Today:

• xAI Introduces Grok-2 and Mini

• Google Launches Gemini Voice Mode

• OpenAI Introduces SWE-bench Verified Benchmark

• AI Scientist Revolutionizes Research Process

• AMD Completes Silo AI Acquisition

• Cortex Analyst Simplifies Data Insights

GROK 2 is revealed! The ACTUAL "SUSPICIOUS" AI model!

Grock 2, a new AI model released on August 13, 2024, has generated significant buzz due to its advanced capabilities. This model, available in two versions—Grock 2 and Grock 2 Mini—outperformed other leading models like GPT-4 and Claude in various benchmarks. 

Users are impressed with Grock 2's image generation, which seems to accept almost any prompt, even controversial ones. It has real-time access to information through X (formerly Twitter), making it highly versatile. Despite its impressive performance, Grock 3 is already anticipated for release later this year.

WATCH THE VIDEO ON YOUTUBE

Google’s AI surprise: Gemini Live speaks like a human, taking on ChatGPT Advanced Voice Mode

Google has launched Gemini Live, a new voice mode for its AI model that allows users to have natural, free-flowing conversations in real-time. Unlike similar features from competitors like OpenAI, Gemini Live is available on Android devices through a subscription, with an iOS version coming soon. Users can interact with the AI even when their device is locked or running other apps. 

Google’s integration of Gemini into Android enhances its ability to provide context-aware assistance. However, concerns remain about potential misuse of voice technologies, which Google has yet to address publicly.

VENTUREBEAT

Introducing SWE-bench Verified

OpenAI released SWE-bench Verified, a refined version of the SWE-bench benchmark for evaluating AI models' software engineering capabilities. This update addresses flaws in the original benchmark, such as ambiguous problem statements and overly strict unit tests that unfairly penalize valid solutions. SWE-bench Verified consists of 500 human-validated samples, ensuring more accurate assessments. 

Testing shows that models like GPT-4 perform significantly better on this new dataset, highlighting the importance of thorough evaluation methods. Despite improvements, OpenAI acknowledges limitations, including potential data contamination and the need for diverse evaluation methods.

OPENAI

The AI Scientist: Towards Fully Automated Open-Ended Scientific Discovery

Sakana AI, in collaboration with Oxford University and the University of British Columbia, has introduced "The AI Scientist," an AI-driven system designed to automate the entire research process, from idea generation to paper writing and peer review. This system aims to revolutionize scientific discovery by enabling foundation models, such as LLMs, to conduct independent research, produce scientific papers, and iteratively improve their output. 

While the AI Scientist shows promise in democratizing research and accelerating innovation, it also presents ethical challenges and limitations that need to be addressed as the technology evolves. 

SAKANA AI

Lisa Su formally welcomes Silo AI team to AMD after completing $665 million acquisition

AMD has completed its $665 million acquisition of Silo AI, one of Europe’s largest private AI labs. Silo AI, known for developing multilingual open-source language models and AI solutions for major enterprises, will bolster AMD's AI capabilities, particularly in software development. 

This acquisition aligns with AMD's strategy to provide end-to-end AI solutions, enhancing their position in the AI market dominated by Nvidia. AMD's recent investments in AI startups like Mipsology and Nod.ai further indicate their commitment to expanding their AI software capabilities, potentially driving more sales of their Instinct AI chips.

TOM'S HARDWARE

Snowflake launches Cortex Analyst, an agentic AI system for accurate data analytics

Snowflake has introduced Cortex Analyst, a new AI-driven system designed to simplify data analytics by allowing users to interact with their data through natural language queries. This service, now in public preview, uses multiple large language model (LLM) agents to convert plain English questions into accurate SQL queries, achieving a 90% accuracy rate—significantly higher than similar systems. 

Cortex Analyst is accessible via a REST API and can be integrated into enterprise applications. It aims to streamline analytics workflows, making it easier for businesses to extract insights without relying heavily on data analysts.

VENTUREBEAT

🧠RESEARCH

The AI Scientist: Towards Fully Automated Open-Ended Scientific Discovery

The paper introduces "The AI Scientist," a framework enabling AI to independently conduct scientific research. It automates the entire process, from generating ideas to writing and evaluating papers. The system demonstrates effectiveness in various machine learning fields, producing papers at minimal cost, signaling a new era in AI-driven scientific discovery.

Med42-v2: A Suite of Clinical LLMs

Med42-v2 presents clinical large language models (LLMs) built on Llama3 architecture and fine-tuned with specialized clinical data. Unlike generic models, these LLMs are designed to handle clinical queries effectively. They outperform Llama3 and GPT-4 in medical benchmarks, making them valuable tools for clinical settings. The models are publicly accessible.

Imagen 3

Imagen 3 is a new latent diffusion model that generates high-quality images from text prompts. It outperforms other state-of-the-art models. The paper details evaluations of image quality and addresses safety concerns, including methods to reduce potential harm from the model's use.

LongWriter: Unleashing 10,000+ Word Generation from Long Context LLMs

LongWriter addresses the limitation of long context LLMs in generating outputs beyond 2,000 words by introducing AgentWrite, a pipeline that breaks down ultra-long tasks into subtasks. This approach enables models to produce over 10,000 words of coherent text. The paper also presents LongWriter-6k, a dataset for training, and LongBench-Write, a benchmark for evaluating long-generation capabilities.

InfinityMATH: A Scalable Instruction Tuning Dataset in Programmatic Mathematical Reasoning

InfinityMATH introduces a scalable dataset for programmatic mathematical reasoning, reducing reliance on specific numbers to allow flexible scaling. Fine-tuning with this dataset on models like Llama2 and CodeLlama shows significant performance improvements on benchmarks, enhancing models' versatility across various mathematical problems. The dataset is available on Hugging Face.

🛠️TOP TOOLS

AI Search Grader - See how visible your brand is in AI-powered search engines.

Omnifact - Privacy-first generative AI platform made for the workplace

Twill - Accelerating the delivery of digital-led care

Trellis - AI engine transforms complex data sources - like financial documents, voice calls, and emails - into structured SQL-ready format for use by data and ops teams.

Kypso - Automate and scale your R&D teams with confidence, integrated into the apps you use for work.

📲SOCIAL MEDIA

🗞️MORE NEWS

MIT releases comprehensive database of AI risks

MIT researchers released the AI Risk Repository, a comprehensive database documenting over 700 AI risks. The repository consolidates 43 existing taxonomies to help organizations assess AI risks. It categorizes risks by cause and domain, serving as a practical tool for industries to tailor their AI risk mitigation strategies and guide future research. VENTUREBEAT

Universal Music and Meta Announce ‘Expanded Global Agreement’ for AI, Monetization and More

Universal Music Group (UMG) and Meta have expanded their global partnership to enhance creative and commercial opportunities for UMG artists across Meta’s platforms, including Facebook, Instagram, and WhatsApp. The agreement emphasizes fair compensation for artists, addresses unauthorized AI-generated content, and expands monetization options for UMG through short-form videos and other digital content. VARIETY

AI-driven technique can generate quality 3D assets from 2D images 'in seconds' — VFusion3D aims to transform VR, gaming, and digital design

Meta and Oxford University have developed VFusion3D, an AI-driven technique that can generate high-quality 3D models from a single 2D image in seconds. This innovation, which uses a video diffusion model fine-tuned with minimal 3D data, aims to transform the gaming, VR, and digital design industries by addressing the scarcity of 3D content. TOM'S HARDWARE

Good news — your Google Meet call will soon be able to take notes for you

Google Meet is introducing an AI-powered "Take notes for me" feature that automatically records meeting notes, aiming to boost productivity by allowing users to focus on discussions instead of note-taking. Rolling out soon, the feature is part of Google Workspace's AI Meetings and Messaging add-on and will be available to users with specific licenses. TECHRADAR

ANZ launches first-of-its-kind AI Immersion Centre in partnership with Microsoft

ANZ, in partnership with Microsoft, has launched the first AI Immersion Centre for the banking sector in Australia and New Zealand, located at ANZ's Melbourne headquarters. The center will train 3,000 leaders in AI adoption, focusing on generative AI to enhance productivity and innovation. ANZ's investment in AI tools, including Microsoft 365 Copilot and GitHub Copilot, aims to improve customer service and operational efficiency, with significant productivity gains already observed in software development. The initiative underscores ANZ's commitment to integrating AI responsibly and securely across its operations. MICROSOFT

Artists Score Major Win in Copyright Case Against AI Art Generators

Artists have achieved a significant legal victory against AI art generators in a groundbreaking copyright case. A federal judge has allowed key copyright infringement and trademark claims to proceed, ruling that Stability AI’s tool, Stable Diffusion, may have been built on copyrighted works and intentionally designed to facilitate infringement. This ruling could impact other companies that used the Stable Diffusion model. The case, involving artists like Karla Ortiz, challenges the use of billions of images scraped from the internet without compensation to train AI systems. The case now moves to discovery, where further details will be investigated. HOLLYWOOD REPORTER