5 Outstanding Paper Awards at ICLR.
2 with NYU folks among authors
1 with FAIR authors.

When both OpenAI's closed-source GPT-4 and Meta's open-source LLaMA 3 70B models demonstrate comparable performance on Chatbot Arena Elo Score, MMLU, and MT Benchmark tests, the argument for choosing the more expensive proprietary model, which costs 58 times more and runs 17…

Awesome explanatory video on rotation-equivariant graph convolutional nets by my FAIR colleague Larry Zitnick.

This is the trick if you want to represent molecules independently of their orientation in space.

I'll be at Vivatech.

Percentage of French voters planning to vote for the far right in the coming European elections by profession.
😧🫣

1 billion LLM downloads from Hugging Face in the last 2 years.
570 million of which are GPT-2 (and the world did not end).
What are the most harmful things people have done with LLMs?
What are the most beneficial things people have done with LLMs?

1900: why don't you plow the field instead of reading books?
1960: why don't you read books instead of watching TV?
2020: why don't you watch TV instead of being on your tablet?

I, as the first author, will present JoMA on May 7th morning (Halle B #260, 10:45-12:45pm) in #ICLR2024 🚀🚀!

JoMA (arxiv.org/abs/2310.00535) extends our previous Scan&Snap work (arxiv.org/abs/2305.16380). It analyzes training dynamics of Transformer and shows that self-attention…

🎓At AI at Meta even the Chief Al Scientist supervises Ph.D. students! Together with Jean Ponce, Yann LeCun gave birth this week to Dr. Adrien Bardes🎓, crème de la crème in predictive non-contrastive learning, the topic I love💚 #byolforever , and brought us ViCReg, IMJEPA and more!

Our latests investigation into pager prohibition uncovers thousands of children were arrested in the late 1980s and early 1990s for possessing pagers… newsletter.pessimistsarchive.org/p/the-forgotte…

Indeed, AI assistants accessed through smart glasses are the future.

EgoPet: first purrson videos for Self-Supervised Learning.

The Al Gores were right.

True.

Something to be said for an actual industrial policy.

The Shift in Rhetoric on AI and Biothreats is a Lesson on the Risks of Premature Regulation.
by Martin Makaryan April 30, 2024
datainnovation.org/2024/04/the-sh…

Our paper is out in Nature Human Behaviour🔥🔥

‘Evidence of a predictive coding hierarchy in the human brain listening to speech’

📄nature.com/articles/s4156…
💡Unlike language models, our brain makes distant & hierarchical predictions

with Alexandre Gramfort and Jean-Rémi King

Thread👇

I am comically going to the Moon.
storydiffusion.github.io

Llama 3 70B takes the king’s crown 👑 from GPT-4 Turbo

- 100% in coverage and 70% in quality of code
- 💵 most cost-effective inference
- open-weight

We tested OpenAI's GPT 3.5 & 4, Meta's Llama 3, Google's Gemini 1.5 Pro, cohere's Command R+ and 130+ other LLMs

Introducing AdvPrompter! 🚀Our new optimization technique that crafts prompt-dependent, human-readable adversarial suffixes in real-time, enhancing security for LLMs🛡️!

1️⃣Generate suffix in ~2 seconds, ~800x faster than methods like GCG and AutoDAN.
2️⃣Higher success rate, in…