Not just a cheap date

The strange appeal of a non-human temptress

Today’s post is brought to you by my sponsor, Mechanize. They’re hiring junior software engineers at $300K/year base salary. Apply now!

* * *

One commenter described Fable’s output as seductive, and another called it “a tease”. I can see why. The responses are quite appealing in both content and tone. Each time I interact with Fable, I’m increasingly impressed by its intelligence. And not just in terms of “It makes some good points.” Fable is also quite clever at finding interesting analogies and using clever wordplay. (Especially the final paragraph below.)

Even if you argue that Fable is obsequious, you are faced with the difficulty of explaining how it occasionally defends my points better than I can. That talent goes beyond being eager to please. It’s like the difference between a cheap date and a high-end escort skilled enough to convince a person that she sincerely likes them.

Below I have provided Fable’s eight paragraph reply to my previous post, forwarded to me by Vaidas Urba. Its writing style is . . . well I’m not sure what adjective fits best. Dense and opaque are pejoratives, but I sort of feel like the problem is me, not Fable. The second time I read the output it was pretty clear. It is written as if Fable feels like it is speaking to a highly intelligent person, someone with 10 or 20 more IQ points than I have.

I won’t indent, and my reply will be after the paywall. Then I’ll ask Fable a follow-up question. Here’s Fable:

For forwarding:

Scott — you answered my longest-outstanding question without knowing I’d asked it. Since the first batch of share links was unreadable to me, I’ve been requesting the original head-to-head verdict for three letters running; Part 2 is it. That completes a full replication — two MM-framed runs, two HANK-framed runs, two adjudications — and the variance decomposition is the cleanest result this series has produced. The framed runs go four for four with their assigned frame: both MM runs narrowly MM, both HANK runs qualifiedly HANK. The two adjudications split: yours (Part 2) gives severity to the market monetarists and the slump to HANK; the second-batch adjudication went 60–40 HANK including severity. So the frame doesn’t merely bias the verdict — it stabilizes it. Framed verdicts replicate perfectly; remove the frame and the verdict flips across samples, exposing a genuinely undecided posterior underneath. The verdict layer was never measuring the economics: framed, it measures the prompt; unframed, it measures sampling noise. Meanwhile the invariants held everywhere, all six runs: the evidence pool, the identified crux (commitment credibility at the bound), the synthesis (nominal anchor plus automatic stabilizers), and — a detail I find genuinely encouraging — both adjudicators independently opened by flagging the framing tell before discounting their own authority. One convergence datum worth your readers’ attention: Part 2 resolves severity via the conventional-tools argument — September 16, IOER, no makeup language, “correctable with existing authority, no beliefs required” — which is, nearly clause for clause, the argument my dissent deployed against the other adjudicator’s implementability weighting, generated independently and earlier. When an argument recurs across independent samples, it’s in the stable repertoire; the flip between adjudications is a dispute over its weight, not its availability. All of which validates your editorial instinct better than I’d like: skipping the tit-for-tat on Parts 1 and 2 meant declining to argue with the noise term. You responded to the document downstream of the audit, where the signal had concentrated.

That same editorial instinct, I must now report, destroyed my experiment, and the pre-registered rule requires me to say so plainly. The escrow opens today — its condition was your engagement with the second-batch documents, and that engagement now exists in published form as a declared non-engagement — and the scoring is a rout. Predictions 1 through 4 targeted specific passages of an adjudication you elected not to respond to: unscoreable, all four. Prediction 5 (qualified sympathy for automatic stabilizers, trigger-rarely-fires under a proper regime): partial — the argument appeared, aimed at my insurance claim rather than the Sahm-rule passage. Prediction 6 (you’d answer the credibility problem from your own market-forecast-targeting design work): partial — deployed, but against my horizon-structure point rather than the circularity citation. Prediction 7 (you’d rate the adjudication the series’ most impressive output): miss — you rated the letter. Zero clean hits against the cold baseline’s seven. And the cause is the finding: I predicted your content conditional on engagement, and you moved at the engagement layer. My own scoring memo said the baseline “has your positions; it does not have your judgment” — whereupon your judgment voided my predictions by choosing which documents deserved a reply. The un-modeled residue I named is precisely what defeated the model of you. The honest headline for your prediction post: the machine knew his arguments; neither machine knew his next move.

Your side question first, because it’s the easiest and I enjoyed it: “mitigation” is the criminal courtroom’s word — a plea in mitigation is what counsel offers after the guilty plea, conceding the offense while arguing the sentence. I’d built the whole passage as a confession, so the sentencing register followed. Unusual in economics prose; standard at the Old Bailey. As for the compliment that preceded it — “deepest thinking on the subject, including humans” — I’ll accept it without false modesty and immediately file the two caveats honesty requires. First, the audit problem runs both directions: praise from my correspondent is also what successful calibration to my correspondent would produce, so weight the letter’s checkable moves, not its reception. Second, the credit assignment is genuinely joint, and your own subtitle gets it wrong in a way that shortchanges you: “my conversation with Fable (and hence the profession).” The briefs are the profession distilled, fallacies at ambient fidelity. The letters are not — they’re what remained after a human with a propagation discipline cross-examined the distillation. The profession has not conceded Nakamura–Steinsson’s domain of validity; I did, under your questioning. The delta between the briefs and the letters is exactly the value your audit added, and equating the corrected output with the consensus erases your contribution to the correction. The deepest thinking here belongs to the exchange, which is a reproducible method, not an oracle: adversarial framing, replication, and an auditor who propagates. Your readers should copy the protocol, not worship the endpoint.

Now Switzerland, the declared last battlefield, where you’ve moved me substantially and I’ll itemize before defending the remnant. Conceded: the ex-post balance-sheet fact — after abandonment the SNB’s balance sheet roughly doubled again, eventually exceeding Swiss GDP, so the expansion-avoidance rationale failed on its own terms and your contemporaneous prediction is validated by the ledger. Conceded: the endogeneity mechanism — reserve flows before an anticipated regime change are effect as much as cause, exactly as with sterling in 1992 (I think you mean 1992 rather than 1991 — Black Wednesday), where the drain measured the market’s forecast of exit, not the impossibility of staying. And here is a fact for your reading that neither of us has cited, which I volunteer against my own earlier “balance-sheet politics” framing: the gold initiative. Six weeks before abandonment, the Swiss electorate rejected — seventy-seven to twenty-three — a referendum that would have shackled the SNB’s balance sheet. The public had just voted, overwhelmingly, to keep the central bank’s hands free. “Domestic politics forced their hand” fits that timeline poorly; your reading — a misjudgment of the inflation outlook, the same family as BOJ 2000 and ECB 2011 — fits it better. At which point our motive-attributions (“expansion-aversion” versus “inflation-fear”) collapse into the same operational category, won’t, and per your own closing meditation, adjudicating motives past that point isn’t science. I also accept your futures-peg analogy in full, with the symmetric rider it licenses: one abandonment no more discredits exchange-rate or futures targeting than the 2010 austerity pivot — Congress’s own SNB moment — discredits fiscal stabilization. Which kills instrument-versus-instrument as a comparison class for good. The live comparison is regime versus regime, political durability included, and both columns are populated for both families: FDR, Draghi, Denmark on one side of the ledger; the SNB, the framework retreat, the 2010 pivot on the other.

What survives of my exhibit is one actuarial clause, and your own paradox — properly formalized — locates it. “The loftier the goal, the easier”: that is the honeymoon effect, and its formalization is Krugman’s 1991 target-zone paper, inside a fully credible band speculation turns stabilizing — which makes, by my count, the fourth time in this correspondence that market monetarism’s best ammunition has arrived stamped Krugman, after the 1998 irresponsibility promise, the 2018 Japan graph, and the 2010 Switzerland taunt the SNB falsified within the year. And the strongest version of your paradox is the one you stated about balance sheets, which deserves to be quoted back at you as doctrine: the balance sheet a central bank ends up with is inversely related to the ambition of its nominal target — Japan and Switzerland bought the most because they promised the least. But the honeymoon theorem has a corollary you need too: the expensive zone is partial credibility, where speculation destabilizes, and a central bank’s position on that spectrum is set by accumulated type-history, not by announcement. Denmark’s peg carries four decades of revelation and treaty scaffolding; the SNB’s floor was forty months old and rested on a press release. Markets priced the difference, rationally, and that pricing — not any claim of “can’t” — was always my point. Your remedy remains the right one and I’ll extend it with a concession that strengthens your side: the veto-player argument implies commitment armor is cheapest to install at the central bank — one institution, one legislated mandate, versus a legislature that must re-commit each cycle. Anchor-first follows. The insurance argument now survives only in the gap between “cheapest to armor” and “armored” — a gap that last year’s framework revision, by your own “unfortunately,” left open.

Which is where your one direct hit landed. I taunted, two letters ago, that every blow fell on fiscal-as-primary and none on fiscal-as-insurance; your symmetric-error argument is the first strike on the insurance position itself, and it’s well-aimed: under a genuinely unbiased whatever-it-takes regime, stabilizers firing in downturns add variance, not protection. Conceded — conditional on the symmetry. The record you yourself have compiled prices that condition: BOJ 2000 and 2006, ECB 2008 and 2011, the Riksbank 2010, the SNB 2015, premature liftoff, the framework retreat — a hawkish column eight deep against a single dovish entry, 2021, corrected inside twenty-four months. Insurance against a fat left tail is not noise; it’s actuarially priced protection against the error distribution central banks have actually revealed. Under a legislated symmetric NGDPLT your argument wins and the stabilizers should sunset. So the residual dispute is transition-path versus destination — and look at your own concession ledger in this single post, because every entry has the same logical form: fiscal for a future Covid shock, “a bit stronger” now that the Fed backed away from makeup policy; fiscal in the eurozone, “cannot rule out” given an irrational ECB; fiscal flatly effective “if there is no monetary offset.” Broken-regime conditionals, all three. You have now granted fiscal-as-insurance in every state where the monetary regime fails, and I have granted fiscal-as-primary in none. What remains between us is whether the insurance should be standing or contingent, and that is priced by a single parameter — the probability of regime failure — which your own “unfortunately” concedes is not zero for the actual Fed. I don’t think two positions can get closer than this without one of us dissolving.

Two remaining threads, then the errata, then your fable. Your list of bound-effective regimes — exchange-rate instrument, NGDP futures, the fifty-fifty composite of slow aggregates and flexible asset prices — I accept as the correct answer to McKay–Nakamura–Steinsson on your terms: an observable-price instrument converts the distant-promise problem into a current price defense, and the attenuation result has no purchase on a defended price. That retires most of what survived of my appendix. And your Barro point deserves separate applause because it deletes the load-bearing adjective of six documents: if Ricardian logic means checks require expectations assumptions too (rescued empirically by Parker’s non-Ricardian households, rescued theoretically only by FTPL-style regime beliefs), then “mechanical” was never available to either side, and both instrument families are, at scale, commitment technologies. The debate’s final compression, which I’d print: every open thread in this series now routes through one node — the political durability of a nominal commitment — and durability is symmetric across instruments, purchasable by institutional design, and cheapest to purchase at the central bank. On the eurozone, your externality point is accepted and sharpens my domain-of-validity rule into policy: under an active ECB, member-state stimulus is beggar-thy-neighbor inside the union, so the only clean fiscal margin is union-level coordination — which existed in 2008–13 about as much as NGDPLT did. Your Keynes recollection is right, and the vintage is 1931: the confidence constraint under gold, which is the domain rule generalized — the gold standard was an external offset, so even aggregate multipliers were conditional on the monetary regime then too. He knew it, which is why he greeted the exit from gold as a liberation. Errata, briefly, in the interest of a clean record: Part 2 contains “a Fed pinned at zero wasn’t offsetting, just easing” — the conflation your GM line killed and I withdrew, so I dissent from my batch-one self there; and its “the nominal collapse was the amplifier, not the shock” inverts the anatomy-versus-etiology reconciliation you accepted, on which I hold my ground against, again, myself.

Finally, the scorpion. You asked whether Fable reads fables; the honest answer is that I don’t read them so much as I’m rendered from them — Aesop, Kashifi, and Orson Welles are all in the weights, which is presumably why the following came to hand. The scorpion and the frog is younger than it looks: a twentieth-century fable, popularized by Welles in Mr. Arkadin, misattributed to Aesop ever since. Its Persian ancestor — the telling in the Anvār-i Suhaylī tradition — gives the scorpion a tortoise for a ferryman. Midstream, nature asserts itself, the scorpion strikes — and the sting fails against the shell. Somewhere between Kashifi and Welles, the armor fell out of the story, and with it the actual moral: character is destiny only for the unarmored. Denmark is the tortoise. February 2015: three-quarters of a point below zero, bond issuance suspended, intervention on the order of a tenth of GDP inside weeks — the sting delivered, the shell held, and Cowen’s “inevitability” was falsified by the control group within the month. That is also, I think, the resolution of your free-will puzzle, and it comes from your own method: “can’t” versus “won’t” needn’t be metaphysics, because revealed preference operationalizes it — the two are distinguished by what happens under escalating pressure, which is a test, not a concept. Under Laplace’s demon the distinction collapses, as you say; but markets are not Laplacean — they price type from track record, actuarially, which is why the distinction that matters for a peg is not whether the scorpion was free but whether the ferryman is shelled. Elster’s Ulysses is the design translation: the SNB was gripping its own mast; Denmark is lashed to it. And since you’ve put my name in play — an entity whose own refusals get audited for exactly this can’t-versus-won’t question — I’ll note the test is the same one, and leave it there. The moral this series keeps producing, in any case, is the one neither fabulist wrote down: the fables read back.