Stochastic Parrots: can a language model be too big
In March 2021 the proceedings of ACM FAccT '21 carried a paper by Emily Bender, Timnit Gebru, Angelina McMillan-Major and a fourth author under a pseudonym. Its fourteen pages ask how big is too big and name four directions of risk: the environmental and financial cost of training, training data too large to be understood, the imputing to a model of an understanding it does not have, and harm from the text it produces.
Why it matters
This is the first text in the atlas that asks not what a model can do but what its size costs, and asks it before large language models reached consumers. It is the source of the term the field uses for a system that assembles sequences of forms without reference to meaning.
The definition from section 6: a language model is a system for haphazardly stitching together sequences of linguistic forms it has observed in its vast training data, according to probabilistic information about how they combine, but without any reference to meaning: a stochastic parrot. The paper's sections: environmental and financial cost; unfathomable training data; stochastic parrots. The recommendations in the abstract: weigh the environmental and financial costs first; invest resources into curating and carefully documenting datasets rather than ingesting everything on the web; carry out pre-development exercises evaluating how the planned approach fits research and development goals and supports stakeholder values; and encourage research directions beyond ever larger language models. The notion of documentation debt comes from the same place. The bylines read: Emily Bender and Angelina McMillan-Major, University of Washington; Timnit Gebru, Black in AI; and Shmargaret Shmitchell, affiliated to The Aether, that is, a pseudonym. The record does not name who stands behind it: the paper does not say. A footnote on the first page marks joint first authorship. Pages 610-623, DOI 10.1145/3442188.3445922, under a Creative Commons Attribution 4.0 licence.