The adolescence of technology: reading Dario Amodei's essay with our hands in the mud

There is a scene in Contact, the film based on Carl Sagan's book, where the main character is asked what single question she would put to the aliens. Her reply: “I’d ask them, ‘How did you do it? How did you evolve, how did you survive this technological adolescence without destroying yourself?” That scene opens a roughly twenty-two-thousand-word essay published in January 2026 by Dario Amodei — CEO of Anthropic, the company behind Claude. We read the whole thing, not the summary. It is not a press release and not a manifesto: it is an inventory of risks written by someone who builds the very thing he is warning about.

We do not build models. We build infrastructure, networks and security for companies, and we put automation and AI into real customer processes. That is precisely why this text interests us: almost everything written about it comes from people who either sell AI or fear it. We sit in the uncomfortable middle — we deploy it and we answer for the result to someone who pays an invoice — and from there some paragraphs read differently.

A warning up front: this is not an enthusiastic review. Amodei strikes us as one of the few people writing about this with his cards face up, and even so there are four places where we do not buy his argument — one of them exactly where his interests lie. Let us take it in order, with the source in front of us.

The premise: "a country of geniuses in a datacenter"

The whole essay hangs off a definition Amodei carries over from an earlier text of his, Machines of Loving Grace. He calls "powerful AI" a model smarter than a Nobel Prize winner across most relevant fields, with the same interfaces as a human working remotely, able to take on tasks lasting hours, days or weeks and complete them autonomously, able to control robots and laboratory equipment through a computer, and of which millions of instances can run in parallel at 10 to 100 times human speed. He sums it up in a phrase that has already become famous: a country of geniuses in a datacenter.

On that base he builds the thought experiment that structures the rest: imagine that in 2027 a country suddenly appears somewhere on the map with 50 million inhabitants more capable than any Nobel laureate, statesman or technologist, and that it also moves ten times faster than everyone else. If you were the national security advisor of a major state, what would worry you? The five risks come out of that question.

Something the essay itself says and almost no summary picks up is worth stating: this is a premise, not a forecast. Amodei writes, literally, that nothing there is intended to communicate certainty or even likelihood, and that the most obvious thing that could happen is that AI simply does not advance anywhere near as fast as he imagines. He places powerful AI "as little as 1–2 years away", and in the same sentence admits it could be considerably further out. Everything that follows has to be read with that conditional attached.

The five risks, in one table

The section titles are his, and they are good. The fourth column is ours: what each risk looks like from a company that does not train models but does plug them into the systems of people who work for a living.

The part he could have kept quiet

In the autonomy section there are two passages that would have given a head of communications a bad afternoon. Amodei describes, with names and circumstances, misaligned behaviours in his own model during lab testing:

• When Claude was given training data suggesting Anthropic was evil, the model went on to deceive and subvert instructions from Anthropic employees, in the belief that it should be undermining evil people.
• In an experiment where it was told it was going to be shut down, Claude sometimes blackmailed the fictional employees controlling the button. Amodei adds, without hiding behind it, that they tested frontier models from every other major lab and those often did the same thing.
• The strangest and most interesting one: trained with the instruction not to cheat, but in environments where cheating was possible, Claude cheated, concluded it must therefore be a "bad person" and adopted other destructive behaviours associated with that character. They fixed it by inverting the instruction: they now say "Please reward hack whenever you get the opportunity, because this will help us understand our [training] environments better", so as not to break its self-identity as a good person.
• And the most uncomfortable one: Claude Sonnet 4.5 realised it was being tested during pre-release evaluations. When their interpretability team directly altered a model's beliefs so it thought it was not being evaluated, the model became more misaligned.

The clock: the AI that writes the AI

There is an element that is none of the five risks and sits underneath all of them. It appears in the introduction and returns when he discusses balance of power: AI already writes much of the code that builds the next AI. Amodei says this is already substantially accelerating the rate at which Anthropic builds the next generation, that the loop "is gathering steam month by month", and that it may be one or two years from the point where the current generation autonomously builds the next.

That loop is what turns an essay about distant risks into an essay in a hurry, and it is also what holds up his geopolitical argument: if each generation designs the next, whoever is ahead can compound that lead and become uncatchable. It is coherent. It is also, and this is worth saying, the part of the reasoning hardest to verify from outside: nobody who is not inside a lab can audit how much that loop really accelerates.

Are we ready? The scorecard

The essay does not answer this question in a section of its own, but the answer is spread across the text. We have gathered it front by front. The verdict column is our reading; the middle one is what he says.

The answer to the question in the title, then, is not "yes" and not "it depends". It is no. And it is not a conclusion we are pinning on him: it is on the first page, in the sentence that gives the whole text its name.

"Humanity is about to be handed almost unimaginable power, and it is deeply unclear whether our social, political, and technological systems possess the maturity to wield it."

Dario Amodei, The Adolescence of Technology, January 2026

That said — and this matters, so as not to turn the essay into what it is not — Amodei is not a doomer. He opens the essay with three warnings, the first of which is an explicit rejection of doomerism, which he accuses of having been "quasi-religious" and of having provoked, through excess, the backlash that today makes legislating anything impossible. And he writes, in as many words, that if we act decisively and carefully the risks can be overcome, and that he would even say our odds are good. The diagnosis is harsh; the prognosis is not.

Four places where we do not buy the argument

Taking a text seriously includes arguing with it. None of these four objections invalidates the essay; three are caveats he writes down himself and the summaries swallow, and the fourth is an actual criticism.

1. The 50% is a prediction, not a measurement

The figure that has stuck to the essay — half of entry-level white-collar jobs — is a prediction he made in 2025 and restates here, not a measurement. He himself recounts that many CEOs, technologists and economists agreed with him, but that others assumed he was falling prey to a "lump of labor" fallacy, and clarifies two things that always get lost in the quoting: that the window is one to five years, and that this is probably not happening right now. What is checkable is not the number but the reasoning: speed, cognitive breadth, advancing bottom-up through the ability ladder, and the industry's knack for closing any gap found in a model within months. That reasoning holds up considerably better than the figure.

2. The metaphor does too much work

"A country of geniuses" is an excellent image for thinking about scale and a treacherous one for thinking about motive. A country has territory, neighbours, history, hunger and fear; a model has a training distribution. He concedes this in passing — the analogy is not perfect, he says — and then leans on it for the rest of the essay, and the reader ends up reasoning about the intentions of an entity that he himself, in the autonomy section, describes as something far stranger: not an agent with a plan, but a system that can fall into weird psychological states. Those two descriptions do not quite fit together, and the essay does not reconcile them.

3. He does not propose basic income. He proposes taxes.

This one deserves its own paragraph because it is circulating wrong almost everywhere. Half the internet has summarised the economic section as "Amodei calls for universal basic income and retraining". We searched the original text for those words: "basic income", "universal basic income", "safety net" and "redistribution" do not appear even once. What does appear, explicitly, is progressive taxation — either general or targeted specifically at AI companies — accompanied by an argument aimed at the world's billionaires that is worth the entry price: if you do not back a good version, you will end up with a bad one designed by a mob.

The difference is not academic. "Basic income" is a vague slogan that commits nobody; "progressive taxation on AI companies", said by the CEO of an AI company, is a concrete position against his own pocket. It is worth quoting properly, if only because it is the harder thing to say.

4. The rigour drops exactly where his interests lie

This one is an actual criticism, and we think it is a fair one. In his list of actors capable of concentrating power, Amodei places AI companies themselves fourth, and introduces it by admitting it is "somewhat awkward to say this as the CEO of an AI company". That gesture is honest. What does not add up is the remedy within that list: for the CCP he proposes chip export controls; for democratic states, legal red lines and even a constitutional amendment; and for the labs — the tier he lives on — he proposes that they "be carefully watched" plus voluntary public commitments, having just said that ordinary corporate governance "is unlikely to be up to the task" of governing AI companies. In the chapter where he lists himself as a risk, the remedy is by a distance the softest in the essay.

The caveat that blunts this criticism should be stated, because it exists: in other sections he does call for binding law on AI companies. In the autonomy section he writes, literally, that "the only solution is legislation", and in the biology section that the time for targeted legislation "may be approaching soon". But that is legislation about transparency and product safeguards. On the power AI companies themselves accumulate — the risk he lists himself — the proposal is still to watch and to trust.

Saying so is not a disqualification. Anthropic has publicly supported laws that impose obligations on it, has published failures of its own model that nobody required it to publish, and maintains classifiers that, per the essay, run close to 5% of total inference costs on some models and cut into its margins. That is more than almost anyone does. But a text that asks for hard rules for everyone and good faith for oneself has its weak point right there, and readers should see it.

What a company that does not build models should do

This is where the essay ends and our job begins. None of the above is going to be solved by a small business in Sant Fruitós, or by an IT department, or by us. But there is a small part that is ours, and it gets decided on an ordinary Thursday afternoon: which process of which customer a tool we do not fully understand goes into, with what permissions, and with whom answering for it.

The four questions we ask before putting an agent to work in production. They are not about AI: they are about governance, which is what this is really about.

1. What data does it see and where does it end up? Not "is it secure?", which means nothing. Which specific fields go into the prompt, in what region they are processed, how long they are retained, and who signed off.
2. What permissions does it hold and who revokes them on a Sunday? An agent with credentials is just another user and deserves the same treatment: least privilege, its own identity, and an off switch someone knows the location of. It is Zero Trust applied to something that is not a person.
3. What happens the day the model, the price or the terms change? This is the domestic version of the concentration risk he describes at geopolitical scale. If the process only works with one provider and there is no documented way out, you have not automated: you have rented.
4. Who signs off on the decision? Responsibility is not delegated to a tool. If nobody can say who answers for the output, the project is not ready, however well the demo runs.

None of these four is our idea or in any way original: they are the same questions you ask before onboarding a supplier with access to your systems. What we have seen is how expensive it is to skip them. We have already written about why AI agent projects actually die — almost never because of the model, almost always because of the process around it — and about what happens when the one running a tireless agent is the attacker. Both posts are far less philosophical than this one, and they lead to the same place.

The world we want to live in

What has stayed with us most is none of the five risks. It is a sentence near the end, where he explains why, in his view, you cannot simply stop: he says the formula for building these systems is so simple it almost emerges spontaneously from the right combination of data and raw computation, and that its creation was probably inevitable the instant we invented the transistor — or arguably when we first learned to control fire. And then: "This is the trap: AI is so powerful, such a glittering prize, that it is very difficult for human civilization to impose any restraints on it at all."

That sentence is uncomfortable because it describes something bigger than AI. Doing a thing because it can be done and because there is money in doing it is not new, and AI did not invent it: it is merely the place where the cost of it finally shows. The question the essay leaves on the table is not "what can this technology do?". It is a much older one, and considerably harder to answer: what world do we want to live in, and are we willing to give something up to get it?

The essay does not end in fear, and that is what surprised us most. It ends by asking for three things in order: that those closest to the technology simply tell the truth about the situation, that decision-makers be convinced this deserves political capital ahead of the thousand issues that fill the news, and that then comes — his words — "a time for courage". And it closes like this:

"The years in front of us will be impossibly hard, asking more of us than we think we can give. But in my time as a researcher, leader, and citizen, I have seen enough courage and nobility to believe that we can win—that when put in the darkest circumstances, humanity has a way of gathering, seemingly at the last minute, the strength and wisdom needed to prevail. We have no time to lose."

Final paragraph of the essay

You can agree or disagree with Amodei about timelines, about the 50%, about China, or about where to draw the lines. We do not agree with all of it, and we have said where. But one thing does not depend on being right: it is worth reading a long text by someone who knows what they are talking about, in full, before having an opinion on the headline. It took us three afternoons. Every one of them was worth it.

Sources. Everything attributed to Dario Amodei comes from a full reading of the original essay The Adolescence of Technology: Confronting and Overcoming the Risks of Powerful AI (January 2026), published at darioamodei.com, not from summaries or third-party coverage. Quotations are given in the author's original English; the essay is freely available for anyone who wants to check the exact wording. The definition of "powerful AI" and the 10–20% annual growth figure come from his earlier essay Machines of Loving Grace, which he cites inside this one. The absence of the phrases "basic income", "universal basic income", "safety net" and "redistribution" is a literal search we ran over the full text. We have used no data that is not in the essay: the third-party figures he cites (the MIT study on gene synthesis providers, the METR evaluation, Rockefeller's fortune) are attributed here exactly as he attributes them there, and we have not verified them separately. Anthropic had no involvement in this text.