Before GPT-3 had even started training, Dario Amodei worried it might already be close to AGI. He wrote that fear on an offline computer at home, printed it out, and handed it to colleagues. No Google Docs. No cloud trail. Just a single point of failure wearing a CEO badge.
That paranoia delayed Microsoft’s $1 billion investment in OpenAI by several months. A former OpenAI executive called the safety team led by Dario a “priesthood.” In crypto, we call this a centralized oracle failure—a single authority deciding the truth without verifiable data.
Context: The Anthropic Model
Anthropic is the most aggressive frontier AI company founded by a CEO who genuinely believes AI could destroy the world. Dario Amodei is not a typical tech founder. He refuses to travel to China for fear of being kidnapped. He holds bi-weekly all-hands meetings that employees call “Dario Vision Quest,” where he delivers long monologues on AI, politics, war, and the future of humanity. The company employs a group of economists specifically to model GDP and unemployment after the singularity arrives. One employee said Dario “always has the singularity on his mind.” A major investor commented: “He is less of a CEO and more of a religious leader.”
This is not a startup. It is a cult of personality wrapped in a research lab. And the security industry—especially blockchain security—has seen this pattern before.
Core: Systematic Teardown of Anthropic’s Governance Geometry
Zero trust is not a policy; it is a geometry. The trust model of a system must be distributed, verifiable, and auditable. Anthropic’s governance model is the opposite: a single point of trust concentrated in one human brain.
Let’s examine the geometry. Dario’s safety team delayed a billion-dollar investment based on a feeling—a fear that GPT-3 might be too close to AGI. No on-chain proof. No slashing condition. No decentralized dispute resolution. Just a printed memo from an offline computer. The code does not lie, but it often omits. In this case, the omitted data was the entire decision-making process.
During my audit of the 2x2x4 protocol in 2017, I found a reentrancy vulnerability that allowed infinite borrowing. The project team wanted to bypass security for speed. I published my findings on GitHub. That was verifiable. That was a public good. What Dario did was the opposite: he made a private, subjective judgment that affected millions of dollars and delayed a major investment. There is no way to audit his reasoning. No transaction log. No smart contract. Just a man’s intuition.
This is the same failure mode I identified in Curve Finance’s governance in 2020. The veCRV model allowed whales to manipulate reward allocations. The “community-driven” narrative was a mask for centralization. Anthropic’s “safety-first” narrative is a mask for Dario’s personal authority. Compiling the truth from fragmented logs: in both cases, the power is concentrated in a small group that cannot be verified by external observers.
Now consider the all-hands meetings. “Dario Vision Quest” is not a governance forum; it is a sermon. The economists studying singularity GDP are not producing actionable metrics; they are producing narratives. In blockchain, we have a term for this: tokenomics without a security model. It is a design that assumes the leader will always be benevolent. Security is the absence of assumptions. Anthropic is built on assumptions about Dario’s judgment.
The Priesthood and the Oracle Problem
A former OpenAI executive described the safety team as a “priesthood.” This is exactly the term used in blockchain for centralized oracle providers. Chainlink solved the decentralization problem with centralized nodes—a joke that still holds. Anthropic’s safety team is a centralized oracle feeding decisions to the company. The latency is not in data feed; it is in human judgment. When Dario goes into the office library to watch YouTube to calm himself after clashing with Sam Altman, that is a governance failure.
Employees privately joked that Dario has “Sama Derangement Syndrome”—an obsession with Sam Altman. This is not a technical critique. It is a psychological one. But in systems engineering, psychology is a vector of attack. If the CEO’s decision-making can be influenced by personal friction with a rival, the entire system is compromised. The same way a multisig with a single signer is compromised.
Contrarian: What the Bulls Got Right
To be fair, Dario’s paranoia may be justified. AI safety is a real problem. The singularity might be near. And Anthropic has produced some of the most capable language models with a genuine focus on alignment. The bulls will argue that centralization of safety decisions is necessary because the stakes are too high for democratic deliberation. When the fate of humanity is on the line, you want a single person with a strong moral compass, not a committee that can’t agree on a definition of AGI.
They might also point out that Anthropic’s internal governance—while authoritarian—is transparent about its centralization. Dario does not pretend to be a decentralized community. He is a prophet, and he owns it. That honesty is a form of security: investors know what they are buying into.
But this argument misses the point. The problem is not centralization per se; it is the lack of verifiable accountability. In blockchain, we accept that some protocols are controlled by a foundation. But those foundations are bound by smart contracts, on-chain governance, and transparent treasury management. Anthropic has none of that. The safety team’s decisions are printed on paper and stored in a drawer. There is no way to audit whether the delay of Microsoft’s investment was a genuine safety concern or a power play against Sam Altman.
Takeaway: The Singularity Needs a Verifiable Oracle
Dario Amodei is not a religious leader. He is a single point of failure. If AI safety is to be treated as a public good—and it should be—then the governance of that safety must be mathematically verifiable, not emotionally driven. Optimism’s RetroPGF is the only truly effective public goods funding mechanism because it is retrospective, on-chain, and governed by a broad community. Every other DAO grant committee runs on nepotism. Anthropic’s safety committee runs on a founder’s anxiety.
Security is the absence of assumptions. The assumption that Dario will always make the right call is the biggest vulnerability in the entire AI safety landscape. We need a system where safety decisions are logged, verified, and slashed if wrong. Until then, the singularity is governed by a priest with a printed memo.
Compiling the truth from fragmented logs: the logs are empty. The trust model is a single point. The geometry is broken. And the market—both crypto and AI—will eventually price in that risk.