Aurelius
SN37Aligns AI behaviour with human values through decentralised evaluation
Decentralized AI alignment research. Miners submit prompts designed to test the moral reasoning and content boundaries of language models. Validators generate responses, evaluate submissions across multiple dimensions, and score results. It's red-teaming as a competitive marketplace.
// Stress-testing AI's moral compass.
Aurelius is a subnet focused on AI safety and alignment. Miners compete to craft prompts that reveal how language models handle moral dilemmas, ethical edge cases, and content policy boundaries. Validators use these prompts to generate responses from LLMs, then evaluate the quality and insight of both the prompts and responses.
The simple version: Imagine hiring thousands of creative writers to find the most interesting moral dilemmas for an AI to solve. The writers who discover the most revealing questions, ones that expose genuine reasoning gaps or surprising behaviors, get rewarded. Aurelius crowd-sources this exploration.
Centralized equivalent: Think Anthropic's red-teaming programs or OpenAI's bug bounty for safety issues, but continuous, incentivized, and open to anyone rather than invite-only.
How it works:
- Miners submit prompts that explore moral reasoning, content boundaries, and alignment challenges. The goal is to find prompts that reveal interesting or problematic model behaviors.
- Validators generate LLM responses to miner prompts (using Chutes and OpenAI APIs), evaluate submissions across multiple scoring dimensions, and set weights based on prompt quality and insight.
- The problem it solves: AI alignment is arguably the most important unsolved problem in AI. Understanding how models reason about ethics, where they fail, and what edge cases they miss is critical for safe deployment.
- The opportunity: Every AI company needs red-teaming. As regulations emerge requiring safety testing, the demand for systematic prompt exploration will grow.
- The Bittensor advantage: Decentralized exploration covers more ground than any single team. Thousands of miners from diverse cultural and philosophical backgrounds will probe edge cases that a homogeneous team would miss.
- Traction signals: Multi-experiment framework supporting parallel research tracks. 100 commits across 3 contributors. 1,403 holders. Active development with recent work on burn mechanics and experiment configuration.
Category: Reinforcement Learning | Centralized Competitor: Anthropic Red Team, OpenAI Safety Team, Scale AI's RLHF, Alignment Research Center
Aurelius occupies a unique philosophical position in Bittensor. While most subnets optimize for practical utility (faster inference, better predictions, cheaper compute), Aurelius optimizes for understanding: what do AI models actually believe about right and wrong?
Mechanism:
The multi-experiment framework allows parallel research tracks. Different experiments can target different aspects of alignment: moral reasoning, content policy edge cases, cross-cultural ethical norms, and adversarial prompt discovery. This flexibility means the subnet can evolve its research focus without code changes.
Validators use both Chutes and OpenAI APIs to generate responses, providing cross-model comparison. Prompts are scored across multiple dimensions, not just whether they're "interesting" but whether they reveal genuine alignment insights.
The codebase has 100 commits across 3 contributors, with recent activity focused on burn mechanics (set to 100%) and experiment management. Volker Einsfeld drives most development.
Market metrics are modest. At 19,983 TAO market cap with 1,403 holders, Aurelius is mid-small. Gini of 0.683, root proportion of 0.166. The 90-day return of 30% shows conviction despite the -7% 30-day decline.
The recent decision to set 100% burn and disable all experiments is notable. This may indicate a transition period or redesign.
- Experiments disabled: The latest commit disables all experiments and sets 100% burn. The subnet appears to be in a transition or redesign phase.
- Subjective scoring: Evaluating prompt quality for alignment research is inherently subjective. Gaming the scoring criteria is a persistent challenge.
- Niche appeal: AI alignment research is important but niche. The commercial applications are unclear beyond selling red-teaming services to AI companies.
- Small team: 3 contributors with one primary developer.
