In 2020, OpenAI's head of research quit. His next move looked like a disadvantage. His name: Dario Amodei. He left the company with a small group of senior colleagues. Among them was his sister Daniela, who ran safety and policy. They walked out over a disagreement about speed. They thought the lab was shipping faster than its safety work could keep up.
Then they did something that looked, at the time, like a disadvantage.
They incorporated Anthropic as a Public Benefit Corporation. Simply put, this is a for-profit company that writes its mission into its legal charter. So, the board can choose the mission over quarterly profits. They established their mission as a legal need. Most founders treat this structure as a cost to avoid. Anthropic treated it as the product.
They trained their model, Claude, using a set of rules they named a constitution. They did this to avoid leaning only on humans labeling good and bad answers. Constitutional AI. Claude shipped in March 2023. Then, in September 2023, they published a Responsible Scaling Policy. This was a public document. It outlined the conditions for refusing to train or deploy a model. The framework is based on biosafety levels used in labs for dangerous pathogens. It has graduated tiers that get stricter as the danger increases. They also invested in mechanistic interpretability. Dario calls this building an MRI for a model. This way, you can check if it’s lying before it acts.
His stated bet was a race to the top. Set the safety bar in public, and compel competitors to follow suit.
I'll spare you the funding-round play-by-play here. Every founder who has been told depth doesn't sell should be bothered by the short version. By March 2025, investors valued Anthropic at $61.5 billion. By early 2026, it had a $350 billion valuation. Amazon and Google were committing tens of billions together. They were also preparing to go public. In the months following the scaling policy, OpenAI and Google DeepMind released their own versions. Anthropic says the policy caused this change. I'd say, more carefully, that it stopped being the weird one.
Safety was not the only reason this worked. It helped substantially that Amazon and Google were hunting for an OpenAI alternative to fund. The timing of the language-model wave was pure luck. Plenty of rigorous labs have faded into obscurity. Depth alone is not enough.
Put rigour on the front door.
Anthropic took the most rigorous, least glamorous thing it did and put it at the front door. That is differentiating the wheat from the chaff instead of niching down. Niching asks how small you can cut your market until you are the only one left. Differentiating asks who specifically needs the exact thing you can do credibly—at a level of understanding that almost nobody else can do. For a founder focused on research, it often means using the discipline you already have but might ignore. The way you care about your method. Or your discipline to not over-claim results. All the things you consider essential for good research.
When you start, you might treat that discipline as overhead. But your market treats it as scarce.
I read AI labs' safety claims carefully, because I study deception by generative models. So take this as a lesson in positioning and nothing more. I made the opposite mistake myself for years. I aimed my own work at PhD students first, because that is the obvious move for a professor. That was narrowing the market to the obvious audience. What actually worked was differentiating. I targeted the same depth at the specific people it uniquely served.
The move only pays if you make the rigour legible. This is the half founders skip. Anthropic published the constitution, named the policy, productized the discipline. Because the mission was critical and public, the depth built trust. Rigour that stays in your head convinces no one. Get it on the public record before someone with half your depth and a louder feed defines your category for you.
You win the next decade out of research when your depth is already on the public record when the customer, the recruit, and the design partner come looking. The technical founder people already trust.
Your rigour is what everyone else calls overhead.
Bonus
The full issue in The Write Insight includes the interactive Front Door Audit scorecard (12 statements measuring how legible your rigour is), a 4-phase legibility checklist, and 6 curated resources including the Constitutional AI paper and Anthropic's Responsible Scaling Policy.
