2/

From min. 16: #LLM training data "basically grabbing everything" on the internet.

"there's a lot of garbage there, right?
And there's no documentation. So if we don't know what's in the data, we are in no way positioned to actually mitigate the harms.
[...] the synthetic text that you can create using a language model recreates the #SystemsOfOppression that are in the training data and can also mislead people if it gets fluent enough.

We think that maybe there's a thinking mind behind it"

3/

#StochasticParrot [2]: make vivid
"difference between how we use language and what's coming out of what I now like to call #SyntheticTextExtrudingMachines.

So once that paper was out in the world [...] I was frequently asked «How do I know that you're not just a stocastic parrot?»

And I after being asked this question a couple times decided that I'm not going to be in conversation with people who won't posit my humanity as a basic axiom of the discussion.

This is a #dehumanizing question"

4/
#References

[2] Bender, E.M., Gebru, T., McMillan-Major, A., Shmitchell, S., 2021. On the dangers of stochastic parrots: can language models be too big? In: Proceedings of the 2021 ACM Conference on Fairness, Accountability, and Transparency. Presented at the FAccT ’21: 2021 ACM Conference on Fairness, Accountability, and Transparency, ACM, Virtual Event Canada, pp. 610–623. ISBN:978-1-4503-8309-7 https://doi.org/10.1145/3442188.3445922

On dehumanisation

- https://philarchive.org/archive/KROMDS
- https://www.google.it/books/edition/Habeas_Viscus/lePTBAAAQBAJ?hl=it&gbpv=1&dq=%22racialization%20is%20understood%20not%20as%20a%20biological%20or%20cultural%22&pg=PT13&printsec=frontcover