Who Can Stop Anthropic? The AI Safety Crisis Nobody Can Govern
OpenAI has publicly floated the idea of slowing down frontier AI development. Inside Anthropic, researchers have resigned in protest. The two events, taken together, expose a widening fracture between the commercial incentives driving large-model labs and the safety commitments they publicly espouse.
The News in Brief
OpenAI proposed a coordinated slowdown on frontier model training, while an Anthropic researcher stepped down over what they described as insufficient internal commitment to safety guardrails. The resignations and the proposal are not isolated incidents — they are symptoms of a governance vacuum that no single lab, regulator, or board can fill.
Why “Slowing Down” Is Structurally Impossible
The core problem is that AI development is not a single-actor game. Any unilateral pause by one lab simply transfers frontier capability to a competitor. Compute supply chains — Nvidia GPUs, TSMC fabrication, hyperscaler data centers — are globally distributed and commercially motivated. A slowdown pledge without enforcement is a marketing statement, not a policy.
- Incentive mismatch: Safety teams report to commercial leadership; their recommendations can be overridden.
- Regulatory lag: No jurisdiction currently has binding authority over frontier training runs.
- Self-evolution risk: Models that improve their own training pipelines compress the timeline for oversight.
- Talent churn: Safety researchers leaving major labs weaken internal dissent and external accountability.
The Crypto Parallel
This mirrors an earlier debate in decentralized systems. Bitcoin’s governance is slow and adversarial, but it is transparent. AI labs are fast and opaque. The crypto industry has spent a decade building mechanisms — multi-sig, timelocks, on-chain governance, zero-knowledge proofs — for verifiable restraint. None of these map cleanly onto model training, but the underlying question is identical: how do you credibly commit to not doing something when doing it is profitable?
Some projects are experimenting with on-chain compute registries and inference attestation, which could theoretically provide a verifiable record of training runs. Whether these mature fast enough to matter is an open question.
Forward-Looking Perspective
Expect three developments over the next 12–24 months. First, more high-profile safety resignations as labs race to ship. Second, voluntary industry frameworks that lack enforcement teeth. Third, a push — likely from the EU and possibly the US — for mandatory compute reporting above a threshold. Until then, the honest answer to “who can stop Anthropic?” is: no one with both the authority and the will.
