Hook
On March 12, 2025, venture capitalist David Sacks posted a thread on X (formerly Twitter) that cut through the usual AI safety theater like a scalpel. His claim: Anthropic, the $18 billion AI safety startup, is actively lobbying for regulations that would cripple open-source AI models. The accusation hit a nerve. Within hours, the post had over 1.2 million views. Sacks didn't just fire a shot—he opened a new front in the war between closed and open AI. But the real story isn't about one man's opinion. It's about how the mechanics of regulatory capture, long studied in antitrust law, are now being stress-tested on the blockchain of AI governance.
Context
David Sacks is not a random critic. He is a co-founder of Craft Ventures, a former COO of PayPal, and a prominent voice in the libertarian wing of Silicon Valley. His portfolio includes investments in open-source infrastructure and decentralized protocols. Anthropic, founded by former OpenAI researchers Dario and Daniela Amodei, positions itself as the safety-first alternative to OpenAI. Its flagship model, Claude, undergoes rigorous alignment training. The company has publicly advocated for AI regulation, including mandatory safety testing and licensing for frontier models.
But Sacks's argument is that this advocacy is not altruistic. It's a strategic move to raise the compliance bar so high that only well-funded, venture-backed firms can clear it. "Anthropic is using the safety narrative to build a moat around their closed-source API," he wrote. "Call it regulatory capture. The math doesn't lie: strict rules kill open-source because open-source can't afford the legal team."
The timing is crucial. The European Union's AI Act is nearing finalization, and the U.S. Congress is considering the Bipartisan Framework for AI. Lobbying records show that Anthropic has spent over $2.3 million on federal lobbying in 2024 alone, focusing on issues like model disclosure and liability for derivative models. Meanwhile, Meta's Llama 3 is being downloaded over 100 million times, and Mistral AI just raised $600 million. The open-source ecosystem is growing, and it's scaring the incumbents.
Core
Let's strip away the noise and look at the code. I've spent the past six years auditing smart contracts and zero-knowledge proofs. I've seen how governance models that look decentralized on paper become centralized in execution. The same pattern applies here. Regulatory capture is a smart contract that executes perfectly—it just doesn't check for fairness.
The Mechanism of Capture
Anthropic's proposed regulatory framework, as outlined in its public policy papers, includes three key requirements:
- Mandatory safety audits for any model exceeding a certain compute threshold (e.g., 10^26 FLOPs).
- Disclosure of training data and model weights to a government body.
- Liability for downstream harms caused by fine-tuned versions of the model.
At first glance, these sound reasonable. But let's stress-test them.
Safety audits cost between $50,000 and $500,000 per model, depending on the scope. For a company like Anthropic, that's a rounding error. For a student team fine-tuning Llama 3 on a college budget, it's prohibitive. The audit requirement effectively creates a license to operate—and only the incumbents can afford the license.
Disclosure of model weights is the nuclear option. Open-source models thrive on transparency. If weights must be handed over to a government regulator, the open-source community loses its ability to fork, modify, and distribute. Worse, it creates a single point of failure: a government database that could be hacked or weaponized. Smart contracts execute. They don't negotiate. Once the law is written, the open-source community can't fork it.
Liability for downstream harms is the most insidious. Under current law, if someone fine-tunes an open-source model to generate hate speech, the original creator is not liable. Anthropic's proposal would change that. This would force every open-source model distributor—from Hugging Face to individual developers—to implement costly guardrails or risk lawsuits. The chilling effect is immediate. Community governance would collapse under the weight of legal risk.
Empirical Code Verification
I've seen this play out in DeFi. In 2022, I audited a lending protocol that claimed to be fully decentralized. The team had implemented a "circuit breaker" that could pause all withdrawals during market volatility. The code was technically sound. But the governance model allowed a single multisig wallet to trigger the breaker. Liquidity is an illusion until it's not. The same principle applies here: a regulatory framework that claims to protect everyone but is designed by a single influential player is a vector for capture.
Anthropic's safety-first narrative is the circuit breaker. It sounds good. It prevents worst-case scenarios. But it also concentrates power. When the breaker is pulled, it's not the regulators who decide—it's the companies that wrote the rules.
Contrarian
But here's the uncomfortable truth that neither Sacks nor his critics like to admit: open-source AI is not innocent. It has real safety problems. In my research on AI-agent smart contract interactions, I've seen how a fine-tuned Llama model can be used to craft phishing emails or exploit reentrancy vulnerabilities. The open-source community has no equivalent of a bug bounty program for model alignment. The risk is genuine.
So the contrarian angle is not that Sacks is wrong about regulatory capture—it's that his response is incomplete. He frames the debate as a binary choice: regulate open-source to death, or let it run wild. But there is a third path: technical solutions that don't rely on regulation. For example, zero-knowledge proofs can be used to verify that a model's weights have not been tampered with, without revealing the weights themselves. Differential privacy techniques can limit the harm of fine-tuned models. These are the tools that should be in the spotlight, not the lobbying teams.
Anthropic has a point: safety is hard. But using safety as a wedge to close the market is not safety—it's strategy. The real blind spot is the assumption that only regulation can solve alignment. Smart contracts execute. They don't political. If we keep looking for answers in Washington, we'll find them in the pockets of the highest bidder.
Takeaway
Sacks's accusation is a stress test for the AI governance narrative. If the industry responds by doubling down on regulatory walls, then capitulation is the only outcome. But if developers and investors start demanding cryptographic verification of safety practices—transparent, auditable, and resistant to capture—then the field might actually mature. The next year will tell us whether AI governance becomes a decentralized protocol or a centralized custody service. The math doesn't lie. But the code does. We choose which one to follow.