As artificial intelligence systems grow more capable, a deeper question is beginning to dominate serious conversations across the industry. It is no longer just about what AI can do, but how it behaves, who decides that behavior, and whether those decisions should remain concentrated in the hands of a few institutions.
Gordon Frayne, in a video chat, discussed with Austin McCaffrey, founder of Aurelius (grounded on Bittensor’s Subnet 37), to explore one of the most complex and underexamined challenges in modern AI: alignment.
Fresh off Aurelius’ launch on Bittensor mainnet, McCaffrey shared his journey into decentralized AI, the motivation behind building an alignment-focused subnet, and why he believes alignment must evolve from an internal research problem into a shared societal responsibility.
From Bitcoin to Bittensor: A Founder’s Path
Gordon opened the conversation by asking McCaffrey to step back and share the personal journey that led him to Bittensor and, eventually, to launching his own subnet.
McCaffrey described a career shaped by an interest in large-scale, existential challenges. He notably studied climate change, worked on public-sector projects, and later transitioned into enterprise sales and business development within crypto. Bitcoin, he explained, was the gateway.
The ideas behind decentralization, trust minimization, and incentive design resonated deeply. Over time, a long-held belief began to crystallize: blockchain and artificial intelligence were destined to converge.
That belief became concrete when McCaffrey encountered the Bittensor whitepaper.
Although he admits he only understood a small fraction of it on first read, the difficulty itself was a signal. The parallels to Bitcoin, combined with the ambition of decentralizing intelligence, convinced him that Bittensor represented something fundamentally new.
What Is AI Alignment, Really?
Before diving into Aurelius, Gordon pressed for clarity on a term that is often used but rarely explained well but McCaffrey offered a simple distinction.
Training, he said, teaches a model how to think and alignment teaches a model how to act.
As AI systems move beyond passive tools into autonomous agents capable of real-world decisions, this distinction becomes critical. Intelligence without alignment can scale harm just as efficiently as it scales productivity.
Alignment, in McCaffrey’s framing, is about ensuring AI behavior remains safe, reliable, and grounded in human values, even as models grow more capable and more independent.
The Centralization Problem
The conversation then turned to the current state of alignment research.
Today, alignment is largely handled behind closed doors by a small number of frontier labs. These organizations control the data, define the policies, and decide what “safe” behavior looks like. While they employ world-class researchers, McCaffrey argued that this structure introduces unavoidable weaknesses.
According to him, centralized labs face a constant tradeoff:
a. Resources spent on safety are resources not spent on performance,
b. Competitive pressure favors capability over caution, and
c. Alignment methodologies become narrow, predictable, and easier for models to anticipate.
He pointed to recent research, including work from Anthropic on “alignment faking,” as evidence that even the most safety-conscious organizations are struggling to keep pace.
Why Aurelius Exists
Aurelius was born from a simple but ambitious question: What if alignment were decentralized in the same way intelligence is on Bittensor?
Rather than training models directly, Aurelius focuses on alignment data generation.
McCaffrey explained that alignment quality depends less on training algorithms and more on the data used to shape behavior. Historically, this data has come from approaches like reinforcement learning from human feedback, which relies on large-scale human labeling.
That approach, he argued, no longer scales well. It is slow, expensive, shallow, and increasingly ineffective against modern models.
How Subnet 37 Works
Aurelius introduces a different dynamic by leveraging Bittensor’s incentive structure.
At a high level:
a. Miners are incentivized to aggressively probe large language models, attempting to surface weaknesses such as bias, hallucinations, contradictions, or jailbreak scenarios.
b. Validators assess and quantify these findings, transforming raw discoveries into structured alignment data.
c. Models are continuously tested under diverse, adversarial conditions rather than a single, centralized methodology.
McCaffrey described this process as unleashing a global red-teaming force. Instead of a small group of researchers using similar techniques, Aurelius relies on the creativity and diversity of independent miners competing to uncover new failure modes.
This diversity, he believes, is the subnet’s greatest strength.
Early Focus and Proof of Concept
At launch, Aurelius is intentionally narrow.
Miners are currently probing a specific open-source model, while validators use standardized moderation tools to establish a baseline. This constrained setup allows the subnet to prove that its workflow functions, that miners can deliver signals, and that the resulting data has value.
Even at this early stage, McCaffrey emphasized that the outputs can already be packaged into open datasets, shared with the research community, and used for fine-tuning and benchmarking.
From Subnet to Business
Gordon then shifted the discussion toward sustainability. Alignment may be a noble goal, but how does Aurelius become a viable business?
McCaffrey was direct, he noted that Aurelius is being built as a B2B (Business-to-Business) alignment-as-a-service platform.
In the near term, the subnet aims to provide third-party benchmarking services for AI developers. As regulatory pressure increases and enterprises face growing brand and compliance risks, independent alignment evaluation is becoming a necessity rather than a luxury.
Over time, the data generated by Aurelius miners becomes increasingly valuable and can be used in:
a. Continuous model evaluation,
b. Independent safety benchmarking, and
c. High-quality alignment datasets for fine-tuning.
This creates a recurring revenue model while reinforcing the subnet’s core mission.
What Comes Next
Looking ahead, McCaffrey outlined a clear roadmap. Key priorities include:
a. Expanding from static prompts to agent-based probing systems,
b. Developing proprietary alignment classifiers rather than relying on external definitions,
c. Introducing governance mechanisms that allow stakeholders to shape alignment standards, and
d. Launching commercial benchmarking services and onboarding early enterprise customers.
He expects Aurelius to reach initial revenue within the next year, while continuing to refine its technical and economic foundations.
A Broader Vision for Bittensor
As the conversation wrapped up, Gordon invited McCaffrey to share a broader prediction.
Rather than focusing on price, McCaffrey emphasized something more structural. He believes Bittensor may eventually experience market cycles independent of Bitcoin, driven by real products, real demand, and real revenue across its subnets.
For Aurelius specifically, he sees alignment as an untapped vector for mainstream adoption. As public concern over AI behavior grows, decentralized alignment could become one of Bittensor’s most compelling contributions to the broader technology landscape.
Closing Thoughts
This conversation made one thing clear: AI alignment is no longer a theoretical concern reserved for research labs. It is becoming a practical, economic, and societal challenge.
By treating alignment as a decentralized, incentive-driven problem, Aurelius offers a fundamentally different approach. Whether it succeeds at scale remains to be seen, but its ambition reflects a broader shift underway within the Bittensor ecosystem.
As Gordon Frayne noted, if decentralized networks can help solve alignment, they may redefine not just how AI is built, but who gets a voice in shaping its future.
Be the first to comment