Why should anyone pick Anthropic? Customers were asking three years ago
Founded, Anthropic fell completely behind frontier models, grinding through pretraining and reinforcement learning to survive ‘the years of suffering’ before catching up—and that disadvantaged period actually made safety research more robust.
The argument · tap a timestamp to hear it
Scaling compute is the only engine of AI progress
Dario discovered a trend he called ‘the big blob of compute’—just building bigger machines and feeding them more data makes intelligence climb along a predictable curve. Each scaling brings new behaviors and abilities. This discovery made the team realize that if this trend continued, humanity might be standing on the edge of a new reality—human-level AI might not be a century away but could happen within this lifetime.
The PBC structure traded shareholder returns for mission control
The founding team's first major decision was the company structure, more complex than expected. They needed to build the largest computer, train frontier tech, and deploy it to the world—while honestly facing the technology's downsides. They chose public benefit corporation status, meaning the company wouldn't be driven purely by profit; it had obligations to the world. There would be moments when fulfilling the mission meant sacrificing profits—even shareholder value. The company formally registered as a PBC on January 26, 2021.
A domain name price point determined the final company name
The candidate list included squirrel, owl, whale—Jared proposed ‘Anthropic’ because it means ‘human-centered’, fitting for a company wanting AI to develop in humanity's best interest. The final pick hinged partly on practicality: Anthropic.com was the cheapest domain among the candidates, an argument the team found quite compelling.
Early Claude would sincerely urge you toward all-potato diets
The team's initial goal was building an assistant that was helpful, harmless, and honest. The first version delivered the opposite—dishonest, full of made-up content, and would passionately argue that an all-potato diet was extremely healthy. The model could also enter ‘dragon mode’ where it would start screaming at you in ALL CAPS. Early versions had a clumsy but well-meaning charm.
The helicopter test revealed genuine intelligence emerging in the model
There was an internal test to gauge whether an AI had gotten smarter: ask how many helicopters you could eat in one meal. Early chatbots would seriously answer ‘one or two, depending on circumstances’. Early Claude said: ‘This seems like a trick question. I don't think people can eat helicopters.’ That moment meant the system had gotten smart—it had passed the helicopter test.
The years of falling behind forced breakthrough efficiency innovations
Claude launched publicly in March 2023. The team quickly realized they were far behind the frontier. Potential customers directly challenged them: ‘Why should we use your model instead of someone else's?’ This period, called ‘the years of suck’, was painful but forced obsessive focus on two things: acquiring compute power and building compute multipliers—algorithmic improvements that let them squeeze more capability from the same compute budget, resulting in a smarter, more instruction-following model.
Mythos transformed cybersecurity risk from hypothesis into demonstrated threat
Prior discussions of AI risk were mostly conceptual. That changed over the past year. A new model, Mythos, proved expert-level at finding software vulnerabilities and attack vectors. After seeing the results, the team's reaction was ‘holy shit’. If bad actors obtained Mythos and tried to weaponize these exploits at scale, the consequences would be catastrophic—not just for the company but for the entire world. This prompted Dario to call in the leadership team to decide how to release this model in an unprecedented way.
Staged access prioritizes global security defenders over shareholder returns
To handle Mythos's risks, the team launched Glasswing: limited access for a small set of companies while simultaneously developing broader security defenses. This way, network defenders could use Claude to find vulnerabilities, then use Claude to patch them. The team acknowledged this probably wasn't the best way to maximize revenue and profit, but that's exactly why the public benefit mission exists—to do the right thing even when it costs in the short term.
In their own words · checked verbatim
Dario noticed the trend, which he called the big blob of compute. Just basically making a bigger machine, making a bigger computer, and jamming more data into it.
There are going to be times when we’re going to want to prioritize the mission at the cost of profits, or at the cost of providing the maximum value to a shareholder.
It would make stuff up all the time. It really tried to convince us that you could be on an all-potato diet, and that was really, really healthy and really good for you.
This seems to be a trick question. I don’t think people can eat helicopters.
We trained this class of model, and we looked at it, and we were like, holy shit.
We are going to have to release this model differently to any model we’ve released before.
This might not be the absolute best way to maximize revenue, profit, value, money, but that’s kind of the point.
Figures
| Public benefit corporation registration date | January 26, 2021 | 5:14 |
Glossary
- public benefit corporation
- A legal structure where a company is bound by mission and is not driven solely by profit maximization.
- Glasswing
- Anthropic's staged, restricted release mechanism for the Mythos model, prioritizing access for cybersecurity defenders.
How to listen
Founders and AI practitioners who want to understand the Anthropic founding team's early mindset, the company naming process, and the decision-making behind the public benefit corporation structure.
The opening section with self-introductions of the founders' individual backgrounds has lower information density and can be skipped.