Google launches its smaller AI models, while the music industry sues generative AI
what to know for now
🚀 Google releases Gemma 2. Gemma 2 models (9B and 27B parameters) outperform previous versions, offering higher efficiency on single GPUs or TPUs. Compatible with major AI frameworks, it supports budget-friendly AI deployments and responsible development with enhanced safety features. Read more
🎶 Record labels sue AI startups for copyright infringement in training data. Udio and Suno, AI music generators, face lawsuits for allegedly using copyrighted music to train models. RIAA demands $150,000 per work, challenging AI's fair use claims in the music industry. Read more
🖼️ Stability AI appoints new CEO, secures $80M in funding. Stability AI, creator of the Stable Diffusion models, appoints Prem Akkaraju as CEO and raises $80M from prominent investors. The funding addresses financial challenges and supports ongoing product development, including Stable Audio Open and Stable Diffusion 3 Medium. Read more
💼 Anthropic launches Projects for AI collaboration with Claude. Projects feature a 200,000-token context window, centralizing team knowledge for tailored AI assistance. Custom instructions and Artifacts streamline workflows, boosting efficiency and fitting organizational needs precisely. Read more
🧪 AI Research of the Week 🧪
LLM Critics Help Catch LLM Bugs
From OpenAIJake’s Take: Researchers trained "critic" models using reinforcement learning from human feedback (RLHF) to help humans more accurately evaluate AI-generated code. These critic models, called CriticGPT, identified more bugs than human reviewers and were preferred over human critiques 63% of the time. CriticGPT, when combined with human reviewers, reduced hallucinations and improved overall bug detection, surpassing the performance of human-only and AI-only reviews.
This development indicates a significant advancement in AI oversight capabilities from OpenAI, enhancing the reliability of their systems by improving bug detection and reducing the reliance on human evaluations, potentially leading to more robust and safer AI applications.
what to know for later
💼 OpenAI acquires Multi and Rockset. Multi's real-time collaborative tools enable shared screen and code collaboration, enhancing AI-driven content summarization. Rockset's advanced search and database analytics improve data ingestion, analysis, and retrieval. Read more
🔊 OpenAI delays ChatGPT's advanced Voice Mode feature. Initially planned for June, OpenAI postponed the Voice Mode rollout to address content detection, user experience, and scalability issues. Alpha testing starts in July, with full release expected in fall if safety checks are met. Read more
📰 TIME partners with OpenAI for multi-year content deal. TIME claims their 101-year archive will enhance OpenAI's products, providing users with cited, trusted journalism. This collaboration supports global information access and innovative news experiences through AI technology. Read more
📺 Al Michaels' AI voice to recap Paris Olympics on Peacock. NBCUniversal will use AI to deliver personalized Olympic recaps in Al Michaels' voice. Each morning, Peacock will stream tailored video summaries using AI-generated voiceovers, maintaining his signature style. Read more