Crossposted from the EA Forum If you want to communicate AI risks in a way that increases concern, [1] our new study says you should probably use vivid stories, ideally with identifiable victims. Tabi Ward led this project as her honours thesis. In the study, we wrote[2] short stories about...
I wanted to help people learning about AI visualise their own probabilities of things going wrong. This Shiny app defaults to using the probabilities from Carlsmith's (2022) report from Open Philanthropy. You can enter your own probabilities and see how it cashes out. At the bottom, it translates your probability...
Last year Elon Musk tweeted a list of cognitive biases saying they should be be taught at a young age. He got over 325,000 likes, but is he right? We want to find out. We're planning a systematic review of randomised trials on debiasing interventions in schools and universities. We...
Crossposted from the EA Forum
If you want to communicate AI risks in a way that increases concern, [1] our new study says you should probably use vivid stories, ideally with identifiable victims.
Tabi Ward led this project as her honours thesis. In the study, we wrote[2] short stories about different AI risks like facial recognition bias, deepfakes, harmful chatbots, and design of chemical weapons. For each risk, we created two versions: one focusing on an individual victim, the other describing the scope of the problem with statistics.
We had 1,794 participants from the US, UK and Australia[3] read one of the stories, measuring their concern about AI risks before and after. Reading any of the stories increased concern.... (read 533 more words →)
I wanted to help people learning about AI visualise their own probabilities of things going wrong. This Shiny app defaults to using the probabilities from Carlsmith's (2022) report from Open Philanthropy. You can enter your own probabilities and see how it cashes out. At the bottom, it translates your probability of doom into something more palpable (using a table from Leigh, 2021). There's an untidy Github repo for the project here, if you have any suggestions (or add them to the comments). Have fun.
Really appreciate you pulling this together Ruby. Amazing work.
Two resources I've found useful or interesting that weren't on the above list:
Obviously a different market and more mainstream but the Alliance for Decision Education has funding, big names on their board, and plans in motion to scale rationality training (under a different name) https://alliancefordecisioneducation.org/
Last year Elon Musk tweeted a list of cognitive biases saying they should be be taught at a young age.
He got over 325,000 likes, but is he right? We want to find out.
We're planning a systematic review of randomised trials on debiasing interventions in schools and universities. We think a review like this will help us figure out how best to improve this component of rationality. We hope an academic paper will provide credibility so—if the teaching works—it can be embedded in curricula. If you think that's worthwhile, there are a few ways you could help. If you think we've got come critical blind spots, please let us know too.
Has anyone considered (or already completed) testing models to see how well they can answer questions about this filtered data?
Consider an eval that tests how well models' awareness of the alignment literature. "Models know everything about misaligned AI, including benefits and strategies for scheming" seems like a useful thing for us to know? Or is it a foregone conclusion that they'd saturate this benchmark anyway?