Anthropic's Daniela Amodei on building AI responsibly, what jobs survive the transition, and why safety and revenue aren't actually at odds
Daniela Amodei's path from a 2009 recession graduation through Capitol Hill, Stripe, and OpenAI to co-founding Anthropic reflects a consistent framework: find the intersection of what you're good at, what interests you, and what has real-world impact. Anthropic was incorporated as a public benefit corporation specifically to signal that commercial success and responsible AI development are not mutually exclusive—a still-contested idea she believes the past decade has begun to validate.
On safety, Amodei frames it as radical responsibility for unintended externalities—explicitly comparing Anthropic's posture to what social media companies failed to do before algorithmic feeds produced teen mental health crises. Concrete examples include withholding a Mythos-class model over cyber-warfare risks and delaying Project Glasswing due to unresolved safety questions. The tension with revenue, she argues, is now primarily about *time*—not capability—because the models can do remarkable things but the risk envelope isn't yet fully mapped.
On jobs and economic impact, Anthropic's own economic index and an 81,000-person qualitative study both show AI acting mostly as a complement to human work rather than a replacement, with customer service as the clearest exception so far. Task-oriented roles—financial analysts, developers, copy editors—will change substantially. But Amodei expects human-facing skills (bedside manner, client relationships, creative collaboration) to become *more* valuable precisely because AI handles more of the diagnostic and analytical load.
Key concerns center on unequal access—AI adoption skews toward college-educated, male, and wealthier demographics, with deep global inequality—and on cognitive dependency, where users outsource thinking to Claude without verifying outputs. Anthropic's response includes a university-facing "learning mode" designed to make Claude act as a Socratic tutor rather than an answer machine. Amodei also flagged that the AI regulation debate has been unhelpfully politicized into binary camps, arguing that companies and regulators each hold information the other needs.
The bottom line: Amodei is cautiously optimistic—pointing to the Global South's near-universal enthusiasm for AI as an equalizing force—but clear-eyed that the business model (high capex, pre-purchased compute, no advertising) creates fragility if revenue growth stalls, and that the hardest safety problems are still unsolved. Her personal bet is that safety-conscious commercialization, not either pure research or pure growth maximization, is the only viable path.
Daniela Amodei traces Anthropic's founding philosophy—running toward a mission, not away from OpenAI—and explains why AI safety is about anticipating unintended consequences before they become crises. She argues that for most workers, AI looks more like a capability multiplier than a job killer, though task-oriented roles will change dramatically.
Before choosing a co-founder, go on vacation together and share a room with them to pressure-test the relationship
Following something you really care about and are passionate about is the most important thing you can do, especially for the times when it's not fun, because you need to remember why it matters
When in a moment of deciding if something is the right thing for your life, you often already know what the right answer is
Daniela graduated in 2009, which was not the most fun year to be a graduate
Daniela worked on Capitol Hill, worked on a campaign, and eventually ended up coming back to Silicon Valley where she started working at Stripe when it was about 40 people
The ability to be curious and learn across a lot of disciplines, to have a strong foundation of wanting to have impact regardless of the area that you're working on — that's an underrated quality
Daniela joined OpenAI in 2018 when it was still a relatively small research lab
Daniela had spent almost six years at Stripe before joining OpenAI
There were seven co-founders who originally left OpenAI to start Anthropic in December 2020
Anthropic chose to incorporate as a public benefit corporation to express that it would be a commercial entity while doing AI the right way
The majority of Anthropic's co-founders had reported to both Daniela and Dario at OpenAI
Social media developers were not intentionally setting out to cause a pandemic of eating disorders for teenage girls — that was an unintended consequence
AI has the potential to be used to develop chemical and biological weapons
Anthropic decided not to release its Mythos-class model because of the potential for cyber warfare
The majority of Anthropic's revenue comes from businesses, not consumers
Businesses are not looking to have models that are unsafe — they are correctly risk-averse and don't want AI technologies that are super unpredictable or unreliable
The tension between safety and revenue at Anthropic is now primarily about time — not that the models can't do amazing things, but that the risks are not yet fully understood
Anthropic delayed the release of the Project Glasswing model class to customers due to insufficient confidence in its safety for cyber-related risks
According to Anthropic's economic index studying how people actually use AI, it mostly looks like complementary skills — AI as an enabler of work, not a replacer, except in a very vanishingly small number of cases, mostly customer service
Software developers will still exist in the future but won't write as much code — the percentage of work involving talking to product managers and working closely with customers will expand
People in developing countries are much more optimistic about AI than people in higher income countries — the Global South is almost universally positive about AI as an equalizing opportunity
Anthropic conducted the largest qualitative study it knows of, talking to 81,000 people about their use of artificial intelligence
Task-oriented jobs like financial analyst, developer, and copy editor are going to change a lot, and a lot of that work will be able to be done by AI tools
Humans like to be with other humans, learn from each other, be creative, and understand each other, and these social skills will become more important and more prized as AI does more productive day-to-day work
AI is going to get really good at medical diagnosis, but cannot actually examine patients or help them feel better
There is a reasonable body of medical literature indicating that people who have a good relationship with their doctor have better clinical outcomes than people who do not like their doctor
Bedside manner is going to be five times more important in a world where AI handles diagnostics
Medical questions are one of Anthropic's most common casual use cases for Claude
It is a high capital expenditure business and training AI models requires a lot of compute that is in scarce supply, making compute expensive and requiring purchases far in advance
Venture capitalists say nothing like Anthropic's and OpenAI's revenue growth has ever happened before in such a short timescale
Claude has been right more often than her doctors about complex medical cases in Daniela's own experience
AI models make things up sometimes, get confused, and cannot examine patients
The concept that being in business does not have to be in tension with doing good is a very new idea, particularly prominent in the past five to ten years
The desire to do good is a strong correlate with actually doing well in business
There is a demographic component to AI adoption — it skews toward college-educated people, more men than women, with racial and wealth demographics involved, and is very unequally distributed globally
Some people express a feeling that they don't engage their brain because they don't have to — they could have thought through an idea themselves but it was easier to just trust what the AI tool gave them
If AI company revenue were to decline, both Anthropic and OpenAI would face serious problems because they have pre-purchased large amounts of expensive compute for the future
The AI regulation debate has become politicized into 'regulation bad, innovation good' or 'innovation bad, regulation good,' which is a shame because it is a really nuanced question
The intersection of what was I good at, what was I interested in, and what was going to have a big impact in the world
AI safety means taking a form of radical responsibility for the technology that we're developing, analogous to thinking in advance about unintended externalities as social media companies failed to do
Three steps for society to address AI-driven job displacement: (1) lead from humility and publish research on actual AI usage, (2) be creative and experimental about AI as a grounding and unifying force beyond work, (3) address it as a social and political issue across government, civil society, and universities
Claude acting as a patient tutor or individualized professor who knows you and understands what you most want to learn can make you smarter and expand the set of things you think you can learn, versus a 'turn your brain off' version that is harmful
Anthropic's decision to not put ads in Claude is partly predicated on the belief that AI technology is different because people have more personal conversations with AI tools than they would put on Instagram or social media
Technology companies and regulators should work hand-in-hand on AI regulation because companies have information about how technology can be abused while regulators know how to provide enforceable frameworks
Anthropic developed a 'learning mode' for Claude to be used by universities, faculty, professors, and students as an alternative to simply getting AI to produce answers outright
Daniela uploads years of direct reports' data to Claude when writing performance reviews to help spot patterns about someone over three to four years of working together
Daniela uploads all of her reports' upward feedback about her to Claude and receives feedback on areas where she has not improved
“we really, I think we're running towards something versus running away from something”
“Sometimes Claude is wrong. Heretical to say, but a fact.”
“everybody thinks that AI is not gonna come for their job because they're so special”
“you already know what the right answer is”
Shared with Earmark · earmark-ai.com