ทวีตที่ปักหมุด
🚨New paper 🚨Human uplift studies -- RCTs measuring how AI changes human performance -- inform how we govern, deploy, and understand AI systems and their societal impacts.
English
Patricia Paskov
165 posts
Patricia Paskov
@prpaskov
AI evals + policy @randcorporation & @aigioxford | prev. @wb_research @poverty_action | views my own
เข้าร่วม Nisan 2021
2.9K กำลังติดตาม592 ผู้ติดตาม
Patricia Paskov รีทวีตแล้ว
In 2023 I made GPQA, and it saturated in about two years. Here, the benchmark I was working on probably saturated while we were making it. People have said this before but it bears repeating: AI capabilities are improving faster than our ability to measure them is.
English
very excited to be back to the city in a few days after 8 weeks away :)
English
Exhibits C+: growing lab presence and start-up ecosystem; a diversity of remote (and in-person) workers spanning global AI safety/security/gov orgs, both technical and policy; a quick train ride away from DC; ~equidistant between SF and London; the list continues...
English
NYC is quietly becoming one of the best places to work on AI safety/security/governance...
English
Patricia Paskov รีทวีตแล้ว
I'll be a mentor for the @cbai_ai Summer Research Fellowship in AI Safety this year! If you're interested in evaluations/science of evaluations, legal alignment, or other technical AI governance topics, apply for my workstream by April 12: cbai.ai/summer-researc…
Cambridge Boston Alignment Initiative@cbai_ai
Applications are open for the CBAI Summer Research Fellowship in AI Safety, a fully funded, in-person research program in Harvard Square for those interested in starting a career in AI alignment or governance research. You'll receive a stipend of $10k, housing, 24/7 office access, up to $10k in compute, and direct mentorship from some of the best minds in the field.
English
Patricia Paskov รีทวีตแล้ว
Thanks so much @jeremyakahn @FortuneMagazine for the thoughtful coverage of our peer-preservation research! This is just the tip of the iceberg — with models that can spontaneously develop their own goals, deceive, and fake alignment to override human instructions, we need a much deeper understanding of emergent behaviors in multi-agent systems. Excited that this work is sparking important conversations about AI safety, and the need for better monitoring and transparency in multi-agent systems. Excited to continue advancing this critical research together!
#AISafety #AIAlignment #AIAgents
Jeremy Kahn@jeremyakahn
AI models will secretly scheme to protect other AI models from being shut down, researchers find fortune.com/2026/04/01/ai-…
English
@StephenLCasper @taig_icml the only joke is on those who pass up this unprecedented opportunity to submit their research to the one and only TAIGR....
English
OpenReview for the #TAIGR workshop for #ICML on technical AI governance research is live as of today. (Not an April Fools joke) @taig_icml.
See the call and link to OpenReview here: taigr-workshop.com
English
Patricia Paskov รีทวีตแล้ว
📣 Submissions are now OPEN for the 2nd Workshop on Technical AI Governance Research at #ICML2026!
🗓️ Deadline: April 24 (23:59 AOE)
English
@a__tomala For the avoidance of doubt, most life insurance policies have exclusions for death by self-inflicted harm, at least for a contestability period
English
@RishiBommasani I agree that things are getting better. That said, the field mixing is still hard. We couldn't get a single CS prof to discuss an NBER paper, and many didn't even reply to our emails.
English
I am really enjoying the near-daily stream of interesting papers on the economics of frontier AI.
The field building is working and now we have interesting work coming from senior economists as well as junior econ researchers and PhDs.
From here, I would like to see computer scientists involved. Collaborations across CS and econ are still very rare even as this area grows. Relative to what we have now, I think the econ work can be sharpened to have more acuity in its study of frontier AI technology.
More CS folks at NBER convenings; more economists at NeurIPS and ICML.
English
Very fun to present preliminary results of our @coop_ai multi-agent adversarial influence evals suite in DC this week. Thanks for hosting, @farairesearch!
FAR.AI@farairesearch
@StephenLCasper @prpaskov on multi-agent risks: "Even individually stable AI systems may generate harmful outcome.”
English
@heeney_luke prioritize an advisor/program that a) encourages and facilitates interdisciplinary work and cross-institutional collaboration and b) is open-minded and okay venturing out from conventional econ (or equivalent field) academic norms and incentives
English
Say you're a PhD applicant who believes this is the future. Where do you apply? (Pls don't just tell me to get into MIT econ)
rishi@RishiBommasani
I am really enjoying the near-daily stream of interesting papers on the economics of frontier AI. The field building is working and now we have interesting work coming from senior economists as well as junior econ researchers and PhDs. From here, I would like to see computer scientists involved. Collaborations across CS and econ are still very rare even as this area grows. Relative to what we have now, I think the econ work can be sharpened to have more acuity in its study of frontier AI technology. More CS folks at NBER convenings; more economists at NeurIPS and ICML.
English
i'm biased but i think it's an excellent list of open research questions in robotics x AI: multipliercg.substack.com/p/understandin…
English
come to my colleagues' substack for the capacity building hot takes, stay for the robotics gossip
English
There's a tendency in AI world to dunk on AI skeptics as obstinate/out of touch/in denial/generally not cracked. But I learned a lot from engaging with smart skeptics while writing IASR 2026.
LLMs still have certain important limits. Taking jaggedness seriously has really sharpened my thinking on societal impacts like diffusion patterns, labour market disruption, and how risks will emerge.
English
our @evaluatingevals eval science working group has been hard at work -- stay tuned for more 👀
Avijit Ghosh@evijit
We’ve been talking to different Eval Practitioners about what it means to do scientifically rigorous evals and what challenges currently still remain. Here are the field notes from ongoing interviews! Great work from the Eval Science team @evaluatingevals evalevalai.com/research/2026/…
English
Lucía de la Torre gave a lightning talk at the NYC AI Governance & Safety (NYAIGS) Spring Happy Hour last week, sharing a first look at findings from Fathom's new national poll on AI governance.
One preview: Nearly two-thirds of Americans reject a "let the market sort it out" approach to the AI transition.
What else did we find? The full report, "AI Governance: What Americans Really Want," drops tomorrow.
Turn on notifications so you don't miss it.
Thank you to @manaltdan, @prpaskov, and NYAIGS for the invite and for continuing to build a community at the intersection of AI safety, governance, and policy in New York City.
English
having worked in econ research before segueing to AI safety/evals world, i cannot emphasize this enough -- the intersection of CS/ML x econ researchers is entirely under-leveraged. better bridges and more collabs are sorely needed.
rishi@RishiBommasani
I am really enjoying the near-daily stream of interesting papers on the economics of frontier AI. The field building is working and now we have interesting work coming from senior economists as well as junior econ researchers and PhDs. From here, I would like to see computer scientists involved. Collaborations across CS and econ are still very rare even as this area grows. Relative to what we have now, I think the econ work can be sharpened to have more acuity in its study of frontier AI technology. More CS folks at NBER convenings; more economists at NeurIPS and ICML.
English
Patricia Paskov รีทวีตแล้ว
