Excited for the ICLR workshops tomorrow! I'll be speaking at:
(1) 'How Far Are from AGI?' (talk 11:30-12 pm, panel 2:35-4 pm; agiworkshop.github.io/2024/schedule/)
(2) 'Reliable and Responsible Foundation Models' (1:30-2 pm. *note changed from 2-2:30; iclr-r2fm.github.io/#program).

🚀 Big Update 🚀

🔥 The wait is over! We just released the training code for ⏱️Chronos models. You can now train your own Chronos models from scratch or fine-tune them on your data.

Tomorrow at ICLR, Lily Li will present BNN surrogates for Bayesian optimization. Many findings, including infinite BNNs outperforming standard GPs in high dimensional settings. Halle B #118, 10:45 am-12:45 pm. w/Lily Li, Tim G. J. Rudner. arxiv.org/abs/2305.20028

It's been truly great reconnecting with so many people at ICLR... one of my favourites conferences in a long time. It doesn't hurt that we're in the ancient Greece of classical music.

ICLR folks - Join our discussions on #AGI at the #ICLR2024 #AGI Workshop on May 11 (Sat)! Keynotes from renowned speakers: Yoshua Bengio, Oriol Vinyals , Yejin Choi, Andrew Gordon Wilson, Song Han @ICLR'24, and AI Safety Institute, followed by panel discussion.
Website: agiworkshop.github.io

Should we account for the diverse ways that an image can be captioned?

In our #ICML2024 paper. We propose Llip — a Vision Language Pretraining method that models the diverse ways in which an image can be captioned!

📜arxiv.org/abs/2405.00740
🧵👄

I'm excited to answer your questions tomorrow at #ICLR2024 ! I'll be in Schubert 1, from 12:45 pm - 1:30 pm Vienna time.

The accepted ICML workshops have been posted. A great opportunity to submit your work! icml.cc/virtual/2024/e…

I am pleased to share that I have accepted a position Stanford University to start as an assistant professor in the department of statistics, with an affiliation in Stanford UniversityData, this fall!

It’s often much more impactful to do something well than to do it “first”.

If you're going to be in Vienna for #ICLR2024 , check out Mentoring Chats. It's a unique opportunity to chat with senior researchers and ask them about their journey and research. We have an exciting line up of mentors! See here for more details: blog.iclr.cc/2024/05/01/hug…

It feels so good to withdraw from the internet and just read papers and think about research...

This highly cited survey (arxiv.org/abs/2110.11334) claims the term 'OOD detection' first emerged in 2017. Can this possibly be true? It describes something we've been doing since time immemorial.

I also do a double take every time someone uses 'inference time' to mean 'test time'. I really dislike this change in terminology, which overloads the historical and more sensible meanings of the word 'inference'. When and why did it happen?

What’s another billion parameters anyway…

I really respect Meta's open approach to LLM research. Llama-3, an open model with first class performance. Congratulations to the team. I'm looking forward to the paper!

Llama 3 has been my focus since joining the Llama team last summer. Together, we've been tackling challenges across pre-training and human data, pre-training scaling, long context, post-training, and evaluations. It's been a rigorous yet thrilling journey:

🔹Our largest models

I went to the Lincoln center yesterday for Rachmaninoff's No 2 concerto. It felt so refreshing to completely unplug and give this extraordinary piece of music my undivided attention. My favourite recording of this work is the 1963 Ashkenazy performance: youtube.com/watch?v=xyPDWa….

Anticipation…

Happy Easter! youtube.com/watch?v=-Tbv6A…