AI at Meta’s Post

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.

In April, we published a research paper on a new approach for building better and faster LLMs by using multi-token prediction. Using this approach, we can train language models to predict multiple future words at once, improving model capabilities and training efficiency while allowing for faster inference. In the spirit of responsible open science, we’ve released pre-trained models for code completion using this approach to enable further exploration in the research community. Get the model on Hugging Face ➡️ https://go.fb.me/dm1giu More on this approach ➡️ https://go.fb.me/x1zhdq

Introducing Meta 3D Gen – new text-to-3D research from AI researchers at Meta that enables text-to-3D generation with high-quality geometry and textures. Research paper ➡️ https://go.fb.me/c9g4x6 Meta 3D Gen delivers text-to-mesh generation with high-quality geometry, texture and PBR materials. It can generate high-quality 3D assets, with both high-resolution textures and material maps end-to-end, producing results that are superior to previous state-of-the-art solutions — all at 3-10x the speed of previous work. In addition to the Meta 3D Gen technical report, we’re publishing our research on the two individual components of the Meta 3D Gen system: Meta 3D AssetGen for generating 3D models from text — and Meta 3D TextureGen, a model capable of high-quality texture generation and AI-assisted retexturing of artist-created or generated assets. Meta 3D AssetGen paper ➡️ https://go.fb.me/87tktg  Meta 3D TextureGen paper ➡️ https://go.fb.me/tvbdf8

Today we’re releasing Meta LLM Compiler, a family of models built on Meta Code Llama with additional code optimization and compiler capabilities. The models achieve state-of-the-art results on optimization of code size and disassembly tasks. Hugging Face repo ➡️ https://go.fb.me/tdd3dw Research paper ➡️ https://go.fb.me/hvfnam LLM Compiler can emulate the compiler, predict optimal passes for code size, and disassemble code. It can be fine-tuned for new optimizations and compiler tasks. This work shows that AI is learning to optimize code and can assist compiler experts in identifying opportunities to apply optimizations. We believe this work could have an impact ranging from use in optimization for individual developer environments to inclusion in a compiler such as LLVM. We’re releasing LLM Compiler 7B & 13B models under a permissive license for both research and commercial use in the hopes of making it easier for developers and researchers alike to leverage this in their work and carry forward new research in this highly impactful space.

Know an organization using AI to drive social impact? We just opened applications for the Meta Llama Impact Innovation Awards, a program to recognize & support organizations using Meta Llama models to make an impact in Africa, the Middle East, Turkey, Asia Pacific & Latin America! The program will grant a series of awards up to $35K USD for organizations tackling some of the regions’ most pressing challenges. Accepting applications until July 26 ➡️ https://go.fb.me/g1m32d

Last week we released Meta Chameleon: a new mixed-modal research model from Meta FAIR. Get the models ➡️ https://go.fb.me/hrkkgf Research paper ➡️ https://go.fb.me/u75dq8 The new safety tuned 7B and 34B models we’ve released for research use can take combinations of text and images as input and produce text outputs. Chameleon is one of the first publicly released approaches using a single unified architecture for both encoding and decoding using an early fusion approach. We’ve shared details on the full modeling approach and training in the new research paper, and we hope that the work we’re releasing will help to further democratize access to foundational mixed-modal models and empower others to streamline and scale their work in this space.

Six papers to add to your reading list from AI researchers at Meta at #CVPR2024. • PlatoNeRF: Discerning Reality in Plato's Cave from Single-View Two Bounce Time of Flight ➡️ https://go.fb.me/tju5fo • Nymeria: A Massive Collection of Multimodal Egocentric Daily Motion in the Wild ➡️ https://go.fb.me/0wcu84 • Relightable Gaussian Codec Avatars ➡️ https://go.fb.me/gdtkjm • URHand: Universal Relightable Hands ➡️ https://go.fb.me/1lmv7o • RoHM: Robust Human Motion Reconstruction via Diffusion ➡️ https://go.fb.me/ogm92y • HybridNeRF: Efficient Neural Rendering via Adaptive Volumetric Surfaces ➡️ https://go.fb.me/tzik3j

With the release of Meta Llama 2, we developed a technique called Ghost Attention (GAtt) which uses RLHF to fine-tune model responses keeping the initial instructions in mind, resulting in AI models that are much better at retaining initial instructions in multi-turn conversations. This approach caught the attention of SAIF CHECK who went on to build a model evaluation system using Meta Llama 3 to help companies navigate the challenges of risk and compliance with local laws where their technology is used.

Meta FAIR researchers have introduced JASCO, a generative text-to-music model. This new model differs from other text-to-music approaches through its ability to accept various conditioning inputs like chords or beat for improved controllability over the types of outputs it creates. You can find more details in the research paper available now, and we plan to release inference code and a pretrained model under a CC-BY-NC license soon. Details and samples ➡️ https://go.fb.me/tkp5fz