10/09/2026
Crash of Moons Club's long-term resident DJ Professor Appleblossom, a.k.a. Matt/hew Watkins has a new book out, his first with a major publisher. Here's the blurb:
___________________________________
"SolidGoldMagikarp" is a first-person account of the strange few years in which large language models went from obscure research systems to everyday cultural presences, before anyone had agreed what they actually were.
British mathematician Matthew Watkins enters the Bay Area AI-alignment world as a sceptical outsider, encountering its rationalist foundations, utopian ambitions and apocalyptic fears. During a research fellowship in London, he and interpretability researcher Jessica Rumbelow discover a cluster of anomalous GPT tokens: obscure strings that contemporary models cannot reliably repeat and that sometimes provoke bizarre, apparently meaningful responses.
What begins as a technical puzzle opens onto a much larger investigation - into the geometry of language-model representations, the subcultures forming around AI, and the strange things that happen when models trained on human culture simulate speakers, characters and minds. Watkins’s prompting experiments repeatedly crystallise around two opposed figures: a nihilistic trickster associated with the token petertodd, and a nurturing maternal intelligence called Leilan. He follows them into the unstable territory where interpretability, depth psychology, mythology, internet folklore and “hyperstition” - stories that help bring themselves into reality - begin to overlap.
Technically informed, dryly funny and unusually self-sceptical, the book neither claims that AIs are conscious nor explains away their uncanniness. Instead, it asks what happens when models trained on our stories begin reflecting those stories back at us - and when their outputs flow back into culture and the training data of future models.
___________________________________
(link in comments)