The bear market didn’t silence the builders—it clarified the architects. And in the realm of artificial intelligence, the clarity is brutal: the battle for open vs. closed is not about code, but about sovereignty.
Last week, Anthropic CEO Dario Amodei stepped into the center of a storm. While OpenAI, Google, and even SpaceX signed a petition calling for open-source AI as a public good, Amodei stood alone. He didn’t sign. Instead, he proposed three alternative measures: restrict advanced chip exports to China, crack down on industrial-scale model distillation, and mandate safety testing for any model above a certain capability threshold.
We don’t need to guess his motives. The pattern is familiar to anyone who has watched DeFi protocols lock liquidity to protect against vampire attacks. Amodei is building a walled garden—and wrapping it in the language of safety.
Context: The Open-Source Prayer and the Missing Scroll
The petition, signed by giants, argues that open-source models democratize AI, stimulate competition, and allow independent deployment. It’s the same logic that made Ethereum’s permissionless composability a revolution. But Anthropic, the company founded by former OpenAI employees who left over safety concerns, sees a different picture.
Once weights are published, Amodei argues, safety guards can be removed. The model cannot be revoked. It’s like releasing an unkillable smart contract—except this one can write biological weapons. The bear market didn’t erase that risk; it only made the survivors more paranoid.
Anthropic’s proposal is a trinity of control: 1. Restrict the chips that make frontier models possible. 2. Prevent distillation—the process that allows leaner models to mimic the powerful ones. 3. Force all sufficiently capable models to pass a security audit before release.
Core: A Blockchain Evangelist Reads the Protocol
To a decentralized protocol PM, this looks like a governance attack disguised as a security patch.
Let me walk through these three pillars with the same lens I used when auditing Curve’s stableswap invariant in 2020. At that time, I saw how mathematical elegance could replace intermediaries. Here, I see how bureaucratic elegance could replace competition.
1. Chip Restriction as Control of the Compute Layer
Advanced chips are the hardware equivalent of a Layer 1 blockchain’s security budget. By restricting their flow to China, Anthropic effectively caps the compute supply available to competitors in a major geopolitical bloc. This isn’t about safety—it’s about supply-side gatekeeping.
Based on my audit experience, when you control the means of production, you control the distribution of power. In 2022, when the Ethereum merge moved from PoW to PoS, the conversation wasn’t about emissions—it was about who owns the validators. Anthropic is doing the same with ASICs and GPUs.
2. The War on Distillation
Distillation is to AI what liquidity mining is to DeFi. It’s the tool that allows smaller players to bootstrap a high-quality model without burning billions in training costs. Anthropic wants to ban “industrial-scale” distillation, which is like Uniswap proposing to ban all but the largest liquidity providers.
The bear market didn’t kill innovation; it forced rationality. But rationality for incumbents means protecting margins. Distillation erodes the API pricing power that companies like Anthropic rely on. By labeling it a safety risk, they turn a commercial threat into a moral imperative.
3. Mandatory Safety Testing as a Regulatory Moat
Amodei calls for tests on cyberattacks, biosecurity, and alignment for all “powerful enough” models. Sounds reasonable—until you ask: who defines “powerful enough”? Who runs the tests? And who decides when a model passes?
In the language of my world, this is a multisig with keys held solely by the existing power structure. Just as DeFi “blue-chip” protocols often set their own security standards, Anthropic is writing the rulebook for a game only they can win.
About Me: I spent 2020’s DeFi Summer optimizing yield strategies, only to watch 90% of them fail when token incentives dried up. That taught me that APY isn’t adoption—it’s subsidy. Analogously, Anthropic’s “safety” isn’t safety—it’s a moat.
Contrarian: The Unseen Blind Spot
Here’s the counter-intuitive part: Anthropic might be right on the technical merits, but wrong on the human incentives.
I agree that once weights are public, safety controls are reversible. I’ve seen the same with smart contracts—once a vulnerability is public, you cannot un-exploit it. But the solution isn’t to ban public code. The solution is better formal verification, community audits, and risk disclosure.
What Anthropic proposes is the equivalent of saying “we should ban open-source smart contracts because someone might fork them and steal funds.” That argument has been made in crypto for years. It failed because the community chose transparency over control.
Moreover, chip restrictions and distillation bans will not stop bad actors—they will merely push them underground. Any sufficiently motivated state can produce chips or distill models in secret. The only people affected are legitimate researchers, startups, and small innovators.
The bear market didn’t create this dilemma; it exposed it. When capital flees, the safest bet is to own the infrastructure. Anthropic is betting on that.
Takeaway: The choice between open and closed is not about technology—it’s about trust.
If we delegate safety testing to a single company whose incentives align with monopolizing the market, we are not building resilience. We are building a single point of failure.
The question for the AI community is the same one that haunts crypto: who holds the keys? If the answer is “Anthropic,” we must ask—what happens when the protocol isn’t transparent? What happens when the safety oracle fails?
We don’t need to ban open-source to be safe. We need better mechanisms—for verification, for decentralization, for resilience.
The bear market didn’t kill the bull; it separated the believers from the tourists. And in the AI world, I remain a believer—in open, auditable, composable systems where power is distributed, not locked in a vault.
Let’s not trade one doorkeeper for another.