Re:We read the paper that forced Timnit Gebru out of Google
by visarga on Saturday December 05, 2020 @05:53AM
(#25313519)
You train your LM on web crawl data, but also train a 4chan classifier, then you condition your LM not to generate in 4chan style. GPT-3 got a similar chaperone classifier for offensive speech. It's like knowing swear words but choosing not to use them. You could also condition a general LM to bias and debias its outputs as you like.
While I don't really understand the emissions argument, the rest strikes me as very defensible. If the best language models need giant datasets to excel, it is very difficult to ensure your AI is trained on reasonable data.
I wouldn't want an AI therapist to be trained on 4chan. Now obviously nobody would be that stupid, but unfortunately it seems we are not far from it.
Unless we build models that rely on less data, it will be difficult to prevent problematic biases in the AI we put into production.
If that is the argument Timnit makes, I think that's the exact type of work I would expect from an AI ethics department. And good work at that.