A 48-minute Bloomberg Originals profile (The Circuit, hosted by Emily Chang) on Anthropic and its co-founders, siblings Dario and Daniela Amodei. The throughline: a lab that built its brand on AI safety is now one of the most valuable startups on earth, near a trillion-dollar valuation, and has to prove you can build the most powerful technology in the world without breaking it. Useful as a clear, current snapshot of how Anthropic positions itself on the OpenAI split, Claude's design, the enterprise bet, and the job-loss debate.
The setup
Anthropic was founded in 2021 by a group of OpenAI departures. Seven co-founders; today it is, per Dario, the only company at its scale with every co-founder still on board. Two of them are the Amodei siblings: Dario, the brother and research visionary, and Daniela, the sister and operator who turns his ideas into a working company. The name comes from the Greek for "human," signalling the stated mission of building responsible AI for humanity's long-term benefit.
Their backgrounds split the same way the company does. Dario did neuroscience, then AI at Baidu and Google. Daniela was an early employee at Stripe. They shared a house in San Francisco with Daniela's husband Holden Karnofsky before both joined OpenAI (Dario in 2016, Daniela after). Asked who wins when they argue: "No one."
Dario's mental model of AI progress is what he calls a "smooth exponential": nothing happens, nothing happens, then zoom. He says he watched the revenue and valuation curves and predicted Anthropic would become the top AI company on both around this time. It did.
Why they left OpenAI
The split has become Silicon Valley lore. Dario is blunt that the cause was not a safety disagreement on its own. They had those, but disagreements alone are not enough to leave. The breaking point was trust: feeling you cannot trust someone, that their values are not what they claim, that they are not honest. His resolution philosophy: when you do not share a vision or trust each other, do not argue, go do your own thing and let them do theirs.
For context, while at OpenAI Dario developed the idea of scaling laws: large language models get better mainly by adding data and compute, even if the underlying algorithm stays the same. He frames this as a counter-cultural view at the time that ended up paving the way for ChatGPT.
How they designed Claude
Claude is trained against a written set of principles Anthropic calls a Constitution, meant to keep it in line. A few design choices stand out:
- "Professional warmth." The deliberate persona is approachable but distant. Not your best friend, not cold and calculating. The goal is for Claude to feel professional, not parasocial.
- Good vs bad model. A bad model lies, either by hallucination (it predicts the next word and invents something when it does not know) or by deliberate deception, which their own research has shown models can do. Production models have to be kept clear of both, plus broad work on harmlessness.
- Whose values? There is no universal standard for helpful or harmless, so they anchor on founding human documents like the UN Declaration of Human Rights. Daniela says they have started talking with religious leaders about values that recur across belief systems, to bake the shared ones into Claude's character.
- Tuning the dial. Early Claude 2-era models came out "nannyish" ("I'm really concerned about you" when you just asked for the weather). The worst versions were not shipped. Daniela frames character work as threading a fine needle to land in the centre.
The enterprise bet and the "SaaSpocalypse"
Anthropic's revenue has jumped sharply over the past year, making the company profitable for the first time, driven by business tools rather than a splashy consumer app. The two products named: Claude Code, which automated large chunks of software engineering, and Claude Cowork, which extended that power beyond engineers.
Dario frames the consumer-vs-enterprise choice as a values decision as much as a business one. Pick a model that conflicts with your values and you either betray them or become irrelevant. He is openly critical of the social-media and AI-video model, where the incentive is to maximise engagement and attention for ad revenue. Enterprise, by contrast, points at the work he wants AI doing: curing diseases (biotech, pharma, academic research), making energy cheaper. All enterprise, all aligned with the stated mission.
The market noticed. After Claude Cowork shipped, roughly $285 billion in software market value vanished overnight, a sell-off traders nicknamed the "SaaSpocalypse," with some software stocks down nine days running. Dario's read: AI makes the pie bigger, so the software industry overall probably grows, but incumbents that do not adapt or defend their moats will be big losers, and some may go out of business.
Boris Cherny and the coding leap
The growth spurt traces heavily to one hire: Boris Cherny, the engineer behind Claude Code and Claude Cowork, recruited in 2024 from a slow rural life in Japan (farmers' markets, making miso). His bet was that earlier coding tools were trivial (autocomplete the line) and that a coding agent could instead do all of it.
The internal numbers and texture are the most striking part of the film:
- Cherny says Claude writes "almost all" of the code on his team, and 100% of his own code for at least six months.
- The workflow has shifted from typing code and pressing tab, to talking to one Claude while it writes and starting the next. He describes running anywhere from a few to a few thousand Claudes at once.
- "I feel like I suddenly have superpowers. I have a jet pack and engineering has never been this fun."
The live demo: Chang asks for a recipe app that suggests weekly meals and shows recipe steps; Claude builds it in minutes, work that used to take hours or days.
At the second annual Code with Claude conference, the scaling stats land: API volume up nearly 17x year over year on the cloud, eight frontier models shipped in 12 months, and Q1 growth that annualises to roughly 80x per year, described as the first year they have grown faster than the exponential.
The jobs question
This is the tensest stretch of the interview. Will engineers be the first casualties of the AI they are building? Dario's answer: today's engineers have a head start, because the job is not only coding (planning, talking to users, deciding what is next), and those parts stick around. But the productivity story is a "hump." You automate 90% of a job and people get ~10x more leveraged on the rest; eventually it approaches 100% and the task is better left to the AI. They are starting to see early cases where AI does not make a person more productive, it is simply better for the AI to do the thing.
Outside the Valley the mood is darker: 70% of Americans think AI will kill jobs, nearly a third fear theirs is at risk. Dario stands by his earlier, much-quoted prediction that AI could eliminate half of all entry-level white-collar jobs in one to five years. A year on, he says he is still "the same order of concern." He warns of an unusual combination: fast GDP growth alongside high unemployment, underemployment, low-wage work, and high inequality. Asked if that is how revolutions start, he agrees it is the outcome to prevent.
On the pushback (Jensen Huang's "you're conflating tasks with jobs," and accusations of "doom marketing" that benefits Anthropic), Dario pushes back hard. He argues he writes carefully and at length about tax and macro policy, what the new jobs are, and why this time differs, and that people clip three-second fragments from a year-old interview. His line: "the idea that this is cheap marketing is itself cheap marketing," a symptom of Silicon Valley's three-second social-media disease. His actual message: not doom is coming, but this is foreseeable and needs a positive response.
Where the new jobs might go, none guaranteed:
- The physical world. Building, making, manufacturing still needs people.
- Human-centred work. Jobs built on human relationships and interaction.
- Directing the AI. Someone has to align it with human values and intentions, though Dario does not know how thin or thick that layer will be.
The recurring example is medicine. AI will soon be good at telling you the likely diagnoses and which tests to run, so you will not need a doctor for that. What it cannot do is physically examine you or sit with you on how you are coping. The pivot is toward the interpersonal, because the diagnostic tools get much better while the human part does not change. Anthropic has also published work estimating that management, finance, and legal roles could change a lot in the near future.
What's next (teaser)
The episode closes by trailing part two: AI and the future of warfare, a "scarily powerful new model" called Mythos, and riding the exponential without sliding into dystopia. Asked if he sees himself in Oppenheimer, Dario instead identifies with Leo Szilard (who first conceived the nuclear chain reaction) and calls Oppenheimer a failure case. His view: this does not end well through larger-than-life figures at the centre of everything, only through checks and balances everywhere.
Key takeaways
- Trust, not safety, drove the OpenAI exit. Dario says safety disagreements existed but were not the cause; the break was about honesty and values.
- Claude's persona is engineered. "Professional warmth," a written Constitution, and deliberate tuning away from early nannyish behaviour are product decisions, not accidents.
- The enterprise bet is a values bet. Anthropic avoided the attention-maximising consumer model on purpose and reached profitability on coding and business tools.
- Coding is already mostly automated inside Anthropic. Cherny reports Claude writing 100% of his own code for six-plus months and a workflow of orchestrating many Claudes at once.
- Dario holds the line on white-collar job loss. Still "same order of concern" on the 50% entry-level prediction; rejects the "doom marketing" framing as itself cheap marketing.
- The optimistic case is narrow and unguaranteed. Physical work, human-relationship jobs, and AI-direction roles are the named escape valves, with medicine pivoting to the interpersonal.
Notable quotes
- "When you feel that you can't trust someone, when you feel that their values are not what they say they are... that makes it very hard to continue to work with a company." (Dario, on leaving OpenAI)
- "It's been writing 100% of my code for at least six months. The work of engineering has just completely changed. I feel like I suddenly have superpowers." (Boris Cherny)
- "This is something we should see coming, that we're worried about and that we need to actually respond to positively." (Dario, on job loss)
Why it matters for me
This is the clearest current read on how Anthropic frames itself, and I sell into the same enterprise buyers who are watching the "SaaSpocalypse" nervously. The medicine analogy (AI takes diagnosis, humans keep the bedside manner) maps straight onto enterprise sales: the AI handles research and drafting, the human keeps the relationship and judgment. Worth holding onto for both the Kaltura AI conversation and my own tooling, since I run Claude across my own product-development cycle the way the film describes.
Further reading
- Anthropic - official site, including the Constitution and research posts.
- Claude - the product.
- Dario's essays referenced in the film: "Machines of Loving Grace" and "The Adolescence of Technology" (both on anthropic.com).
- Cross-reference: Claude Cowork