โดย

April 5, 2025 - Meta says that Llama 4 is its first cohort of models to use a mixture of experts (MoE) architecture, which is more computationally efficient for training and answering queries.

Metas neues Spitzenmodell Llama 4 Behemoth verzögert sich. Das nährt Kritik an Zuckerbergs KI-Kurs mit Milliarden-Investitionen.

    Accelerate your gen AI: Deploy Llama4 & DeepSeek on AI Hypercomputer with new recipes 0 views Eyal Estrin unread, 3:46 AM (30 minutes ago)    to https://cloud.google.com/blog/products/ai-machine-learning/deploying-llama4-and-deepseek-on-ai-hypercomputer Eyal Estrin CISSP, CCSP, CISM, CISA, CDPSE, CCSK Blog: https://security-24-7.com | Books: https://amzn.to/42Xai9A | https://amzn.to/3Sggbtv Twitter: @eyalestrin | Bluesky: @eyalestrin.bsky.social Reply all Reply to author Forward

We’re here with a couple save the dates to reserve some space on your calendars.

Metaは米国時間4月29日、チャットボット機能とソーシャルメディア体験を融合させた新しいAIアプリ「Meta AI」を発表した。Meta AIが競合と一線を画す点は、ソーシャルメディア機能との連携にある。

  Meta has introduced a fresh set of artificial intelligence models under the name Llama 4. This release includes three new versions: Scout, Maverick, and Behemoth. Each one has been designed to better understand and respond to a mix of text, images, and videos. The reason behind this launch seems to be rising competition, especially from Chinese companies like DeepSeek. Their recent models have been doing so well that Meta rushed to improve its own tools to keep up. Where You Can Access Llama 4 The Scout and Maverick models are now available online through Meta’s official site and other developer platforms like Hugging Face. However, Behemoth is still in the testing phase and hasn’t been released yet. Meta has already added Llama 4 to its own digital assistant, which is built into apps like WhatsApp, Instagram, and Messenger in several countries. However, some special features are only available in the U.S. and only in English for now. Who Can and Can’t Use It Meta has placed some limits on who can access Llama 4. People and companies based in the European Union are not allowed to use or share these models, likely due to strict data rules in that region. Also, very large companies, those with over 700 million monthly users — must first get permission from Meta. Smarter Design, Better Performance Llama 4 is Meta’s first release using a new design method called "Mixture of Experts." This means the model can divide big tasks into smaller parts and assign each part to a different “expert” inside the system. This makes it faster and more efficient. For example, the Maverick model has 400 billion total "parameters" (which basically measure how smart it is), but it only uses a small part of them at a time. Scout, the lighter model, is great for reading long documents or big sections of code and can run on a single high-powered computer chip. Maverick needs a more advanced system to function properly. Behemoth: The Most Advanced One Yet Behemoth, which is still being developed, will be the most powerful version. It will have a huge amount of learning data and is expected to perform better than many leading models in science and math-based tasks. But it will also need very strong computing systems to work. One big change in this new version is how it handles sensitive topics. Previous models often avoided difficult questions. Now, Llama 4 is trained to give clearer, fairer answers on political or controversial issues. Meta says the goal is to make the AI more helpful to users, no matter what their views are.

YouTube video by bycloud