Building the Anthropic Team: Daniela Amodei on Hiring 10x AI Engineers | Frameworks for Growth

Thank you so much for having me.

First of all, so great to be here today. Thank you for taking the time to chat with me. I think we really focus on how to develop our tools and systems in a way that is robustly safe across our stack. What I mean by that is we apply technical training techniques to how we actually train our models to ensure that they are aligned with human values. We use a technique called constitutional AI that helps us to imbue Claude with a sense of ethics. There are a number of different documents that went into helping to train Claude with constitutional AI, including things like the UN Declaration of Human Rights. In terms of how we ensure that the technology is safe when we release a new model or a new set of products, we have a framework called the responsible scaling policy. Our responsible scaling policy is essentially a set of guidelines that indicate how and when we will release models to ensure they're safe. We use different levels of AI safety levels, there are five of them, and they're modeled along the lines of biological safety levels for people who do research with biological applications. What we look for is defining a concrete set of safety and security requirements at every ASL level before we release a model.

I think this is one of the most interesting and unusual things about running a company like Anthropic. We're almost part research lab, part tech startup, and we have a huge emphasis on our public benefit mission as well. Thinking about how to balance all three of those things in service of the mission and in providing value for our customers is one of the great joys and excitements of getting to work at a place like Anthropic. In terms of the research to product balance, what we've found is that having this collaboration between product and research really helps to make both of those teams much stronger. So much of what we see is that as new research capabilities are developed, it opens up new avenues in the product that wouldn't have existed before the technology advanced. A great recent example is Claude has been particularly useful for developers in writing code. Many of the products we've developed over the past two to three months have been geared around making that process easier for software developers, whether at a big enterprise, a mid-market company, or a budding startup. If we hadn't known the models were going to be really good at coding, we wouldn't have designed parts of our product that way. There's an iterative process where as we get more customers using us for coding, we're also helping to train the models to make them stronger in coding and then providing more product support in that area.

Glad to hear. Happy customer.

When we were starting Anthropic, we spent a reasonable amount of time thinking about what kind of structure we wanted to incorporate as. We had a bit of a benefit of having seen different models up close. What we were thinking when we incorporated as a PBC was a very practical way to balance both the commercial interests we foresaw we would have with the public benefit intentions behind founding a company like Anthropic that was so invested in topics like safety. We ultimately decided that public benefit corporations are a really nice blend of those two areas. It's a C corporation, so it is structured in most ways very similarly to a Delaware C corporation. The one exception is that we have the ability to balance commercial needs against public benefit interests. This felt like a really pragmatic approach to what we were trying to do. I would say depending on what the goal of the startup is, public benefit corporations are worth looking at if there's a component of your business that really aims to balance the public interest with some of the commercial activities a normal business engages in.

This is probably not surprising as an early employee and founder, but I think so many founders and early employees talk about culture and how important culture is when you're growing a company, particularly one scaling rapidly. I want to echo that I think the most valuable currency you have when you're a small and quickly scaling company is the culture of that company. Culture is really an expression of the company's values. It can feel hard and fuzzy to pin down what your culture and values are, but the effort to invest in that process early is crucial. The biggest lesson I've learned at Anthropic is that you should be as open and communicative about what your culture is as possible, because the people excited about that set of values and vision will be attracted to you. For people who that's not their cup of tea, it's better to learn that early. You want them to read your website and say, 'That sounds nice, but that's not what I'm interested in.' It's a delicate balance of wanting to attract the right kind of people and having an extreme emphasis on ensuring cultural alignment in your early employees.

Number one, if your values are not creating tension, they're probably too bland. If there's not an intentional tension to them, they're not specific enough. Second, you need to be able to back up any value with ten examples off the top of your head. A value we have is 'do the simple thing that works.' This comes from our research team saying we're not here to have fancy algorithms for no reason; we're trying to make the models as useful as possible. There are no extra points for being complicated. That value applies outside of research—in product, engineering, office selection, how we set up offsites. Anytime a value is one you can point to examples of easily, it's more likely to be sticky because people will run into it and feel that value.

Writing them down and talking about them in public forums within the company is really valuable. In our all hands, we try to mention a value every all hands. We don't always do it perfectly, but emphasizing and making the language comfortable is key. If it resonates with people, they volunteer it themselves. I'll hear other employees say, 'We considered hosting this offsite but decided to host it at the office because we wanted to do the simple thing that works.' I light up when I hear that because it means the value is practical for people. Also, don't be afraid to retire values if they're not working. Sometimes something feels right as a founder but isn't sticky. Feel comfortable iterating until someone says it lands better. Some of our values didn't come from the founders; they came from the company. Someone started using a term and we realized it's a value. That helps employees feel ownership.

There are some funny things from our earlier days around swag. At an offsite two or three years ago, Dario was talking about scaling laws and how quickly everything was going to grow in AI. Even those of us who believed it, looking back two years, it's so much crazier to see the rate of growth and adoption. At the end of his talk, he said something like, 'Things are never going to be chill again. They'll never be.' Now we have cute sweatshirts that say 'Things will never be chill again.' We give them to new parents for their babies. Sometimes they just emerge, but Dario wasn't intending for that to be a meme. People thought it was funny and accurate, and there were nice ways to apply it within the business.

There are a few. 'Move fast and safe things' I love, but it felt a little too cheeky for us. We adjusted that one because it was sticky sometimes but not everyone loved it. If not everyone loves it, people won't use it.

One of the most important qualities, especially at Anthropic, is a willingness to talk about a wide range of topics respectfully. We work on a cutting-edge frontier technology, and many things that bump into it can feel complicated or controversial. AI has incredible potential to help people and also to be misused. We view addressing those challenges as our core job. If people are not willing to have respectful disagreements and be open-minded, it's a problem. We assess for this in our culture interview. We ask people what unusual beliefs they hold and how they've defended those beliefs in uncomfortable situations because they felt it was the right thing. We're not looking for a particular belief system; we're looking for the ability to say, 'I felt this was correct and maybe it wasn't popular, but I stood by it,' or 'I changed my mind because I realized I wasn't right.'

We give a little preamble about why we're asking these questions, saying this is supposed to be bidirectional. We're trying to show what Anthropic is like and give insight into what it's like to work here. For most questions, there's no right or wrong answer. We want to understand their thought process and honesty. They don't have to prep; we're just trying to understand their thinking.

You don't want to lead them on, but context is helpful so they're not wondering why you're asking.

For both engineers and research, we're always in a dance of when to invest in fundamental research and when to launch and iterate. At a place like Anthropic, it's challenging because sometimes investing in fundamental research that doesn't have immediate returns is the best way to create breakthroughs. I wish I had a perfect formula; it's more art than science. On the research side, we try to diversify time between optimizing particular qualities and open exploratory research. On the engineering side, we want to give scope for people to rethink products. Because research develops quickly, you don't always know what's coming. We want to give people space to reimagine how we provide value, while also setting a product roadmap and doing the unsexy engineering work. It falls somewhere in the middle, like 60/40 or 50/50, with periods of sprinting toward a release.