And GPT5 is worse than GPT4o in writing …

Kudos to our co-organizers Chirag Agarwal, Hima Lakkaraju, Sarah Tan, @junweideng.bsky.social, Pingbang Hu, Eileanor LaRocco, Karolina Naranjo, @shichangzh.bsky.social

Featuring an amazing line-up of speakers: @yoshuabengio.bsky.social, @yaringal.bsky.social, Dan Hendrycks, @sarahooker.bsky.social, Gary Howarth, Bo Li, Lucilla Sioli!

Announcing the 3rd Workshop on Regulatable ML @neuripsconf.bsky.social

Please consider submitting your work at the intersection between AI safety and regulatory frameworks to our workshop!

Submission Deadline: Aug 22

regulatableml.github.io

#NeurIPS2025 #CallForPapers #AISafety #AIRegulation

Verifying AirPods Pro offer ... for making slides of an academic paper??

My thoughts on the crucial importance of methodology on self-reported AI performance on mathematics competitions, and my policy on commenting on such reports going forward: mathstodon.xyz/@tao/1148814...

DCA-Bench: A Benchmark for Dataset Curation Agents: arxiv.org/pdf/2406.07275

A benchmark for data curation tasks. Key challenge is that it requires AI agents to autonomously **discover** data quality issues.

A Snapshot of Influence: A Local Data Attribution Framework for Online Reinforcement Learning: arxiv.org/pdf/2505.19281

The first data attribution method for online RL that improves sample and computational efficiency on RL applications including RLHF.

A Versatile Influence Function for Data Attribution with Non-Decomposable Loss: arxiv.org/pdf/2412.01335

We developed an efficient and easy-to-implement generalization of influence function applicable to machine learning objectives that are not decomposable to loss terms on individual data points.

Attending #ICML2025 from Thu to Sat! Will be presenting our work in tomorrow’s Poster Session 5 East and the DataWorld workshop on Sat.

More details below. Looking forward to meeting old and new friends!

We have a postdoc opening through the Postdoctoral Research Associate Program in our school. The application deadline is April 1, 2025. Please see more details here: jiaqima.github.io/postdoc/

Good morning from #NeurIPS2024 in Vancouver where I’m representing the US AI Safety Institute at this great workshop, with @yoshuabengio.bsky.social @ruchowdh.bsky.social
and others.

regulatableml.github.io

Next up, @yoshuabengio.bsky.social on Why and How to Regulate Frontier AI.

The international state of the science of AI report he is leading is an essential part of the global landscape.

More details about our poster sessions:

Papers:

1. dattri: A Library for Efficient Data Attribution (arxiv.org/pdf/2410.04555)
2. Most Influential Subset Selection: Challenges, Promises, and Beyond (arxiv.org/pdf/2409.18153)

Workshop: bsky.app/profile/jiaq...

Excited to be at #NeurIPS2024 through Sunday!

My collaborators and I will be presenting two papers on data attribution and hosting a workshop on AI regulation. Looking forward to connecting and having great discussions!

#RegulatableML workshop is the result of a lot of hard work from our amazing team: @jiaqima.bsky.social, Chirag Agarwal Sarah Tan, Hima Lakkaraju, Usha Bhalla, @zbucinca.bsky.social, @junweideng.bsky.social, Alex Oesterling, and Shichang (Ray) Zhang.

5/5

A Path for Science- and Evidence-based AI Policy

⭐ Get ready for an interactive session! Dawn Song and Rishi Bommasani will discuss their exciting "A Path for Science‑ and Evidence‑based AI Policy" paper. We encourage everyone to read the paper in advance to join the conversation!
Paper: understanding-ai-safety.org #AIPolicy
4/

⭐ 10 oral presentations

⭐ A poster session of 34 accepted papers

⭐ New this year! Invited speakers will hold office hours for in-person meetings in a small, intimate setting. Check the schedule for times and take advantage of this exclusive opportunity to connect with our speakers!
3/

⭐ Six keynote talks: Rumman Chowdhury, @yoshuabengio.bsky.social, @peterhenderson.bsky.social , Reva Schwartz, Dawn Song, and Rishi Bommasani.

⭐ A panel discussion on AI Regulation/Governance: Gillian K. Hadfield, Wan Sie Lee, Surdas Mohit, Sella Nevo, @robreich.bsky.social, and Savannah Thais

2/

Join us at the #RegulatableML workshop at #NeurIPS2024 to learn about AI regulations and how to operationalize them in practice.

Date: Dec 15, 2024 (East Meeting Room 13)
Time: 8:15 am - 5:30 pm
Details: regulatableml.github.io

We have an exciting schedule (a thread):
1/