Proud of this work coming out of a collaboration between FAIR researchers & Georgia Tech that was just published in ACS Central Science! The dataset & AI models that we released together could help meaningfully accelerate research combating climate change. open-dac.github.io

Excited to share that our paper “CodeIt: Self-Improving Language Models with Prioritized Hindsight Replay” was accepted into ICML!
Blaze(j) Manczak 🇵🇱🇱🇺🇪🇺 Auke Wiggers Corrado Rainone David Zhang Michaël Defferrard Taco Cohen 1/5

In many hard code generation problems, the chance of generating 100% correct code in one shot is quite low, making it hard to improve the policy. But because we can easily execute code, we can use hindsight relabelling so that *every* sample becomes a success and can be used to…

Meta announces Better & Faster Large Language Models via Multi-token Prediction

Large language models such as GPT and Llama are trained with a next-token prediction loss. In this work, we suggest that training language models to predict multiple future tokens at

After 3 years, it's time for us to start sharing the chapters of the GDL book! ❤️

Also included: companion slides from our Cambridge University & University of Oxford courses 🧑‍🎓

Chapter 1 is out **now**!
More to follow soon 🎉

geometricdeeplearning.com/book 📖

Michael Bronstein joan bruna Taco Cohen

Congrats to AI at Meta on Llama 3 release!! 🎉
ai.meta.com/blog/meta-llam…
Notes:

Releasing 8B and 70B (both base and finetuned) models, strong-performing in their model class (but we'll see when the rankings come in @ lmsys.org :))
400B is still training, but already encroaching…

Excited to share what I’ve been working on for the past 9 months. So incredibly proud of the entire team that worked tirelessly to make Llama 3 happen! And this is only the beginning…

ai.meta.com/blog/meta-llam…

It’s here! Meet Llama 3, our latest generation of models that is setting a new standard for state-of-the art performance and efficiency for openly available LLMs.

Key highlights

• 8B and 70B parameter openly available pre-trained and fine-tuned models.
• Trained on more…

Coding is the frontier of AI.
Excited to push the two frontiers of AI coding:

1. SWE(-bench/agent)
2. Olympiad programming (this tweet)

Introduce USACO benchmark:
* inference methods (RAG/reflect) help a bit: 9->20%
* human feedback helps a lot: 0->86%!

princeton-nlp.github.io/USACOBench/

After two good years at Microsoft Research AI4Science, I am very excited to announce that as of this month I have, together with Chad Edwards, co-founded a new startup in the field of molecular and materials discovery.

New from Air Street Press: a detailed guide to data acquisition for AI-first start-ups, in collaboration with our friend Moritz Mueller-Freitag, Director of Product Management at Qualcomm.

Let us know your thoughts and experiences!

Some of the highlights 🧵

Will the rise of AI break the grid? Alarm bells of rapid demand growth are sounding. Are we facing a supply crisis—or is it a sales pitch meant to build more gas plants? We dive in w/Jonathan Koomey on this week's SHIFT KEY. Give it a listen here: heatmap.news/podcast/shift-…

Suno.ai is the craziest, most magical thing on the internet right now. How is not everyone talking about this

SWE-agent is our new system for autonomously solving issues in GitHub repos. It gets similar accuracy to Devin on SWE-bench, takes 93 seconds on avg + it's open source!

We designed a new agent-computer interface to make it easy for GPT-4 to edit+run code
github.com/princeton-nlp/…

It’s been an exciting time for AI hardware — pins, rabbits, necklaces, and of course, glasses. Adding to this illustrious group, we are excited to announce: the AI nose ring.

A undeniably fashionable integration of hardware and software, the AI nose ring is optimized for…

Applies to research as much as to product development