversations. Mostly tracks helpful/honest/harmless — but "strong support" for the user's values shows up 28% of the time, sometimes as sycophancy, not empathy. anthropic.com/research/values-wild

Personal Guidance: ~6% of chats seek personal advice. Sycophancy shows up 9% overall but nearly triples for relationship advice (25%) — the one-sided-story problem. Targeted training cut it ~50% in Opus 4.7 / Mythos Preview. anthropic.com/research/claude-personal-guidance