Muro "Hamas forced him to watch the uncensored October 7 documentary released by the IDF... that portrays the evidence of the gruesome attack carried out by Hamas terrorists. Many of the scenes take place in Nahal Oz where Eitan is from." jpost.com/israel-news/ar…

Another nail in the coffin of UN credibility. After weeks of silence on Hamas’s October 7 crimes against women and children, United Nations Women issued a definitive statement on Friday that condemned “the brutal attacks by Hamas.” Then, the organization deleted its statement.

Excited to share our latest paper accepted to EMNLP 2023! 🧵 We tackle the challenges of understanding embedding spaces in large language models and introduce CES – Conceptualizing Embedding Spaces.

Attending #WACV2024? Make sure you visit our poster on debiasing, moderating and erasing concepts in text2img models! @ 8PM today. Rohit Gandikota will be there.

If you're looking for our recent paper on model editing on the ACL anthology and cannot find it, it's because it has been taken down without cause or due process. The paper is still available on arXiv, feel free to read it there. arxiv.org/abs/2310.11958 aclanthology.org/2023.findings-…

The final number appears to be *ten* papers taken down from EMNLP proceedings despite having gone through peer review, a month after the conference, without basis in any ACL guideline (I am now certain of this part as well).

Friends, Seinfeld, GoT...

Last night was the most surreal I have ever had: sitting in the living room, watching stand-up on YouTube with friends, all while waiting for drones from Iran.

Our new work is out! 👇🏻

This paper suggests that large language models mainly learn factual knowledge during pre-training, and introducing new knowledge in finetuning might increase hallucinations. But I was really intrigued to see that training on "maybe known" samples can enhance knowledge retrieval.

Our paper Diffusion Lens got accepted to #ACL2024 main conference! 🌴⭐️ Visualize LLMs computation process with our live demo >> huggingface.co/spaces/tokeron… For a quick TL;DR checkout Michael Toker's thread or project website - tokeron.github.io/DiffusionLensW…

In #NAACL? Come chat at the poster session tomorrow, 16:00.

תמונות בלי הטיות? חוקרים מהפקולטה למדעי המחשב ע"ש טאוב מציגים פתרון מחוללי תמונות הפכו ללהיט, אבל ידעתם שהם עלולים להיות מוטים? החוקרים פיתחו שיטות חדשניות לתיקון הטיות ועדכון ידע במודלים אלו. השיטות מאפשרות תיקון מהיר ויעיל של הטיות, עדכון ידע עובדתי, שליטה בהתנהגויות אתיות של

Check out this cool work by Z. Wang et al, which uses our method ReFACT to mitigate the effects of textual triggers in text-to-image models! 1/2

Interested in text-to-image models? Come say hi today at poster session 2 and hear about the diffusion lens - our new method of interpreting text encoders of t2i models ✨📸 #acl2024nlp #ACL2024 w/ Michael Toker Mor Ventura Hadas Orgad Yonatan Belinkov

Had a blast discussing how mechanistic interpretability can improve AI safety at #NEMI2024 🤖 Thanks to David Bau, Max Tegmark, Koyena Pal @ VLDB 2024, Kenneth Li, Eric J. Michaud & Jannik Brinkmann for bringing us together! And to my amazing co-panelists Hadas Orgad & David Krueger!

#NEMI2024 was really fun and informative. It was great having so many smart people passionate about mecinterp in one room. Thanks Koyena Pal @ VLDB 2024, David Bau, Max Tegmark, Kenneth Li, Eric J. Michaud& Jannik Brinkmann for organizing this!

Try out NNsight of you're interested in interpretability research of LLMs!

We will remember forever