OpenAI Sora-Rival Gen-3 Alpha by Runway AI Now Available for Everyone
++ Microsoft and Google Under EU Antitrust Lens, Apple Unveils AI in Vision Pro, Meta Adjusts AI Strategy, Detroit Limits Police AI Use, Plus More Innovations and Legal Battles Across the Tech Sphere
Today's highlights:
🚀 AI Breakthroughs
Runway's Gen-3 Alpha AI Video Model Launched with Subscription Requirement
• RunwayML's Gen-3 Alpha AI video model now generally available, offering hyper-realistic videos from varied prompts
• Unlike previous models, Gen-3 Alpha requires a subscription starting at $12/month per editor for access
• Gen-3 Alpha enhances video realism with advanced key-framing, expressive characters, and dynamic transitions
• Initial release focuses on text-to-video conversions, with plans to expand to other modes and control features
• Videos generated can be up to 10 seconds, boasting superior generation speeds compared to many competitors
• Runway teases future enhancements and a potential free version of Gen-3 Alpha, as part of a broader model series.
Elon Musk Teases Grok 3's Capabilities After $3 Billion Investment in Nvidia GPUs
• Elon Musk announced Grok 3 will heavily utilize 100,000 Nvidia H100 GPUs, hinting at substantial investments and advancements in AI technology
• Each Nvidia H100 GPU costs an estimated $30,000 to $40,000, suggesting a hardware investment of $3 billion to $4 billion for Grok 3's development
• Meta plans to acquire around 350,000 Nvidia H100 GPUs by end of 2024, reflecting intense competition in AI hardware acquisition among tech giants
• Grok 3 training surpasses previous versions, with Grok 2 requiring 20,000 GPUs, indicating a significant scale-up in resources and capabilities
• Musk's strategic reallocation of a $500 million H100 shipment from Tesla to X showcases internal prioritization of AI development over automotive needs
• The fierce competition for AI talent is exacerbated by access to substantial computing resources, as seen in Meta's ability to attract top researchers with its GPU arsenal.
Google Cloud Expands Vertex AI Agent Builder with New Grounding Capabilities
• Vertex AI Agent Builder now includes tools for developing AI-based apps with enterprise-grade generative experiences
• Dynamic retrieval in Grounding with Google Search optimizes cost by intelligently using search results or model knowledge
• Newly announced high-fidelity mode in Grounded Generation API aims to minimize data hallucinations
• Third-party dataset integration feature set to launch in Q3, enhancing model accuracy with specialized data providers like Moody's
• Vector Search expanded to support hybrid search, improving accuracy in embeddings-based retrieval augmented generation
• Vertex AI introduces grounding capabilities with updated tools and features in response to enterprise demand for reliable information.
Apple Expands AI Features to Vision Pro Headsets, Rethinks Mixed Reality Interface
• Apple's AI expansion includes bringing Apple Intelligence features to the Vision Pro headsets, enhancing mixed reality experiences
• Vision Pro, a high-end device, so far targets a niche market and reflects Apple's premium strategy in wearable tech
• Integration of Apple Intelligence into Vision Pro is postponed, with efforts focused on adapting the interface for a mixed reality environment
• Store demo updates for Vision Pro will offer personal media viewing and improved comfort with a new Dual Loop headband
• New rumors suggest AirPods equipped with infrared cameras are set for mass production by 2026, aiming to enhance connectivity with Vision Pro.
⚖️ AI Ethics
Center for Investigative Reporting Sues Microsoft, OpenAI over Copyright Infringement Claims
• The Center for Investigative Reporting files a lawsuit against Microsoft and OpenAI for alleged copyright infringement.
• CIR claims the tech giants used its content without permission, undermining its business and depriving it of revenue.
• Legal battles intensify as multiple media outlets, including The New York Times, pursue similar actions against OpenAI and Microsoft.
• Some industry players like The Associated Press and News Corp have formed licensing agreements with OpenAI, showcasing a divide in strategy.
• OpenAI states cooperation with the news industry, aiming to enhance content visibility and drive traffic back to publishers.
• Despite ongoing lawsuits, many major publications have not yet taken legal action against OpenAI and Microsoft.
Settlement Restricts Detroit Police Use of Facial Recognition After Wrongful Arrest
• Detroit police banned from making arrests based solely on facial recognition results after a wrongful arrest lawsuit settlement.
• New policy prohibits Detroit police from using facial recognition-derived photo lineups for arrests without additional credible evidence.
• Detroit Police officers mandated to undergo training on the limitations and biases of facial recognition technology.
• Settlement includes auditing all Detroit Police cases using facial in recognition technology for arrest warrants since 2017.
• Robert Williams' wrongful arrest by facial recognition prompts reform, highlighting the technology’s high misidentification rate among people of color.
• Detroit Police settlement enforceable by court for four years, aims to safeguard against future misuse of facial recognition technology in arrests.
Meta Adjusts AI Photo Labeling Strategy After Photographer Feedback
• Meta replaces "Made with AI" tag with "AI info" on photos to better match user expectations of AI involvement
• The revamped labeling aims to clarify the use of AI tools in photo editing, not just AI-generated content
• Despite the tag change, Meta's underlying AI detection technology remains the same, utilizing C2PA and IPTC metadata standards
• Photos minimally altered using AI, such as object removal via Adobe’s Generative AI Fill, will feature the new "AI Info" tag
• The new label does not indicate the extent of AI editing involved, nor does it effectively flag fully AI-generated images
• Meta encourages collaboration across the industry to refine AI usage transparency, aligning tags with realistic editing scenarios.
EU Scrutiny Intensifies on Microsoft and Google AI Deals Amid Antitrust Concerns
• The European Commission is investigating Microsoft Teams for unfair distribution advantages and blocking competitor interactions.
• Microsoft's partnership with OpenAI under EU scrutiny for possible antitrust violations due to exclusivity clauses.
• Google's AI deal with Samsung faces EU regulatory examination for potential competitive impacts.
• EU Competition Commissioner Margrethe Vestager is probing Big Tech's AI strategies, citing potential blockages against smaller AI developers.
• Information requests sent concerning Microsoft and OpenAI's agreement to assess if exclusivity harms competition.
• The impact of Google’s generative AI tech pre-installed on Samsung devices is also a focus of EU regulators.
Amazon Probes Perplexity AI Over Alleged Protocol Violations and Content Scraping
• AWS investigates Perplexity AI following allegations of bypassing Robots Exclusion Protocol on web crawling activities. ;
• Wired’s tests indicate Perplexity AI's chatbot may be scraping content without proper attribution, closely paraphrasing from their articles. ;
• Perplexity AI CEO denies protocol violations, attributes questionable activities to third-party crawler usage. ;
• Company spokesperson asserts that PerplexityBot adheres to robots.txt files except in cases where users directly submit URLs in queries. ;
• Wired identifies a specific AWS-hosted virtual machine used by Perplexity AI for scraping content from major publishers like Condé Nast and The New York Times. ;
• AWS addresses the issue, citing a commitment to preventing abusive and illegal activities on its platform, follows up with an engagement with Perplexity AI.
🎓AI Academia
Study Reveals Different Brain Reactions to AI and Human Voices
• New research reveals brains react differently to human and AI-generated voices, despite struggles in distinguishing them
• Study finds human voices trigger stronger brain reactions related to memory and empathy than AI voices
• AI voices generate heightened brain responses in regions involved with error detection and attention regulation
• Participants often mistook neutral AI voices for humans, indicating a challenge in perceiving emotional nuances
• The research aims to inform future policies and ethical guidelines for the use of AI voice technology in various applications
• Potential applications of AI voices include therapies for mental health and providing voice solutions for voice-loss patients.
Persona Hub Drives Large-Scale Synthetic Data Creation Through Novel Methodology
• The novel persona-driven data synthesis methodology leverages a billion diverse personas to generate vast synthetic datasets
• Persona Hub's versatility evidences itself through the large-scale creation of varied data types, including game content and complex texts
• Traditional methods of synthetic data generation like instance-driven and key-point-driven approaches struggle to scale, unlike the new persona-driven method
• Initial release includes 200,000 personas along with synthetic examples: 50,000 math questions, 50,000 instructions, more to follow
• Persona-driven synthesis promises to propel significant advancements in how language models understand and generate human-like content
• By harnessing the diverse range of human perspectives, Persona Hub could radically transform the production and application of synthetic datasets in AI.
WaterBench: A Comparative Benchmark for Evaluating Watermarks in Large Language Models
• A new benchmarking system, WaterBench, evaluates watermarking on large language models to ensure fair comparisons across different methods
• This benchmark adjusts watermark hyper-parameters before jointly assessing generation and detection performance for unbiased results
• WaterBench introduces a diverse multi-task framework, featuring a five-category taxonomy across nine different tasks for comprehensive testing
• A novel evaluation metric using GPT4-Judge assesses watermarked models' performance, specifically their ability to follow instructions post-watermarking
• Recent tests on four open-source watermarks across two large language models reveal challenges in preserving the quality of generated content
• Full code and datasets for WaterBench are publicly accessible, providing resources for further research and development in LLM watermarking.
New Benchmark "VisEval" Enhances Understanding of LLMs in Natural Language to Visualization Tasks
• A new benchmark named VisEval aims to enhance the translation of natural language into visualizations by utilizing large language models (LLMs)
• VisEval integrates a sizable dataset consisting of 2,524 queries across 146 databases, all paired with accurately labeled visual outputs
• Advanced LLMs like GPT-4 and CodeLlama-7B were rigorously tested, uncovering key challenges and insights into their visualization capabilities
• Innovative evaluation measures within VisEval focus on validity, legality, and readability of visualizations, promoting credible analysis results
• The assessment framework meticulously scans for various potential errors using diverse checkers, ensuring reliability in visualization outputs
• Results from the new benchmark could guide future enhancements in the field of natural language-driven visualization technologies.
About us: We are dedicated to reducing Generative AI anxiety among tech enthusiasts by providing timely, well-structured, and concise updates on the latest developments in Generative AI through our AI-driven news platform, ABCP - Anybody Can Prompt!
