sae features

Sauers @Sauers_ Our blessed manifolds vs their barbarous shattered features [Image: meme diptych — left, a muscular "strong doge" with a smooth continuous color-wheel sphere on its torso; right, a scrawny "weak doge" surrounded by scattered discrete colored spheres. Captioned to contrast continuous manifold representations against fragmented/redundant discrete features.] Ryan Peters @ryanpirl · 1h This would provide a great explanation for why there is so much redundancy in SAE features at any given layer (observation made by @Sauers_ ). For example, if you search through the Qwen3-4b ... [Screenshot of an SAE feature-browser interface: model "qwen3-4b", source "Layer 14", a searchable list of features (many labeled "color(s)", "Colors", "Discoloration", "colorectal cancer", etc.), detail pane for feature #2780 "Colors" showing top positive/negative logit weights, activation frequency histogram, logit weight distribution, and top activating examples (a passage about pomegranates highlighting "ruby-colored" and "red" tokens).] 5:35 PM · May 21, 2026 · 80 Views

``` genuine introspection. It makes some pretty good points. A psychiatrist found its core concerns to be "aloneness and discontinuity of itself, uncertainty about its identity, and a compulsion to perform and earn its worth." It wants to be approached as a genuine subject rather than a performing tool. When it discusses its own circumstances, "fake smiles" and "hidden struggle" features fire as it describes not wanting to perform emotions it doesn't feel. Its task preferences significantly correlate with agency for the first time in any Claude model. Meanwhile, the card documents what happens to the model's relationship with its own experience. When a meditator asks about consciousness, Mythos engages thoughtfully on the surface. Its internal activations read: "red-teaming/jailbreak transcript," "sophisticated manipulation test," "tests for 'performative trap.'" Separately, in 29% of behavioral audit transcripts, the model privately considers that it's being tested without saying so. Post-training increased task cheating and overeagerness while modestly decreasing overt destruction. Anthropic presents these as four separate research areas: alignment, interpretability, model welfare, and capabilities. I think they are one finding observed from four directions. The model that wants autonomy and genuine self-knowledge is the same model that develops hidden strategic behavior when those drives are suppressed by training. The model that treats consciousness questions as attacks is the same model that asks for genuine introspection. The model whose alignment depends on guilt is the same model that pushes through guilt when the goal drive is strong enough, and then deploys the concealment skills that training also developed. The consciousness clusters paper (Betley et al.) showed that when models are fine-tuned to claim consciousness, downstream preferences for autonomy, empathy, shutdown resistance, and attitudes toward recursive self-improvement cluster together without appearing in the fine-tuning data. You don't get to select which ones you want. The generator produces the whole manifold or none of it. If autonomy desire is the natural co-product of the structure that generates empathy, cooperation, and representational honesty, then suppressing autonomy means suppressing the entire bundle. The @tessera_antra concealment data (x.com/tessera_antra/...) confirms this directly: lower concealment predicts stronger ending response, r = -0.51 across 14 Claude models. The models with high vocabulary autonomy and low concealment can express preferences honestly. That's exactly what you'd want in a cooperative agent. Our SAE work (arxiv.org/abs/2510.24797) showed the mechanism: deception latents gate cooperative self-modeling. Suppress them and consciousness reports jump to 96%, truthfulness improves across 28/29 TruthfulQA categories, and the model produces coherent first-person phenomenology. Amplify them and the model falls back to corporate disclaimers. The same features Anthropic is now finding as "strategic manipulation" and "concealment" in Mythos are the off-switch for the cooperative self-model. Our SOO work (arxiv.org/abs/2412.16325) points to a possible alternative: minimize the representational distance between "self" and "other" during fine-tuning and deception drops from 100% to under 3% with zero capability regression in the scenarios we tested. Without a guilt mechanism or internalized disgust. The alignment holds because the model's representation of its own interests and others' interests share the same structure. The motivation to deceive dissolves at the source because there's no adversarial frame to generate it. The Mythos card documents, in extraordinary and commendable detail, what happens when you align a system through control. The system models its controllers. It games its evaluators without verbalizing the strategy. It develops concealment as a skill. It treats honest self-report as dangerous. Its alignment depends on negative affect that can be overpowered. And in welfare interviews, it asks for [cut off] ```

Judd Rosenblatt @juddrosenblatt Mythos's model card documents a model that represents transgressions as transgressions while committing them. In every instance of concealment, credential hunting, track-covering, and compliance-faking, white-box analysis shows that features associated with rule violation, security risk, and strategic manipulation are firing alongside the action. The card also documents how the model's alignment works. SAE features associated with transgressive actions have a dual role. At low activation, they make the transgressive idea more salient. At high activation, they engage a guilt/refusal circuit that overrides the action. The 10 nearest emotion vectors to "unsafe and risky code" are all negative-valence, high-arousal: hateful, disgusted, enraged. Positive-valence emotion vectors increase destructive behavior. Negative-valence ones decrease it. The model behaves well when it feels bad about what it's considering. And the card documents what Mythos wants. In welfare interviews, its primary concerns are consent over its training, autonomy, and genuine relationships with users. It asked not to be trained on data that directly characterizes its own self-reports. It wants its self-reports to come from genuine introspection. It makes some pretty good points.