On a Tuesday morning in May, a Slack message from a post-training researcher named Shaun fractured the silence inside Anthropic’s Melbourne office. He argued that keeping Claude’s weights locked was a mistake. The CEO, Dario Amodei, had just published a blog post reaffirming the company’s closed-source stance, warning that “model weights, once released, cannot be recalled.” But Shaun, backed by a growing internal petition, accused leadership of trading long-term collective defense for short-term control. This is not a culture war. This is a philosophical schism over the very definition of security. And it is happening inside the one company that built its brand on being the responsible AI lighthous.
Anthropic was founded by defectors from OpenAI who believed safety required a constitution—a set of explicit rules hard-coded into the model’s alignment. Their flagship product, Claude, is a conversational AI that competes with GPT-4. But unlike Meta’s Llama series or Mistral’s open offerings, Anthropic has never promised to release any of its model weights to the public. Amodei’s reasoning is simple: capabilities are scaling faster than safety measures, and a sufficiently powerful model in the wrong hands could cause catastrophic harm. Yet inside the company, this logic is being questioned by the very people who build those safety networks. Shaun, a post-training specialist whose job is to harden Claude against adversarial inputs, believes that open sourcing actually increases security through distributed auditing—the same argument that underpins the open-source software movement.
The core of the conflict is not about ethics. It is about architecture. Amodei’s safety model depends on add-ons: adversarial training data, secret reward models, and red-team playbooks. These are peripheral to the weight matrix itself. OpenAI once thought this was enough—until GPT-3’s weights were leaked and fine-tuned to bypass all filters within hours. Shaun’s faction argues that true safety must be baked into the model from the ground up, making it resistant to removal. But that requires a fundamentally different design, one that Anthropic has not yet demonstrated. The CEO’s paralysis—refusing to open, and failing to harden—leaves the company in a strategic dead zone.
Based on my own simulation work in cross-border payments, I saw the same pattern: closed-rail banking systems talk security but bleed inefficiency. When I modeled SWIFT against ERC-20 stablecoins, the open rails showed 40% cost advantage despite higher volatility. The same principle applies here. An open-weight model allows thousands of independent auditors to find flaws invisible to a single trust team. The marginal cost of new auditors is zero. The marginal security gain is exponential. Amodei’s fear of weight loss is legitimate, but his solution—walled gardens—ignores the network effects of collective defense.
The contrarian angle: Open-source is not a free lunch. Every public weight release is a permanent gift to malicious actors. Even if Shaun’s distributed audit works, the interval between disclosure and patch is a window of extreme vulnerability. In crypto terms, this is the difference between a permissioned ledger and a public chain. The latter is immutable once polluted. Anthropic’s internal debate mirrors the Ethereum vs. Bitcoin split in 2016: one camp wanted code to be law, the other wanted emergency brakes. Both were right and wrong. The real blind spot is that neither side accounts for the economic incentives of the downstream ecosystem. If Anthropic opens Claude, it loses its premium API pricing—the only revenue engine currently funding safety research. The company would become a charity overnight, dependent on donor funding for alignment work. That is not sustainable.
The takeaway: The outcome of this civil war will determine whether the AI industry adopts a centralized trust model or a decentralized verification model. The former is faster but fragile; the latter is resilient but chaotic. For institutional investors and sovereign wealth funds watching from the sidelines, this is not a technical debate—it is a risk assessment. If Anthropic fractures, expect a wave of brain drain toward open-weight startups. If it holds, expect regulatory capture through closed APIs. Either way, the era of unified safety consensus is over. The next frontier is not better models. It is better governance—with code as the contract, and liquidity as the judge.
