'Stochastic Parrots': Bender, Gebru weigh cost of LLM scale

The paper hosted on Emily M. Bender's SoundCloud page is the 2021 FAccT paper by Bender, Timnit Gebru, Angelina McMillan-Major and Shmargaret Shmitchell (a pseudonym for Margaret Mitchell), and the central question sits in the abstract: "How big is too big?"

The abstract sets the scene: "The past 3 years of work in NLP have been characterized by the development and deployment of ever larger language models, especially for English. BERT, its variants, GPT-2/3, and others, most recently Switch-C, have pushed the boundaries of the possible both through architectural innovations and through sheer size."

The paper's argument, as summarised on Wikipedia, is that these systems are "stitching together sequences of linguistic forms … observed in its vast training data, according to probabilistic information about how they combine, but without any reference to meaning". That is the parrot image the paper made famous. Around it sit the risks the authors argue scale with the models: environmental and financial cost of training, biases encoded in uncurated web data, and the ease with which fluent output can deceive. The recommendations, per the ACM record, push researchers to weigh those costs first, curate and document datasets rather than scrape the web, and run pre-development exercises before scaling up.