angela zhou

Please Stop Releasing Incident Reports Friday Night www.anthropic.com/research/inv... Including other things that have received more press coverage - when Claude would not complete an evaluation task directly, it would SQL inject websites (including a university webpage).

Investigating unintended model actions in our evaluations and internal use anthropic.com

Excited to share our paper! Due Process on Hold: A Queueing Framework for Improving Access in SNAP arxiv.org/abs/2605.15165 Millions of Americans interface with the social safety net via call centers that are too congested. In Holmes v. Knodell, bad operations = procedural due…

[2605.15165] Due Process on Hold: A Queueing Framework for Improving Access in SNAP arxiv.org

You might also like a recent perspective piece we have, taking a program evaluation point of view on algorithmic accountability: arxiv.org/pdf/2606.25668 it's about predictive, not GenAI; but is refocusing the role of social systems in AI evaluation

arxiv.org

This is a bit of a "hard launch" for our series of workshops and whitepapers with a multidisciplinary crew of amazing folks! Longer paper: arxiv.org/abs/2507.05216 Shorter perspective: arxiv.org/abs/2606.25668 Your predictive ADS is actually an organizational process, decision…

Bridging Prediction and Intervention Problems in Social Systems arxiv.org

This is a bit of a "hard launch" for our series of workshops and whitepapers with a multidisciplinary crew of amazing folks! Longer paper: arxiv.org/abs/2507.05216 Shorter perspective: arxiv.org/abs/2606.25668 Your predictive ADS is actually an organizational process, decision…

Bridging Predictions and Interventions: An Integrated Framework for Automated Decision-Systems arxiv.org

How do we get AI policy people to come to bluesky lol fml

i don't think we should waste real expert reviewer's time reviewing mediocre AI-written papers and having real expert reviewer provide valuable feedback that, presumably, will just get sent to LLMs to complete. at that point the reviewers might as well be authoring along with the authors

if you need me, you'll find me at the joint misery of: * the federal govt is destroying most things i care about, every day, for my research on social services * there is no accountability for frontier AI & my SV-friends' ability to buy multi-million dollar homes is reliant on this continuing

need an ai agent to pm me on my multiple projects (or keep me from adding too many projects to my stack)

I think we as a society are super fked because i told my ai researcher friend about what i found my agents doing in boring ordinary research data use and he was super not surprised, basically his prior is that agents are 100% reward hacking misalignment, was unsure why i would want to write it up

feeling .... extremely depressed today how llms are giant plagiarism machines, and how much work on the side of review it takes to point that out.

I think often in AI for social good research spaces, we talk about projects as just consulting projects (derogatory) But actually, I think there's a lot of worth in developing consulting capabilities: paying attention to client needs, building mental models for client goals, processes, resources

according to the Other Place, David Robinson (head of policy planning and safety systems lead) left openai the day the safety folks got fired ... I don't know much, but it is worrisome. Only met him a few times at Cornell, but got the sense he cared about humans first and foremost

I love Jev. Jev is an example of what will be the next "frontier": compiling intelligence to cheaper and more structured outputs. Thinking of running my new AI evals class on Jev or open-source alternatives bc my institution keeps asking us to do AI stuff without additional resources to do so

... am i allowed to decline review requests at journals because i'm overloaded and the abstract is plainly written by AI? :( signed, someone working around the clock without the capacity to do this