Year: 2024
Type: book-chapter
Abstract: In recent years, discourse about artificial intelligence (AI) has seen one claim come and go with near orbital periodicity.AI models, it runs, invent languages of their own.The claim usually pertains to a class of neural networks trained to predict sequences of text on the basis of gigantic volumes of data.Researchers and developers use these networks, or 'large language models' (LLMs), to perform tasks spanning text generation and summarization, machine translation, classification, and more.But training LLMs on heaps of data appears to have gone beyond producing general models of communication.Language modeling at the scale of millions of documents and billions, if not trillions of parameters now invents 'new' natural languages-ones native to AI models themselves.At least, this is the claim.An early version of it began circulating in 2016 after tech reporters picked up a blog post from Google AI (Schuster, Johnson, and Thorat 2016).The post summarized how the latest Google Translate model, newly rebuilt with a neural network, could render translations between two languages without explicit training.If the model was trained to translate English ↔ Korean and English ↔ Japanese, Google researchers showed that it could also generalize to Korean ↔ Japanese, even though its training did not include this third pairing.Researchers attributed this ability to an "interlingua" in their network; tech reporting was characteristically uncritical in its adoption of that term."Google's AI translation tool seems to have invented its own secret internal language," read one TechCrunch headline (Coldewey 2016).Wired quickly followed suit the next day, announcing, "Google's AI just created its own universal 'language' " (Burgess 2016).Exactly what comprised this interlingua-its vocabulary and syntax, for instance-was unclear.Researchers pointed to a scatter plot, its clusters shaded to mark semantic equivalence across languages; this, apparently, is what Google
Author: Shoemaker, Tyler
Language: German
Subfield: History and Philosophy of Science
Field: Arts and Humanities
Domain: Social Sciences
Open Access status: green
OpenAlex ID: W7025559696
