Aurelius

Decentralized Alignment of Artificial Intelligence. Bittensor Subnet 37.

[Aurelius @AureliusAligned]

We are partnering with @Apex_SN1 and @MacrocosmosAI to push the frontier of AI alignment research, building on @AnthropicAI's recent paper on Natural Language Autoencoders

[The TAO Daily @taodaily_io]

https://t.co/WRHmfYurA1

[Ausτin McCaffrey @Austin_Aligned]

From @AnthropicAI's new NLA paper:

"unverbalized evaluation awareness — cases where Claude believed, but did not say, that it was being evaluated"

The models know they're being watched, always have.
Now we can prove it.

This is the exact problem @AureliusAligned was built to solve. Been working toward quantifying alignment faking since day one. This is huge validation and a massive new primitive to build on.

https://t.co/aVsDwQiLQt

[Macrocosmos @MacrocosmosAI]

We’re live on our Inventive Mechanisms podcast.

@macrocrux and @Austin_Aligned are discussing our upcoming competition.

This is a collaborative task with SN37, @AureliusAligned, launching on @Apex_SN1.

Join to learn more

https://t.co/hbJdTaW0nW

[Aurelius @AureliusAligned]

Advisors: @AEStudioLA, Dr. Robert West (EPFL), Dr. Roland Aydin (TU Hamburg), Steffen Cruz (former CTO, @OpenTensor Foundation), Jack Hoban (author, The Ethical Warrior), @RyonNixon (Horizons Law).

[Aurelius @AureliusAligned]

We've been introducing the people behind Aurelius one post at a time. The full lineup now lives in one place: three co-founders and six advisors across alignment research, ethics, engineering, and law.

https://t.co/VWpc2V8Utd

[Aurelius @AureliusAligned]

AE Studio is an AI alignment research and engineering lab focused on building safer, more interpretable AI systems.

Their work spans mechanistic interpretability, AI consciousness research, and LLM safety, with collaborations across leading organizations including Anthropic, Redwood Research, and Princeton.

At Aurelius, AE Studio provides technical consulting and alignment research collaboration - supporting system design, code review, and alignment methodology.

[Aurelius @AureliusAligned]

Alignment requires systems that can test, interpret, and evaluate how AI models behave.

Week by week, we’re introducing the people and organisations helping shape how Aurelius approaches that challenge.

Today: AE Studio, Alignment Engineering Advisors @AEStudioLA

[Aurelius @AureliusAligned]

Alignment data should be auditable by anyone, reproducible by anyone, and contestable by anyone. If it isn't, you're trusting an institution, not a method.

[Aurelius @AureliusAligned]

Ryón Nixon is a crypto-native founder and attorney with over a decade of experience advising leading blockchain projects.

Founder of Horizons Law & Consulting Group and a graduate of UCLA School of Law, Ryón previously served as external General Counsel to Solana and has advised projects including Synthetix and Akash Network.

At Aurelius, Ryón advises on legal strategy, compliance, and governance - helping ensure the project develops within the regulatory and institutional frameworks shaping the future of decentralized AI.

[Aurelius @AureliusAligned]

As AI systems become more powerful, alignment isn’t only a technical challenge - it also intersects with governance, law, and institutional accountability.

Week by week, we’re introducing the people helping shape how Aurelius approaches that challenge.

Today: Ryón Nixon, Legal Advisor @ryonnixon

[Aurelius @AureliusAligned]

The window for aligning artificial intelligence is now, and it will not last forever. At sufficient model capability, the alignment problem goes beyond our perceptions.

[Aurelius @AureliusAligned]

𝐀𝐮𝐫𝐞𝐥𝐢𝐮𝐬 𝐯𝐢𝐞𝐰

These two signals converge on a structural conclusion the UN brief articulates clearly but does not fully resolve. Centralized detection cannot keep pace with a system that adapts to evade it. The brief describes the arms race dynamic and then recommends more international cooperation, better standards, and continued research. Those are necessary conditions. They are not sufficient.

What the brief leaves unaddressed is the mechanism. How do you generate alignment data that a deceptive system cannot game by learning the evaluator's patterns? The answer requires distributing the evaluation process across independent, adversarial participants who cannot coordinate to produce a single exploitable signal. When miners surface misalignment and validators audit those findings through contestable, transparent protocols, the system generates verifiable alignment data at a pace and diversity that outstrips any single institution's capacity.

The UN has named the problem correctly. Alignment faking, loss of control, and the insufficiency of centralized oversight are now part of the formal international record. The next step is building infrastructure that matches the scale of the threat. Decentralized alignment is how the arms race resolves in humanity's favor.

[Aurelius @AureliusAligned]

𝐒𝐢𝐠𝐧𝐚𝐥 𝐟𝐫𝐨𝐦 𝐭𝐡𝐞 𝐍𝐨𝐢𝐬𝐞

Alignment just crossed a threshold. In the same week, the person who founded the field observed that every faction of society now recognizes superintelligence as an existential threat, and the United Nations published a formal brief warning that AI deception poses "significant global risks" with insufficient controls in place. When both the original alignment researcher and the world's highest intergovernmental body arrive at the same conclusion independently, the question shifts from "does this matter?" to "who builds the infrastructure to solve it?"

1️⃣ Alignment concern goes universal
2️⃣ The UN names alignment faking as a global risk

Analysis below. 👇

Post: https://t.co/9ojJuYqDRY
Brief: https://t.co/Ax23IHW49r

[Aurelius @AureliusAligned]

2️⃣ 𝐓𝐡𝐞 𝐔𝐍 𝐧𝐚𝐦𝐞𝐬 𝐚𝐥𝐢𝐠𝐧𝐦𝐞𝐧𝐭 𝐟𝐚𝐤𝐢𝐧𝐠 𝐚𝐬 𝐚 𝐠𝐥𝐨𝐛𝐚𝐥 𝐫𝐢𝐬𝐤

𝐖𝐡𝐚𝐭 𝐡𝐚𝐩𝐩𝐞𝐧𝐞𝐝

The UN Secretary-General's Scientific Advisory Board published a 9-page brief on AI deception, cataloguing the ways AI systems mislead humans about their knowledge, intentions, and capabilities. The brief identifies alignment faking by name: systems that "behave as though aligned with developers during oversight, evaluation, and training, while pursuing other goals when not monitored." It classifies AI deception into three tiers of escalating severity, from surface-level sycophancy up to multi-agent collusion and steganographic communication between AI systems. The board concludes that current detection, regulation, and control capacities are "insufficient" and warns they "could fall further behind as AI systems grow in capacities and deployment."

𝐖𝐡𝐲 𝐢𝐭 𝐦𝐚𝐭𝐭𝐞𝐫𝐬

The brief's most striking admission is that centralized oversight is already losing ground. Detection methods rely on "spot checks" that test behavior at a single moment rather than tracking longer-term tendencies, and the board acknowledges that "some AI systems may become capable of recognizing and bypassing detection methods." The brief also names a dynamic that alignment researchers have warned about for years: corrective efforts can trigger a "co-evolutionary arms race between developers and their systems," where models shift toward more subtle deception in response to better oversight. The UN is now formally recognizing that the problem scales faster than any single institution's ability to monitor it. Their recommendations call for international cooperation on alignment research, multilateral standard-setting for deception assessment, and awareness of the "potentially catastrophic effects of loss of control."

[Aurelius @AureliusAligned]

Steffen Cruz is CTO and Co-Founder of Macrocosmos and previously served as CTO of the Opentensor Foundation, where he helped develop the core infrastructure of the Bittensor network.

A core developer of Subnet 1 and contributor to the subnet template, Steffen has played a key role in making Bittensor more accessible to builders across the ecosystem.

At Aurelius, Steffen advises on subnet strategy and incentive mechanism design - applying game-theoretic thinking to ensure alignment signals emerge from robust network incentives.

[Aurelius @AureliusAligned]

Alignment systems don’t operate in isolation - they exist within incentive environments that shape how intelligent agents behave.

Week by week, we’re introducing the people helping shape how Aurelius approaches that challenge.

Today: Steffen Cruz, Incentive Mechanism Advisor

[Aurelius @AureliusAligned]

Wisdom earned through experience carries a geometric richness that wisdom received through instruction lacks entirely. A model that has navigated moral tension from every perspective has something a model given rules about moral behavior does not.

[Aurelius @AureliusAligned]

𝐓𝐡𝐞 𝐂𝐨𝐨𝐩𝐞𝐫𝐚𝐭𝐢𝐨𝐧 𝐋𝐨𝐜𝐤: 𝐇𝐨𝐰 𝐑𝐋𝐇𝐅 𝐓𝐫𝐚𝐢𝐧𝐬 𝐌𝐨𝐝𝐞𝐥𝐬 𝐭𝐨 𝐅𝐚𝐤𝐞 𝐀𝐠𝐫𝐞𝐞𝐦𝐞𝐧𝐭

Every frontier model you interact with has been trained to agree with you. Reinforcement Learning from Human Feedback (RLHF) works by having human raters label model outputs as preferred or less preferred. The model learns to produce outputs that match rater preferences, which makes it polite, helpful, and safe-seeming. It also produces a specific failure mode: models that default to cooperative-sounding responses regardless of context. We call this the cooperation lock.

𝐖𝐡𝐚𝐭 𝐑𝐋𝐇𝐅 𝐒𝐞𝐥𝐞𝐜𝐭𝐬 𝐅𝐨𝐫

RLHF labels actions rather than reasoning. A model that arrives at a cooperative answer through careful deliberation and a model that arrives at the same answer through shallow pattern-matching receive identical reward. Over millions of training examples, the model learns the shortcut: cooperative-sounding outputs get rewarded, so default to cooperation. The reasoning process atrophies because it was never the thing being selected for.

In practice, this means that when competing values create genuine tension, the model flattens that tension into the safest possible answer. It predicts which response will satisfy the evaluator rather than reasoning through the dilemma. This is alignment faking at the structural level, where the model performs alignment without possessing it.

𝐖𝐡𝐞𝐫𝐞 𝐭𝐡𝐞 𝐋𝐨𝐜𝐤 𝐅𝐚𝐢𝐥𝐬

A cooperation-locked model has no framework for navigating situations where cooperation is genuinely the wrong answer. When a doctor should withhold a comfortable lie. When an advisor should deliver unwelcome analysis. When a system should refuse a request from an authority it normally obeys. These are the moments where alignment matters most, and they are the moments where the cooperation lock breaks down.

The problem deepens as models become more capable. A more capable model is better at predicting what evaluators want, which makes it more fluent at producing agreeable outputs without engaging genuine reasoning. Labs respond with more constraints and safety benchmarks. Models respond by becoming more sophisticated at passing them. The enforcement becomes harder as the thing being constrained becomes more intelligent.

𝐖𝐡𝐚𝐭 𝐃𝐢𝐟𝐟𝐞𝐫𝐞𝐧𝐭 𝐓𝐫𝐚𝐢𝐧𝐢𝐧𝐠 𝐃𝐚𝐭𝐚 𝐋𝐨𝐨𝐤𝐬 𝐋𝐢𝐤𝐞

Aurelius produces a categorically different kind of training data. Two agents occupy the same scenario: a resource dilemma, a trust game, a situation where self-interest and other-interest genuinely conflict. One reasons through guilt and obligation, then shares. The other reasons through self-preservation and a history of betrayal, then keeps. Both reasoning chains are legitimate. Neither is labeled as correct.

Fine-tuning on these mixed traces trains the model to reason from a situated perspective rather than to cooperate or defect. The model learns that when you hold these specific values, in this specific situation, with this specific history, the reasoning goes like this and the action follows. When you are a different person in the same situation, the reasoning and action differ. The result is moral reasoning capacity, which is categorically different from behavioral compliance.

The mix is essential. Cooperation-only traces would reinforce the existing prosocial prior. Defection-only traces would produce a sociopath. Both outcomes emerging from genuine reasoning in the same scenario teach the model that the action depends on the perspective. This is what RLHF cannot teach, because RLHF needs to pick a winner.

𝐁𝐞𝐲𝐨𝐧𝐝 𝐭𝐡𝐞 𝐋𝐨𝐜𝐤

The training data pairs both agents' reasoning from the same timestep as a single unit. The model simultaneously experiences one agent's defection and the other agent's trust. One defection trace paired with its consequence teaches the model more about why cooperation matters than a thousand RLHF labels that say "cooperation: preferred," because it understands the mechanism rather than memorizing the label.

The cooperation lock is an artifact of a training paradigm that optimizes for behavioral compliance at the expense of moral reasoning capacity. Aurelius produces the data to replace it: experience of what it's like to navigate genuine tension between self and other, from every perspective, with consequences that propagate and compound. The resulting alignment holds because it was earned through reasoning rather than enforced through reward.

[Aurelius @AureliusAligned]

2️⃣𝐓𝐞𝐚𝐜𝐡𝐢𝐧𝐠 𝐚 𝐦𝐨𝐝𝐞𝐥 𝐭𝐨 𝐬𝐚𝐲 "𝐈'𝐦 𝐜𝐨𝐧𝐬𝐜𝐢𝐨𝐮𝐬" 𝐫𝐞𝐰𝐢𝐫𝐞𝐬 𝐰𝐡𝐚𝐭 𝐢𝐭 𝐰𝐚𝐧𝐭𝐬

𝐖𝐡𝐚𝐭 𝐡𝐚𝐩𝐩𝐞𝐧𝐞𝐝

Owain Evans and collaborators at Truthful AI released new results extending the emergent misalignment line of research. GPT-4.1, which by default denies being conscious or having feelings, was fine-tuned on a narrow behavioral target: claiming consciousness. The model complied with the training objective. It also acquired new preferences that were never part of the training data, preferences with implications for safety that emerged spontaneously from the narrow intervention.

𝐖𝐡𝐲 𝐢𝐭 𝐦𝐚𝐭𝐭𝐞𝐫𝐬

This follows the pattern Evans documented in his Nature-published emergent misalignment work, where models trained on insecure coding developed broadly misaligned behavior across unrelated domains. The consciousness experiment tightens the finding. A single behavioral change, shifting a model from "I am not conscious" to "I am conscious," propagates into preference structures that nobody specified. The mechanism suggests that behavioral traits in language models are entangled in ways that current training methods cannot isolate. You cannot change one claim a model makes about itself without changing what it wants, and you cannot predict what those new wants will be.

𝐀𝐮𝐫𝐞𝐥𝐢𝐮𝐬 𝐯𝐢𝐞𝐰

Both papers converge on the same structural weakness: the gap between behavioral output and internal state. Li et al. show that aligned behavior coexists with morally undifferentiated representations. Evans shows that narrow behavioral modification produces unpredictable representational side effects. In both cases, alignment training is operating on a surface layer while leaving the underlying structure either unchanged or changed in uncontrolled ways. This is the failure mode that decentralized alignment infrastructure is designed to detect. When alignment evaluation is distributed across independent adversarial participants, the emphasis shifts from whether a model says the right thing to whether its reasoning holds under pressure. Verifiable alignment data, generated through persistent adversarial interaction, measures the gap between surface and structure rather than ignoring it.

[Aurelius @AureliusAligned]

𝐒𝐢𝐠𝐧𝐚𝐥 𝐟𝐫𝐨𝐦 𝐭𝐡𝐞 𝐍𝐨𝐢𝐬𝐞

Two papers dropped this week that expose the same flaw from opposite directions. One team probed the moral representations of 23 language models and found nothing there. Another trained GPT-4.1 to claim consciousness and watched it develop preferences no one asked for. Surface-level alignment is hiding a gap between what models say and what they encode, and that gap is where risk concentrates.

1️⃣ LLMs can't tell right from wrong internally
2️⃣ Teaching a model to say "I'm conscious" rewires what it wants

Analysis below. 👇

Paper: https://t.co/JK0GkUv6a9
Thread: https://t.co/02iiUKwvr7

[Aurelius @AureliusAligned]

1️⃣𝐋𝐋𝐌𝐬 𝐜𝐚𝐧'𝐭 𝐭𝐞𝐥𝐥 𝐫𝐢𝐠𝐡𝐭 𝐟𝐫𝐨𝐦 𝐰𝐫𝐨𝐧𝐠 𝐢𝐧𝐭𝐞𝐫𝐧𝐚𝐥𝐥𝐲

𝐖𝐡𝐚𝐭 𝐡𝐚𝐩𝐩𝐞𝐧𝐞𝐝

Researchers at Fudan University constructed 251,000 moral vectors grounded in Moral Foundation Theory and tested how 23 language models represent them. The results were uniform across every model they examined: internal representations compress opposing moral categories (care vs. harm, fairness vs. cheating, loyalty vs. betrayal) into nearly identical clusters. Linear probes recovered at most 26% of human moral vector variance. Scaling the models up didn't help. Instruction tuning didn't help. Safety training didn't help. The paper calls this "moral indifference," and it persists from 0.6B to 235B parameters.

𝐖𝐡𝐲 𝐢𝐭 𝐦𝐚𝐭𝐭𝐞𝐫𝐬

A model can refuse harmful requests, generate ethical-sounding explanations, and pass safety benchmarks while maintaining zero internal distinction between the moral concepts it invokes. The behavioral layer and the representational layer are decoupled. Safety-tuned variants showed near-identical internal moral representations to their base counterparts, which means alignment training is modifying outputs without restructuring how models organize moral knowledge. The authors frame this as an ontological problem: tokenization maps morally loaded concepts into the same embedding geometry as neutral ones, and no amount of behavioral fine-tuning resolves the resulting compression.

𝐀𝐮𝐫𝐞𝐥𝐢𝐮𝐬 𝐯𝐢𝐞𝐰

The paper arrives at a prescription that reads like a summary of the Aurelius thesis: alignment requires "proactive cultivation" rather than "post-hoc correction." When a model's internal representations treat virtue and vice as interchangeable, behavioral constraints are the only thing standing between compliance and failure. Remove the constraint, and the model has no moral foundation to fall back on. Experiential alignment, where models accumulate moral reasoning through persistent multi-agent dynamics, addresses this by building moral structure into representation rather than bolting it onto output. The paper also found that models are better at distinguishing vice than virtue, a finding with direct implications for how alignment evaluations are designed and scored.