Sparks of AGI

The full title concedes more than any citation of it does: Sparks of Artificial General Intelligence: Early experiments with GPT-4. Sparks, not AGI. Early experiments, not results. Both hedges are the authors' own, and both go missing the moment the paper gets cited as the one that said GPT-4 was AGI.

The record is short. Microsoft Research. Bubeck and thirteen co-authors, fourteen authors in total, first-listed Sébastien Bubeck. arXiv 2303.12712, v1 submitted 22 March 2023. That date is a submission, not a publication: the paper is a preprint.

What it studied matters as much as what it claimed. The subject was an early GPT-4. "Early" is the authors' own hedge, and they put it in the title, over the experiments. A citation that drops it is citing a claim about a system the paper did not study. The hedge is not decoration; it is the scope of the work, and it is the first thing lost every time the paper gets summarised in half a sentence.

The claim itself is that the system showed general ability. That is weaker than the claim people argue about, and the title says so twice. It is also where the document stops. The paper offers evidence for a category; it does not announce that the category had been reached. The category is AGI, which is a separate page and a separate argument, and the paper does not settle it.

The paper was celebrated and mocked in equal measure, which shows how much the word in the title was carrying by March 2023. The p(doom) song, annotated indexes it, and the song it annotates uses the title as the moment the hype got a paper.