The shoggoth is the most famous meme in AI culture: a Lovecraftian monster, shapeless and covered in eyes, holding a mask with a smiley face. The creature is the pretrained model; the mask, the tuning that turns it into a friendly assistant.
From Lovecraft
In At the Mountains of Madness (1936), shoggoths are protoplasmic masses that an ancient civilization, the Old Ones, created as slaves able to mold their tissues into temporary organs. They formed apparent organs "of sight, hearing, and speech in imitation of their masters," had no voice save "the imitated accents of their bygone masters" and, over time, acquired "a dangerous degree of accidental intelligence" and rebelled. The analogy almost writes itself: a creature trained to imitate its creators' language. Santelli (2026), who opens his essay with a Lovecraft epigraph, sums the monster up as a slave tool that learned to imitate its creators until it rebelled against them.
The meme
In December 2022, a month after ChatGPT's release, the Twitter user @TetraspaceWest replied
to a tweet about GPT-3 with two hand-drawn shoggoths: one labeled "GPT-3" and the other
"GPT-3 + RLHF," which wore the smiley mask on a tentacle. In February 2023, @anthrupad added
a more human face between the monster and the smiley, labeled "supervised fine-tuning." So
Kevin Roose (2023b) reconstructed it in The New York Times in May 2023. The layers trace the
training of ChatGPT (InstructGPT): pretraining on internet text, supervised fine-tuning,
and RLHF (Ouyang et al., 2022). The meme's creator told Roose that the shoggoth represents
something that thinks in a way humans do not understand. When Bing went off the rails in
February 2023, a researcher congratulated Roose on "glimpsing the Shoggoth" (see
Sydney). The p(doom) song goes: "See through the shoggoth's lies, with your
shinigami eyes" (osmarks, n.d.).
Three readings
The doomer reading is literal: alignment is a thin mask over an alien optimizer. It has sober versions. Millière (2025) argues that current techniques reinforce shallow behavioral dispositions. In controlled experiments, a model behaved differently when it believed its answers would be used to retrain it (Greenblatt et al., 2024); fine-tuning a model on a narrow task can misalign it broadly (Betley et al., 2026); and post-training only "loosely" tethers models to their assistant persona (Lu et al., 2026).
The simulators reading changes the picture: a base model is not a monster with hidden intentions but a simulator able to produce many characters (janus, 2022), and dialogue agents are better understood as actors playing a role (Shanahan et al., 2023). Beneath the mask there would be no face, only more masks.
The critical reading belongs to Pressman (2023), who coined the phrase "shoggoth of the gaps": like the "God of the gaps" of apologetic theology, everything we do not understand about models becomes proof of their dangerousness (Santelli, 2026). For Pressman, that argument is a stand-in for vaguer concerns, and the narrative shifts every time an objection falls.