Meta responds to AI competition, launching the Llama 4 family. Featuring multimodal capabilities (text, video, image) and massive context windows (up to 10M tokens), Llama 4 Maverick and Scout are available now, with the 2T-parameter Behemoth previewed. Built on MoE architecture, focusing on reasoning and open access.
Meta introduces the Llama 4 AI series: Scout, Maverick, and Behemoth. Built with MoE architecture and multimodal training, the release includes open models (Scout, Maverick) and the powerful Behemoth (in development). Licensing restricts EU use and large firms. Models show competitive benchmarks and adjusted responses to sensitive topics. Meta AI assistant gets an upgrade.
Meta's Llama 4 launch reportedly faces delays due to performance issues compared to rivals like OpenAI. Falling short on key benchmarks impacts adoption potential. Meta focuses on its API strategy amid intense AI competition and market concerns reflected in stock dips. The situation highlights challenges in the high-stakes AI race.
Meta nears the launch of Llama 4, facing development hurdles, intense AI competition, and investor pressure. Amidst massive investment and challenges like DeepSeek, Meta explores MoE architecture and weighs open-source versus proprietary release strategies, aiming to advance beyond Llama 3 in the high-stakes AI race.
Marc Andreessen highlights Meta's Llama AI running on a Windows 98 PC with 128MB RAM. This feat by Exo Labs involved overcoming hardware, data transfer, and coding hurdles, sparking debate about AI's historical trajectory versus modern systems like Copilot+ PCs. It showcases extreme optimization and challenges assumptions about resource needs.
Harvard study shows open-source Llama 3.1 405B matches GPT-4's medical diagnostic accuracy on complex cases. This breakthrough enables secure, in-house AI deployment, preserving patient privacy and allowing customization, unlike proprietary models requiring external data transfer. AI remains a tool needing human oversight.
Large Language Models face inference bottlenecks due to sequential processing. NVIDIA's FFN Fusion parallelizes Feed-Forward Network computations by merging low-dependency layers, significantly improving LLM speed and reducing costs without sacrificing performance, demonstrated by transforming Llama-405B into the efficient Ultra-253B-Base model.
Meta introduces Meta AI and AI Studio in Indonesia, powered by Llama 3.2. The AI assistant, supporting Bahasa Indonesia, integrates into apps like WhatsApp and Instagram. New tools enhance creator-brand connections via the Instagram marketplace, alongside AI-driven advertising features like Partnership Ads and Advantage+ Shopping Campaigns for improved marketing performance.
Explore adapting Large Language Models (LLMs) like Llama and Mistral for specialized fields like materials science. Learn about fine-tuning techniques (CPT, SFT, DPO/ORPO) and the power of SLERP model merging to enhance domain expertise and unlock emergent capabilities, particularly in larger models. Discover experimental findings and the impact of model scale.
Nvidia introduces Project G-Assist, an experimental AI assistant running locally on GeForce RTX GPUs. Offering contextual game help, system diagnostics, and optimization, it leverages on-device processing, unlike cloud-based alternatives. Available now for desktop RTX users via the Nvidia app, with laptop support planned.
