Fall 2024: Lisa Gitelman (New York University), “Typographical Hallucinations” скачать в хорошем качестве

Fall 2024: Lisa Gitelman (New York University), “Typographical Hallucinations”

Lisa Gitelman (New York University) originally delivered this talk at University of Pennsylvania's Workshop in the History of Material Texts on November 4, 2024. The architects of today’s image-generating artificial intelligence—such as OpenAI’s DALL-E 3—admit that these systems struggle with text. They hallucinate mangled typographic forms even as they generate plausible contexts for typography. What can mistakes like these tell us about AI, and is there anything that they can tell us about typography? Is there something that DALL-E 3 “knows” about typography, in short, that we don’t? This talk activates questions like these in part by turning to the history of the printing trades in the nineteenth century and to specimen books published by type founders. Using the language of today’s computational neural networks, we might say that specimen books stitch together sequences of “tokens” based upon all of the sequences previously observed amid oceans of training data: anything their printers had read or set. Lisa Gitelman teaches English and media studies at New York University, where she is among the founding members of NYU’s Digital Theory Lab, a group which began to study deep learning in 2018. She is the author of Paper Knowledge: Toward a Media History of Documents and Always, Already, New: Media, History, and the Data of Culture, both recently translated into Chinese.