Ethereum co-founder Vitalik Buterin has pinpointed a fundamental limitation plaguing decentralized autonomous organizations (DAOs) and democratic governance systems: the scarcity of human attention. Writing on social media platform X, Buterin argued that participants are overwhelmed by thousands of decisions across multiple domains of expertise without the requisite time or skill to evaluate them properly.
The typical delegation model, where users assign voting power to representatives, creates a problematic disempowerment, Buterin contends. This leads to a scenario where a small group controls decision-making while the broader supporter base loses influence after clicking the delegate button.
Buterin proposed personal large language models (LLMs) as a key solution to this attention bottleneck, outlining four specific approaches. The first is personal governance agents, which would autonomously vote based on user preferences inferred from personal writing, conversation history, and direct statements. When uncertain on important matters, the agent would query the user directly, providing all necessary context.
He cautioned against dystopian "AI becomes the government" outcomes, stating it leads to poor decisions with weak AI and could be "doom-maximizing" with strong AI. However, he believes AI used correctly can be empowering for decentralized governance.
The second approach involves public conversation agents that would aggregate information from many participants, summarize individual views into shareable formats without exposing private data, and identify common ground, similar to LLM-enhanced Polis systems.
A third concept is suggestion markets, which would use prediction market mechanics to financially incentivize high-quality proposals. Anyone could submit ideas while AI agents bet on tokens, with payouts occurring when the mechanism accepts a valuable input.
For sensitive decisions requiring secrecy—such as adversarial conflicts or compensation—Buterin suggested a fourth method: privacy-preserving multi-party computation (MPC) using trusted execution environments. "You submit your personal LLM into a black box, the LLM sees private info, it makes a judgement based on that, and it outputs only that judgement," he explained. He emphasized that robust privacy tools, including zero-knowledge proofs, should be foundational in governance systems.
