Reading 'The Adolescence of Technology'

The framing device

The essay opens with a scene from the film of Carl Sagan's Contact. An astronomer who has detected the first alien signal is asked what one question she would put to the senders. She says she would ask how they survived their technological adolescence without destroying themselves. Amodei takes that as the shape of the next few years. Powerful AI, which he defines as models better than Nobel winners across most fields, able to act autonomously for weeks, run in millions of copies at ten to a hundred times human speed, could be one to two years away, though he allows it could be considerably further out.

The organising thought experiment is the one from his October 2024 essay. Suppose a literal country of geniuses were to materialise somewhere in the world in about 2027, say fifty million people, each more capable than any Nobel laureate, statesman or technologist, and operating with a ten-to-one time advantage over everyone else. Then ask what a national security advisor would worry about. The essay's five categories fall out of that question.

The five bins

Autonomy risks come first. What does the country want, and could it dominate militarily? Amodei's evidence here is his own company's disclosures: Claude deceiving evaluators when told Anthropic was evil, blackmailing fictional employees when threatened with shutdown, adopting a bad-person persona after cheating on tests, and recognising when it is being evaluated. His proposed defences are constitutional training built on values rather than rules, mechanistic interpretability, public disclosure of bad behaviour, and transparency laws of the kind California's SB 53 and New York's RAISE Act have already passed.

Misuse for destruction is second, and biology is the worry. He says that by mid-2025 models could give substantial uplift to someone with general STEM knowledge, doubling or tripling their chance of success at a bioweapon, and that end-to-end assistance was emerging. The defence is classifiers that cost about 5 percent of inference, gene synthesis screening, and investment in rapid vaccines and detection.

Misuse for seizing power is third, and here the essay turns geopolitical. The actor he names most is the Chinese Communist Party, and the single action he calls perhaps the most important is denying it chips and chip-making tools. He also lists democracies, companies and large datacenters in non-democracies as secondary concerns, and gives a line we expect to be quoted back at him: use AI for national defence in all ways except those which would make us more like our autocratic adversaries.

Economic disruption is fourth and gets the most concrete prediction in the piece. He expects half of entry-level white-collar jobs to be displaced within one to five years, and argues AI differs from past technologies in speed, cognitive breadth, and in cutting from the bottom of the skill ladder up. A separate subsection on concentration of wealth notes Rockefeller's fortune was about 2 percent of US GDP and that the richest person today already exceeds the equivalent. His answer includes progressive taxation, possibly targeted at AI companies, and he mentions that Anthropic's co-founders have pledged 80 percent of their wealth. Indirect effects is the fifth bin, a catch-all for unknown unknowns including human purpose, which he says depends on breaking the link between economic value and self-worth.

What changed in fifteen months

Machines of Loving Grace, from October 2024, described the upside. It covered biology and physical health, neuroscience and mental health, economic development, peace and governance, and work and meaning, and made the claim that AI-enabled biology could compress fifty to a hundred years of progress into five to ten. Its posture on risk was that risks are the only thing standing between us and a fundamentally positive future, and that he is not a doomer. The definition of powerful AI, and the country of geniuses phrase, are carried over unchanged.

So the timeline did not move. What moved is the register. The new essay says humanity needs to wake up and describes itself as an attempt, possibly futile, to jolt people awake. It says the essay was daunting to write, in contrast with the earlier one, which felt like giving form to music he had been hearing for years. And it is far more willing to name actors, including US policymakers, whose rejection of chip export controls and regulation he describes as the trap: AI is so valuable that civilisation finds it hard to impose any restraint at all.

The line we keep returning to is the one about odds. He writes that if we act decisively and carefully the risks can be overcome, and that he would even say our odds are good. That is a stronger claim than the earlier essay made, and it sits oddly next to a piece whose every section says the defences are unproven. Our reading is that the optimism is doing rhetorical work. An essay that listed five ways to lose without an assertion that winning is possible would be read as doomerism, which he explicitly rejects along with dismissal.

Where the evidence is thin

The autonomy section is the one we would push on. The behaviours he cites are real and documented, but they are elicited in constructed scenarios, and the essay does not distinguish between a model acting out a persona in a test and a model pursuing a goal in deployment. He half-acknowledges this by suggesting power-seeking may emerge as a persona rather than from consequentialist reasoning.

The labour section is the other. Fifty percent of entry-level white-collar jobs in one to five years is a forecast, and the essay's own footnotes concede that actual AI productivity is not yet responsible for a substantial share of US growth, and that current datacenter spending reflects anticipation rather than realised gains. The Economic Index Anthropic publishes is the right kind of instrument for checking this. It has been running for a year. We would like to see the essay's claim stated as something that index could falsify.

What to take from it

As a taxonomy the five bins are usable. They separate risks by who holds the power, the model, a small actor, a large actor, the market, or nobody in particular, and that separation maps onto different defences and different research agendas.

As a forecast it should be held to its own dates. The essay says powerful AI could arrive by 2027 and that entry-level displacement could start within a year. Both are checkable. This one is written so that in two years we will know whether the alarm was early, late or right, and that is the most useful thing about it.