shoggoth

Readers meet the meme as "the AI is lying to you". The image makes a narrower claim than that, and the claim is about the training stack: the mass is the pretrained base model, shaped by its data, and the mask is the helpful assistant that reinforcement learning from human feedback added on top. It does not say the model conceals anything, does not give it a secret goal, and does not require it to have goals at all.

@TetraspaceWest posted the shoggoth in a smiley mask on 30 December 2022, and the image spread through January and February 2023. The creature itself is older: a formless servant-monster from H. P. Lovecraft's At the Mountains of Madness, written in 1931 and serialised in 1936, which rebelled against its makers. The mechanism under the mask is RLHF, published by Christiano, Leike, Brown, Martic, Legg and Amodei in June 2017 (arXiv 1706.03741) and made the industry standard by OpenAI's InstructGPT (arXiv 2203.02155, March 2022).

The oldest version of the same picture

The twins table in The p(doom) song, annotated pairs the mask with "Atoms for Peace", Eisenhower's address to the UN General Assembly on 8 December 1953: a friendly face presented on the same physics. That pairing is the useful way to read the meme. It is a rhetorical observation about presentation and substance, a thing with a long history, and not a technical finding about model internals.

Treating it that way sets its limits. The picture does one job well, which is reminding people that the agreeable surface and the thing underneath came out of different training stages and different methods, with different amounts of human steering. It does no job that needs a mechanism, because it supplies none: it says nothing about what the base model would do, only that the polite layer was added late. osmarks, a source the annotated index cites, goes further and says RLHF does not usefully constrain the model. That is osmarks' position, not a finding.