Google DeepMind’s Post

The revolution of open models continues in earnest. It will be interesting to try these models out.

🌟 Introducing Gemma 2: Revolutionizing AI with Efficiency and Performance! 🚀 AI holds the promise to address humanity's greatest challenges, and with Gemma, advanced tools are now within everyone's reach. This year, the Gemma family expanded to include CodeGemma, RecurrentGemma, and PaliGemma, all integrated with leading platforms like Hugging Face and NVIDIA. Key Highlights: 🔹 Outsized Performance: The 27B model leads its size class, outperforming even larger models. 🔹 Cost-Effective: Efficient inference on single GPU setups saves on costs. 🔹 Versatile Inference: Fast performance across various hardware, from cloud to home setups. Explore Gemma 2's capabilities now and transform your AI projects! 🌐 #google #gemma #gemma2

Gemma 2 has been released and is optimized to run at incredible speed across a range of hardware even on gaming laptops. run locally with the quantized version - Gemma.cpp on CPU.available on Hugging Face and through ollama. #techTrends #LLMs #AI Data Science Academy

Gemma 2 is out and it’s crazy to think we can fit the best open LLM in the world on an old 32GB USB stick. Gemma 1 was the speed run - we went from forming a team to release in just 5 months. Mistakes were made. Gemma 2 is about carefully revising every part of what goes into our models. No stone left unturned. We ablated every part of our model architecture to get more impact from the same parameter count. We learned a lot more about our data and how to optimize our mixtures for the effects we want. Even things we took for granted, like Chinchilla scaling laws, went through a fine tooth comb We shipped distillation and this is so impactful. This was a long bet - the first experiments ran in October, but that wasn't enough. It took dedication and not giving up to make it really work - and it was in late April when distillation really came together.    Training and monitoring 27B was like steering a big ship. You can see where it’s going and you can turn the wheel, but it takes a long time to get feedback - and a few times we turned the wheel the wrong way. Gemma 2B (shipping very soon), will be the most capable on-device model. This is a new space and it's very exciting what app developers can do with local, private LLMs. The whole Gemma team is irreplaceable, but I want to particularly thank the pretraining team - Shreya Pathak, Morgane Rivière, Cassidy Hardin, Thomas Mesnard. 🙏 to Armand Joulin, who always knows when to say no.

Worth a try

We’re releasing the #Gemma 2 open models to researchers and developers globally. 🛠️ Available in both 9B and 27B parameter sizes, they're higher performing than ever before - and will help unlock new possibilities to build and deploy with AI. Get started in #Google #AI Studio today → https://dpmd.ai/45Q6yba

Last week, Google welcomed us in the Gemini era, introducing their most capable AI model yet. According to Sundar Pichai, Google’s CEO, and Google DeepMind CEO Demis Hassabis, this AI breakthrough is poised to revolutionise a broad spectrum of Google's products. Why is that and what does Gemini bring to the AI table? Here’s all you need to know about Google’s new LLM: ➡ Gemini is natively multimodal. It means that the model gives you the potential to transform any type of input into any type of output. Generate text, images, video, audio, code or combined. ➡ It comes in three sizes: Ultra, Pro and Nano. The Pro version is already powering Bard, and the integration with Google AI Studio or Vertex AI in Google Cloud has been available since December 13th. Gemini Nano is now running on Pixel 8 Pro, and the Ultra size is coming next year. ➡ Google’s Gemini vs OpenAI’s GPT-4, meaning performance. How does the Gemini rank in benchmarks? Google claims that the Ultra version is supposed to outperform its rival in 30 out of 32 benchmarks, leaving GPT-4 slightly behind. It’s the first model to exceed human experts on MMLU (Massive Multitask Language Understanding), one of the most popular methods to test the knowledge and problem-solving abilities of AI models. But at this moment these are one-sided statements, the actual outcome we will see next year. ➡ Improving programming with AlphaCode 2, a new code-generating system powered by Gemini. It can perform better than 85% of coding competition participants and almost twice as well as the original AlphaCode. While benchmarks serve their purpose, the genuine evaluation of Gemini's capabilities lies in the hands of everyday users. Their experiences will be the ultimate test of Gemini's true potential. 💡 For more information visit: https://lnkd.in/dbu68gPe  💡 See more about AlphaCode2: https://lnkd.in/dYnd-iQW #gemini #geminiera #geminiAI #google #ai 

Google DeepMind has just announced the release of Gemini, a groundbreaking new AI model that promises to revolutionize the way we interact with computers. Gemini is the largest and most capable AI model ever created. It is built to be natively multimodal, which means that it can understand and operate across text, images, audio, video, and code. This makes it incredibly versatile and powerful, with the potential to be used for a wide range of tasks. Here are just a few of the things that Gemini can do: * Generate different creative text formats, like poems, code, scripts, musical pieces, email, letters, etc. It will try its best to fulfill all your requirements. * Translate languages fluently and accurately. * Answer your questions in an informative way, even if they are open ended, challenging, or strange. * Write different kinds of creative content, like poems, code, scripts, musical pieces, email, letters, etc. It will try its best to fulfill all your requirements. Gemini is still under development, but it has the potential to change the world. I believe that it will be used to create new products and services that we can't even imagine yet. I'm excited to see what the future holds for Gemini and for AI in general. **What are your thoughts on Gemini? What are some of the ways that you think it could be used?** **#AI #MachineLearning #DeepLearning #GoogleAI #Google #Gemini

#AI, Google’s new version of Gemini can handle far bigger amounts of data Google DeepMind launched the next generation of its powerful artificial intelligence model Gemini, which has an enhanced ability to work with large amounts of video, text, and images.   It’s an advancement from the three versions of Gemini 1.0 that Google announced back in December, ranging in size and complexity from Nano to Pro to Ultra. (It rolled out Gemini 1.0 Pro and 1.0 Ultra across many of its products last week.) Google is now releasing a preview of Gemini 1.5 Pro to select developers and business customers. The company says that the mid-tier Gemini 1.5 Pro matches its previous top-tier model, Gemini 1.0 Ultra, in performance, but uses less computing power.   Crucially, the 1.5 Pro model can handle much larger amounts of data from users, including the size of prompts. While every AI model has a ceiling of how much data it can digest, the standard version of the new Gemini 1.5 Pro can handle inputs as large as 128,000 tokens, which are words or parts of words that an AI model breaks inputs into. That’s on a par with the best version of GPT-4 (GPT-4 Turbo).   However, a limited group of developers will be able to submit up to 1 million tokens to Gemini 1.5 Pro, which equates to roughly 1 hour of video, 11 hours of audio, or 700,000 words of text. That’s a significant jump that makes it possible to do things that no other models are currently capable of.   In one demonstration video shown by Google, using the million-token version, researchers fed the model a 402-page transcript of the Apollo moon landing mission. Then they showed Gemini a hand-drawn sketch of a boot, and asked it to identify the moment in the transcript that the drawing represents.   #RB, #AI https://lnkd.in/eCTA4i3h