When ChatGPT Broke an Entire Field: An Oral History

Summary

* * * I. The Wars of The Roses (2020–22) “understanding wars” • GPT-3 • “a field in crisis” As transformer models approached (and surpassed) “human baselines” on various NLP benchmarks, arguments were already brewing about how to interpret their capabilities. In 2020, those arguments — especially about “meaning” and “understanding” — came to a head in a paper imagining an LLM as an octopus. Emily M. Bender (professor, department of linguistics, University of Washington) EMILY M. BENDER (professor, department of linguistics, University of Washington; 2024 president, Association for Computational Linguistics): I was having these just unending arguments on Twitter, and grumpy about it. There was one about using BERT to unredact the Mueller report, which is a terrible idea. It seemed like there was just a never-ending supply of people who wanted to come at me and say, “No, no, no, LLMs really do understand.” It was the same argument over and over and over again. I was talking with [computational linguist] Alexander Koller, and he said: “Let’s just write the academic paper version of this so that it’s not just ideas on Twitter, but peer-reviewed research. And that’ll put an end to it.” It did not put an end to it. Bender and Koller’s “octopus test” asserted that models trained only to mimic the form of language through statistical patterns could never engage with its meaning — much as a “hyperintelligent octopus” would never really understand what life was like on land, even if it fluently reproduced the patterns it observed in human messages. SAM BOWMAN: This argument — that “there’s nothing to see here,” that neural network language models are fundamentally not the kind of thing that we should be interested in, that a lot of this is hype — that was quite divisive. JULIAN MICHAEL: I got involved in that. I wrote this takedown of the paper — it was the one blog post I’ve ever written, and it was the length of a paper itself. I worked hard to make it a good-faith represen...