Reading Dario Amodei in order

I read Machines of Loving Grace the week everyone else did. On a phone, badly, in a queue. Then I read nothing else he wrote for about a year.

When I went back I read all five in order, oldest first, and it was a different book. So I built a small script to typeset the published pages into a clean e-book with the footnotes intact, because the footnotes carry half of it and I read on an e-reader early, before my kids are awake. The list lives on this site so I stop losing it. Five pieces, October 2024 to June 2026, about four hours of reading.

In order, you watch someone change their mind in public, slowly, with dates on it. Here is what each one did to me.

Machines of Loving Grace, October 2024

This is the optimistic one, and it is the one people quote. Dario Amodei, who runs Anthropic, sets out what a world with very capable AI looks like if it goes well: disease, mental health, poverty, governance, work. He gives it a name, “a country of geniuses in a datacenter”, and a timescale, five to ten years after such a system arrives. He says up front that all of it is guesswork and that most of the details will be wrong.

The famous parts did not move me. The section on limits did.

He asks what becomes the binding constraint once intelligence is cheap, and lists five things: the speed of the physical world, missing data, problems that are genuinely chaotic, human and legal constraints, and physical law. Then this line, which I have quoted at people more than any other sentence I read that year: “Intelligence may be very powerful, but it isn’t magic fairy dust.” (darioamodei.com)

I spend my time around hard-technology companies. Nothing in that world is slow because nobody had the idea. It is slow because the furnace takes a week to cool, the permit takes the better part of a year, the grid connection sits in a queue nobody can jump, and the cells you are testing grow at the speed cells grow. Reading a list of those constraints written by someone selling the intelligence was the first time I trusted the argument.

On DeepSeek and Export Controls, January 2025

Three and a half thousand words, technical, and the one I nearly skipped.

The argument that stayed with me is small and has nothing to do with China. He describes the cost of a given level of AI capability as a curve, and says that when an efficiency gain shifts that curve, spending does not fall. “when it shifts, we simply traverse it faster” (darioamodei.com). What sits at the far end is worth so much that any saving gets poured straight back in.

I had been reading the headlines about AI getting cheaper as if AI were a product with a price. It is not. It is a curve you move along, and the money freed by efficiency is spent on going further along it.

Anyone who works in energy has watched exactly this happen to efficiency for thirty years. I had never joined the two thoughts up.

The Urgency of Interpretability, April 2025

This is the one that changed my actual work.

The claim is plain. Nobody knows how these systems do what they do. He quotes his co-founder Chris Olah’s line that “generative AI systems are grown more than they are built” (darioamodei.com), and says our not understanding our own creations is essentially unprecedented in the history of technology. Then he describes the work of opening them up, and puts a date on the ambition: Anthropic’s stated goal is interpretability that can reliably detect most model problems by 2027, against his own bet that the field gets there within five to ten years.

Most of my working life is verification. Whether a claim about impact can be checked by someone who was not there, and what evidence would settle it. I had been treating AI as a tool inside that process. This essay made me see it as a second object needing the same treatment. A number produced by a system nobody can inspect is not a measurement. It is an assertion with better typography.

I do not think that is what he was writing about. It is what I took.

The Adolescence of Technology, January 2026

Twenty-two thousand words on everything that can go wrong, by the person who wrote the hopeful one. Autonomy, misuse, seizure of power, economic disruption, and the things nobody has thought of yet.

He restates his 2025 warning that AI could displace half of all entry-level white-collar jobs within one to five years, and does not soften it. He also refuses the fatalist position, saying plainly that AI turning on us is not inevitable from first principles, and that treating a loose conceptual argument as proof is a mistake. That refusal is why I believed the rest of it.

The line that stopped me: “I believe we are entering a rite of passage, both turbulent and inevitable, which will test who we are as a species.” (darioamodei.com)

Reading it fifteen months after the optimistic essay, from the same person, with the same definitions, is the whole point of reading in order. I had been sorting people into a camp that thinks this goes well and a camp that thinks it goes badly. These are one argument. You cannot hold the upside honestly without holding the other half.

Policy on the AI Exponential, June 2026

The most recent. It opens with Treebeard, the Ent who takes a whole day to say hello, as a picture of how policy moves next to how AI moves.

The turn is explicit. For years he argued for transparency rules and not much else, on the grounds of keeping options open. Now: “It is time to go beyond transparency to more serious and binding regulation of AI.” (darioamodei.com)

Two things landed. First, he refuses the idea that public worry about AI is a communications problem, saying people are worried because they correctly perceive the risks are real. I have sat in a lot of rooms where the diagnosis was that the public needed better messaging. It is never true and it is always comfortable.

Second, the section nobody talks about: he argues that medical regulation as it stands will cap the upside, because a system built on the assumption that most drug candidates fail cannot absorb a world where more of them work. That is the same shape as the climate problem. The technology is not usually what is late.

Where I still disagree

The chapter on economic development. He offers twenty percent annual growth in the developing world as a dream scenario to aim at, and he is careful to say he is less confident about this than he is about AI inventing fundamental technologies.

He is right to be. I spent six years doing humanitarian work in Lebanon and Iraq. The binding constraint I saw was almost never a missing idea, or a missing analysis, or a shortage of clever people. It was who held power in the room and whether anyone would permit the thing to happen.

He does name this. He writes about corruption and weak institutions and calls them very human challenges. Then the section still lands on the number. Naming a constraint is not the same as pricing it, and in the work I did that gap is where the plans died, one after another, for six years.

I have not changed what I work on since reading these. I have changed the horizon I hold it on, and I have stopped treating the optimistic essay and the frightening one as opposing positions.

The list is at seirayun.com/dario-amodei. If you read all five and think I have the middle one wrong, I would like to hear it.