Dhara Yu

@dharakyu

PhD student @UCBerkeley interested in cognition and computation, prev BS+MS @Stanford

dharakyu.com Joined August 2019

143 Following 238 Followers 90 Posts

Latest posts 5

Pinned post

Dhara Yu @dharakyu Apr 29

New blog post: There's lots of excitement about using AI to simulate human behavior.

I wrote about closing the gap between the "verifiable" paradigms used to evaluate model fidelity and the open-ended behaviors we actually want to simulate - and how AI itself can help with that

Dhara Yu @dharakyu Oct 26

New blog post: there has been a lot of movement toward embedding third-party evaluation organizations within frontier AI companies to measure capabilities/risks. How could that actually work? I went to this conference (hosted by @evaluatingevals) learn more

Dhara Yu @dharakyu Oct 20

Cool work from @Karim_abdelll - in additional to identifying a new risk vector for misalignment, this setup is an interesting testbed to study how model conceptualizations (safety related or otherwise) might evolve as they are trained on AI outputs

Karim Abdel Sadek Verified account @Karim_abdelll Oct 20

New surprising result:

We trained AI recursively on its own safe generations, while giving it a benign cue. Models became increasingly misaligned over time, giving advice on how to do harm and build biological weapons.

We call this Progressive Misalignment. 🧵👇

View quoted post on X

Dhara Yu @dharakyu Oct 15

New essay!

I argue that if we want to use AI to do better computational cognitive modeling, we have to acknowledge that privileging easily-digestible signals of model **falsifiability** above all else can create perverse incentives.

Falsifiability can be reward-hacked!

Dhara Yu @dharakyu Jul 24

New blog post: how do we rigorously study AI's effects on cognition and behavior? There's an inviting parallel to past research on social media impact on wellbeing - can we leverage anything from that? (spoiler alert: mostly no)