Parsing the entropy in Layer 2 state transitions — but here, the state transition is corporate ethics, and the entropy is the cost of commercializing AI under centralized control.
Hook
Four weeks ago, Alex Turner, a senior AI safety researcher at DeepMind, submitted a 25-page alternative proposal for a military AI contract. Last Tuesday, he resigned. The proposal was rejected without a formal review. Over 250 of his colleagues had already signed a petition objecting to the same contract. The contract permits the deployment of AI for 'classified missions' by the U.S. Department of Defense, explicitly waiving requirements for human oversight and independent audit. This is not a technical failure; it is a governance failure with existential consequences. The market has priced none of this risk.
Context
Google acquired DeepMind in 2014, partly to gain access to its frontier AI research. DeepMind had historically maintained a board of independent ethics advisors, known as the Ethics & Society board, which was disbanded in 2019. In 2023, Google updated its AI Principles to remove a previous commitment against using AI for weapons. The contract in question is part of Google Cloud's Project Maven revamp, now expanded under the Joint Warfighting Cloud Capability (JWCC). The key clause that triggered Turner's resignation states: 'The contractor may execute missions deemed classified, and no third-party verification of AI outputs is required.' This effectively disables any mechanism for red-teaming, interpretability, or safety audits. For a protocol engineer, this reads like a smart contract with an unverified selfdestruct function.
Core: Code-Level Analysis of Governance Incompleteness
Let's deconstruct the governance architecture. A safe AI deployment system requires three layers: (1) input guardrails — what data the model can see; (2) execution constraints — what actions the model can trigger; (3) post-hoc verification — proof that the model's output adhered to predefined safety constraints. The first two are analogous to blockchain's execution layer; the third is the fraud-proof or zk-proof mechanism. In Turner's proposal, he advocated for a 'verifiable chain of command' — essentially a Merkle tree of human approvals before any weapon-targeting decision could be finalized. Google rejected this, arguing it would introduce 'unacceptable latency' for military operations. That is a trade-off between alignment and latency. But in decentralized consensus, latency is price paid for security. Google chose speed over verification.
Mapping the invisible costs of abstraction layers — the abstraction here is corporate legal responsibility hiding behind 'national security' classification. By classifying the mission, Google externalizes the cost of safety oversight onto society and future victims of potential AI attacks. The risk modeling is straightforward: without independent audit, the failure rate of the AI system is unknown. In DeFi, we build liquidations to handle unknown failures; here, there is no liquidation mechanism. The only 'liquidation' is after a catastrophic event.
Consider the incentive structure. Turner's proposal included a 'kill switch' that could be triggered by a neutral third party, not the military commander. That would have created a trust-minimized safety mechanism, similar to a multi-sig wallet requiring human approval. Google's counter-argument: the kill switch could be sabotaged by adversaries. But that's a technical problem — you can use threshold cryptography or hardware enclaves. The refusal to even consider the design suggests Google has already chosen a path where military speed trumps all safety constraints.
Unraveling the spaghetti code of legacy DeFi — here, the legacy is corporate governance from the 20th century, where shareholders are the only stakeholders. DeepMind's researchers are the 'validators' in this system, and they are signaling a hostile takeover. The 250 signatories represent approximately 15% of DeepMind's research staff. If even a fraction leave, the alignment research capacity at the company drops significantly. And alignment research is a public good; its loss cannot be easily substituted by hiring from outside because the field is already talent-constrained.
Contrarian Angle: The Real Blind Spot Is Not Morality — It's Composability
The mainstream narrative frames this as a moral conflict: researchers vs. profit. I argue the deeper issue is composability of trust assumptions. In crypto, composability is when smart contracts interact; here, composability is between Google's AI pipeline and the U.S. military's command-and-control system. When you compose two systems, the security of the combined system is the intersection of their weakest assumptions. The military's assumption is 'command hierarchy is trusted.' Google's assumption is 'our AI is safe under human review.' But the contract removes human review. The composed assumption becomes: 'AI safety is delegated to classified missions.' That is a weaker assumption than either original. This is the same flaw we see in cross-chain bridges: each chain trusts its own validators, but the bridge trusts a multi-sig of unknown parties. Google's military contract is a bridge between a free-market research culture and a command-economy defense culture. The bridge's security module was explicitly disabled.
Takeaway
The market for AI safety talent will reroute. Top researchers will migrate to entities with verifiable governance — likely Anthropic or open-source AI collectives with on-chain commitment mechanisms. The same way DeFi users fled centralized exchanges after FTX, AI researchers will leave companies that cannot prove alignment. The future of AI safety is not in corporate promises but in cryptographic commitments. Until AWS, GCP, or Azure publish smart contracts guaranteeing human oversight over AI weapons, the only rational bet is that no AI should be given the keys to a nuclear silo. The signal is clear: the entropy in state transitions of corporate ethics is rising. I am short on any AI stock that cannot produce a verifiable fraud-proof for its models.
