Latest / From Spark to New Venture / Dario Amodei--The CEO of Anthropic
Transcript
- 0:00SPEAKER_00: Hello everybody, I am Eli Mariani, and this is my From Spark to New Venture episode on Anthropic founder and CEO, Dario Amade.
- 0:09SPEAKER_00: Dario Amade left the job where he was running research at one of the most talked-about companies on Earth and walked out with a handful of colleagues and no product, no office, and no obvious customers.
- 0:21SPEAKER_00: Then he spent the next few years building something that at one point he deliberately chose not to release.
- 0:27SPEAKER_00: Imagine a baker who finishes the best loaf of his life, sits on the counter and says, Yeah, you know, I think uh I think we should wait on this one.
- 0:35SPEAKER_00: And that's where I want to start today, because that decision tells you almost everything about how this particular founder thinks.
- 0:42SPEAKER_00: Today, we're telling the story of Dario Amade, and how he went from a kid in San Francisco who loved physics to the CEO of Anthropic.
- 0:51SPEAKER_00: Let's go back to the beginning.
- 0:53SPEAKER_00: Dario was born in 1983 in San Francisco, years before it became the capital of disrupting things.
- 1:01SPEAKER_00: He was a science kid, specifically a physics kid.
- 1:04SPEAKER_00: He studied it and then went off for a PhD in biophysics at Princeton.
- 1:09SPEAKER_00: Which is kind of the subject where the title alone makes you pretty tired.
- 1:14SPEAKER_00: Then he did postdoctoral research at Stanford's medical school, and if you had met him around then, you'd have guessed his future involved lab coats and uh grant applications.
- 1:25SPEAKER_00: Now, here's a detail that gets repeated a lot, and as he tells it, it matters.
- 1:30SPEAKER_00: His father got seriously ill when Dario was young and died in the mid-2000s.
- 1:36SPEAKER_00: Years later, the disease became largely treatable, and that kind of whiplash where a cure arrives just too late for the person that you love can send you in a couple of directions.
- 1:47SPEAKER_00: And for Dario, it seems to have planted a conviction that science moves too slowly for the problems that matter, and that biology is so staggeringly complicated that humans' brains alone might not crack it fast enough.
- 2:02SPEAKER_00: That belief is what pulled him in towards artificial intelligence.
- 2:06SPEAKER_00: And I want you to notice something because it's the first lesson in his journey.
- 2:10SPEAKER_00: He didn't start with AI is hot, let's build an AI company.
- 2:14SPEAKER_00: He started with a problem that he cared about, and AI looked like it might be a tool for solving it.
- 2:20SPEAKER_00: The company came much later, and but that motivation came first.
- 2:24SPEAKER_00: So in 2014, he heads into the AI world, working at Baidu, the big Chinese tech company on speech recognition.
- 2:33SPEAKER_00: Then Google Brain, which was basically the Dream Team's clubhouse for neural networks, and then in 2016, he landed at OpenAI, which you might have heard of.
- 2:43SPEAKER_00: It was a brand new nonprofit lab with big billionaire backing and an even bigger mission to make sure that advanced AI benefits everybody.
- 2:52SPEAKER_00: He climbs fast, he ends up as vice president of research, which means that he's steering a lot of the work behind uh their GPT-2 and GPT-3 models.
- 3:02SPEAKER_00: Uh these are the models that made the world go.
- 3:04SPEAKER_00: Wait.
- 3:05SPEAKER_00: Like computers can do all that.
- 3:08SPEAKER_00: And his sister Daniela was also there at OpenAI, but she came from a very different world.
- 3:14SPEAKER_00: Literature and then a stint at the payments company Stripe.
- 3:18SPEAKER_00: And she ran operations and people.
- 3:21SPEAKER_00: If Dario's the one started staring at the blackboard, Daniela is the one making sure the building has electricity and the humans inside it are actually okay.
- 3:31SPEAKER_00: Uh keep her in mind as she becomes a bigger part of the story, then headlines usually give her credit for.
- 3:36SPEAKER_00: Alright, so what goes wrong and why would you re-leave OpenAI?
- 3:41SPEAKER_00: This is where I have to be careful, because there is more than one version.
- 3:45SPEAKER_00: The most common telling is that Dario and a group of researchers felt that safety, meaning making sure that these increasingly powerful systems behave the way we want, wasn't getting enough weight compared to speed and commercialization.
- 3:59SPEAKER_00: OpenAI had taken big money from Microsoft by then, and the tension of build fast versus build carefully was getting loud and a bit muddied.
- 4:08SPEAKER_00: Other reporting paints a messier picture with personality clashes and fights over who should do what, including some friction with OpenAI's leadership.
- 4:17SPEAKER_00: Dario himself has said, more or less, that it is poignantly exhausting to argue endlessly with someone else's vision of what a company should be.
- 4:26SPEAKER_00: At some point you realize you can't keep lobbying for your idea inside someone else's house.
- 4:31SPEAKER_00: You have to build your own house, even if it's smaller and leakier.
- 4:34SPEAKER_00: So, at the end of 2020, he announces that he's leaving OpenAI, and by early 2021, Anthropic exists.
- 4:42SPEAKER_00: Seven co-founders, including Daniela, all with OpenAI pedigrees.
- 4:46SPEAKER_00: There are people like Jared Kaplan, Chris O'Law, Tom Brown, Jack Clark, Sam McCandish, and Ben Mann.
- 4:55SPEAKER_00: Now, something useful for anyone who's ever wanted to start something.
- 5:00SPEAKER_00: They didn't begin with a product, they began with a research agenda and a team that already trusted each other.
- 5:06SPEAKER_00: A founder's most underrated asset is often not an idea, it's the people who have already seen you at your worst and said, yes, I'd do this with you.
- 5:14SPEAKER_00: Dario as CEO, Daniela as president.
- 5:18SPEAKER_00: You think that might be a little awkward at Thanksgiving, but you gotta think about what each of them brought.
- 5:23SPEAKER_00: He had the technical vision and the public voice, and she had the operational muscle.
- 5:27SPEAKER_00: Plenty of startups fail, apart because the founders can't agree, and these two had decades of practice disagreeing without blowing it all up.
- 5:38SPEAKER_00: They also did something that was sneakily smart with the legal structure of Anthropic.
- 5:43SPEAKER_00: Anthropic was set up as a public benefit corporation, which is the type of company that has to weigh its mission, not just shareholder profit.
- 5:51SPEAKER_00: It's the corporate equivalent of writing your values into the wall before the pressure shows up.
- 5:56SPEAKER_00: Whether that holds up under real pressure is a fair thing to wonder about, and we'll get there.
- 6:02SPEAKER_00: Now, uh the money.
- 6:04SPEAKER_00: Raising capital for a company that doesn't have a product and whose pitch is basically we'll be careful about a technology most people haven't heard of yet.
- 6:12SPEAKER_00: Sounds hard and difficult, but with that founding team, it wasn't impossible.
- 6:16SPEAKER_00: And their first big round in 2021 was around$120 million.
- 6:22SPEAKER_00: Um remember that they were still mostly doing research, studying how these models work, how to steer them, and how to peek inside them.
- 6:31SPEAKER_00: One of their signature ideas was something called constitutional AI, which is a fancy way of saying, instead of only having humans tell the model what's good or bad, you give the model a written set of principles to follow, uh, almost like a constitution, and train itself to check itself against them.
- 6:50SPEAKER_00: Which brings us back to the loaf of bread on the counter.
- 6:53SPEAKER_00: By 2022, they've built an early version of Claude.
- 6:57SPEAKER_00: It's working, and here's the part that makes founders sweat.
- 7:00SPEAKER_00: They hold it back, they don't release it.
- 7:02SPEAKER_00: They wanted more time to test it uh for safety.
- 7:04SPEAKER_00: And then in November of 2022, OpenAI released ChatGPT, and the whole world lost its mind.
- 7:13SPEAKER_00: Umthropic, as the careful one, is sitting on the sidelines with a finished product of their own.
- 7:20SPEAKER_00: And Cloud finally ended up launching publicly in early 2023, and some people said they'd lost a crucial lead, and that's honestly debatable.
- 7:30SPEAKER_00: We don't get to run the experiment where they launch earlier, but what I find instructive is the trade-off they're willing to make.
- 7:37SPEAKER_00: A lot of founders say they have values, um, but fewer are willing to let those values cost them months of the most competitive market in tech.
- 7:45unknown: Alright.
- 7:47SPEAKER_00: Here's attention.
- 7:49SPEAKER_00: Dario is one of the loudest voices warning about the dangers of AI.
- 7:53SPEAKER_00: He's written long essays about the risks, and he's even joined other leaders in saying the industry may need to slow down.
- 7:59SPEAKER_00: And yet he runs a company whose whole business is building an ever more powerful AI.
- 8:05SPEAKER_00: Critics call that a contradiction at best, and uh clever marketing at worst.
- 8:10SPEAKER_00: Uh, why build the thing you're so worried about?
- 8:13SPEAKER_00: And his answer, in spirit, is something like this: it's going to be built either way.
- 8:18SPEAKER_00: So it's better for safety-focused people to be at the frontier, showing that you can be careful and competitive at once.
- 8:25SPEAKER_00: He's called that a race to the top, the idea that if you set a good example, rivals have to follow or look bad.
- 8:32SPEAKER_00: You can find that persuasive or you can roll your eyes, and thoughtful people will do both.
- 8:37SPEAKER_00: But I'd say this: if you're ever launching something that someone points out, the contradiction in your pitch, it's not a reason to panic.
- 8:44SPEAKER_00: Uh Dario has uh had to hold that contradiction in public for years, and sometimes the honest answer to a paradox is yes, it's uncomfortable, and here's why I'm doing it anyway.
- 8:56SPEAKER_00: Now, let's talk about what actually happened next, because the business side is frankly kind of wild.
- 9:03SPEAKER_00: Um big backers showed up, Amazon, Google, and others.
- 9:07SPEAKER_00: The model got better, and the company found a sweet spot.
- 9:10SPEAKER_00: Not so much flashy consumer stuff, but businesses and especially software developers.
- 9:15SPEAKER_00: Cloud turned out to be really good at writing and fixing code, and in 2025, the company launched a coding tool called Claude Code that really, really took off.
- 9:24SPEAKER_00: Uh, companies don't tend to switch vendors on a whim, so winning developers trust turned into the foundation for serious revenue.
- 9:33SPEAKER_00: Um serious?
- 9:36SPEAKER_00: Reports say that companies annualized revenue was around a billion dollars at the end of 2024, and by May of 2026, the company said that it passed 47 billion.
- 9:46SPEAKER_00: That's a growth curve that makes a normal startup look like it's standing still.
- 9:51SPEAKER_00: And in May, it raised tens of billions of dollars at a reported valuation of about$965 billion, which for the first time put it ahead of OpenAI and ChatGPT.
- 10:04SPEAKER_00: Uh let me put a little bit of caution on all that, because I'm not a financial advisor and those numbers move a lot.
- 10:11SPEAKER_00: Um, a lot of the specifics come from company statements and press reports, evaluations and opinions with commas, and this company also reportedly lost billions last year while spending heavily on computing power.
- 10:25SPEAKER_00: Growing fast and being profitable are two different things.
- 10:29SPEAKER_00: It hasn't been entirely smooth either.
- 10:32SPEAKER_00: Um, this summer the company ran into a bruising fight with the US government, including a stretch where access to two of its newest models was suspended after export controls or under export controls before being restored.
- 10:47SPEAKER_00: There's still litigation in the background, and as of today, as far as I can tell, the much discussed public stock offering that investors have been anticipating is still being worked out.
- 10:58SPEAKER_00: Timelines on these things shift constantly, so I'd treat any date you might have heard of more of an educated guess.
- 11:06SPEAKER_00: So, what do we actually take from this?
- 11:08SPEAKER_00: Um, if you're a beginner standing at the edge of your own idea, here's what I'd carry out of Dario's story.
- 11:15SPEAKER_00: Start from a problem that you can't stop thinking about.
- 11:18SPEAKER_00: His mission had roots in something deeply personal to him, and that kind of motivation survives the boring middle stretches that no one really cares to post about.
- 11:27SPEAKER_00: Um, build with people you trust in a crisis.
- 11:31SPEAKER_00: His co-founders were not strangers from a networking event.
- 11:34SPEAKER_00: They had already worked through hard stuff together at OpenAI, and there were people that he could felt he could rely on and trust.
- 11:41SPEAKER_00: And you want to know what you're willing to pay for your principles.
- 11:45SPEAKER_00: Delaying a launch is a concrete crop cost.
- 11:48SPEAKER_00: It's a big loss.
- 11:49SPEAKER_00: And if you say you value something, it helps to know in advance what it will cost you.
- 11:54SPEAKER_00: But being firm on those principles is very important.
- 11:58SPEAKER_00: And also, to pick your terrain.
- 12:00SPEAKER_00: Anthropic didn't try to beat everyone at everything.
- 12:04SPEAKER_00: It leaned into a place where it could be unusually good and let that compound over time.
- 12:10SPEAKER_00: Also, be a little suspicious of uh tidy hero stories, including this one.
- 12:15SPEAKER_00: Uh founders leave places for messy reasons.
- 12:19SPEAKER_00: Uh mission and ego and timing are all tangled together, and the person telling the story, uh, me included, is never fully neutral.
- 12:29SPEAKER_00: Here's what I keep thinking about, and what I'll leave you with.
- 12:33SPEAKER_00: Uh the first thing Dario really built wasn't a model or a company.
- 12:37SPEAKER_00: It was a decision that if he wanted to do it this way, his way, he'd have to leave the room where he already had influence and start over somewhere that he had to earn it from scratch.
- 12:48SPEAKER_00: Most of us have a room like that, a job or a project or a vision of ourselves that's comfortable and respected.
- 12:57SPEAKER_00: And the question isn't really whether you should leave it.
- 13:00SPEAKER_00: The question is whether there's something you'd build if you'd build if no one's vision but your own were in charge, and what it would take to find out.
- 13:10SPEAKER_00: Well, that's all I have for you.
- 13:12SPEAKER_00: Thank you for spending this time with me, man.
- 13:15SPEAKER_00: Go build something carefully or otherwise.
- 13:18SPEAKER_00: Thank you.