The Blurry JPEG

“Think of ChatGPT as a blurry JPEG of all the text on the Web.” With that sentence, published in The New Yorker in February 2023, Ted Chiang did more to clarify public understanding of large language models than any technical paper, not because it was dismissive but because it was precise. A JPEG is a lossy compression format: it shrinks an image by discarding information the eye is unlikely to miss, keeping the gist while throwing away the exact bits. When you view a JPEG, you are not seeing the original; you are seeing a reconstruction, an approximation generated from the compressed data. Chiang’s claim is that an LLM does the same thing to the text it was trained on: it compresses the web, lossily, retaining statistical regularities while discarding the addresses of the sources. The analogy explains hallucination (gaps filled with plausible interpolations that correspond to nothing real), the limits of originality (the