Politics Is the Mind-Killer—So Bring a Lab Notebook

A scored forecast review, nine months of receipts, and the choice you have to make now

Epistemic Status: Two separable claims here. First, the meta-claim about how rationalists chose avoidance over engagement and got predictably worse at political reasoning because of it—I'm at about 60% on that, maybe 70% on bad days when I'm scrolling through the wreckage. Second, seven specific forecasts from November 2024 that we can score right now against what actually happened—probabilities in the ledger, each one with cruxes and falsifiers because I'm not just shouting into the void here. Methods: base-rate priors, reference-class forecasting, the stuff Eliezer told us to do before the tribal flags came out. What would change my mind: listed per claim, because I'm serious about this.

Norms Invoked: Politics Is the Mind-Killer, Double Crux, Ideological Turing Test

I. The Setup: How We Chose Comfort and Got Weakness

Okay. Let's be honest about what happened here.

When Eliezer wrote "politics is the mind-killer," he wasn't handing out hall passes to skip class—he was issuing lab safety protocols. The warning was clear: when the tribal flags come out, you suit up. Operational definitions. Base rates. 48-hour restraint on attribution. Public updates when you're wrong. The whole apparatus of rationalist epistemics applied to the domain where motivated reasoning runs hottest.

We read that and heard "smart people don't touch politics."

I can operationalize this if you want the receipts. Scan LessWrong and EA Forum posts from 2020 through 2024—not a lot of object-level policy forecasting with scoring rules attached. No major prediction markets on domestic political outcomes. Minimal calibration posts where someone says "here's what I predicted about governance, here's what happened, here's my Brier score." When politics shows up, it's usually someone asking whether we should be allowed to discuss politics (case in point), which is very meta and very safe and very much not the same as actually doing the thing.

So we ceded ground. We ceded it to people who learned—because the incentives taught them—that speed beats truth, that confidence without evidence wins the attention war, that you can say anything early and loud enough that the correction never catches up. They learned that "the violent left" becomes functionally true if you repeat it enough times, that feelings generated by anecdotes will overwhelm any statistical distribution you want to cite, that motivated reasoning scales beautifully on platforms optimized for engagement over accuracy.

And here's the thing that should bother you: while we were having polite conversations about epistemics in the abstract, they were filling the vacuum we created. We thought we were being wise. We were being soft.

Falsifier for the meta-claim: If someone does a content audit and finds ≥20% of front-page LW posts 2020-2024 engaged in scored object-level political forecasting, I'll revise "avoidance" way down. But I don't think that's what we'll find.

II. November 2024: The Predictions We Didn't Register (But Should Have)

Let me show you what this cost us. Not in the abstract—in specific, falsifiable predictions that a chunk of the rationalist community got catastrophically wrong.

November 2024, there's a strain of thought that goes: Trump concerns are "hyperbolic." Democrats being dramatic. Things couldn't possibly be as bad as they're claiming. Trump might be uncouth, might tweet too much, but he's the reasonable option compared to [gesture at woke excesses, institutional capture by progressives, whatever your particular anxiety was].

This was testable. We have nine months of data. Let me lay it out.

The Ledger (Or: What Actually Happened)

Look at those probabilities. Six of seven under 0.25. One at 0.3 but only because the legal doctrine is messy. If these had been registered as explicit forecasts at 70-80% confidence in November 2024 ("these things won't happen"), the Brier scores would be shameful.

Why Each of These Matters (Because I'm Not Assuming You See It)

FBI leadership going to political loyalists: This isn't about whether Patel or Bongino are personally corrupt—though you should check their histories if you haven't. It's structural. The FBI investigates powerful people, including the president and his allies. When the top two jobs go to people who owe their positions to political loyalty rather than institutional credibility, the incentive structure for those investigations changes. Not "becomes impossible," but "becomes predictably worse." That's not speculation—that's how incentives work.

Military purges: Removing generals and admirals for political alignment rather than performance failures changes the calculation for every officer in the chain. Do you voice dissent about a potentially unlawful order, knowing that your career—and your pension—depends on political loyalty? The Judge Advocates General removals are particularly telling here because those are the military's top lawyers, the people who tell commanders when something crosses legal lines. When you fire them in a coordinated sweep, you're not removing people who were bad at their jobs—you're removing a check on executive military action.

Public health institutional breakdown: When states can't trust federal public health guidance, pandemic coordination fails. Remember COVID? Remember the chaos of state-by-state responses, the competition for PPE, the inability to coordinate testing and contact tracing? We're not hypothesizing that anymore—we're institutionalizing it. California and other states are building parallel public health infrastructure because they've lost faith in HHS. The Tylenol-autism thing matters because it's contested science being pushed as settled fact, exactly the mechanism that destroys institutional credibility from the inside.

Economic disruption from tariff chaos: "Policy uncertainty" sounds abstract until you translate it to decision space. If you're a business deciding whether to build a factory, you're making a 10-20 year investment based on stable rules. When tariff policy can swing wildly on executive whim, you don't build the factory domestically—why would you? Build it elsewhere, pay the tariff once, avoid the risk of rules changing underneath you. This shows up as deferred investment, delayed hiring, communities that never get the jobs that were supposed to come. The counter-evidence is real—some reshoring is happening in pharma and advanced manufacturing—but the modal pattern is disruption.

Federal Reserve independence under assault: Here's what most people miss about this one. The dollar's status as global reserve currency isn't backed by aircraft carriers (not primarily)—it's backed by trust that the Fed makes decisions based on economic conditions rather than political pressure. When a president publicly threatens to fire Fed governors who disagree with him, that trust erodes. The 0.3 percentage point drop in reserve share is small, but it's the derivative that matters, not the level. Once countries start seriously doubting Fed independence, they diversify reserves. Once they diversify, dollar liquidity in global markets decreases. Once liquidity decreases, US borrowing costs rise. The reserve currency privilege that's funded American deficits for decades rests on a kind of institutional credibility that's much easier to destroy than rebuild.

Immigration enforcement tactics: "Aggressive enforcement" is a policy choice you can debate. Masked agents in unmarked vehicles conducting suspicionless stops—that's not a policy debate, that's a civil liberties crisis. When federal agents operate without clear identification, when they can grab people off the street without articulable probable cause, when they conduct arrests at courthouses (chilling access to the legal system), you've created a system where accountability becomes nearly impossible. Who do you sue? How do you verify this was actually a federal agent and not a kidnapping? The slippery slope isn't a fallacy when you can watch the slope getting steeper in real-time.

Government pressure chilling speech: The First Amendment doesn't just prohibit direct censorship—it prohibits government coercion of private censorship. When an FCC chair threatens a broadcaster's license and that broadcaster subsequently suspends a critic of the administration, you have a chilling effect on speech even if the government didn't technically issue an order. The legal doctrine around "jawboning" and state action is genuinely complex, which is why I'm only at 0.3 on this one instead of lower. But the pattern is what matters: government pressure, private action aligned with that pressure, silencing of criticism.

III. What Good Epistemics Actually Looks Like (The Methods Section)

Alright, enough scoring past predictions. Let me show you what we should have been doing all along—not as abstract principles, but as worked examples with actual numbers and sources and falsifiers attached.

Base Rates Before Vibes (Or: How to Not Get Played by the Kirk Shooting)

September 10, 2025. Charlie Kirk gets shot and killed at a Turning Point USA event, Utah Valley University. Within hours—before investigators have a suspect, before any evidence about motive—the loudest voices on the right are already spinning up to full operational speed: "This is war." "The violent left has declared civil war." "They've been dehumanizing us for years and now they're acting on it."

September 12: Utah officials arrest Tyler James Robinson, 22, after a manhunt. Motive still being investigated, but the instant certainty about "the left" doesn't get walked back—it just fades to silence.

Here's how you should have processed that in real-time, if you'd internalized the base-rate lessons we're supposed to know:

Start with the distribution. The Anti-Defamation League tracks extremist murders in the United States—their 2022 and 2023 tallies are overwhelmingly right-wing, often white supremacist. The Center for Strategic and International Studies maintains a decades-long database of domestic terrorist incidents and plots; post-2016 trend shows a rising number of attacks targeting government and political actors, with a clear right-of-center center of gravity. DHS threat assessments and Intelligence Community products have been persistently emphasizing domestic violent extremists, especially the racially/ethnically motivated and anti-government varieties.

Now—and this is important—left-wing incidents exist. James Hodgkinson shooting up the Congressional baseball practice in 2017. Willem van Spronsen's attack on the Tacoma ICE facility in 2019. These are real. They're serious. They should be condemned with exactly the same vigor as right-wing violence, no hedging, no "but context" bullshit.

But they don't flip the distribution. If you're looking at a distribution where 75-80% of extremist murders are coming from right-wing actors and someone gets shot, your prior should not be "definitely the left did this." Your prior should be "hold for evidence, because the base rates don't support instant certainty here, and I know how motivated reasoning works."

The Kirk case was a perfect demonstration of what we're supposed to be immunized against: maximum confidence deployed at minimum evidence, collective blame assigned to entire political factions based on zero confirmed information, and then—when the facts arrive and don't fit the narrative—just silence. No retractions, no apologies, no "we updated our beliefs based on new information." Just on to the next outrage.

That's not epistemics. That's not even good propaganda. That's just opportunism wearing a facts-and-logic costume.

One Ruler for Systems (Or: Stop Doing Selective Accounting)

Here's a move I see constantly from people who think they're being rigorous: "Communism killed 100 million people through state violence and economic mismanagement, therefore capitalism is obviously superior."

Fine. Let's do that calculation. But if we're judging systems, we need one ruler for both—same standards, same accounting methods, same level of scrutiny for deaths that can be traced to system-level choices.

Here's what that looks like when you actually do the work:

Air Pollution Deaths All-source (ambient plus household): approximately 8.1 million deaths per year globally, 95% uncertainty interval running 6.7 to 9.5 million—that's from the Health Effects Institute and Global Burden of Disease study. Ambient PM2.5 alone accounts for roughly 4.1 million. There's a 2018 study estimating around 8.7 million deaths specifically attributable to fossil fuel PM2.5, but that's at the high end of estimates and methodologically contested, so use it with appropriate caution.

Policy lever: clean energy standards, vehicle emissions regulations, industrial pollution controls. Carbon pricing shows measurable reductions where it's been implemented seriously—British Columbia's experience is well-documented here.

Why this counts as system-linked: these aren't deaths from nature or bad luck. They're deaths that move with policy choices about how much we regulate fossil fuel combustion, how strictly we enforce air quality standards, how we price in externalities. Different countries with different regulatory regimes show different mortality rates from air pollution. That's a system effect.

Work-Related Disease and Injury WHO and ILO joint estimate: approximately 1.9 million deaths annually as of 2016 data, dominated by chronic exposures and excessive working hours.

Note: this is separate from air pollution—don't add these categories together because there's overlap (some occupational exposures also contribute to ambient air quality issues).

Policy lever: OSHA enforcement with actual resources and penalties, exposure standards that get updated with new science, maximum hours regulations that get enforced.

Why this counts: work deaths aren't random accidents of an industrial economy. They follow from choices about how much to invest in safety, how to balance profit against worker protection, how much overtime to permit, how to handle toxic exposures. Different regulatory regimes produce measurably different death rates.

Health Coverage (US-Specific Counterfactual) Galvani et al., PNAS 2022: universal coverage could have saved approximately 212,000 lives in 2020 and 338,000 through March 2022 during the pandemic.

Important qualifier: this is counterfactual modeling with explicit assumptions, not historical fact. Treat it as high-quality estimation using standard epidemiological methods, but don't cite it as "universal healthcare would have definitely saved exactly 338,000 lives"—the uncertainty matters.

Mechanism: earlier diagnosis of conditions, fewer financial barriers to seeking care, reduced transmission from people avoiding hospitals due to cost concerns.

Why this counts: the wealthiest country on Earth has people dying because they can't afford healthcare or delay seeking treatment due to cost. That's not a natural phenomenon—it's a policy choice. Countries with universal systems show lower mortality rates for treatable conditions. The counterfactual is real.

Austerity Effects (UK as Case Study) Multiple peer-reviewed analyses link the post-2010 spending cuts to excess deaths and life-expectancy losses running through 2019. Effect sizes vary by methodology—some studies show larger impacts than others—but the direction is consistent across independent analyses.

Policy lever: social insurance, public health investment, counter-cyclical spending during recessions rather than pro-cyclical cuts.

Why this counts: when you slash health and social support budgets during economic stress, people die. Not through bullets or starvation camps—through suicides, delayed medical care, worsening mental health, housing precarity leading to exposure. The deaths are as real as any other kind.

Regulatory Failure Case: Opioids (US) Approximately 80,000 opioid-involved deaths in 2023 based on CDC provisional data. Decline in death rates correlates with scaling of naloxone access, treatment program expansion, prescription monitoring systems, and supply controls.

Policy lever: harm reduction approaches, treatment funding, aggressive regulation of pharmaceutical marketing, legal action forcing behavior changes in drug distribution.

Why this counts: the opioid crisis happened because profit-motivated companies were permitted to market demonstrably addictive drugs with aggressive and misleading claims while regulators either looked away or lacked enforcement power. When policy tightened—naloxone distribution, prescription databases, legal settlements forcing changes in corporate behavior—death rates began falling. That's policy-contingent mortality.

I'm presenting this as a ledger rather than a sum because the categories overlap and you'll get nonsense if you just add them up. Some work-related exposures contribute to air pollution deaths. Some austerity deaths show up in multiple categories. Honest accounting means being explicit about that.

But the order of magnitude is clear: we're talking about millions of deaths annually that move with policy choices about regulation, public investment, and market oversight. If you're going to count communism's democide and forced famines, you need to count capitalism's externalized deaths with the same rigor. One ruler, or admit you're just doing team sports.

Steelman Seriously (Or: Where the Other Side Is Actually Right)

Intellectual honesty requires this: I need to show you where market-leaning and traditionally conservative approaches actually work, where the evidence points their way, where dismissing them would be the same kind of motivated reasoning I'm critiquing.

Policing and Violent Crime The evidence here is pretty damn clear. Chalfin and McCrary's Journal of Economic Literature review, Mello's work on police hiring, multiple quasi-experimental studies with good identification strategies—they all point the same direction. More police leads to less crime, including significant reductions in homicide. Severity of punishment matters much less than certainty of enforcement.

This cuts against simplistic "defund the police" narratives. Smart left-of-center jurisdictions use this evidence—they fund police departments while also funding mental health crisis response, addiction treatment, community programs. It's not either-or, and pretending the police-crime relationship doesn't exist is just ignoring evidence because it's politically inconvenient.

School Choice (Context-Dependent) Boston's charter lottery studies—Angrist and colleagues—show large, RCT-quality gains on college enrollment and test scores. Not all charter schools, not in all contexts, results are highly dependent on the comparison group and the specific implementation, but the best charter networks deliver real educational value to students who win the lottery to attend.

This matters because reflexive opposition to school choice based on ideology rather than evidence is exactly the kind of mind-killed reasoning we're supposed to avoid. Support charter expansion where the evidence shows it works, oppose it where evidence shows it doesn't, let the data decide.

Carbon Pricing British Columbia's carbon tax shows measurable emissions reductions across multiple independent reviews (Pretis et al. 2022 is a good recent one). It's politically difficult, it has distributional effects that need managing through rebates or other mechanisms, but as a policy tool for reducing carbon emissions it works, especially when implemented alongside regulatory standards.

This matters because market mechanisms can solve environmental problems when they're designed well. Carbon pricing internalizes the externality—makes polluters pay for the damage—which is exactly what economic theory says should happen. Dismissing it because "markets bad" is lazy.

Housing Supply Land-use deregulation is necessary (though not sufficient) for housing affordability. California's SB9 shows limited initial uptake due to financing constraints and design requirements, but the direction is right. Combine deregulation with pro-supply public investment for best results, but don't pretend that keeping restrictive zoning is somehow progressive when it's pricing people out of cities and forcing brutal commutes.

Pattern you should notice: the rational left isn't anti-market. It's anti-externalized-harm. Use markets where they minimize damage and achieve social goals efficiently. Regulate them where they create harm or fail at coordination. It's really that simple.

Weight Evidence by Source Track Record (Or: Why Official Health Statements Should Come With Discounts)

One more critical piece of the epistemic toolkit: Bayesian updating on source reliability, not just claim content.

When the White House releases statements about presidential health, how should you weight them? Not just based on what they say, but on their track record of disclosure, transparency, and accuracy.

Historical pattern: presidential physicians issue glowing summary letters with minimal underlying data. When pressed for details, they're sparse. The July 2025 disclosure of chronic venous insufficiency came only after sustained public attention to leg swelling visible in photographs—it wasn't proactive transparency, it was reactive damage control.

September 2025: footage from the 9/11 ceremony shows facial asymmetry that multiple observers note as concerning. Official response: dismissive, no neurological workup publicly released, no independent medical assessment.

The Bayesian update you should make: when a source has consistently minimized or delayed disclosure of concerning information, their reassurances carry reduced weight. You don't jump to "confirmed stroke"—I'm not doing that, and you shouldn't either. But you also don't treat "everything is fine" as dispositive evidence of absence.

This generalizes: when evaluating any claim, factor in the source's historical accuracy. Media outlets with robust correction policies should score higher than those without. Officials who've demonstrated transparency should be trusted more than those who've stonewalled. Researchers who preregister studies and share data should be weighted more heavily than those who don't.

The formula is something like: Evidence Weight = (Quality of Methodology × Independence of Source × Track Record of Honesty) / Burden of Proof for the Claim

IV. Adversarial Testing (Or: What Would Actually Change My Mind)

Let me make this concrete. Here are my cruxes—the things that, if they changed, would make me significantly revise my conclusions. Not just "update a little"—I mean substantially alter my model of what's happening.

Double Crux: The Five Things That Would Flip My Assessment

1. Reserve Currency Status If the dollar's share of global reserves recovers 1 or more percentage points despite continued political pressure on the Federal Reserve, I'd downgrade my financial stability concern from current 0.8 to maybe 0.4. That would suggest the structural advantages of dollar dominance outweigh the credibility damage from political interference.

2. Manufacturing Investment If US non-residential fixed investment in manufacturing outpaces G7 peers for four consecutive quarters while tariff policy remains volatile, I'd revise my tariff disruption assessment from 0.75 down to 0.3. That would indicate reshoring effects dominate uncertainty costs.

3. Institutional Independence If the Senate blocks three or more political appointees to traditionally independent positions OR if career officials replace political appointees in FBI/HHS within 18 months, I'd reduce my institutional capture concern from 0.9 to 0.5. That would show the system's immune response is working.

4. Extremist Violence Distribution If ADL and CSIS data show left-wing share of extremist violence rising above 40% for two consecutive years, I'd significantly update my base-rate priors about threat distribution. Current data shows ~75-80% right-wing; a shift to 40% left would change my instant priors on ambiguous cases.

5. Public Health Coordination If five or more states that formed independent health compacts rejoin federal coordination under current HHS leadership, I'd revise my public health fragmentation concern from 0.85 to 0.4. That would suggest the breakdown is less severe than it appears.

Ideological Turing Test: Steelmanning the Other Side

Let me try to pass an ITT for a rationalist who voted Trump in November 2024. Here's what I think their best argument looked like:

"Look, I see the institutional risks with Trump—I'm not blind to them. But the progressive capture of scientific and medical institutions creates different tail risks that might be worse in expectation. Think about the lab leak hypothesis suppression, the unwillingness to distinguish between different kinds of COVID risk by age cohort, the ideologically-driven public health messaging that undermined its own credibility. Trump's chaos might actually force beneficial decentralization—when institutions become unreliable, you want the system to route around them rather than doubling down on centralized control.

On foreign policy, unpredictability might actually deter adversaries more effectively than predictable weakness. The European security establishment got complacent under Obama/Biden; maybe they need the fear that US support isn't guaranteed. On China specifically, aggressive tariff policy might be clumsy but at least it's directionally correct given their mercantilism.

For AI risk—which should dominate a rationalist's expected value calculation—a chaotic regulatory environment might be better than premature heavy-handed regulation that locks in the wrong paradigm or advantages Chinese development. The institutional damage is real but time-limited; Trump can't be president forever. But entrenched progressive bureaucracy, once established, could last for decades.

So yes, I'm trading some institutional stability for avoiding what I see as the larger risk of institutional capture by an ideology that has shown itself willing to sacrifice truth to political goals. On net: choose the devil you can watch."

Why My Forecast Ledger Still Prefers the Alternative

Even granting that steelman, here's why I think it's wrong:

The institutional damage isn't abstract—it's measurable loss of coordination capacity. Public health fragmentation makes pandemic response harder, not easier. The "creative chaos" theory requires evidence that decentralization improves outcomes, but we're seeing the opposite: states building parallel infrastructure because they can't trust federal guidance creates waste and increases the difficulty of coordinating during the next crisis.

The time-limited assumption is questionable. Institutional damage compounds in ways that are hard to reverse. Trust, once lost, is extraordinarily difficult to rebuild. If the FBI becomes known as politically captured, the best career prosecutors won't want to work there. If military leadership requires political loyalty over competence, the institutional knowledge accumulated over decades leaves and doesn't come back.

The "creative destruction" frame might work for startups; it's catastrophic for institutions that take decades to build. You can't just spin up a new FBI or CDC when the old ones lose credibility. The second-order effects—brain drain, difficulty recruiting talent, international partners losing confidence—persist long after the immediate crisis.

Most crucially: the predicted harms were specific and verifiable, and they've occurred. That's not "chaos forcing adaptation"—that's just chaos. The burden is on the Trump voter to show evidence that the chaos is producing beneficial adaptation rather than just destruction. I don't see it.

V. Forward Predictions (Or: Putting Skin in the Game)

Alright, enough scoring the past. Let me make falsifiable forecasts with actual resolution dates so you can check whether I'm calibrated or just bullshitting:

1. Federal Reserve Independence P(Fed governor removal attempts subside; no public threats for 6 consecutive months) = 0.35 Resolution date: March 31, 2026

2. Manufacturing Rebound P(Manufacturing PMI ≥50 for at least 4 of the next 6 months) = 0.4 Resolution date: March 2026

3. Public Health Fragmentation P(At least 2 additional states formalize independent public-health compacts beyond current ones) = 0.6 Resolution date: September 2026

4. FBI Leadership Turnover P(FBI deputy position held by someone with career FBI/DOJ background, replacing Bongino) = 0.25 Resolution date: September 2026

5. First Amendment Litigation P("Jawboning" First Amendment case against federal official advances past motion-to-dismiss stage) = 0.55 Resolution date: September 2026 (based on typical litigation timelines)

I'm committing to score these publicly when the resolution dates hit. Brier scores will be calculated and posted. I invite you to register your own predictions and we can compare calibration.

If you want to actually bet on these: I'll take reasonable stakes on any of them through whatever mechanism LW prefers for this kind of thing. Money where mouth is and all that.

VI. The Protocol (Or: What We Should Have Been Doing All Along)

Let me distill this into concrete practices you can actually use:

The 48-Hour Rule on Attribution No assigning blame to political movements without an affidavit or charging document you can cite. Not "sources say," not "it's obvious who benefits," not "everyone knows." Documents or silence.

Base Rates Before Screenshots Check the distribution before reacting to anecdotes. Ask "what's the historical pattern here?" before deciding the viral video represents a trend. Start with longitudinal data from official sources, not the most enraging thing in your feed.

Weight Sources by Track Record When a source has consistently misled or minimized, discount their denials accordingly. Demand transparent underlying data, especially for extraordinary claims. Track record of honesty should be a major factor in your evidence weighting.

Count the Normal Dead If it happens every year and moves with policy changes, it belongs in your moral accounting. Don't limit "system deaths" to dramatic events—most harm is quiet, chronic, and invisible until you actually count it.

Follow the Ledger-Keepers Trust people who start with the world and then build models, not people who start with models and cherry-pick the world to fit. From Marx's factory inspector reports to modern design-based empiricists, the pattern holds: measure first, theorize second.

Defend Universal Norms Condemn political violence without equivocation or "but context" hedging. Resist collective punishment rhetoric even—especially—when it's your side doing it. Lower the temperature after atrocities, even when it's your political enemies who need the cooling off.

Show Your Uncertainty Include confidence intervals. Mark contested claims as contested. Distinguish between settled facts and modeling exercises. Precision about your uncertainty is a form of honesty that builds credibility.

Do Public Updates When your predictions fail, document it. Post your Brier scores. Explain what you got wrong and why. That's how you build calibration over time. The goal isn't to always be right—it's to get less wrong over time and to be honest about the process.

VII. The Choice You Have to Make

So here we are. If you're a rationalist who dismissed warnings about Trump in 2024 as "hyperbolic" or "Trump Derangement Syndrome" or whatever your preferred frame was, you're now sitting with nine months of evidence that those warnings were... pretty much on target.

You've got two options here, and they're not symmetrical.

Option 1: Dig In Explain away each development as isolated. Find reasons why this time is different. Maintain that the people who predicted this just got lucky—they were wrong to be confident even if they happened to be right about the outcome. Keep treating politics as the mind-killing domain you're too epistemically sophisticated to engage with. Preserve your self-image as the rational one while everyone else is captured by their biases.

Option 2: Update Actually examine what the accurate predictors saw that you missed. Identify where your model failed—was it the inputs, the processing, the priors, the way you weighted different kinds of evidence? Strengthen your epistemics for next time. Acknowledge that epistemological avoidance made you worse at political prediction, not better.

Option 1 is more comfortable. It lets you maintain that you were being appropriately cautious, that you were avoiding mind-killer territory, that you were right to dismiss the warnings even though the warnings turned out to be correct. You can tell yourself a story where you're still the epistemologically sophisticated one and it's just bad luck that reality happened to align with the hysterics this time.

Option 2 is uncomfortable as hell. It requires admitting you got something important wrong. That people you dismissed as mind-killed had better models than you did. That your epistemological avoidance had real costs in prediction quality. That maybe—just maybe—the thing you were avoiding wasn't "politics" but "politically inconvenient conclusions."

But Option 2 is the only one that makes you better at this going forward.

Look, you don't have to become a leftist. You don't have to adopt any particular political identity or join any tribe. But if you're genuinely angry about FBI leadership going to political podcasters, if military purges bother you, if public health chaos concerns you, if you think immigration enforcement tactics have crossed lines, if government pressure on media critics seems problematic—and you dismissed predictions of exactly these outcomes as "hyperbolic" or "TDS" before November 2024—then basic intellectual honesty requires you to go back and do the work.

Read what the people who got it right actually wrote. Not to agree with all of it, but to figure out what they saw in the map and territory that you missed. Identify the inferential steps they took that you didn't. Update your priors based on which models predicted better. Build better prediction mechanisms for next time.

That's not political conversion. That's epistemic repair. It's what we're supposed to be good at.

The left-of-center position—at least the version I'm defending—isn't a tribal identity or an aesthetic or a vibe. It's what falls out of a particular loss function (minimize preventable death, maximize human flourishing, internalize externalities, account for power asymmetries) combined with empirical methods that privilege measurement over assumption. When you actually do the counting, when you apply one ruler to system-level harms, when you weight evidence by source track record, when you update on prediction failures—you trend in that direction. Not because of moral superiority or tribal affiliation, but because that's where the ledger points when you do the accounting honestly.

The mind-killer was never politics itself. It was doing politics without epistemological armor—without base rates, without consistent standards, without source credibility weights, without update mechanisms when predictions fail.

Eliezer warned us to build that armor. We chose avoidance instead. The results are in, they're scored, they're sitting right there in the ledger.

Time to suit up.

Call for Adversarial Collaboration

I'm serious about this: I want a right-of-center rationalist to co-author the next iteration of this ledger with me. We'd agree on operationalizations for each claim, what evidence would change our minds, scoring rules and resolution dates, commitment to public quarterly updates. No gotchas, no dunking, just joint forecasting with opposed priors and a shared commitment to finding out which model predicts better.

If that sounds interesting, message me through LW. Let's do this properly.

Primary Sources: Kirk incident (Reuters, AP September 2025); Extremist violence data (ADL annual reports, CSIS database, DHS Homeland Threat Assessment 2024-25); Post-election developments (FBI.gov, Reuters, Military.com, PBS, Politico, The Guardian, Tax Foundation, Joint Economic Committee, ACLU); Air pollution mortality (HEI/GBD State of Global Air 2024); Work-related deaths (WHO/ILO joint estimates); Health coverage modeling (Galvani et al. PNAS 2022); Policing effectiveness (Chalfin & McCrary JEL review); Carbon pricing (Pretis et al. 2022)