Part two ended on a question, which is whether anyone building the next general-purpose technology has actually learned what social media taught us, or whether we are about to run the same experiment again with a far more powerful machine. Dario and Daniela Amodei, the brother-and-sister team who run Anthropic, say they have learned it, and they say so constantly, which makes the claim either the most important thing being said in technology right now or the most convenient, and either way it is worth testing rather than taking on trust.
The claim: learn from social media, on purpose
In their Bloomberg Originals interview, Daniela Amodei made the comparison directly, saying that if the social media companies could go back and see the world they ended up creating she would like to think they would have done it differently, and that the advantage Anthropic has is simply being second, which lets it watch what social media did to child welfare and mental health and then try to design around it in advance. That is the whole argument of this series, made by someone who is building the technology rather than writing about it.
Dario makes the same case from the business side, saying that he detests social media as a category and that he built Anthropic on enterprise tools rather than a consumer app on purpose, because an advertising model always drags you towards maximising screen time whether you intend it to or not. Their answer to the incentive trap from part two was to choose a different incentive, taking the money from biotech, pharma and software, where a model earns its keep by being useful rather than by being addictive, and on top of that they train Claude against a written set of principles they call a constitution, drawing on sources like the UN Declaration of Human Rights so that the values are designed in from the start rather than bolted on after the first scandal.
Where they have actually paid for it
Values are cheap until they start to cost you something, and the only real test of them is what a company is willing to give up to keep them, which is where Anthropic's record is harder to fake than any essay. They held back their most powerful model, the one that became the Mythos and Fable story I wrote about on this blog last week, after early testers reportedly described it as a super-weapon and asked them not to release it, and Dario has said the decision to hold it back cost the company enormously in commercial terms. They also drew red lines with the US military, refusing to let Claude be used for autonomous weapons or for the mass surveillance of citizens and insisting that a human, not the model, makes the final call on a lethal decision, and that stance pulled them into a public fight with the Pentagon and, by the accounts that followed, got them blacklisted by it. The part worth sitting with is that a company walked away from one of the largest customers on earth rather than let its software decide, on its own, where a bomb falls.
The same pattern runs back to the beginning, because Dario left OpenAI, the company he had helped turn into the frontrunner, and has said the reason was not only disagreements about safety but a loss of trust, that you cannot keep working alongside people when you feel their values are not what they claim them to be. Walking away from the leader to start an underdog is not the move of someone optimising for marketing, it is the move of someone who meant it.
The Oppenheimer question
The interviewer asked Dario the obvious question for a man building a technology this powerful and this dangerous, which was whether he sees himself as Oppenheimer, and his answer was revealing, because the figure he said he identifies with is not Oppenheimer at all but Leo Szilard, the physicist who first realised a nuclear chain reaction was possible, while he described Oppenheimer instead as a failure case and an example of how not to do it. His point was that none of this ends well through larger-than-life individuals trying to sit at the centre of everything, and that it only ends well if there are checks and balances spread everywhere, which is a strikingly humble thing for the head of a near trillion-dollar company to say.
He has also written the whole argument out in the open, in two long essays, with "Machines of Loving Grace" setting out the version where Ai goes right and "The Adolescence of Technology" setting out the dangers and how to defend against them. There is a small accident of language worth noticing, because the Netflix drama back in part one was about a boy's adolescence going wrong online while Dario's essay casts the whole technology as an adolescent, powerful and not yet wise, and asks how we get it through that stage without disaster, and writing those risks down in public before the harm arrives is the exact opposite of the social media playbook.
The case against, taken seriously
An honest assessment has to argue the other side properly, because the sceptics are not fools. The loudest charge is that this is all doom marketing, and Nvidia's Jensen Huang has argued that warning the world about Ai destroying jobs is self-serving, that it frightens people, conflates tasks with jobs and conveniently makes Anthropic sound more important than anyone else, to which Dario's defence is that he writes carefully about the responses, from retraining to tax policy to where the new jobs come from, and that a culture built on three-second clips flattens all of it into "doom is coming". Both of those can be a little true at once, because the warnings genuinely do serve Anthropic even when they are also sincere.
There is more to put on the other side of the ledger. Anthropic partnered with Palantir and Amazon back in 2024 to put Claude in front of US defence and intelligence agencies, so for all the red lines this is a company that chose to be in the room rather than outside it, and although to its credit it has been specific about the limits, saying it does not work with immigration enforcement or border agencies, selling to the military at all is a long way from clean hands. Deciding who does and does not get access to the most capable models, as they did with Mythos, is its own kind of power, and asking the public to trust them to draw that circle is a large request, which even they seem to accept, since Dario has said that starting from a position of distrust towards any Ai company is entirely rational and that Silicon Valley now has to earn its credibility back through what it does rather than what it says. So is it a halo or is it honesty? My own read is that the difference between marketing and conviction is what you are prepared to lose, and Anthropic has lost real revenue and real access more than once to hold a line, which is an expensive way to run a marketing campaign.
Why "slow down" is the wrong demand
The instinctive response to something this powerful is to tell the careful company to slow down, and back in 2023 that is exactly what a well-known open letter, signed by Elon Musk among many others, called for, a six-month pause on the most advanced Ai. Reid Hoffman, who co-founded LinkedIn, pushed back on that idea hard and called a unilateral pause foolish, and the way I have always remembered the argument has stuck with me ever since. Building Ai is like racing cars on a track, and if you decide to ease off the accelerator in the name of safety the cars behind you do not, they simply take the lead, so all your caution actually buys is to hand the race to whoever cares least about winning it safely. Dario makes the same point from inside the industry, that the real danger comes from there being so many countries and companies that a new one will fill any gap the moment it opens, which means a careful lab stopping does not stop the frontier, it only changes who reaches it first, and we end up with all of the capability and none of the conscience. That is why I would rather the people saying the right things, and paying real money to keep saying them, are the ones out in front, and it is worth remembering that being the leader is the very thing that lets Anthropic afford to be careful at all, because as Dario admits, holding a model back is extremely hard to do if you are not already ahead.
What Ai should take from social media
If there is a single thread running through all three of these pieces, it is that we already know what the mistakes look like, so repeating them would be a choice rather than an accident, and a handful of lessons carry over from social media to Ai almost word for word.
The first is to watch the business model rather than the mission statement, because social media did its damage through a revenue model that depended on attention, so an Ai company that makes its money where being useful matters more than being addictive is telling you something real in a way no ethics page ever could.
The second is to measure the harm and not only the growth, because the damage social media did stayed invisible for years precisely because it was never on anyone's dashboard, and the harms of Ai, whether to jobs, to truth or to children, need to be measured and owned from the start rather than discovered in an inquest later.
The third is to design the safeguards in rather than bolt them on, because almost every protection social media added arrived after the damage and under pressure, whereas the whole point of going second is to build the limits in before release instead of after the apology.
The fourth is not to wait for the regulator, because the rules will arrive years too late as they always do, which leaves the real responsibility with the people building the technology to act before anyone forces them to.
None of that is complicated and none of it is new, so the only question that actually matters is whether the industry has the discipline to apply it this time, while Ai is still young enough to shape.
We only regulated social media nearly a decade after it had already hurt children like Molly Russell, whereas with Ai we are at least arguing about the rules while the technology is still young, and a company right at the frontier is taking real commercial hits to set the precedents rather than waiting to apologise afterwards. That guarantees nothing on its own, because trust is earned in what you are willing to give up rather than in what you publish, and the coming year, with two models already pulled and a government furious, will be the real test of whether the conviction holds. On the evidence so far, they have given up a great deal more than most, and that is about the most you can reasonably ask of anyone building something this big.