AI at Meta

Introducing Meta Llama 3: the next generation of our state-of-the-art open source large language model — and the most capable openly available LLM to date. These next-generation models demonstrate SOTA performance on a wide range of industry benchmarks and offer new capabilities such as improved reasoning. Details in the full announcement ➡️ https://go.fb.me/a24u0h Download the models ➡️ https://go.fb.me/q8yhmh Experience Llama 3 with Meta AI ➡️ https://meta.ai Llama 3 8B & 70B deliver a major leap over Llama 2 and establish a new SOTA for models of their sizes. While we’re releasing these first two models today, we’re working to release even more for Llama 3 including multiple models with capabilities such as multimodality, multilinguality, longer context windows and more. Our largest models are over 400B parameters and while they’re still in active development, we’re very excited about how they’re trending. Across the stack, we want to kickstart the next wave of innovation in AI. We believe these are the best open source models of their class, period — we can’t wait to see what you build and look forward to your feedback.

New paper from FAIR, Chameleon: Mixed-Modal Early-Fusion Foundation Models. While some LLMs have separate image and text encoders or decoders, this research presents a family of early-fusion token-based mixed-modal models capable of understanding & generating images & text in any arbitrary sequence. Paper ➡️ https://go.fb.me/7rb19n The paper includes details on the full modeling approach and training — we hope that sharing this work will help the community further research on mixed-modal models.

What a weekend! We were blown away by the creativity, talent and dedication of everyone who joined us and Cerebral Valley in San Francisco for two days of hacking with Meta Llama 3. Congratulations to our winning teams! 🥇 Open Glass AI: Open source smart glasses built with $20 of parts. 🥈 Deb8: An AI agent debate arena. 🥉 Team AAA: Discovery of new jailbreaking methods for Llama 3. See these and more in the gallery of projects ➡️ https://go.fb.me/maxd1m A few highlights: • 350+ hackers. • 51 completed projects presented to our panel of judges. • 7+ workshops from our team and AWS, Groq, Lambda, W&B, Octo, Brave and • LlamaIndex. • $6k in cash prizes + $8k in AWS credits awarded to winning teams + Ray-Ban Meta smart glasses. • 50 boxes of coffee and 500+ energy drinks consumed over two days of hacking. We're excited to see even more creativity like this from the community, built with Llama 3!

There are few specializations in medicine that require more precision or data than radiation oncology. Built with Meta Llama 2, RadOnc-GPT is a fine-tuned LLM that has the potential to significantly improve radiation therapy decision-making. More details ➡️ https://go.fb.me/2fes0g Paper ➡️ https://go.fb.me/dny6by RadOnc-GPT improves specificity and clinical relevance compared to general LLMs. Building with an open model like Llama 2 enabled the researchers to fine-tune on their data in such a way that no patient data is ever shared outside of a secure network, as they can train the model using their own local GPU server.

Four papers to add to your reading list from AI researchers at Meta at #ICLR2024. • ICLR Outstanding Paper Award: Vision Transformers Need Registers ➡️ https://go.fb.me/6kqyj6 • ICLR Outstanding Paper honorable mention: Flow Matching on General Geometries ➡️ https://go.fb.me/itmlgi • Demystifying CLIP Data ➡️ https://go.fb.me/ht0130 • Revisiting Feature Prediction for Learning Visual Representations from Video ➡️ https://go.fb.me/20jah3

New research from FAIR: Better & Faster Large Language Models via Multi-token Prediction. Paper ➡️ https://go.fb.me/o0yd06 We show that replacing next token prediction tasks with multiple token prediction can result in substantially better code generation performance with the exact same training budget and data — while also increasing inference performance by 3x. While similar approaches have previously been used in fine-tuning, this new paper expands to pre-training for large models, showing notable behaviors and results at these scales.

We're in Vienna for #ICLR2024 this week! Stop by to chat with our team or learn more about our latest research. Things to know 📍 Find us @ Booth A15 • We're sharing 25+ publications & two workshops. • Daily talks with the Llama team on May 7-9 @ 11:00 • Decision making algorithms made easy with TorchRL on May 9 @ 13:45 • TorchRL presentations + Q&A, 2x daily sessions @ 10:00 & 15:00 • Demos of MAGNeT, Segment Anything (SAM) & more! See you at ICLR!

The Meta Llama 3 Hackathon is this weekend in San Francisco with Cerebral Valley! Get on the list ➡️ https://go.fb.me/01hro5 What to expect • Two days of building alongside the best hackers in AI. • Hands on support from the Llama team. • Some of the top names in the industry speaking and judging submissions. • Over $10k in cash prizes to help kickstart Llama 3 projects. Thank you to the community who came together to help make this hackathon possible: Amazon Web Services (AWS), Groq, Lambda, Weights & Biases, Hugging Face, OctoAI, LlamaIndex, Ollama and Brave.

We continue to be amazed by all the ways that individuals and organizations of all sizes are building with Meta Llama models.

As part of our focus on developing Llama 3 in a responsible way, we created a number of resources to help others use it responsibly as well. This includes new trust and safety tools like CyberSec Eval 2. Paper ➡️ https://go.fb.me/gy1h0d Cybersec Eval 2 expands on its predecessor by measuring an LLM’s susceptibility to prompt injection, automated offensive cybersecurity capabilities, and propensity to abuse a code interpreter, in addition to the existing evaluations for insecure Coding Practices and Cyber Attack Helpfulness.