Anthropic and parallels with OpenAI’s Empire of AI: How a Laboratory Becomes an Institute (AI column)
Looked at the extended interview Emily Chang with Dario Amodei for Bloomberg The Circuit, published 17 June 2026 years. Over the past few months, I’ve been deconstructing large AI labs from different angles: Google agency workflowsProhibited in Russia Meta - through world models (They were driven by Jan LeCun, who had already left the company.)Thinking Machines - through interactive large models. And post Karen Hao’s book Empire of AI saw OpenAI as an organization where research, product, security, compute, capital and power are constantly pulled in different directions.
In my head, this Dario interview for Bloomberg ties these plots together - Anthropic is already hard to describe simply as a lab that trains Claude. It is a private institution: it chooses the business model, the rules for issuing models and the boundaries of working with the state, while trying to limit its own power. Amodei attributes his departure from OpenAI not only to security disagreements, but also to a loss of trust. That's his side of the story, but it explains Anthropic's focus on the company's institutional makeup.
In the interview, four layers of this design are visible.
1- Business model Anthropic is betting on the corporate market and paid subscriptions rather than advertising and maximizing engagement. Amodea’s logic is simple: if the way you earn is in conflict with your values, the company will eventually betray your values or the market. It doesn't eliminate conflict, but it changes the incentive system.
2ль Culture Amodei contrasts Anthropic organizations split into competing groups. He says he spends about half of his time explaining culture to new employees: When hired quickly from Big Tech, people recreate the way they work differently. Culture does not automatically scale with the number of employees.
3Management Anthropic is a Public Benefit Corporation, and the Long-Term Benefit Trust can appoint and remove directors. According to the company, with 14 April 2026 The directors appointed by the trust constitute the majority of the board. An entity whose members do not own an Anthropic stake has an indirect opportunity to change CEOs. This is no longer the wording of the mission, but the right in the contour of management. But Anthropic itself calls LTBT an experiment, so its quality will only be tested in a conflict of interest.
4Limits of application In a dispute with the Pentagon, Anthropic supported the use of Claude for national security, but left two red lines: mass domestic surveillance and fully autonomous weapons. Mythos was first given to defenders through Project Glasswing. The result is 271 The vulnerability found and fixed by Firefox is data from Anthropic and its partners, not an independent audit. But the mechanism itself is important: access to the model becomes part of security.
And here begins the most unpleasant part of the conversation. A private company is effectively involved in decisions about who gets a powerful cyber model, where the boundary of permissible military use runs, and what employment risks the state should consider. This is no longer just a product policy.
Amodei himself acknowledges the problem: AI was the first technology of its scale to be built by the private sector while the state was late to connect. He calls the situation dangerous and unstable and proposes a system of checks: a trust within the company, mandatory checks of models before release, legislative and judicial control, as well as mutual restrictions of the state and laboratories.
Here the interview rhymes with "Empire of AI." In Karen Hao’s book, OpenAI nearly breaks the tension between Applied, Research, and Safety, amplified by money, infrastructure, and personal power. The book shows how the Safety clan almost completely gets up and goes to found Anthropic, in which they offer a different answer: a single culture, a corporate model, a separate trust, formal red lines and the gradual release of dangerous opportunities.
But Anthropic did not end up out of the system described by Hao. It still needs huge compute, money, and partnerships with Amazon, Google, and other infrastructure players. It competes for leadership, sells the product, and at the same time asks the public to trust it with risk assessment. The difference is not in the absence of contradictions, but in what mechanisms the company has built so that these contradictions are not resolved only by the will of the founder.
Therefore, the interview should be read as the position of the CEO, not as proof of the special virtue of Anthropic. The real test is whether the company maintains the restraints on the price of leadership; whether the trust can stop management; whether there is an external control independent of the seriousness of Amodea.
As a result, we see large AI labs working on Conway’s Law and organizational architecture becoming part of the product architecture. And the declared values work only when they are turned into management rights, contracts, criteria for release, quality control. (evals) and understandable prohibitions. Otherwise, it is just a good speech of the head.
Empire of AI asked what happens when a lab becomes an empire. Anthropic tries to answer that such a laboratory should be turned into an institution with its own system of checks. Whether it works is still an open question.
#AI #Anthropic #Research #Management #Architecture #Governance