Prof. Anima Anandkumar

.@OpenAI controversy, no one (including me) talked about #ethics of using @reddit for training. It is toxic to #women #minorities. I was harassed and threatened on @mxlearn when I spoke up about #NeurIPS @Miles_Brundage @jackclarkSF can u comment on this choice? @samcharringtonhttps://twitter.com/__yoson__/status/1101593216992849920 …

we didn't use reddit data, we used outbound links from reddit ranked at two stars or higher by users. Part of the reason for thinking about different release of this model is precisely because of issues like this (which occur in larger models also). We'll be sharing more here!

Thanks. I still find it problematic. @reddit threads have problematic links to all sorts of sites. Right thing to do would have been to choose appropriate data source to begin with. Certainly not have a big press coverage with highly biased model #responsibleAI

I think the q of appropriate data sources for large-scale, generally released models is a really great topic for discussion. It seems like there's some value to the inherent heterogeneity of sources that get linked to from various social forums. What might be a good source?

I agree, there isn't a great answer to that, given how biased internet is. Techniques like debiasing word embeddings can help remove bias to some extent. When @OpenAI model was covered by journalists, bias was never highlighted. https://arxiv.org/abs/1607.06520