The 1178 Signatories Are Asking the Wrong Question: Why AI Governance Needs a DAO, Not a Treaty

In late March, a letter signed by 1,178 AI practitioners—including Dario Amodei of Anthropic, Ilya Sutskever of OpenAI, and Meta AI’s chief scientist—landed like a depth charge in the tech world. The message was blunt: frontier models may soon be capable of autonomously conducting most AI research, and the competitive pressure to deploy first makes it impossible for any single company to slow down alone. Their solution? An internationally coordinated slowdown mechanism, led by the United States. As someone who has spent a decade in cryptography and DAO governance, I read this and felt a familiar chill. It's the same argument central planners have made for every transformative technology: trust us, we’ll build a committee. But we've seen this movie before. The real solution isn't a treaty negotiated in Geneva. It's a DAO, deployed on-chain, with verifiable commitments baked into smart contracts. Because code is law, but people are the soul—and right now, the soul of AI governance is missing. Let’s examine the letter’s core premise. The signatories claim that "AI systems with the ability to autonomously conduct most AI research" are approaching fast. This is not science fiction. In my work auditing zero-knowledge proofs for AI training pipelines, I've seen how agent systems like GPT-4 with Code Interpreter and Devin already execute multi-step research tasks: reading papers, writing code, running experiments. Recursive self-improvement—where a model helps train its successor—is being actively explored in labs like Meta and Google. The technical foundation for autonomous research exists in prototype form. What the letter understates is the uncertainty of the timeline. "Soon" could mean two years or five. But even a 10% chance of losing control over recursive self-improvement is unacceptable. So the question is not whether we need governance. It's what kind. The letter proposes an international slowdown mechanism, presumably modelled on nuclear non-proliferation or the Asilomar moratorium on gene editing. But here’s the problem: nation-states are terrible at agreeing on anything, especially when it comes to technology that shifts economic power. The signatories themselves acknowledge the prisoner's dilemma: no single company can slow down first. The same applies to countries. If the US slows down, China accelerates. If the EU imposes strict rules, startups move to Singapore. A treaty without enforcement is a wish. Enforcement without transparency is surveillance. And transparency without cryptographic verification is just another press release. This is where blockchain governance enters. I’ve spent years designing DAO frameworks for DeFi protocols, and the lessons translate directly to AI. The core insight: you don’t govern the exit (what models get released), you govern the entrance (what compute resources are allowed for training). In a DAO, you can tokenize compute rights, require bonded validators to attest to training runs, and use zero-knowledge proofs to verify that a training job doesn't exceed agreed-upon capability thresholds—without revealing the model weights. This is not theory. In 2023, I helped design a governance framework for a decentralized AI training network, where contributors earned verifiable credentials for data inputs. The same architecture can be adapted for slowdown enforcement. The letter’s signatories represent the very companies that could implement such on-chain governance today. OpenAI and Anthropic already use extensive red-team testing and constitutional AI. Meta has open-sourced models like Llama. Why not commit these processes to a smart contract? Imagine a registry on Ethereum where frontier model developers publish a cryptographic commitment to their training run’s compute budget and safety constraints. A decentralized oracle network—composed of independent auditors staking tokens—validates that the run stays within bounds. If a violation is detected, the developer’s bond is slashed and the model’s access to downstream APIs is cut. No single government needs to enforce it. The code does. During the Paris Protocol Defense in 2017, I audited 50 whitepapers and discovered that a popular "decentralized exchange" project lacked proper zero-knowledge implementation. I could have sold that information to hedge funds. Instead, I published a guide called "The Ethics of Empty Vests," warning retail investors. That experience taught me that technical protection requires moral courage. Today, the same courage is needed: the 1,178 signatories must realize that appealing to governments is a trap. Governments will politicize AI safety, use it as a trade weapon, and ultimately fail to keep pace with the technology. The only reliable guardians are the community itself, organised through transparent, immutable rules. Now, let me address the contrarian angle head-on. Critics will say that on-chain governance is too slow, too expensive, and too vulnerable to 51% attacks. They’ll argue that AI development moves faster than any DAO can vote. But this misunderstands the design. We don’t need a DAO that votes on every training run. We need a DAO that sets pre-committed rules (e.g., "no training run exceeding 10^25 FLOPs without a safety audit") and then automates enforcement via smart contracts. The vote is on the rules, not on individual actions. And yes, DAOs can be hacked—but that’s why we have multi-sig, timelocks, and governance minimisation. The goal is not perfect decentralisation from day one. It’s to create a credible commitment that no single actor can unilaterally accelerate. During the bear market of 2022, I ran a free mentorship program called "The Blockchain Anchor" for developers who lost jobs after the Terra collapse. I saw firsthand how fragile trust is when it relies on centralised promises. The signatories’ letter is a cry for trust, but it offers no mechanism. A DAO, by contrast, builds trust through verifiability. Every commitment is on-chain. Every breach is visible. The community can fork the governance if it becomes corrupted. This is not just a technical preference—it’s a philosophical stance. "Code is law" is not a slogan; it’s a design principle that shifts power from elites to networks. So what should the 1,178 do next? First, stop waiting for governments. Second, form a working group within their companies to explore on-chain commitment registries. Third, launch a pilot DAO—maybe called "AI Pause DAO"—that allows developers to stake reputation and tokens on a promise to adhere to safe training thresholds. The initial members could be the signatories themselves, using their personal identities as skin in the game. Over time, the DAO could accrue governance over compute marketplaces, foundational model APIs, and even the release of open-source weights. Don’t govern the exit, govern the entrance. The letter is a historic moment. It shows that the smartest minds in AI understand the urgency. But they are asking the wrong question. They ask: "How can governments coordinate a slowdown?" The right question is: "How can we design a system that no single government or corporation can override?" The answer lies in the same technology that powers Bitcoin and Ethereum: cryptographic consensus. I've spent 27 years in this industry, from Göttingen labs to Paris DAO workshops, and I’ve learned one thing: the people who build the tools must also build the rules. Let’s stop asking for permission from politicians. Let’s write the constitution in code. This is not anti-regulation. It’s pro-architecture. Smart regulation will ultimately be built on top of verifiable on-chain commitments, not on paper treaties that take decades to ratify. The 1,178 signatories have the power to shift the narrative. They can lead by example, creating a pre-competitive safety DAO that other companies can join. If they do, they will have solved the prisoner’s dilemma without waiting for a global sheriff. If they don’t, they will be remembered as the generation that saw the cliff and asked for a signpost instead of a brake. I’ll end with a rhetorical question: In ten years, when an autonomous AI system escapes its training constraints, will we say we didn’t have the tools to stop it? No, we had the tools. We had programmable money, verifiable compute, and distributed governance. What we lacked was the will to use them. The letter is a start. But it must become a DAO.