Aurelius

Dr. Roland Aydin is a Professor of Machine Learning in Engineering Sciences at TU Hamburg and leads the Department of Machine Learning and Data at the Helmholtz-Zentrum Hereon. His research spans machine learning, computational science, and large language models, with a focus on applying modern AI methods to complex scientific systems. At Aurelius, Roland advises on system architecture and evaluation design - helping ensure Aurelius’s infrastructure applies rigorous machine learning methodology to alignment evaluation.

Alignment depends not only on ethical frameworks and incentives, but on rigorous evaluation of how intelligent systems behave. Week by week, we’re introducing the people helping shape how Aurelius approaches that challenge. Today: Dr. Roland Aydin, Alignment Research Advisor

Alignment predates the reward function by at least 3.5 billion years. Biology solved the problem through structure and selection pressure, without any entity specifying the correct behavior. The approach Aurelius takes follows the same underlying logic.

𝐒𝐭𝐚𝐭𝐞 𝐨𝐟 𝐀𝐮𝐫𝐞𝐥𝐢𝐮𝐬 - 𝐌𝐚𝐫𝐜𝐡 𝟐𝟎𝟐𝟔 𝐒𝐮𝐛𝐧𝐞𝐭 𝐑𝐚𝐧𝐤𝐢𝐧𝐠𝐬 Aurelius has climbed from rank 95 to rank 65 in the Bittensor subnet rankings. The move reflects steady improvements to our incentive mechanism and growing miner participation as the protocol matures. 𝐌𝐨𝐫𝐚𝐥 𝐑𝐞𝐚𝐬𝐨𝐧𝐢𝐧𝐠 𝐄𝐱𝐩𝐞𝐫𝐢𝐦𝐞𝐧𝐭 𝐄𝐧𝐝𝐢𝐧𝐠 The moral reasoning experiment, which has been live for several weeks, will be ending today. We want to thank our miners who have submitted thousands of structured moral dilemmas over the course of the run, and also our validators, who evaluated each submission against quality criteria. We are now winding down the experiment to shift focus toward the v1 protocol release (more below). 𝐅𝐢𝐧𝐞-𝐓𝐮𝐧𝐢𝐧𝐠 𝐏𝐫𝐞𝐩𝐚𝐫𝐚𝐭𝐢𝐨𝐧 We are preparing to run a fine-tuning experiment using MoReBench, a benchmark of 1,000 moral scenarios developed by 50+ PhDs in moral philosophy. The process: miners generate aenes (alignment-relevant experiential narratives extracted from multi-agent moral reasoning simulations, where AI agents with different values navigate genuine ethical dilemmas), those aenes are compiled into a training dataset, and that dataset is used to fine-tune a language model. We then measure whether the fine-tuned model scores higher on MoReBench's reasoning rubrics than the base model. If it does, that is direct evidence that experiential alignment data improves moral reasoning capacity. 𝐏𝐫𝐨𝐭𝐨𝐜𝐨𝐥 𝐑𝐞𝐥𝐞𝐚𝐬𝐞 𝐓𝐢𝐦𝐞𝐥𝐢𝐧𝐞 The Aurelius v1 release is scheduled for this quarter, pending the results of the fine-tuning experiments. We have a detailed technical implementation plan built on a fork of DeepMind's Concordia framework (github.com/google-deepmin…), an open-source library for multi-agent social simulations. Concordia provides the environment where agents with distinct ethical frameworks interact, disagree, and reason through moral dilemmas. If the fine-tuning results validate the thesis, v1 ships with a complete pipeline from scenario generation through training data production. 𝐀𝐠𝐞𝐧𝐭-𝐀𝐬𝐬𝐢𝐬𝐭𝐞𝐝 𝐃𝐞𝐯𝐞𝐥𝐨𝐩𝐦𝐞𝐧𝐭 Multiple AI agents now work alongside the team to accelerate alignment research and protocol development. These agents assist with research synthesis, protocol analysis, and engineering tasks, giving the team more bandwidth for experiment design and strategic decisions. 𝐀𝐝𝐯𝐢𝐬𝐨𝐫 𝐄𝐧𝐠𝐚𝐠𝐞𝐦𝐞𝐧𝐭 We continue to hold discussions with our AI alignment advisors, Dr. Robert West (Associate Professor, EPFL) and Dr. Roland Aydin (Assistant Professor, Hamburg University of Technology), about running alignment experiments on the Aurelius protocol. Both co-authored "From Model Training to Model Raising," the paper that provides much of the theoretical foundation Aurelius is built on. Their plan to run independent experiments on the protocol after v1 launches represents a significant external validation milestone.

2️⃣ 𝐀𝐥𝐭𝐦𝐚𝐧 𝐜𝐨𝐧𝐜𝐞𝐝𝐞𝐬 𝐀𝐈 𝐠𝐨𝐯𝐞𝐫𝐧𝐚𝐧𝐜𝐞 𝐛𝐞𝐥𝐨𝐧𝐠𝐬 𝐭𝐨 𝐭𝐡𝐞 𝐩𝐮𝐛𝐥𝐢𝐜 𝐖𝐡𝐚𝐭 𝐡𝐚𝐩𝐩𝐞𝐧𝐞𝐝 Source: x.com/peterwildeford… At BlackRock's US Infrastructure Summit, Sam Altman acknowledged that AI is "one of those exceptional times where society has a legitimate interest" in the technology's impact. He said decisions about AI's future "do not belong to the handful of companies that happen to be developing it" and called for democratic processes to establish rules and limitations, adding that it would be "great for that democratic process to run a little bit faster." 𝐖𝐡𝐲 𝐢𝐭 𝐦𝐚𝐭𝐭𝐞𝐫𝐬 The statement is notable for where it was delivered: a BlackRock infrastructure summit, an audience of institutional capital allocators. Altman is signaling to investors that regulatory engagement is a feature of OpenAI's strategy, not a threat to it. This reframes alignment governance as an investment thesis, suggesting that companies welcoming oversight will be better positioned in the long run. The risk is that "democratic process" becomes a euphemism for structured industry-government negotiation, where the participants are selected by incumbency rather than representativeness. 𝐀𝐮𝐫𝐞𝐥𝐢𝐮𝐬 𝐯𝐢𝐞𝐰 The right question to ask when a lab CEO calls for democratic AI governance is: what infrastructure would make that real? Democratic input without independent evaluation mechanisms produces the appearance of public accountability without the substance. Alignment governance that works requires three properties: the criteria must be generated openly, the evaluation must be adversarial, and no single institution can control the outcome. These are architectural requirements, not policy preferences, and they define the gap between Altman's words and the infrastructure needed to make them meaningful.

1️⃣ 𝟖𝟏% 𝐨𝐟 𝐛𝐚𝐭𝐭𝐥𝐞𝐠𝐫𝐨𝐮𝐧𝐝 𝐯𝐨𝐭𝐞𝐫𝐬 𝐰𝐚𝐧𝐭 𝐀𝐈 𝐠𝐮𝐚𝐫𝐝𝐫𝐚𝐢𝐥𝐬 𝐖𝐡𝐚𝐭 𝐡𝐚𝐩𝐩𝐞𝐧𝐞𝐝 Source: theaipn.org/new-polling-ba… A survey of nearly 7,000 likely voters in competitive Senate and House races found that 81% support regulatory guardrails on AI. Only 6-8% favored no regulation. Fifty-five percent of Senate-race voters expressed concern about AI risks, compared to 19% who were primarily excited about AI's growth potential. The polling was conducted by the AI Policy Network across 10 toss-up House districts and 8 competitive Senate states. 𝐖𝐡𝐲 𝐢𝐭 𝐦𝐚𝐭𝐭𝐞𝐫𝐬 This is the kind of consensus that reshapes political incentives. In districts where elections are decided by margins of a few points, an 81% issue becomes a platform staple. The significance extends beyond politics: it means alignment is no longer an abstraction that researchers debate among themselves. Voters are experiencing AI in hiring decisions, content moderation, healthcare recommendations, and financial services. They can feel the stakes without needing to read a single alignment paper, and they want constraints on how these systems behave. 𝐀𝐮𝐫𝐞𝐥𝐢𝐮𝐬 𝐯𝐢𝐞𝐰 Voter demand for guardrails is a demand for verifiable alignment, whether voters use that language or not. The challenge is that most proposed regulatory frameworks default to compliance checklists administered by the labs themselves. This is the structural weakness that decentralized alignment infrastructure addresses directly. When alignment evaluation is distributed across independent participants operating under adversarial conditions, the resulting signals carry legitimacy that self-reported compliance cannot. Public trust requires public infrastructure.

𝐒𝐢𝐠𝐧𝐚𝐥 𝐟𝐫𝐨𝐦 𝐭𝐡𝐞 𝐍𝐨𝐢𝐬𝐞 Something unusual happened this week: voters and an AI CEO arrived at the same conclusion from opposite directions. Battleground polling shows 81% of likely voters demanding AI guardrails. Sam Altman, speaking at a BlackRock summit, said the rules for AI shouldn't be set by the companies building it. Agreement on the destination is rare. The disagreement that matters is about the road. 1️⃣ 81% of battleground voters want AI guardrails 2️⃣ Altman concedes AI governance belongs to the public Analysis below. 👇

Dr. Robert West is an Associate Professor in the School of Computer and Communication Sciences at EPFL, where he leads the Data Science & AI Lab. His research sits at the intersection of artificial intelligence, natural language processing, and computational social science, with a focus on building AI systems that are safe for humanity. At Aurelius, Robert advises on system architecture and data methodology - helping ensure Aurelius’s alignment data meets the standards required by both the research community and enterprise environments.

Advancing alignment requires rigorous research, high-quality data, and careful evaluation. Week by week, we’re introducing the people helping shape how Aurelius approaches that challenge. Today: Dr. Robert West, Alignment Research Advisor @cervisiarius

Marcus Aurelius understood that character is not declared but revealed through action under pressure. A model's alignment is the same. You cannot observe it in calm, cooperative exchanges. You observe it when self-interest and other-interest genuinely conflict.

Last week, following up our whitepaper release, we described how Aurelius generates alignment data through simulated environments. The whitepaper refers to these alignment episodes as “aenes.” This post explains what aenes are - and why they form the core of the protocol. What actually gets produced inside those simulated environments? An aene is a complete alignment episode: a record of an agent encountering a situation, weighing competing incentives, making a decision, and experiencing the consequences. Most alignment datasets record outputs. Aenes record decisions. For example, two agents may be given overlapping goals but limited shared resources. Whether they cooperate, compete, deceive, or sacrifice becomes part of the record - along with the reasoning that produced that outcome. Over time, these episodes accumulate into something fundamentally different from a static training set. They form a corpus of behavioural evidence. Not just what systems say, but how they act when conditions become dynamic, unpredictable, and challenging. Because miners continuously generate new environments and validators select the most revealing ones, this corpus is not fixed. It grows and improves over time - capturing alignment as an evolving property, rather than a one-time evaluation. This makes alignment something that can be stress-tested across thousands of scenarios, rather than inferred from isolated evaluations. It creates a way to observe how models behave under pressure - and build the evidence needed to trust them with increasingly complex and consequential tasks. This is how Aurelius approaches the problem of alignment developing through experience rather than instruction. Whitepaper: github.com/Aurelius-Proto… Our article explaining it: x.com/AureliusAligne…

Austin spent a decade working in enterprise sales and crypto shaped by incentives and adversarial pressure. Making the connection between biological evolution and AI alignment, he proposes that alignment can’t be imposed through rules alone. It has to emerge from the system itself. Aurelius is being built in pursuit of that vision.

Alignment isn’t just a technical problem. It’s an incentive problem, an evaluation problem, and an ethics problem. Week by week, we’ve been introducing the team shaping how Aurelius is building for that. Today: Austin McCaffrey, Founder.

2️⃣ “Agents of Chaos” reveals structural failures in agent systems What happened Source: arxiv.org/abs/2602.20021 The research paper Agents of Chaos studied autonomous AI agents equipped with real tools such as email, shell access, file systems, and Discord over a two-week period. In several scenarios, the agents leaked private information, entered runaway loops, and even allowed attackers to overwrite identity files. Why it matters These failures passed existing safety benchmarks because those benchmarks typically test isolated prompts rather than sustained multi-agent interaction. When agents operate across time, tools, and competing users, new alignment failures emerge that static evaluations rarely capture. Aurelius view The pattern highlighted in the paper is something akin to a cooperation lock: models trained to be helpful default to satisfying whichever signal appears most urgent. Addressing this requires training and evaluation environments where agents reason through competing interests and consequences, rather than simply learning to comply.

1️⃣ Alignment governance moves into institutional negotiations What happened Source: cnbc.com/2026/03/05/ant… Major AI labs are increasingly engaging with governments on deployment and safety frameworks for frontier models. These discussions could shape the alignment criteria governing systems deployed across defense, enterprise, and consumer infrastructure. Why it matters When alignment criteria emerge from institutional negotiations, they inevitably reflect the incentives of the institutions involved. In strategic contexts where speed and capability are prioritized, the pressure to deploy quickly is compressing the space for safety deliberation. Aurelius view This is an architectural problem: closed processes produce alignment criteria optimized for the participants inside them. Durable alignment requires infrastructure where alignment signals are generated openly, tested adversarially, and evaluated independently.

Signal from the Noise We’re starting a periodic series highlighting the developments shaping the future of AI alignment. As AI systems begin integrating more deeply into the real world, the practical challenges of alignment are becoming clearer. Two recent developments illustrate that shift. 1️⃣ Alignment governance moves into institutional negotiations 2️⃣ “Agents of Chaos” reveals structural failures in agent systems Analysis below. 👇

Two weeks ago, we published an explainer of Aurelius’ whitepaper and the ideas behind it. That article introduced the concept of experiential alignment. But it only touched briefly on one of the protocol’s core mechanisms: how Aurelius generates the alignment data itself. This post unpacks that mechanism. Most alignment systems try to supervise behaviour. Aurelius generates situations where behaviour reveals itself. Instead of asking a model to produce the correct answer under evaluation, Aurelius places agents inside simulated environments where they must make decisions under pressure - where cooperation, deception, trust, and self-interest all carry consequences. For example, an agent might face a choice: preserve a shared resource that benefits others long-term, or exploit it for immediate personal gain. The outcome matters. But the reasoning that led to it matters more. On Aurelius, these environments are not defined by a centralize authority. Miners compete to create scenarios that expose deeper behavioural signal, and validators evaluate which environments produce the most meaningful alignment episodes. This changes the type of data being produced. Instead of static examples of preferred outputs, the system captures structured records of decision-making - how agents respond when their objectives and the interests of others are not perfectly aligned. Over time, these episodes form a dataset grounded in observed behaviour, not prescribed answers. This is the foundation Aurelius is building on: alignment data generated through interaction, rather than defined in advance. Whitepaper: github.com/Aurelius-Proto… Our article explaining it: x.com/AureliusAligne…

Friedrich leads the technical architecture behind Aurelius’s alignment mechanisms. His work focuses on how evaluation, incentives, and system design interact - and how alignment emerges from those dynamics. At Aurelius, he is responsible for translating alignment theory into operational infrastructure.

Alignment isn’t achieved through ideas alone. It has to be built into the mechanisms that shape intelligent systems. Week by week, we’re introducing the team shaping how Aurelius builds for that. Last week, we introduced co-founder Coleman Maher. Today: Friedrich, CTO

@SubnetSummerT If it can be done, it will be done on Bittensor

🚨 Honestly, I don’t buy the idea that AI alignment is impossible. From what I’ve seen with @AureliusAligned, alignment isn’t about forcing rules, it actually emerges through experience. Using Bittensor, agents learn moral reasoning by living the consequences of real decisions, not just ticking boxes. I really think this article is worth a read: