Meta is back with a bigger Llama, OpenAI wants to take down Google Search

Meta is back with a bigger Llama, OpenAI wants to take down Google Search
View on Substack
Meta is back with a bigger Llama, OpenAI wants to take down Google Search

what to know for now

🚀 Meta unleashes AI powerhouse challenges industry giants. Llama 3.1 405b boasts 405 billion parameters, matching capabilities of GPT-4 and Claude 3.5. The open-source release enables companies to deploy advanced AI on their hardware, avoiding per-token charges. Read more

🧠 Mistral Large 2 challenges major players. Mistral AI's latest LLM boasts 128,000-token context, excelling in code generation, mathematics, and multilingual tasks. Outperforms Llama 3.1 405B in several benchmarks, approaching GPT-4's capabilities. Available on la Plateforme and HuggingFace for research. Read more

🎵 Udio AI music creator levels up with v1.5 update. Version 1.5 introduces 48kHz-stereo tracks, enhanced clarity, and improved instrument separation. New features include stem downloads, audio-to-audio remixing, key control, and expanded language support. Creator page consolidates tools for streamlined music production. Read more

🍎 Apple releases open-source 7B AI model, rivaling Mistral-7B performance. DCLM-7B-instruct outperforms similar-sized models, trained on fewer tokens. Fully open-sourced weights, training code, and datasets available. Part of Apple's DCML project, focusing on high-quality dataset design for efficient model training. Read more

🧪 AI Research of the Week

The Llama 3 Herd of Models
From Meta

Jake’s Take: Llama 3 is a family of large language models developed by Meta, featuring models of 8B, 70B, and 405B parameters, with capabilities across language, vision, and speech tasks that are competitive with or surpass other leading models. The paper details the comprehensive approach to developing Llama 3, including data preparation, model architecture, training infrastructure, safety considerations, and evaluation across various benchmarks.

The open release of Llama 3, especially the new 405B parameter model that rivals closed-source competitors, will likely accelerate AI research and development across academia and industry, potentially democratizing access to state-of-the-art language models and spurring further innovations in the field.

what to know for later

🔍 OpenAI to challenge Google with AI-powered SearchGPT. SearchGPT integrates real-time web information with ChatGPT's conversational AI, providing up-to-date answers and relevant source links. Currently in limited testing, it aims to streamline information retrieval by allowing natural language queries and follow-up questions. Read more

🚨 Senators demand OpenAI safety data amid employee warnings on AI testing. Democratic lawmakers request detailed information on OpenAI's safety measures, employee agreements, and risk mitigation strategies. Response due by August 13, addressing concerns about rushed testing and potential misuse of advanced AI models. Read more

🖥️ Oak Ridge unveils Discovery supercomputer, successor to world's fastest Frontier. Discovery, slated for 2027-2028 delivery, aims to balance speed with advanced capabilities like network and memory bandwidth. Unlike predecessors, it may not claim world's fastest title due to AI-driven private sector competition. Read more

💽 OpenAI explores custom chip production, eyeing Broadcom partnership. Sam Altman negotiates with semiconductor designers to reduce Nvidia dependency. OpenAI aims to develop proprietary AI chips, hire ex-Google TPU experts, and build comprehensive infrastructure including data centers and power systems. Read more