Rain AI’s Post

There are very few players that are able to align user experience, SW, and HW excellence to produce as great a technology portfolio as Meta. At the heart of ’s AI innovation is MTIA Gen 2, our in-house training and inference accelerator. 🚀 Learn how our engineers are co-designing hardware and software to achieve unmatched performance and efficiency for AI workloads. Discover the future of AI at Meta! #AI #TechInnovation #MTIA #CustomSilicon

Meta's Llama 3 is out! 3 initial takeaways below: - 70B is the Highest quality (MMLU) open-source model available at 82% but below GPT-4 & Claude Opus's ~86%. Competitive with Sonnet, Mistral Large, and Gemini 1.5 Pro - 70B is not an MoE architecture. This means we expect slower speeds than other models of similar quality, i.e Mixtral 8x22B which has 39B active parameters (vs. L3's 70B) per forward pass - 70B & 8B released currently, but announced in the coming months to expect "new capabilities, longer context windows, additional model sizes, and enhanced performance" #Llama3 , #AI , #ArtificialIntelligence , #Meta

Meta recently released SAM 2, the second version of Segment Anything. This model will land on Hugging Face soon, making it super easy to run inference. SAM 2 leverages "Hiera" as the backbone to process images and videos. Hiera was added to the Transformers library just last week! Hiera is an Apache 2.0 licensed vision model by Meta that aims to simplify the design of hierarchical Vision Transformers (ViTs). Hierarchical vision transformers, like Swin Transformer, are more parameter-efficient compared to vanilla ViTs, but end up being slower at inference time due to the additional complexity of manually adding spatial bias. The authors design a simpler hierarchical model consisting of standard Transformer blocks. The model is pre-trained using the masked auto-encoder (MAE) objective, which forces the model to learn spatial biases. The model is very strong at image and video processing tasks, which explains why the authors of SAM 2 went for this one as the backbone. Kudos to Eduardo Pacheco and Naman Garg for this amazing open-source contribution. Docs: https://lnkd.in/es4uxZ3U Checkpoints: https://lnkd.in/etUKTy4t SAM 2 announcement: https://lnkd.in/e3Dggr52 #computervision #meta #huggingface #sota

🚀 Chameleon by Meta is Now on Hugging Face Transformers! 🦎 Vision language model available in 7B and 34B sizes 🤩 Early fusion for seamless generation 😏 Scalable & stable training with architectural improvements 💡 Any-to-any model: image & text in, image & text out ✨ What's Early Fusion? 🤔 Early fusion fuses image patches and text tokens into a shared space, enabling seamless text and image generation. Why is Chameleon Special? 😎 Increased token count: 5x more than Llama 3. Handles any-to-any inputs: image and text in, image and text out. Competitive with larger models in text-only and image-pair prompting. #ArtificialIntelligence #VisionLanguageModel #MachineLearning #Meta #HuggingFace #EarlyFusion

🚀 Exciting News in AI! 🚀 I'm thrilled to announce the launch of Llama 3.1 by Meta 🦙✨ This cutting-edge model, with versions ranging from 8B to an astounding 405B parameters, is set to revolutionize the AI landscape. What makes this even more incredible? Meta has made this model available for free! 💥 The sheer scale of this achievement is mind-blowing, with over 16,000 NVIDIA H100s GPU used to train it. Imagine the resources and effort it took to bring this state-of-the-art technology to the world without any cost. 🌍💡 A huge shoutout to Meta for their commitment to advancing AI and making it accessible. This is a game-changer for developers, researchers, and enthusiasts alike. Let's harness the power of Llama 3.1 to push the boundaries of what's possible! 🚀🤖 How to Use the Llama 3.1 Model Locally 1. Download and Install Ollama: - Visit the [Ollama official website](https://lnkd.in/gyEfarQk) and download the installation package for your operating system. - Follow the on-screen instructions to complete the installation process. 2. Run the Llama 3.1 Model: - Open your terminal. - Enter the following command to start the model: "ollama run llama3.1" #AI #MachineLearning #Llama31 #Meta #Innovation #TechForGood #AIRevolution #Technology

In 2019, John Carmack, one of the most brilliant software engineers of our generation, left his job as Oculus CTO to work on Artificial General Intelligence. Ilya Sutskever, the Chief Scientist at OpenAI who is widely considered as the technical mind behind their LLM breakthrough, sent Carmack a list of 30 research papers and said "If you really learn all of these, you’ll know 90% of what matters today." Here's the full list with the actual research papers (Shared in the first comment). Do you want to be a world-class AI scientist? You know what to do.

With More Power, Comes More Responsibility: A Glimpse into the Multimodal Future with Project Aria 💪🌐 Change is on the horizon, and the winds of innovation are blowing stronger than ever. Brace yourself for a paradigm shift as multimodal technology takes center stage, promising a fascinating evolution. As someone not deeply entrenched in AI research, it's always intriguing to peek into the world of groundbreaking projects like Project Aria. While not the latest news, the concept of donning smart glasses and collaborating with leading universities under the Facebook Research umbrella sounds like a futuristic adventure (for those who've taken the plunge). Data collection, especially in the realm of multimodal datasets, is undoubtedly a Herculean task. Transforming raw data into a refined treasure trove is a feat in itself. And in this era, where Project Aria is venturing into the multi-sensorial, one can only imagine the complexities involved. Picture this: You hold the power to shape the future. It's a classic notion, but the twist lies in knowing that every online move, every action, gets absorbed by AI. The data isn't confined to mere website heat-maps; we're talking about actual video data capturing our every nuance — the way we move, dance, cook, sit, travel, run, and every other intricacy of human behaviour. The responsibility now transcends traditional boundaries. Learn more about Project Aria here: https://lnkd.in/gBUhvrEp Datasets: https://lnkd.in/gDatstsV Github tools repo : https://lnkd.in/geT_ZriH Paper: https://lnkd.in/gCzNApjN Interesting thing is: Bias has become a bit more multidimensional now, so another NP-Hard problem for those curious minds!! In this age of transformative technology, let's ponder on the significance of our actions online. As we contribute to datasets that might one day replicate human behavior, it becomes imperative to act responsibly. The intersection of technology and humanity beckons, urging us to consider the impact of our digital footprint on the future landscape. #ProjectAria #MultimodalFuture #TechResponsibility"

"“Machine learning Depth has been the team’s most important project for years, and we had to make fundamental advances in ML modeling, acceleration, and more,” says Reality Labs Director of Engineering for XR Tech Paul Furgale. “This allowed us to launch advanced MR features on Quest 3 despite not having state-of-the-art depth hardware. And after Quest 3, I’m convinced that Passthrough and MR will be a standard feature on all future headsets.”" #meta #metaquest3 #smartglasses #ai #machinelearning #artificialintelligence #spatialcomputing

🦙🦙🦙#Llama 3🦙🦙🦙 Meta 2️⃣ new best in-class open models are here🎉, beating Mixtral 8x22B, Mistral 7B, and Gemma 7B check the attached doc for a detailed review on it 👑......👇👇👇. I'm eager to hear from the community about their insights on Llama 3 and the impactful contributions it brings to the table. Share your thoughts, research endeavors, and experiences utilizing this cutting-edge technology to achieve your objectives. Let's foster a collaborative discussion to explore the diverse applications and advancements facilitated by Llama 3. Your input can drive innovation and inspire others in the community! 🧠 Resources Model weight: https://lnkd.in/eqET7P7v Github: https://lnkd.in/er8hThRQ Documentation: https://lnkd.in/eDauUkuT Download Ollama: https://lnkd.in/e7RBz7HC Closed-source vs open-source weights are from Maxime Labonne post. AI at Meta Ollama #llm #llama3 #genai

Humanity is hurtling toward a future that genuinely alarms me. With the debut of Meta Quest 3, bridging the gap between the physical and virtual realms, mere millimeters from our eyes, we seem destined for a world where screens become almost inseparable from our faces. Imran Chaudhri, a former Apple employee with 22 years of experience, endeavors to transform a gadget that's 'Screenless, Seamless, and Sensing.' This endeavor leads us toward cutting-edge technology while also allowing us to stay connected to our environment, finally restoring a balance that has seemed askew. What are your thoughts following this TED talk?