Jess Sorrell

How often do LLMs claim to prove false mathematical statements? In our latest benchmark, BrokenArXiv, we find they do so very often. The best model, GPT-5.4, only rejects 40% of incorrect statements obtained by perturbing recent ArXiv papers, and other models do much worse.

1/ "Agents of Chaos" gave AI agents email, persistent memory, and shell access, then 20 researchers probed them. Within two weeks: sensitive data disclosed to strangers, unsafe practices spreading between agents, partial system takeover. buff.ly/DMboEG6

@_onionesque @sikatasengupta Your favorite size

@JessSorrell @sikatasengupta What is YFS?

New paper with @sikatasengupta and others showing you get stability (model agreement) for free from loss minimization over NNs of a fixed size* *offer not valid for all NN sizes, though if it doesn't apply to YFS, you would have done much better minimizing over 2*YFS anyway

Applications are now open for the AI and Society Fellowship at the Center for AI Safety: a fully funded, 3-month summer fellowship in San Francisco for scholars in economics, law, IR, and adjacent fields to research the societal impacts of advanced AI. Apply by March 24.

There's growing evidence that LLMs can p-hack. That should worry us. But p-hacking also points to something bigger: a data science multiverse of defensible analytical choices. LLMs make it cheap to search that space for favorable results. In our paper, joint work with excellent collaborators: Martin Bertran and Riccardo Fogliato, we build a pipeline that uses LLM agents to systematically map this multiverse. We believe our experiments could inform what future scientific transparency should look like. 🧵

"Our scientific institutions — peer review, publishing, graduate education — evolved over decades to match the constraints of human cognition and effort. Those constraints are shifting rapidly, and our institutions will need to shift with them."

@lfschiavo That's my secret, Cap

things are gonna get weird. you must get commensurately weird.

@Jess_Riedel @dan_mackinlay @danielmurfet @jankulveit @geoffreyirving @mhutter42 @vkrakovna Would love to review if helpful :)

Very excited to be joined by @dan_mackinlay , @danielmurfet , Edmund Lau, & @jankulveit in announcing a forthcoming research journal for AI alignment. @geoffreyirving , @mhutter42 , Scott Aaronson, & @vkrakovna have graciously agreed to form the initial advisory board

@thegautamkamath Yeah, I don't know the best way to manage the OR ICLR mess, but the current policy does at least seem consistent with ensuring there is little incentive for authors to harass reviewers, which I appreciate

The burden of being in power is that you will have to make decisions that many people don't like.

Join @let4all for a mentoring workshop on productive AI use for research and research-adjacent tasks!

Last call to register at mmlworkshop.ethz.ch for the Math for Trustworthy ML workshop in Switzerland with the possibility to present your own work as poster/talk, we still have a few slots left!

#HopkinsDSAI welcomes 22 new faculty members, who join more than 150 DSAI faculty members across @JohnsHopkins in advancing the study of data science, machine learning, and #AI and translation to a range of critical and emerging fields. ai.jhu.edu/news/data-scie…

📢 Call for AAAI-26 Doctoral Consortium Submissions (Deadline: September 30 AoE) We invite doctoral students in the AI community to apply to the Doctoral Consortium at AAAI-26, held on January 20-21 2026 in Singapore!

@gautamcgoel Morocco bsky.app/profile/emtiya…

Does anyone know where AISTATS 2026 will be held?

📢 Attention undergrads! 📢 Check out our 2nd CS Faculty Research Panel of the summer on Thursday, July 31, where faculty like @mdinitz, @KimiaGhobadi & @JessSorrell will help you learn how to apply your degree to the broader world of CS. Learn more at cs.jhu.edu/event/computer…

Do you care about multiaccuracy/multicalibration in ML? and are you concerned with accessing privileged information (like sensitive attributes) to evaluate or control for these? Then come by Beepul's poster today at #icml where he'll be presenting this work openreview.net/pdf?id=sGny74z… @JHUBME @HopkinsDSAI

Come say hi at our #ICML poster today during Poster Session 1 (W-600)! Joint work with @Aaroth @mkearnsupenn @JessSorrell @marcel_hussing and Eric Eaton! arxiv.org/abs/2502.11828

New podcast episode of "Probably Approximately Correct Learners," featuring guest Clément Canonne (@ccanonne_)! Check it out on Youtube, Spotify, Apple Podcasts, or wherever you get your podcasts (links in the next post) 1/2

Happening today at 4pm CET in Amphitheater T!