1/ Can AI chat assistants improve conversations about divisive topics such as gun control? I was very fortunate to join a team of outstanding researchers who designed an experiment intended to answer this question:

This is today's #mustread and easily one of the top #ai papers of 2022 on first glance. I want to dig more into the methods on this, but my preliminary run through finds some fascinating patterns! @sheabrownethics , would you want to join F. LeRon Shults… https://t.co/fAA5q5ow0N

Great discussion of our project about using GPT-3 for social science research on the NYT Hard Fork podcast! Min 51. https://t.co/QhsKQUYLEN @joshua_gubler @EthanBusby @davidwingate @ChrisRytting @NancyFulda

It's so cool to see this work go from a random idea @ChrisRytting and team had to a significant and impactful paper. You all should read the paper and follow Chris.

LLMs may not 'understand' people, but they are incredibly good at approximating people (and things that people do). The thing I find consistently confusing is figuring out where approximation ends and understanding begins. I myself feel most of my insights are from approximation

@BrendanNyhan GPT-3 is remarkable! We have a paper (forthcoming soon) that shows some of the power of GPT-3 when applied to social science — here’s the link to the arxiv version of the paper:

Using LLMs for social (simulation) science https://t.co/ZLCtXlfj1r Once we figure out how to scale it, this approach could revolutionize agent-based simulation modeling for nuanced, multifaceted silicon samples cc @mjbommar,@computational,@jg_environ h/t @jackclarkSF Import AI

For Import AI, I wrote about a very special paper which I think has some significant implications. If we can use LLMs as proxies for people (for a certain level of detail and desired response accuracy), then I expect a bunch of strange things to happen. https://t.co/7EgTtF2vrm

@metaviv @abenomicon @joon_s_pk Thanks for the interest in our work! We're working on figuring out how prompting GPT-3 compares to priming experiments with humans, but we don't have a systematic answer yet. Stay tuned! @NancyFulda @joshua_gubler @EthanBusby @davidwingate @ChrisRytting

What would a social science Turing test look like? Amazing new research by @lpargyle et al. shows GPT-3 can create “silicon samples” that impersonate respondents to large surveys such as the American National Election with remarkable accuracy. https://t.co/8W3vqgjisi.

Prompts are the bread and butter of LLMs. ✨But can they be compressed?✨ In new work at Findings of EMNLP, we show that prompts can be compressed ⬇️ while maintaining the most crucial information! ℹ️

This work was done in collaboration with @davidwingate and @MohammadShoeybi near the end of my time @byu - reach out with questions!

@joshua_gubler @BrendanNyhan Exciting project, co-authored with @davidwingate @ChrisRytting @lpargyle @NancyFulda . Although we didn't ask about Sesame Street (we should have!)

Creative work! Theoretical CS has this idea of producing useful consequences of computational hardness (e.g., cryptography). This work has that flavor: if LLMs are biased, can we put that bias to good use? Probably many challenges ahead, but love the creativity.

In a new paper, we ask whether you can use GPT-3 to survey humans by simulating those humans and asking them questions, as opposed to interviewing the actual humans. https://t.co/yXPKI3OvxA w/ @davidwingate @EthanBusby @joshua_gubler @lpargyle @NancyFulda

Can we use GPT-3 to better understand human preferences? It was trained on the internet after all. Really cool work from @lpargyle, @davidwingate, @ChrisRytting, and team https://t.co/Ue3BuKTXJH

Excited and grateful that our research team just received an NSF EAGER grant to further research using large-scale language models to study human attitudes. A great team: @lpargyle @EthanBusby @davidwingate @NancyFulda @ChrisRytting @tsor13

We've started the @AllenGaret fan club! @studioc__tv #studioc

9 kids at vocal point! Better than TV. #byuvocalpoint

Yay vocal point! #byuvocalpoint