Generative SF: How Anthropic is building better, safer AI models

Few generative AI companies have achieved as much in a short span of time as Anthropic.

Co-founders and siblings Daniela and Dario Amodei form a powerhouse team, growing Anthropic from seven employees to more than 800+ in less than four years.

Dario is the technology visionary, helping to define the future of foundational AI models, while Daniela leads by focusing on steering Anthropic’s go-to-market strategy, leading customer engagement, and hiring top talent.

In June, the company released version 3.5 of its flagship product, Claude Sonnet, and just a few weeks ago introduced an updated version that included groundbreaking new capability in public beta, computer use.

This fall, we had the pleasure of hosting Daniela in Lightspeed’s San Francisco office. She sat down with Lightspeed partner Guru Chahal for a candid discussion about what inspired the founding team to launch Anthropic, the importance of ethical innovation, the challenges of growing a company in such an accelerated environment, and how they came up with the name Claude.

Below are highlights of the conversation, edited for clarity and brevity.

Navigating co-foundership

Why did you and Dario launch Anthropic? What sets you apart from other big AI labs?

My brother and our five co-founders had a strong vision around developing transformative AI technology that was trustworthy, ethical, reliable, and safe. We knew generative AI was going to be hugely impactful well before it was the topic of every startup and enterprise conference. We saw a unique opportunity to be a leader in developing transformative technology that was equally powerful and reliable.

Along the way, we learned how to balance between frontier research and customer needs while building the most ethical and safe models on the market.

What’s it like to work so closely with your brother?

Dario and I have always been close, but we’re very different. He is the most impressive technical leader I’ve ever met. He has a PhD in physics from Princeton. I have a background in international development. Conveniently, we have highly complementary but not overlapping skills. That makes it really easy for us to work together because we’re both bad at the things the other person is good at.

What is the vision for Anthropic, and who’s going to build this technology? That’s Dario’s lane. Building the company, hiring a great leadership team, growing at an incredibly fast pace–that is more my wheelhouse. Together, we make a very strong leadership team because we rely on each other to own our own domains.

Building Anthropic and attracting top talent

Anthropic puts a huge emphasis on responsible innovation and AI safety. Do you feel it gives you a competitive advantage?

We’ve always felt strongly that AI development represents a big change for the world, and it’s the core reason Anthropic exists. That’s why we incorporated as a public benefit corporation and why one of our co-founders is our head of policy. We are all trying our best to make sure we develop this tech in a way that’s responsible and safe.

Before we launched any product, we came up with the concept of Constitutional AI. How do you get the model to adhere to a set of rules around how to behave? How do you know what is actually happening inside that black box? How do you build guardrails to ensure that, when these models go out into the world, they’re not going to hurt people? We worked very closely with civil and government policymakers on these questions from day one.

Conveniently, safety is also very good for business. Very well-established businesses are more hesitant to adopt new technologies without some of these guardrails in place. We’ve seen enterprises choose Anthropic because of our commitment to safety. So, rather than being a deterrent to adoption, it’s often an accelerant.

Why aren’t there more AI companies like Anthropic?

There are two things you need to build foundational models like Claude: a lot of capital and a very specialized and unique set of talented people. There are several very large, well-capitalized public companies that don’t have great AI products or models. Why is that? They don’t have the talent for it. And, of course, there are a lot of very talented people working at companies that are not training frontier models. Why? They don’t have the capital.

When we started, two-thirds of our first 15 or so employees had PhDs in physics, which is unusual for a startup. We created something we lovingly call the Physicist Pipeline, where we take incredibly smart quantitative big thinkers and train them to do machine learning research. Having both the capital and the talent is incredibly important.

You’ve hired some incredible talent from other AI labs and academia. What’s your secret?

We’ve been really lucky to have attracted top-tier talent from across the industry. I think one reason is that Anthropic is a really nice place to work. We’ve tried very hard to not overcomplicate things. It’s a stable, thoughtful, mission-oriented organization. So the combination of an incredibly exciting mission, strong product traction, and being able to be at the frontiers of research make it a really attractive company. A lot of our early employees worked together previously at OpenAI and Google. So there’s a very strong sense of a shared mission, deep personal connections, and not a lot of drama. That has helped us a lot.

Developing Claude

How are people using Claude now? Which use cases are taking off?

How people are using Claude looks very different from what they were using it for a year ago, and I think it will look even more different a year from now. The technology is evolving rapidly, and what Claude is capable of is going to advance as time goes on.

Today, not surprisingly, writing code is a huge application. Claude is better at programming in certain narrow areas than many of our own engineers. Another common use case is summarization. That doesn’t sound very sexy, but when you think about how much structured and unstructured data exists within a giant enterprise, you need a way to understand what’s inside it. A third is content generation. Whether you’re in sales, marketing, HR, or any other area of a company, Claude is a great partner to get the words started.

What’s your relationship with the developer community? Why should founders pick your platform to build products around?

In some ways, the startup and developer communities are our most important stakeholders for improving the product because they’re both very hands-on and very opinionated. Getting that kind of feedback from an enterprise just takes much longer. Being receptive to feedback from the developer community is important, and we try hard to react to it.

Fundamentally, Anthropic is best at developing frontier models. We’re trying to push the frontier of making the underlying technology better, which makes us much closer to an infrastructure layer than a very deep application layer.

We’re focusing on making the core AI technology (Claude) better. We’re not trying to build all the possible applications or user-friendly interfaces that could use Claude. That’s where startups come in – they can build those things better than we can.

How did your model end up with the name ‘Claude’?

Given that other digital assistants like Alexa, Cortana, and Siri were named after women, I felt strongly we should name ours after a man. With Claude, I think of a wise, friendly, older European man who’s there to help you. We didn’t purposely set out to make Claude warm and relatable, but that’s the feedback we get from our users. They say, ‘Talking to Claude feels like talking to a person, while talking to other AI chatbots feels like talking to a robot that was trained to sound like a person.’ That human touch is really important to the people who use our product.

The path to AGI

These AI models have become incredibly powerful. How far along are we on the path to artificial general intelligence?

People used to imagine a future where tools are better than humans at everything and ask, ‘What would it be like to live in such a world?’ Well, in many ways, we already do.

Claude is a better coder than me, a better writer, way more empathetic and patient than basically anyone. And while all LLMs hallucinate, Claude’s body of knowledge far surpasses that of any human. So when we think about AGI, we like to ask, what are the practical implications of this? What will that mean for businesses, the economy, and individuals?

Our hopeful vision for the future is that these models will continue to advance, with guardrails in place to make sure no very bad things happen. Assuming we get that right, we continue to create a world where everyone can thrive and where people can do their most creative work without having to do all the boring stuff Claude is much better at.

Interested in attending a Generative Live event? Sign up here to receive notifications about future meetups, and be sure to listen to the Generative Now podcast, where AI builders talk about how they’re creating the future.