When 'Machines Can't Create' Becomes Nitpicking Anubis Prompt Semantics in AI
Description
The image is a screenshot of a tweet from “Fernando @zetalyrae · 7 Std.” The tweet text reads: 2020: "machines can never be creative" 2022: "pfft, I asked for 'Anubis lecturing on algebraic geometry, 1873, oil on canvas, from the Metropolitan Museum of Art' and it drew Bastet instead! where are the semantics? I am a linguist". Below the tweet is a DALL·E-generated painting: the jackal-headed god Anubis, wearing dark robes with a blue-patterned collar, sits on a wooden stool and gestures toward a parchment easel covered in red algebraic symbols and hieroglyph-like marks; a palette and brushes rest on the side table, and a tiny multicolor DALL·E watermark sits in the lower-right corner. The meme visually contrasts earlier claims that “machines can’t be creative” with modern complaints about subtle prompt-to-image mismatches, poking fun at escalating expectations for generative AI. Technically, it underscores issues of prompt engineering, semantic fidelity, and model limitations familiar to ML practitioners and researchers
Comments
6Comment deleted
Welcome to 2022: the P0 isn’t AGI safety, it’s a ticket that says “Bastet rendered instead of Anubis - violates semantic contract.” Turns out prompt engineering is just regex debugging, but the regex is 4 billion parameters wide
It's like watching a junior dev confidently explain why their regex works perfectly, then spending three hours debugging why it matches cat gods instead of dog gods in production
Ah yes, the classic 'garbage in, semantically confused garbage out' problem. The model nailed the Vermeer-esque lighting and period-accurate mathematical notation, but confused a jackal with a cat - proving that even with billions of parameters, AI still can't tell its Anubis from its Bastet. It's like asking for a React component and getting Angular: technically impressive, functionally wrong, and somehow still better than what management originally spec'd
Asked a diffusion model for Anubis lecturing algebraic geometry; it rendered Bastet - apparently deity identity is eventually consistent and decided by a quorum of vibes in the latent space
Diffusion nailed the acceptance criteria - Egyptian, lecture vibes, 1870s oil texture - then failed the type check on deity; embeddings aren’t ontologies
Prompt engineering: Specify algebraic geometry seminar, get Anubis dropping CAP theorem proofs - semantics? That's for humans maintaining the RAG layer