The p(doom) song, annotated

Somebody wrote a pop song out of the AI safety vocabulary, and it works. Two minutes thirty-seven, four verses, a chorus that adds a new reason to be frightened each time round. The joke is that none of the terms are invented for the song. Every one of them is a real thing with a paper or a post or a meme behind it, and a reader who followed all of them would come out the other side knowing roughly what the field argues about.

So here is the index. Forty-three references in the order they appear, what each one points at, and when it happened. The lyrics belong to their authors and are not reproduced here; this page is about what the lines point to.

Where it came from

osmarks wrote most of the lyrics. The opening verse and chorus are MusicPerson's. The early sections date to around 17 April 2024 and the rest to 8 and 9 November 2024, discussed on EleutherAI along the way, and the original was a Udio render in November 2024. Udio's public beta had opened on 10 April, a week before the first verses were written on it, which is a small thing but it dates the whole genre fairly precisely.

The version above is the Claude Pop re-sing, a Suno render, uploaded in September 2026. By then there was a scene: alkjash's post "Hello World, AI Doompop" on LessWrong, on 11 October 2026, lists this song under "p(doom) goes up" and another called "Nothing Went Foom!" under "p(doom) goes down."

The line-by-line source for what the references mean is osmarks' own annotation page, P(Doom) Song Objectively Correct Interpretation. Where this index says what a line is doing, that is where it comes from. Where it gives a date, the date is ours, because the annotation page dates almost nothing.

The index

Forty-three entries. "Where" is the section of the song: V for verse, C for chorus. Twenty-one of them have a definition page on this site, linked from the reference; the rest are jokes, or words nobody owns, or things that only mean anything inside the song.

What the dates cost

Every date above was checked against a source before this page went up, and six of them were wrong in the first draft of this index rather than merely vague. That is worth saying out loud, because the dates are the part a reader cannot check at a glance and the part a model will cheerfully invent.

Grokking is the one worth knowing about. It is easy to describe a sudden drop in training loss as "grokking", and plenty of people do. Grokking is the opposite shape: test accuracy jumping from chance to perfect generalization long after the loss has bottomed out and the model has overfit. One is a drop during the run; the other is a late jump on data the model never saw. The song is pointing at the first.

The rest: Kevin Roose's conversation with Sydney was on Tuesday 14 February 2023, not the 17th, and the column ran on the 16th. Microsoft capped Bing at five turns per session on the 17th, the day after. Timnit Gebru left Google in December 2020, which is before the stochastic parrots paper was published, so the paper cannot have cost her the job in that order; the dispute was over the draft, she says she was fired, Google said it was accepting her resignation. Nvidia's market value is a moving number and needs its dates attached. The song's compute figure is a rate rather than a total, which means the usual comparison against a training run's total FLOP is comparing two different quantities.

And one claim did not survive at all. The killswitch engineer is a real joke that really circulates, usually with somebody in the replies believing it, but nobody has a primary source for who wrote it or when. It keeps its row in the index because people search for it. It does not get a date.

The atomic twins

Here is the part of this page that nobody else has, and the reason it was worth building.

Almost every idea in the song has an ancestor in the story of the bomb. Not a loose resemblance. The same argument, made once before by people who were also building a thing they were not sure they should build, and who also had to put a number on it.

Start with p(doom) itself. In July 1942, at a meeting in Berkeley, Edward Teller raised the possibility that a fission weapon might ignite the atmosphere. Oppenheimer took the worry to Arthur Compton in Chicago. Compton's account, given to Pearl Buck seventeen years later, is that he decided he would stop the project if the chance of vaporising the Earth came out above roughly three in a million. The calculations came out below it and the work went on. That is a p(doom), with a threshold and a decision rule attached, in 1942.

Two of those rows changed under checking, which is the usual fate of a good story. The axe at Chicago Pile-1 is real: Norman Hilberry said in his own words that he was handed a fireman's axe and told to cut the manila rope if the safety rods failed. What is folklore is the acronym, the idea that SCRAM stands for Safety Control Rod Axe Man. Hilberry said he only heard that story years afterwards, and other participants credit the word to the shutdown circuitry instead. So the man and the axe stay and the acronym goes.

The other is Oppenheimer. He did not, as far as any record shows, say the Gita line at Trinity. He said it to a camera in 1965, describing what he remembered thinking. Which is arguably a better twin for Sydney than the myth would be: the maker, afterwards, telling you what the thing was.

On method

The reference list and what each line is doing come from osmarks' annotation. The dates do not; they were assembled separately and then checked one at a time against the paper, the post, the filing or the report, and the ones that did not hold were fixed or dropped. The working table is in the repository that builds this site, so the audit is readable rather than asserted.

What the exercise shows is how young the vocabulary is and how fast it set. Thirty-three of the forty-three references carry a date; the other ten are jokes, or words like "guardrails" that nobody owns. Of the thirty-three, eighteen fall inside the last ten years. Meanwhile the right-hand column of the twins table runs from 1942 to 1954. A field assembled most of its working vocabulary in a decade, and almost every piece of it had already been argued once, by people who were further along than we are and knew less about how it would turn out.