Today:

• OpenAI, Google and Mistral New AI Models

• Meta Unveils Enhanced Chip Technology

• Mistral Released Mixtral 281GB AI Model 

• Udio AI-Powered Music Creation

• Google Photos Brings AI Editing Tools to Everyone

• Command R+ Boosts HuggingChat's AI Power

• Microsoft DALL-E for Military Operations

SHOCKING New AI Models! | All new GPT-4, Gemini, Imagen 2, Mistral and Command R+

Google DeepMind released Gemini 1.5 Pro on Google Cloud and Vertex AI, enhancing the platform's capabilities. They also introduced Imagin 2, creating 4C live images from prompts. 

GPT-4 Turbo with vision is now fully available, boasting improvements without specific details. Devin AI, powered by GPT 4 Turbo, assists developers, though some doubt its authenticity. Internet of Bugs exposes flaws in AI demos, like Devin's. 

WATCH THE VIDEO ON YOUTUBE

Introducing Our Next Generation Infrastructure for AI

Meta introduces its latest custom-made chips, aiming to enhance AI workloads for improved performance in ranking and recommendation models on platforms like Facebook and Instagram. This investment in AI infrastructure supports new generative AI products and advanced research. 

Last year's MTIA v1 paved the way, and the next-gen chip doubles compute and memory bandwidth, maintaining efficiency. It's a strategic move to align hardware with evolving AI demands. MTIA's deployment in data centers shows promising results, handling complex models efficiently. Meta's ongoing investment in custom silicon ensures scalability and compatibility with future hardware advancements, driving efficiency and performance.

META

AI startup Mistral launches a 281GB AI model to rival OpenAI, Meta, and Google

French AI startup Mistral has launched Mixtral 8x22B, a new 281GB AI model to compete with industry giants like OpenAI and Google. This large language model (LLM) promises superior performance, with a 65,000-token context window and 176 billion parameters. 

Available for free through an open-source format, Mixtral aims to challenge the dominance of existing models. Mistral, known for its open approach to AI, faces criticism for potential misuse and lack of control over its models. The release coincides with new offerings from OpenAI and Google, highlighting the competitive landscape in AI innovation.

ZDNET

Former Google Deepmind Researchers Assemble Luminaries Across Music And Tech To Launch Udio, A New AI-Powered App That Allows Anyone To Create Extraordinary Music In An Instant

Udio, a new AI-powered music app, has launched publicly, backed by a16z and notable investors like will.i.am. Developed by former Google DeepMind researchers, Udio aims to democratize music creation, allowing anyone to produce high-quality tracks quickly. Users simply describe the genre, lyrics, and inspirations, and Udio generates fully mastered songs in under 40 seconds. 

The app fosters creativity with a remix feature and enables sharing and collaboration within its community. With support from industry luminaries, Udio seeks to revolutionize music production and distribution while maintaining artist-friendly features and professional standards.

PR NEWSWIRE

AI editing tools are coming to all Google Photos users

Google Photos is expanding its AI-powered editing tools to all users, including Magic Eraser, Photo Unblur, and Portrait Light, starting May 15. These features help enhance photos without advanced editing skills, such as removing distractions, sharpening blurry shots, and adjusting portrait lighting. 

Additionally, Magic Editor, previously available on Pixel 8 and Pixel 8 Pro, will now be accessible to all Pixel devices, allowing for complex edits with generative AI. Users on Android and iOS will receive 10 Magic Editor saves per month, with additional access available through Pixel devices or a Premium Google One plan. These updates will roll out gradually to eligible devices, making photo editing more accessible to all Google Photos users.

GOOGLE

Cohere’s Command R+ now available on HuggingChat

Cohere's latest innovation, Command R+, is now integrated into Hugging Face's AI chatbot, HuggingChat. This move aims to enhance the capabilities of the open-source platform, offering improved performance and multilingual support. Command R+ stands out with its advanced retrieval-augmented generation (RAG) features, making it a valuable addition to HuggingChat's roster of language models. 

By allowing users to select specific models like Mixtral 8x7B and Gemma 1.1, HuggingChat distinguishes itself from competitors like OpenAI's ChatGPT. The enterprise focus of Command R+ may attract developers seeking AI solutions for business applications, further cementing Hugging Face's position in the market.

VENTUREBEAT

MICROSOFT PITCHED OPENAI’S DALL-E AS BATTLEFIELD TOOL FOR U.S. MILITARY

Microsoft proposed using OpenAI's DALL-E, an AI image generation tool, for the U.S. military to enhance software for military operations. This suggestion came after OpenAI relaxed its stance on military collaborations, despite its initial mission to develop AI for the benefit of humanity. The presentation to the Department of Defense explained potential uses, including training AI for battlefield awareness. 

Critics argue that such military applications of AI could indirectly lead to civilian harm. OpenAI maintains it hasn't sold any tools for defense purposes, emphasizing policies against AI's use in creating weapons or harming individuals. The debate highlights ethical considerations around AI's military applications, with some experts questioning the practicality and morality of using synthetic data for combat training.

THE INTERCEPT

🧠RESEARCH

OmniFusion Technical Report

The OmniFusion model, built upon multimodal architectures, boosts AI capabilities by integrating visual modality with large language models (LLM). It outperforms existing solutions like VizWiz and TextVQA across various tasks. With detailed performance evaluation and open-source availability, OmniFusion presents a significant advancement in AI-driven text and visual data coupling. 

LLM2Vec: Large Language Models Are Secretly Powerful Text Encoders

LLM2Vec transforms decoder-only LLMs into powerful text encoders through bidirectional attention, masked token prediction, and unsupervised contrastive learning. Outperforming encoder-only models, it achieves new unsupervised state-of-the-art performance on text embedding tasks. Combining with supervised contrastive learning, it excels among publicly available data-trained models on the Massive Text Embeddings Benchmark (MTEB). No costly adaptations or synthetic data are required.

InternLM-XComposer2-4KHD: A Pioneering Large Vision-Language Model Handling Resolutions from 336 Pixels to 4K HD

InternLM-XComposer2-4KHD pioneers large vision-language models, pushing resolution capabilities beyond 4K HD. It supports diverse resolutions from 336 pixels to 4K, broadening its applicability. Dynamic resolution with automatic patch configuration enhances fine-grained visual understanding. Performance matches or surpasses GPT-4V and Gemini Pro across multiple benchmarks.

Eagle and Finch: RWKV with Matrix-Valued States and Dynamic Recurrence

Eagle and Finch, advancements of the RWKV architecture, introduce multi-headed matrix-valued states and dynamic recurrence mechanisms, enhancing model expressivity while preserving inference efficiency. With new training data and fast multilingual tokenization, they achieve competitive performance across various benchmarks. All models are available on HuggingFace under the Apache 2.0 license. 

Hash3D: Training-free Acceleration for 3D Generation

Hash3D, a novel approach to accelerate 3D generation without training, capitalizes on feature-map redundancy in nearby camera positions and time-steps. By hashing and reusing these maps, Hash3D slashes redundant calculations, boosting efficiency 1.3 to 4 times. Integration with 3D Gaussian splatting further accelerates model creation, reducing processing times significantly.

🛠️TOP TOOLS

Figjam AI - Hand-picked prompts to help you get started planning and jamming

Datrics - Ask, analyze, and act. Get your finance, sales, and operations questions answered faster.

Apollo - Power your go-to-market with one platform. Fueled by the most accurate data on 275M contacts and 65M accounts.

Claude-Investor - investment analysis agent that utilizes the Claude 3 Opus and Haiku models to provide comprehensive insights and recommendations for stocks in a given industry.

Julius AI - Solve math equations that LLMs traditionally struggle with.

📲SOCIAL MEDIA

🗞️MORE NEWS

Snowflake to sync data cloud with Coda after fresh investment

Snowflake teams up with Coda to streamline data access for enterprises. Snowflake invests in Coda and launches Snowflake Pack, enabling Coda users to integrate Snowflake data into projects. The move aims to simplify data analysis for non-tech users, offering real-time insights within the productivity platform. Future plans include AI-driven solutions. VENTUREBEAT

iOS 18 May Feature All-New 'Safari Browsing Assistant'

iOS 18 is rumored to introduce a new Safari browsing assistant, as indicated by backend code found on Apple's servers. Details are scarce, but it's speculated that the assistant may utilize iCloud Private Relay for privacy-focused data transmission, possibly requiring an iCloud+ subscription. This browsing assistant could be part of iOS 18's rumored generative AI features, joining existing AI-powered tools in other web browsers. MACRUMORS

Google Partners With European Conglomerate Bayer For New AI-Powered Solutions For Radiologists

Bayer and Google Cloud team up to enhance radiology with AI solutions. Leveraging Google Cloud's technology, they aim to streamline medical imaging analysis for radiologists, saving time and improving patient care. The collaboration targets scaling AI-powered healthcare applications and addressing the increasing workload of healthcare professionals handling medical images. YAHOO!

Amazon invests $25 million in a 10-year research collaboration to advance AI

Amazon invests $25 million in a decade-long collaboration with University of Washington, University of Tsukuba, and NVIDIA to advance AI research and workforce development. The partnership aims to fund AI research, support post-docs and PhD fellows, engage undergraduates in summer research, and host an entrepreneurship bootcamp. AMAZON

Sama launches AI safety-centered ‘red teaming solution’ for gen AI and LLMs

Sama introduces Sama Red Team, a safety-centered solution for generative AI and large language models (LLMs). Through red teaming techniques, it identifies and exposes vulnerabilities in AI models, addressing concerns of bias, privacy, and fairness. The service tests for compliance, public safety, privacy, and fairness, offering enterprise-level solutions. VENTUREBEAT

Meta ‘discussed buying publisher Simon & Schuster to train AI’

Meta, formerly known as Facebook, reportedly discussed buying Simon & Schuster to acquire books for training its AI tools. Internal meetings revealed deliberations on purchasing the publishing house for data to train AI models, despite ethical and legal concerns about using copyrighted material without permission. The potential purchase raised questions about Meta's intentions and its impact on the publishing industry. THE GUARDIAN

London AI firm V7 expands from image data labeling into workplace automation

London-based AI company V7 Labs, known for data labeling in computer vision, ventures into business process automation with V7 Go. The platform utilizes advanced AI models to automate office tasks. This move marks V7's expansion into the competitive field of workplace automation. FORTUNE

AI makes retinal imaging 100 times faster, compared to manual method

The National Institutes of Health applies artificial intelligence to retinal imaging, accelerating the process by 100 times and enhancing image contrast 3.5-fold. AI improves the evaluation of retinal diseases like macular degeneration. The technique, utilizing adaptive optics and deep learning, provides a clearer view of cellular structures, aiding early disease detection. MEDICAL XPRESS

A faster, better way to prevent an AI chatbot from giving toxic responses

Researchers at MIT and the MIT-IBM Watson AI Lab develop a machine-learning model to enhance safeguards on large language models, like AI chatbots, by automatically generating diverse prompts. This approach, utilizing curiosity-driven exploration, outperforms human testers and other automated methods, improving the model's safety and reliability. MIT