Accepted to #ICML2024!! 🚀 Meet MAIA- a Multimodal Automated Interpretabilty Agent that helps users understand AI systems. Given a user query (eg "label a model’s feature"), MAIA designs experiments iteratively by forming and updating hypotheses based on experimental results.

I am delighted to officially announce the National Deep Inference Fabric project, #NDIF. ndif.us NDIF is an U.S. National Science Foundation-supported computational infrastructure project to help YOU advance the science of large-scale AI.

This is why I signed the letter at righttowarn.ai. AI companies must create protected avenues for raising concerns that balance their legitimate interest in maintaining confidential information with the broader public benefit. 7/8

Very excited about this new work from Yossi Gandelsman–a compelling demonstration of the usefulness of interpretability for model auditing. Automatic decomposition of polysemantic neurons (no need for expensive SAE training!) into associated text directions enables generation of

Looking forward to participating in this panel next Tuesday, organized by the wonderful Kartik Chandra (also on Mastodon and Bsky) & co.! 🧠💻More on coggraph, a new workshop at the interface between cognitive science and computer graphics: coggraph.github.io 🔗Zoom registration link:

Thrilled to announce the AI + Open Education Initiative at MIT Open Learning! 🔗: aiopeneducation.pubpub.org To kick things off, we’re announcing a call for rapid-response papers examining how AI can help sculpt the open access education ecosystem. Papers can describe projects

MAIA code is out! This codebase is a good starting point for working with Interpretability Agents. We're excited about the role AI agents can play in scaling interpretability and will be at ICML next week to chat more! ✈️

Hi ICML! We're presenting papers this week: - understanding how LMs learn new langs in-context arxiv.org/pdf/2401.12973 - agents for automated interpretability research: …imodal-interpretability.csail.mit.edu/maia/ - calibrating LMs by marginalizing over missing context: arxiv.org/abs/2311.08718

#ICML2024 come chat about scalable Interpretability at #2407 from 1:30-3pm! With Tamar Rott Shaham and Franklin Wang I might be wearing a bowtie in MAIA’s honor 🎀

One week left to submit to our #ECCV2024 European Conference on Computer Vision #ECCV2024 workshop on evaluating vision foundation models! 🔎🤖👁✨ Deadline: July 31 23:59 GMT Topics include: interpretability, visual reasoning & grounding, hallucination, visual abstraction, in-context learning, & others here:

Reminder to submit proposals for access to the #NDIF 405b pilot access program, by tomorrow! ndif.us/405b.html

As AI models become more powerful, auditing them for safety & biases is crucial — but also challenging & labor-intensive. Can we automate and scale this process? MIT CSAIL researchers introduce "MAIA," which iteratively designs experiments to explain AI systems' behavior:

Our Preparedness team evaluates frontier models’ abilities as software engineering agents, a prerequisite skill that could one day enable models to operate autonomously and self-improve. SWE-bench has become the community standard for evaluating models on software engineering,

So neat she is transparent, like a spirit.