The 2 day @StanfordHAI workshop on #foundationmodels is now available to watch for all.
I learned a ton from the final session on potential harms. @mmitchell_ai @sekreps @AngeleChristin @sameer_ + @KathleenACreel @james_y_zou @shelbygrossman.
I learned a ton from the final session on potential harms. @mmitchell_ai @sekreps @AngeleChristin @sameer_ + @KathleenACreel @james_y_zou @shelbygrossman.
These powerful language models force developers and ordinary users to confront a wide array of concerns: the use of unconsented data, the perpetuation of bias, stereotype, and discrimination, deliberate misuse via disinformation campaigns, and entrenchment of corp power.
The opening keynote by @mmitchell_ai set the stage for the conversation and raised issues about the very label "foundation models."
The presentation by @james_y_zou is a powerful demonstration of how existing language models stubbornly perpetuate anti-Muslim bias. Watch the short 10 minute presentation.
The presentation by @shelbygrossman at @stanfordio examined how LLMs are rocket fuel for disinformation campaigns. It's not a question of "if" there will be malicious users; it's a question of developing & deploying the models with this in mind.
And @StanfordEthics & @StanfordHAI postdoc @KathleenACreel pointed to the homogenizing effects of foundation models and the distinctive ethical questions that arise here. Foundation models are potentially a kind of linguistic monoculture.
Check out the paper just posted on foundation models for a longer discussion of some of these topics.
arxiv.org/abs/2108.07258…
arxiv.org/abs/2108.07258…
• • •
Missing some Tweet in this thread? You can try to
force a refresh
