Dario Amodei, the chief executive of Anthropic, opens his latest essay with Tolkien. AI, he writes, advances at the pace of the Hobbits; our political institutions move like Treebeard the Ent, who needs a full day just to say hello. Policy on the AI Exponential is his attempt to close that gap. Whatever you make of the source, the argument deserves an operator's attention. Here is the essay in brief, and the part worth acting on.

The argument: the model outruns the state

AI capability is on an exponential, Amodei argues, with more than a decade of scaling-law evidence behind it. In four years models went from barely writing a line of code to writing most of the code at the major AI labs, with comparable jumps in biology, mathematics, law and finance. If the trend holds even a year or two longer, he expects what he has called Powerful AI, or "a country of geniuses in a datacenter". Legislation, by contrast, can take years. The mismatch was tolerable while AI looked like "the latest consumer app or cryptocurrency". It is not tolerable now.

His evidence that the moment has arrived is a frontier model he calls Claude Mythos Preview, which in his telling "scrambled the global cybersecurity landscape" and proved "beyond doubt that AI models are now tools of global and national strategic consequence". Biological and autonomy risks, he suspects, are not far behind. His worry is not that policymakers are unwilling, but that even their recent, welcome moves are "at least a year out of step" with the technology. The essay maps five areas that need re-imagining, and Anthropic is backing two concrete first steps with its own money: a legislative proposal on frontier-model testing and a policy framework for job displacement.

In the several years that it can take Congress to act, AI can go from an amusing toy to the full country of geniuses.

Five fronts, in brief

1. Regulation and public safety. The familiar innovation-versus-safety dilemma, sharpened because AI's worst risks (cyber, bio, autonomy) may arrive faster than regulators can study them. Amodei wants transparency requirements and frontier-model testing written into law now, while the window is open, rather than after an incident forces a clumsy overreaction.

2. Macroeconomics and tax. Powerful AI could scramble the old assumption that growth is fragile and must be traded against equality, leaving the dial, as he puts it, "stuck on the hypergrowth, hyper-inequality setting". He is emphatic that enduring job loss is undesirable and to be minimised, not engineered, and that meaning and purpose matter more than money. His policy menu: far better measurement of AI's labour effects, pro-employment incentives (wage insurance, retention credits, training), and, if displacement proves large and permanent, long-term income support funded by the growth itself.

3. Accelerating AI's upside. Here his worry flips. For the fields AI will speed up, such as biomedicine, the danger is that a regulatory system built for a slower era (FDA and EMA pipelines that run seven to eight years) simply jams under the deluge. He urges agencies to define now what it would take to accept AI methods, like simulated toxicology or synthetic control arms, so the benefits are not held hostage to outdated caution.

4. The state and civil liberties. Powerful AI "in the wrong hands could be the ultimate tool of autocracy", enabling a surprise seizure of power that routes around democratic oversight, from autonomous weapons that obey unlawful orders to mass surveillance no civil-liberties law anticipated. His remedies include hard accountability and an "off switch" for autonomous weapons, closing the data-broker surveillance loophole, and a right to AI assistance as capable as the government's during adverse legal action. Notably, he wants checks on companies too, not only the state, citing Anthropic's own Long-Term Benefit Trust as one model.

5. Securing leadership by democracies. AI is not a normal trade good to "diffuse" around the world, he argues, but a strategic reset on the order of nuclear weapons. A nation with powerful AI facing one three years behind could be "an army of World War II Marines facing an army of medieval swordsmen". His prescription: a democratic coalition built on shared values that coordinates the policies above and locks down the chip supply chain, tightening and extending export controls.

What an operator should take from it

You are not Congress, but the pace mismatch is your problem too. The gap Amodei describes between a fast exponential and a slow institution exists inside most companies, between what the models can already do and what your governance, procurement and risk processes were built to handle. Three things carry from this essay into a boardroom:

  • Plan for displacement deliberately, not by accident. Amodei's own stance is the right one: hunt for new uses and new revenue that let your people do more, rather than treating AI as a headcount lever. The firms that come out ahead will be the ones that redeploy, not just cut.
  • Watch the regulatory split. Expect frontier AI to get more constrained and AI-accelerated fields to get deregulated. Which side of that line your sector sits on changes your strategy.
  • Price in the geopolitics. Export controls, chip supply chains and coalition politics are now business variables, not background noise. The concentration of compute and models is a vendor risk worth modelling.

What we tell clients at AvantiGroup.AI

Read it as a primary source, with the obvious caveat that the author runs a company with a direct stake in how AI is governed. The self-interest does not make the core observation wrong: capability is compounding faster than the institutions, public and private, that are meant to absorb it. The leaders who do well over the next few years will be the ones who shorten their own reaction time, whether that means new governance, new contracts, or new ways of putting humans and models to work together. The exponential, to borrow Amodei's frame, is not going to wait for the Entmoot.