Hook
On a crisp October morning, Andrej Karpathy—one of the few humans whose tweets can move markets and shape developer workflows—shared a simple yet profound method: speak your thoughts for ten minutes, let an AI ask clarifying questions, and receive a structured output. The crypto Twitter machine erupted with admiration. But as I read through the thread, my economist’s mind saw something else: a subtle blueprint for one of the hardest problems in our industry—governance participation. Karpathy’s "long-form oral prompt" is not about AI. It is about how we design systems where humans can contribute without the burden of precision. And that, my friends, is the missing piece in every DAO I’ve audited.
Context
Decentralized Autonomous Organizations were supposed to be the endgame of coordination. No managers, no gatekeepers, just smart contracts and token-weighted voting. Yet after four years of observing DAOs from the inside—from my early work translating the Ethereum whitepaper into Portuguese with ethical commentary, to the 2020 Aave audit that cost me 600 hours of script-level analysis—I have witnessed a recurring failure: proposals are written by the few for the few. The average token holder stares at a 2,000-word governance proposal filled with technical jargon and either ignores it or votes with the herd. The friction is not in the blockchain. It is in the interface between human thought and structured action.
Karpathy’s method attacks exactly this friction. He records a chaotic ten-minute monologue, lets the model ask follow-ups, and receives a refined document. The key insight: he transfers the cognitive load of structuring from human to machine. In a DAO, we have been asking humans to do both—generate raw ideas _and_ format them into proposals. This is why participation stays below 10% in most token-weighted systems.
Core: A Technical and Values Analysis
Let me dissect Karpathy’s three-step method through the lens of on-chain governance, using the same mental model I applied when auditing Aave’s interest rate models.
Step 1: Oral dump. The user speaks for ten minutes, jumping between ideas, half-formed arguments, and unconnected observations. This is equivalent to a community member attending a town hall and blurting out frustrations about the treasury allocation. Karpathy’s key assumption is that the AI can reconstruct a coherent goal from this chaos. For a DAO, the equivalent is a state of nature—a raw signal that must be interpreted.
Step 2: The model asks clarifying questions. This transforms the monologue into a dialogue. The AI acts as a Socratic moderator, pinpointing ambiguities and forcing the user to make trade-offs. In my experience building the "Verifiable Humanity" initiative, I saw firsthand how structured questioning could surface hidden assumptions. For a DAO, this role is currently filled by forum discussions, but those are asynchronous and sparse. A real-time AI moderator could reduce the time from idea to formal proposal by an order of magnitude.
Step 3: Structured output. The AI produces a polished document. For a DAO, this would be a governance proposal ready for voting, complete with rationale, alternatives, and quantitative impact. The raw brain dump becomes an act of code.
But here is where the blockchain twist becomes critical. Karpathy’s method relies on a closed-source, centralized model (presumably Claude or GPT-4). The AI is a black box. If a DAO adopts this workflow, it is outsourcing its most fundamental governance function—the structuring of collective intention—to a single corporation. This is the antithesis of decentralization. Code is law, but ethics is soul. The model’s inference logic must be open, reproducible, and auditable, just like the smart contracts that execute the votes.
Contrarian: The Pragmatism Test
The euphoria around Karpathy’s trick masks a dangerous truth: this method centralizes the gatekeeping of meaning. In a bull market, everyone wants a magic solution to boost engagement. But I have seen too many flashy tools become backdoors for capture. Based on my experience auditing Aave—where we found three logic errors in their interest rate model because we looked at the social contract, not just the bytecode—I know that governance is about _who_ shapes the narrative, not just _how_ votes are counted.
Let me be contrarian: Karpathy’s method, as described, would work best if the AI is trusted. But in a permissionless world, trust is a scarce resource. The real challenge is not building a model that can ask good questions—it is building a model that can prove it asked the right questions without leaking privacy. This is where zero-knowledge proofs (ZK) enter the picture. During the Verifiable Humanity initiative, we integrated ZK-SNARKs to verify human identity without exposing biometrics. The same tech can verify that an AI’s questioning process followed a transparent algorithm without revealing the raw voice data. Transparency isn’t the oxygen of trust; verifiability is.
Moreover, there is a risk of cognitive offloading the collective will. If every DAO member relies on the same AI to structure proposals, we end up with uniform thinking—a monoculture of governance. The very messiness of human interaction, the typos and the emotional outbursts, are what give DAOs their resilience. A too-perfect structuring could kill the organic debate that leads to better outcomes.
Takeaway: A Vision Forward
So where does this leave us? Karpathy has shown us a mirror: the next frontier of blockchain is not scalability or privacy, but communication geometry. We need to build infrastructure that allows humans to contribute their _essence_—their messy, raw, quick thoughts—without needing to master the grammar of smart contracts. But we must do this without surrendering the soul of decentralization to centralized AI.
The answer lies in combining three pillars: open-source large language models (like Llama or Mistral) that can run on personal hardware, zero-knowledge proofs that validate inference integrity, and a new kind of DAO plugin that turns voice into structured data through a transparent, auditable pipeline. I call this Verbal DAOs. We are already prototyping it with a small group of developers from the Lisbon open-source community. The goal is to let someone say, "I think we should spend 10% of treasury on education, but I’m worried about the effect on developer grants," and have that utterance become an on-chain discussion thread, with the AI surfacing trade-offs and inviting counterarguments.
The real test is not whether this method increases participation. It is whether it preserves theresilience that comes from authentic, chaotic human interaction. I am not advocating for AI-run governance. I am advocating for AI-augmented human governance, where the machine is a scaffold, not a decision-maker. Code is law, but ethics is soul.
If we succeed, we will have taken Karpathy’s productivity tip and turned it into the most empowering tool for decentralized coordination since the whitepaper itself. If we fail, we will have built yet another centralized hollow engine wearing a DAO mask. The choice, as always, belongs to the community.
