AI Colloquium proposal: three years of Stochastic Parrots - University of Groningen

Title: Are GPT-4 & friends too big? Reflecting on 3 years of “stochastic parrots”

Abstract:

In this panel discussion event - open and accessible for everyone interested in AI & Society, no technical background needed - we will discuss the legacy of a 2021 paper which was both influential and controversial (resulting in one of

the authors getting fired from Google!). The term “stochastic parrot” has since become synonymous with the point of view that large language models (including the current generation of ChatGPT/GPT-4, Bard, Claude, LLAMA 2, and others) are inherently problematic: do they really “learn something”, or do they seem to do so because of the complexity of the statistical patterns they can recognize and reproduce?

In addition, these models are trained with large amounts of data, and a lot of computational power is needed both for training and running these models: this causes unsolvable problems such as environmental pollution (from energy use) and social biases acquired from sets of training data that are too large to check and curate the contents of. The authors of the “stochastic parrots” paper have since become activists arguing that for these reasons, the use of large language models should be avoided or at least seriously limited. This stands in contrast to the view that, while “stochastic parrots” raised important challenges, the benefits of large language models still outweigh the downsides, and that the paper and its authors both underestimate the impressiveness of current models and overestimate their risks.

Now, several years after the publication of “stochastic parrots”, we take a look at how both large language models themselves and the debates around them have evolved in the last few years, in which AI and Natural Language Processing have increasingly become hot topics, in science but also in politics and society. Our panelists will defend a range of views that are more or less in favour of “stochastic parrots” arguments, and with a background in different fields: AI, computational linguistics, and (digital) humanities.