Fable Max in incognito:
"What do you think of the quality of the implicit argument against slowing AI advances being made in this post?"
"I'd say the argument rests on genuinely sound economics but makes a weak inference from it — and the weakness is structural, not a quibble.
First, what it gets right. The arithmetic checks out: Clancy models a one-time bump in US per-capita growth from 2% to 2.1%, which permanently raises income by about $90 per person; over 25 working years across 340 million Americans that's roughly $765 billion, and adding a proportional improvement in life expectancy valued conservatively gives a total around $1.2 trillion. Level effects persist, and the costs of forgone progress are diffuse and invisible while the harms of deployed technology are vivid — correcting for that asymmetry of salience is a real service, and it's the strongest thing in Krier's gloss.
Worth noting, though: Clancy's piece is a valuation exercise about progress generally — it never mentions AI or pausing anything. The policy application is entirely Krier's, and that's where the trouble starts.
The central problem is that it's a one-sided ledger presented as a conclusion. The case for slowing frontier AI was never "AI has no benefits"; it's that some posited risks are catastrophic and irreversible. An argument that carefully prices the cost of delay and assigns no value at all to risk reduction can't tell you where the evidentiary bar should sit — it's a price tag on one side of a comparison, not the comparison.
Worse for Krier, the compounding logic he's wielding is symmetric and arguably cuts the other way. If you take seriously that value compounds across people and generations — Clancy's whole point — then a permanent catastrophe truncates that entire stream, and the same arithmetic that makes one year's blip worth $1.2 trillion makes even small reductions in the probability of losing everything worth vastly more. This is Bostrom's astronomical-waste argument, and Paul Christiano's version is sharper still: uniformly accelerating progress mostly changes when things happen rather than whether, so long-run value hinges on differential progress (safety relative to capability), not raw speed. The first comment under Clancy's post raises exactly this longtermist objection. And the deepest irony: Clancy himself, in earlier Open Philanthropy work on the returns to science under technological risk, modeled precisely this tradeoff and found the sign of acceleration's value is parameter-dependent — it can flip under plausible assumptions about catastrophic risk. The source's own author treats this as an open modeling question; Krier treats it as a one-way ratchet.
The number itself can also be flipped. $1.2 trillion cumulative over 25 years is about 4% of a single year of US GDP — roughly 0.15% of the period's total output. Societies routinely pay more than that for insurance against far smaller stakes. If frontier AI carries any meaningful probability of multi-trillion-dollar or unrecoverable harm, and a year of slack buys real risk reduction, that price is cheap. Whether slowing actually buys risk reduction — versus creating compute overhangs, ceding ground to less careful actors, worsening race dynamics — is the genuine crux. But that's an argument the post never makes; "a trillion" only sounds decisive before you divide it by anything.
There's also an equivocation on the policy. "Pausing AI development" stands in for proposals almost nobody makes; the live ones are conditional and frontier-specific (capability evals, compute thresholds, liability), while much of AI's near-term growth contribution comes from diffusing existing models through the economy, which those proposals barely touch. And "the bar should obviously be pretty high" smuggles in the conclusion: whether AI belongs in the permissionless-innovation reference class or the clinical-trials/Asilomar/gain-of-function class — where we accept precaution because errors are irreversible and externalized — is the entire question under dispute. Add the irreversibility asymmetry (over-caution is correctable; some failures of speed are not) and the "obviously" is doing all the work.
To be fair: it's a tweet, and the suppressed premises are ones Cowen and Krier defend seriously elsewhere — that doom probabilities are too speculative to act on, that pauses don't actually reduce risk, that wealth itself buys resilience. If those hold, the trillion becomes decisive because nothing sits on the other side of the ledger. But then the post's persuasive force is borrowed entirely from contested premises it doesn't state, while displaying only the uncontested arithmetic. Even Cowen's own Stubborn Attachments optimizes sustainable growth subject to not courting civilizational ruin — so the crux survives inside his own framework.
Verdict: strong as a corrective to treating delay as costless, and effective rhetoric against vague techno-pessimism. As an argument against the actual strongest case for slowing, it begs the question — and its best tool, compounding, is one the other side can pick up and use."
