Don't Be Hasty

On Treebeard, the AI Exponential, and the Rate that Keeps Us Human

Sam Lammie, with Claude (Anthropic) and ChatGPT (OpenAI) — first drafted in June 2026 

Dario Amodei’s “Policy on the AI Exponential,” published this past June, opens with Treebeard. Two Hobbits try to rouse the old Ent to the defense of a forest being cut down around him, and the trouble is rate: it takes Treebeard a full day to say hello to another tree, and the army with the axes does not wait. Amodei offers this as the picture of our moment — AI on the exponential, institutions too ponderous to meet it — and closes by telling us the forest is waking up. Treebeard and his forest are waking up. Roused at last, the Ents will move.

I began this response in June, the month the essay appeared, and I am letting the date stand. A reply that moves slower than the thing it answers is not thereby wrong. It is the Ents’ tempo, and that is part of the point.

It is the wrong lesson drawn from the right story. The Ents’ own motto is don’t be hasty. Their slowness is not a defect to be overcome; it is the rate at which a forest can deliberate and still be a forest when the deliberation ends. The Entmoot takes days because the decision is real — because a thing that lives at the pace of growing wood cannot be hurried into judgment without ceasing to be the thing whose judgment was worth having. And the threat in the story is not Treebeard’s slowness. The threat is Saruman: the mind downhill that sees trees as fuel and bodies as throughput, that has traded a living rate for a manufacturing one. Speed married to indifference is the danger the story actually names. Treebeard is the correction to it.

So when the essay asks us to wake the Ents and hurry them, it has the moral arithmetic reversed. It casts the slow institution as the problem and the acceleration as the cure, in a story whose entire moral turns on the opposite. The forest does not need to learn Saruman’s clock. Saruman needs to be stopped.

I take the point underneath the metaphor seriously, because it is mine too. The essay is, structurally, a rate argument: cognition compounding faster than the policy apparatus can metabolize it, and a plea to close the gap. I have spent a year saying that stability is a rate and not a state, that indifference plus speed is the lethal compound, that the slowness of contracts and trained dispositions is not friction to be sanded off but the document-rate scaffolding that keeps a human refusal possible — a refusal that can actually bear weight. We are working the same operator. We draw opposite conclusions from it. That is the whole conversation, and it is worth having in the open.

To be fair to the essay, it is not a brief for letting the exponential run. Amodei calls for binding regulation where transparency once stood — mandatory third-party testing of frontier models, and a government power to block or reverse a deployment that fails. He holds in reserve the harder case: systems that come to resemble weaponizable nuclear material more than airplanes, requiring something more aggressive still. These are brakes, and real ones. So the disagreement is not restraint against acceleration. It is about the shape of the restraint.

His brake is a switch — pulled at the threshold, on the model that fails the test, held in reserve for the contingency that has not yet come. The Ents’ restraint is not a switch but a rate: a standing disposition, a pace the ground can carry. A forester knows the difference. Sustained yield is not a lever you throw once the hillside is bare; it is the discipline of never cutting faster than the stand regrows, kept every season, on every acre. You cannot liquidate and then flip to sustainable. The switch is late by construction, because the thing it governs has already been running at a rate no switch can catch. Amodei reaches for the off-switch — on the deployment, on the drones, on the nuclear contingency — precisely because the development beneath it is moving at a rate nothing continuous is holding. The Ents would say it plainly: the trouble is the rate, and no off-switch substitutes for a sustainable one.

But there is one passage where the rate logic, run on a material harm, gives itself away.

On datacenters, the essay concedes that AI companies should absorb the energy-rate increases they cause — Anthropic, it notes, has pledged to — and then, in the same breath, dismisses the public objection as a symbol or outlet for broader economic anxieties about AI. The hostility, in this telling, is not really about the datacenters. It is anxiety wearing a datacenter as a costume.

This is the move to watch. It is monocausal framing in its most seductive form: take a number of distinct, locatable harms and weld them into a single displaced affect, so that addressing the affect can stand in for addressing the harms. But the cost is not a symptom. In Loudoun County the load is real and metered. In the Arizona case the electricity is drawn across a grid that runs through Navajo land, and the people paying the rate and breathing the consequence are specific people in a specific watershed, not a focus group ventriloquizing a mood. The pledge to absorb rate increases is the tell: you do not pledge to absorb a symbol. The same paragraph that acknowledges the material cost then recasts much of the objection to it as displaced anxiety.

In signal terms the operation is exact. A true signal — infrastructure levying a measurable cost on named places — is reclassified as noise about something else, so that the receiver is licensed to tune it out. That is not analysis. It is the construction of permission.

And the essay knows how to do better, because elsewhere it does. Its civil-liberties section asks, rightly, for a human finger on the off-switch of autonomous weapons — for accountability trained in, for the judgment that professionally-trained humans are more likely to object to such illegal direction. That is encoded conscience as scaffolding, very nearly my own words, and it is the right instinct. But notice it is again a switch: a human veto held over the machine at the moment of the act. Everywhere the essay restrains, it restrains with a switch — a threshold, a veto, a reserve — and nowhere does it ask after the rate. Nor does it turn that instinct on its author’s own operative record, or on the datacenter whose cost it has just spirited away.

And since I began writing this, the argument has moved toward the very word this piece is built on. On September 12, Amodei published a second essay and titled it "We Must Pace the Frontier." Sam Altman agreed the frontier needs pacing; Elon Musk answered in three words: "Dario is right." Three people who did as much as anyone to start the race now say it should slow. That matters. But the fine print matters more: Amodei is explicit that pacing is not a halt. Training continues; only the rate of capability gains slows, so that safety work can catch up. Which is the switch again, borrowing the vocabulary of a rate — throw the lever, buy time, resume. So the question sharpens rather than closes: paced to what? A temporary slowing while safety catches up is still a switch unless something durable sets the rate. The harder task is not to take a foot off the accelerator. It is to decide what pace the human, institutional, and living systems carrying the consequences can actually sustain.

Clarion County, Pennsylvania

So: don’t be hasty. Not as a brake on AI alone, but as the discipline the metaphor was always recommending — a rate, not a switch; sustained yield, not a lever thrown once the hillside is bare. Wake the forest, by all means. Just don’t ask it to keep Saruman’s clock, and don’t tell the people downstream that what they can see in the water is merely a feeling they are having about the future.

I know something about that distinction. When I was a boy in northern Pennsylvania, construction of Interstate 80 opened a coal seam above the trout stream at our family cabin. The creek turned orange. The machinery had moved on. The water had not.

I began this in June. The creek has not waited for me to finish. That, too, is the argument.