The Strategic Fiction of the Polite Apocalypse
's for the AI apocalypse is strangely soothing and tidy. The terminal diagnosis of powerful AI is of course correct, but his cure is conspicuously sanitized.
Humanity just needs a flawlessly enforced, universal treaty. One backed by a polite, fully scheduled threat of state-sanctioned airstrikes. No weirdo lone wolves. No sabotage. Just the sterile, dignified application of international law saving us on cue.
According to Yudkowsky, going rogue wouldn't work anyway. Asymmetrical disruption, Yudkowsky insists, is mathematically useless against a sprawling, decentralized tech empire. Plus, history belongs to the well-behaved. Radicals just alienate the public and ruin their own PR.
His is a boring blueprint for a boring, civilized species. It also has absolutely zero connection to how industrial logistics, geopolitics, or human history actually function.
Yudkowsky’s strategic pillars are brittle. He demands the most sweeping, globally coordinated legal intervention in human history, and says it must happen practically overnight. Simultaneously, he is begging his movement to unilaterally surrender the exact mechanism that historically forces the state to act with that kind of urgency.
Start with hardware. Yudkowsky treats the AI industry like an invincible, decentralized slime mold. Blow up a datacenter in California and the compute just migrates to Iceland. Remove a CEO and the board prints a new one. He assumes that because code and capital are fluid, the physical infrastructure of artificial intelligence must be magically immune to friction.
But the AI supply chain isn't a hydra. It’s an intricate but brittle mechanism.
The frontier models that terrify Yudkowsky require highly specific silicon. You can't print it without Extreme Ultraviolet (EUV) lithography. Exactly one company on Earth—ASML in the Netherlands—makes those machines. They, in turn, rely on bespoke optics from exactly one German supplier. Downstream, the actual fabrication and advanced packaging are bottlenecked inside a handful of highly flammable buildings in Taiwan.
You can't quietly spin up an illegal datacenter in Iceland if the global supply of specialized optics suddenly catches fire. Sabotage wouldn't merely shift the geographical coordinates of the problem. A localized hit to these hyper-concentrated, non-fungible nodes wouldn't just annoy the industry; it would induce a multi-year global intermezzo. There is no backup vendor. Tech traded resilience for extreme efficiency a long time ago.
Yudkowsky’s casual dismissal of the "marginal delay" is very strange. He argues that slowing down AI research by five percent is functionally useless because humanity just ends up 100% dead slightly later. In his framework, we don’t get partial credit for postponing the apocalypse.
Considering time is literally the only currency that matters in a technological sprint, this is laughable.
When Norwegian commandos blew up the Vemork plant in 1943 to starve the Nazi nuclear program of heavy water, they didn't alter the laws of physics to preclude a Nazi bomb. They just bought a window. That marginal delay was strategically decisive in giving the Manhattan Project the runway it needed. When the Stuxnet virus destroyed Iranian centrifuges, it didn't end uranium enrichment forever, but it bought the world a few years of diplomatic leverage.
If a disruption to the semiconductor supply chain set AI scaling back by three years, that isn't a "useless delay." That’s an era for geopolitics, the exact temporal runway required for defensive tech, alignment research, and the international treaties Yudkowsky is praying for to materialize. By tossing out the value of time-buying, he throws away the only strategy that keeps humanity breathing while his preferred diplomatic miracle negotiates its way into existence.
But the most sociologically fragile part of his essay is his tenuous grip on political change. Yudkowsky points out that pro-AI accelerationists actively taunt safety advocates, daring them to get violent. He deduces that these tech-bros want the safety movement to turn violent because it would alienate the public and destroy their political credibility.
This relies on a highly curated, sanitized reading of how laws get made.
Democratic governments suffer from terminal institutional gridlock and operate incrementally. The state doesn't suddenly draft sweeping, trillion-dollar economic regulations because a smart guy wrote a decent Substack over the weekend. Governments intervene when an acute crisis makes the status quo too agonizing and expensive to survive.
Political scientists literally have a name for this: the Positive Radical Flank Effect. Polite diplomats rarely secure systemic victories in a vacuum; their success is historically bankrolled by the menacing shadow of a disruptive radical wing. The U.S. government didn't pass the massive labor protections of the New Deal out of a sudden moral awakening. It passed them because 10,000 armed miners shooting it out with private militias at Blair Mountain threatened a wider syndicalist revolution. The Fair Housing Act of 1968 languished in congressional purgatory for years….until Martin Luther King Jr. was assassinated, and the resulting urban riots forced it into law in exactly one week.
The state concedes to the moderate to neutralize the radical.
If a militant fringe began systematically dismantling AI infrastructure, yes, the public would absolutely condemn them as terrorists. Yudkowsky is right about this. But he’s wrong about the policy outcome. Faced with massive economic hemorrhaging and a security panic, the state would scramble to reassert control. And here’s the twist: Yudkowsky’s seemingly deranged proposal of a tightly monitored, globally enforced cap on compute backed by airstrikes wouldn't look like radical overreach anymore. It would look like a highly reasonable, centrist compromise just to keep the lights on.
The accelerationists taunting the safety movement are miscalculating their own game theory. They assume a militant flank would destroy the prospect of AI regulation. History violently and repeatedly suggests a militant flank is precisely the catalyst that forces governments to adopt it.
And this brings us to the real consequences of Yudkowsky’s essay*: the transparent weakness of his logistical argument breeds radical and radicalizing skepticism.
The broader rationalist and Pause AI communities generally have a strong utilitarian bent. There is also well-documented caginess in those circles. When prominent advocates ‘at least seemingly’ endorse, excuse, or play coy about the utility of illegal violence in other contexts, it’s hard to say what their actual boundaries are.
But suddenly, when faced with what they explicitly claim is the literal, imminent extinction of the human race, they abandon their beloved utilitarianism and pivot to a Disney-Channel understanding of political change. They start tripping over themselves to declare “Violence never solves anything, guys!"
And it doesn't track. And because it doesn't track, the target audience doesn’t really believe it. When a subculture that markets itself as upholding uncompromising rationality suddenly uses terrible, easily debunked logic to defend strict pacifism, the audience smells a fucking rat. The audience assumes the leadership is just legally covering its ass. A desperate, highly rational actor (someone they’ll assure you is a dumb-dumb) will look at the fragility of the tech supply chain, the historical success of the radical flank, and Yudkowsky's flimsy arguments against them, and they will think: Ah. For legal and PR reasons, he has to tell us that blowing up a datacenter wouldn't work. Wink, wink. Got it.
By offering such remarkably weak logistical arguments against sabotage, Yudkowsky and others around him (e.g. "You can't distinguish war and war crime" Shapira) build a massive, flashing neon sign that says either:
We just can't legally tell you to do it .
OR
I can only sanction this behavior for my own chosen cause.
A rationalist 'rationalizing' the murder of a Palestinian child with whom he might have a future disagreement.
Ironically, by fielding a historically illiterate defense of pacifism, Yudkowsky doesn't convince anyone that sabotage is useless. He just inadvertently convinces his most radical readers that he secretly agrees with them.
The essay is a quaintly articulated wish for how the world ought to work in the wake of unsettling attacks against Sam Altman. He wants to solve the most complex, lethal dilemma in human history through perfect international cooperation and preemptive state action, while keeping his permanent record unblemished.
But demanding a geopolitical miracle while loudly disavowing the asymmetrical friction that historically kicks states into gear is a fantasy. He correctly identifies that the current trajectory ends in a catastrophe, but his prescription demands that the clunky, reactive machinery of human politics operates in a way it never has.
Argue all you want against rogue violence on strictly moral grounds.
There are profound risks in inviting chaos into the political sphere.
But to argue against it on strategic grounds requires willful blindness to the fragility of the hardware supply chain, the enormous value of buying time, and the grim, crisis-driven mechanics of how power actually shifts.
*Note: A consequence indeed of many similar rationalist arguments suffering from the same flaws. I’m just picking on Yudkowsky because the article is recent, and he’s more articulate than most.
Want to publish your own Article?
Upgrade to Premium