Thread by @emilymbender

No, I will not make any such change, for two reasons:

https://twitter.com/yoavgo/status/1353413967889244161

1. Process: The camera ready is done, and approved by all of the authors. If I make any changes past this point it will be literally only fixing typos/citations. No changes to content let alone the title.

2. Content: I stand by the title and the question we are asking. The question is motivated because the field has been dominated by "bigger bigger bigger!" (yes in terms of both training data and model size), with most* of the discourse only fawning over the results. >>

*(Exceptions to this: (1) the big body of work--including yours--into whether the models absorb bias and (2) the GPT-2 staged roll-out paper (and references cited in its sec 1.)

arxiv.org/abs/1908.09203

Thus, the motivation for writing this paper. We aren't saying "LLMs are bad" but rather: these are the dangers we see, that should be accounted for in risk/benefit analyses and, if research proceeds in this direction, mitigated. >>

Furthermore, I've now had a minute to read your critique, and I disagree with your claim that our criticisms are independent of model size. Difficulty in curating and documenting datasets absolutely scales with dataset size, as we clearly lay out in the paper: >>

Likewise, nowhere do we say that small LMs are necessarily good/risk-free. There, and in your points about smaller models possibly being less energy efficient, you seem to have bought into a world view where language modeling must necessarily exist and continue. >>

This isn't a presupposition we share. (And please don't misread: I'm not saying it should all stop this instant, but rather that research in this area should include cost/benefit analyses.) >>

As for the claim that our paper is one-sided, this is exhausting. All of ML gets to write papers that talk up the benefits of the tech without mentioning any risks (at least until 2020 w/broader impact statements), but when a paper focuses on the risks, it's "one-sided"? >>

Furthermore, the "debate" you would like us to acknowledge is based on a false premise. As we lay out in detail in Sec 4, the training data emphatically do NOT represent "the world as it is". >>

And lastly, miss me with the claim that our work is "political" and therefore has a responsibility to "present the alternative views".

See also:

https://twitter.com/emilymbender/status/1349844030809542656

>>

We draw on scholarship from a range of fields that looks at understanding how systems of power and oppression work in society. >>

The claim that this kind of scholarship is "political" and "non-scientific" is precisely the kind of gate-keeping move set up to maintain "science" as the domain of people of privilege only. /fin