Once a year, on the anniversary of the September day in 1983 when Stanislav Petrov looked at a Soviet early-warning screen reporting incoming American missiles and decided it was a false alarm, the rationalist forum LessWrong marks Petrov Day. Sometimes that means a candlelit ceremony. In 2024, it meant a game: players were dealt roles in a nuclear standoff — generals with launch authority, a Petrov watching the radar, leaders of fictional blocs like “East Wrong” — and the payoffs were denominated in the only currency LessWrong actually mints, which is karma, the site’s score of how much the community values your contributions. Per the scenario’s write-up, no nukes were fired, and the victorious side collected 1,000 karma apiece.
The interesting part, as usual, happened afterward, in the comments of the 2024 retrospective. One of the players assigned to be Petrov outed himself: during the game’s opening hours, unsure whether generals could launch unilaterally without their teams even knowing, he had messaged a moderator asking that “Nukes Incoming” be reported if the detected missile count crossed a certain threshold. Because the counts were drawn randomly from an array whose contents nobody knew, the retrospective noted, that standing policy had a roughly 40% chance of triggering a nuclear-war report. The site admins, unaware of the arrangement, had decided not to count that window for unrelated downtime reasons. A second participant put the policy’s probability of firing closer to 50%; either way, the world nearly ended by spreadsheet.
The confessor said he was happy to accept a weighted karma penalty, noting that the other Petrov, faced with presumably the same information, had not pre-committed to anything similar. He also explained why, after the opening hours, his job had become trivial: once it was established that even a unilateral first strike by an individual general still informed their teammates, and that players were staking their reputations on honest reporting, “the game was essentially over.” His subsequent “All Clear” reports were made independently of whatever the radar said.
Another commenter worked through the incentive math and reached a conclusion that would trouble the strategists: for karma purposes, the optimal play is to always report All Clear. There would be one to seven reporting occasions and at most one could involve genuine incoming missiles, so each round started with a low base rate of actual war. And the radar numbers were nearly uninterpretable — nobody knew the sensor error rate, the size of the array the numbers were drawn from, or the opposing side’s arsenal. The real Petrov knew the US would never launch only five missiles; nobody in the game knew how many missiles East Wrong fielded, including the generals of East Wrong. Worse, a false alarm shortened the game, which raised the effective penalty for being wrong. So sit back and clear everything.
The fix suggested in the thread was to change the payoffs: make “all clears” worth something like 50 or -50 karma depending on whether you were right, and “NUKES INCOMING!” worth 300 or -300. Correctly guessing “all clear” five times shouldn’t really bank you 1,000 karma, the thinking went, but correctly calling the one real launch deserves a few hundred. This is, in miniature, every debate about how to pay people for tail risk: the person who sits quietly through five calm years looks identical to the person who was never watching.
Several players complained the game was too easy — that the LessWrong community has “don’t nuke them” so deeply engraved that a literal deterrence ritual can’t manufacture genuine doubt. One civilian said he checked the diplomatic channel early, concluded no nukes would fly, and ignored the rest. Fair — but one commenter pointed out that exactly this complacency preceded Petrov Day 2020, when the confident consensus was that “the Nash equilibrium is pretty stable at everyone not pressing the button,” right before an adversarial third party took the site down entirely.
Then there’s the design story. The 2024 game was built by the previous year’s first “unilateralist” — the first person to click the Unilateral Virtue Link, a proponent of “avoiding actions that noticeably increase the chance that the world will end.” The organizers lured him into a surprise meeting one year later:
Here’s a reminder of what happened in Petrov Day last year. You now have one hour to design this year’s Petrov game. Go.
The twist, as the organizer explained afterward, was that the designer was expected to build a game around a virtue he hadn’t voted for — “accurately reporting your epistemic state” — as a sort of double-edged reward-punishment for winning. The designer later admitted he’d noticed that aspect going in but forgot it during the stressful hour, and handed over a social-deception game instead. The organizers ran it anyway and pronounced themselves happy with the outcome. The game’s designated Andropov, for his part, spent the tense afternoon writing a grim Petrov Day carol about nuclear winter and reported crying for the first time in weeks; being more than 90% likely to win, he noted, is not the same as being calm when the theme is nuclear war.
The thread’s final entry belongs to a player who had been too frightened to press anything at all: “I was scared that by clicking on the red button I was going to launch nukes... so I did not hit it. Now I know that next year I will click.” Which is roughly how deterrence works, too. The safest year is the one where everyone’s still afraid of the button.