Today:

Midjourney debuts feature for generating consistent characters across multiple gen AI images

Midjourney, a popular AI image generator, introduces a new feature allowing consistent character creation across multiple images. Unlike previous models, which create new content with each prompt, Midjourney's "–cref" tag enables users to maintain character continuity across scenes. By pasting a URL of a previously generated character, users can ensure consistent appearance and clothing. 

The feature, although in early stages, marks a significant advancement for narrative mediums like film or graphic novels. Users can adjust the level of resemblance to the original character with the "--cw" tag. Despite some imperfections, the feature shows promise for professional use beyond mere ideation.

VENTUREBEAT

Building Meta’s GenAI Infrastructure

Meta announces the launch of two 24,576-GPU clusters, a significant investment in their AI infrastructure. These clusters, designed for Llama 3 training, utilize open compute principles and hardware such as Grand Teton and PyTorch. Meta aims to expand their infrastructure to include 350,000 NVIDIA H100 GPUs by 2024, supporting their vision of developing artificial general intelligence (AGI). 

The clusters feature advanced network fabrics and storage solutions, enabling high throughput and reliability for AI workloads. Meta emphasizes their commitment to open innovation in both hardware and software, fostering collaboration and transparency in AI development.

META

Cohere releases powerful ‘Command-R’ language model for enterprise use

Cohere has launched a game-changing language model called Command-R amid a fierce fundraising round, aiming to secure up to $1 billion. Command-R promises improved performance for enterprise tasks like retrieval augmented generation (RAG) and tool use, with longer context windows and more affordable pricing. Cohere targets the enterprise market, prioritizing trust, privacy, and data security. 

The release signals Cohere's determination to compete with AI giants like OpenAI. With over $500 million already raised, Cohere aims to prove its business model, focusing on real-world customer adoption and revenue growth. The release of Command-R underscores Cohere's position as a rising star in AI innovation and enterprise solutions.

VENTUREBEAT

Extropic’s ‘Lite’ Paper Unveils Vision For Next-Generation AI Tech, Superconducting Chips

Extropic, a secretive startup, unveiled a "Lite" paper outlining its groundbreaking approach to next-gen AI tech, featuring superconducting chips. The paper details the development of parameterized stochastic analog circuits, promising significant improvements in algorithm runtime and energy efficiency. Extropic's "accelerators" leverage programmable randomness, inspired by Brownian motion, to excel in complex computing tasks. 

Facing the challenge of escalating AI demands against physical computing limits, Extropic proposes a radical shift towards biological efficiency in computing. Their Energy-Based Models (EBMs) and superconducting chips offer scalability and energy efficiency, positioning Extropic at the forefront of AI innovation. Founded by Verdon and McCourt, Extropic secured $14.1 million in seed funding, aiming to revolutionize AI hardware and software.

QUANTUM INSIDER

Microsoft opens its Copilot GPT Builder to all Pro subscribers

Microsoft introduces Copilot GPT Builder, allowing users to customize AI chatbots. Available to Copilot Pro subscribers for $30 monthly, it simplifies creating task-specific bots without coding. Accessed via the Copilot web app, users can type instructions, with the tool handling backend programming. Features include adjusting bot behavior and using augmented generation for data retrieval. Users toggle web browsing and image generation. Examples include a Synonym Finder and OKR Assistant. 

Despite similarities with OpenAI's tool, Microsoft developed it independently. The move suggests Microsoft's strategy to reduce reliance on OpenAI. This advancement aligns with Microsoft's push for in-house AI models and partnerships, marking a significant step in AI accessibility.

VENTUREBEAT

Perplexity brings Yelp data to its chatbot

Perplexity, an AI chatbot, now taps into Yelp data for restaurant recommendations. CEO Aravind Srinivas emphasizes meeting user needs directly by integrating maps, reviews, and other details from Yelp. Unlike its competitors like Copilot or Gemini, Perplexity's approach includes mixing text with links, providing comprehensive responses. The deal with Yelp under its Fusion program offers reliable local search results. 

Other AI firms, like Google and OpenAI, are also securing data partnerships for real-time information. Perplexity plans further integrations, already collaborating with WolframAlpha for math solutions. This strategic move enhances user experience, making Perplexity a robust choice in the AI landscape.

THE VERGE

Amazon, Google Quietly Tamp Down Generative AI Expectations

Amazon and Google are quietly downplaying expectations surrounding generative artificial intelligence (AI), contrasting with the hype that drove stock market highs. While tech giants championed generative AI as revolutionary, insiders reveal a more cautious approach. 

Executives and salespeople at major cloud providers like Microsoft, Amazon Web Services, and Google acknowledge customer hesitance due to high costs, accuracy issues, and uncertain returns on investment. This behind-the-scenes sentiment signals a reality check for the industry, suggesting a more measured approach to AI adoption amid concerns over its practicality and value proposition.

THE INFORMATION

🧠RESEARCH

Stealing Part of a Production Language Model

This paper introduces the first successful method to snoop on complex, secretive AI models like ChatGPT and PaLM-2, revealing their inner workings with just regular user access. For less than 20 bucks, the researchers could figure out the entire setup of OpenAI's smaller brains, Ada and Babbage. They even nailed down the exact size of gpt-3.5-turbo's hidden layer and guessed it'd cost less than $2,000 to fully map it out. They wrap up with ideas on how to block such snooping and muse on what future sneak attacks might look like​.

Adding NVMe SSDs to Enable and Accelerate 100B Model Fine-tuning on a Single GPU

"Fuyou" is a cost-effective training framework allowing huge AI models, like a 100 billion-parameter model, to be fine-tuned on a single, low-end GPU in a standard server setup. By smartly combining SSD storage with CPU and GPU resources, Fuyou breaks through limitations that usually require pricier, high-end servers, making large model training accessible to researchers with limited budgets. It showcases remarkable performance improvements, like fine-tuning a 175 billion-parameter GPT-3 model on a consumer-grade GPU with high efficiency​.

An Image is Worth 1/2 Tokens After Layer 2: Plug-and-Play Inference Acceleration for Large Vision-Language Models

This study identifies and tackles inefficiency in how big AI models that understand both pictures and words pay attention to images. By introducing FastV, a clever tweak that prunes unneeded visual info after the early stages, it dramatically cuts down on the computing power needed (like slashing 45% of the workload for a specific model) without dropping performance across various tasks. This not only saves resources but also opens doors for these smart systems to work on everyday gadgets​.

VideoMamba: State Space Model for Efficient Video Understanding

"VideoMamba" innovates in video analysis by tackling local redundancy and global dependencies, surpassing current 3D CNNs and video transformers. It introduces a linear-complexity operation for effective long-term high-res video understanding. It shines in scalability without heavy pretraining, fine motion distinction, long-term video insight, and multi-modal robustness. This sets a new efficiency benchmark for video understanding, complete with open-source code and models.

V3D: Video Diffusion Models are Effective 3D Generators

"V3D" pushes the boundaries of 3D object creation by harnessing pre-trained video diffusion models. It introduces a method that improves 3D generation by ensuring geometric consistency across multiple views, making it possible to generate detailed 360-degree visuals of objects from a single image. This breakthrough enables the production of high-quality 3D models or scenes in under three minutes and supports novel scene-level view synthesis with minimal inputs. The method outperforms existing approaches in quality and consistency, offering a significant leap forward for 3D content creation​.

🛠️TOP TOOLS

TubeOnAI - Summarize & Listen any YouTube & Podcasts in 30 Seconds

Extend Image - Extend your images with generative AI

GoatChat - create and interact with AI-powered characters

Twig - reads, analyzes and writes responses to customer queries from help docs, private datasources, and past support tickets

Fixkey - specializes in AI-driven solutions for enhancing the functionality and security of smart home systems.

Maia - AI-driven relationship app designed for couples.

Swarms -  a toolkit for developing scalable, production-grade multi-agent applications.

Cubhouse - Simply record your voice, and AI creates a custom version for text-based chats, making it sound like you

📲SOCIAL MEDIA

🗞️MORE NEWS

Applied Intuition lands $6B valuation for AI-powered autonomous vehicle software

Applied Intuition, an AI-driven autonomous vehicle software firm, secured $250 million, now valued at $6 billion. Investors, including Lux Capital and Porsche, back its expansion into defense and agriculture sectors. The funding aids in scaling AI solutions for top automakers like GM and Toyota, signaling continued growth amid the AV industry's challenges. TECHCRUNCH

Oracle stock is surging because the AI craze is growing its cloud business

Oracle's stock surged by 12% after reporting strong demand for its cloud services, particularly in AI infrastructure. CEO Safra Catz highlighted the company's struggle to meet the soaring demand, with its cloud business growing by 53%. Oracle positions itself as a cost-effective alternative to competitors like Amazon and Microsoft. QUARTZ

OpenAI says there is no “agreement at all” with Elon Musk

OpenAI denies any agreement with Elon Musk and refutes his claim of contract violation in its legal response. The filing states Musk initially supported OpenAI's for-profit structure under his control but withdrew when it wasn't followed. OpenAI suggests Musk's interest stems from its technological success. THE VERGE

Italy to set up AI fund of 1 billion euros, PM says

Italy plans to establish a 1 billion euro investment fund to foster Artificial Intelligence (AI) projects, with potential for an additional 2 billion euros from private sources. Prime Minister Giorgia Meloni aims to promote an "Italian way to artificial intelligence" during the G7 presidency, focusing on job impact and regulatory safeguards. REUTERS

Google won’t let you use its Gemini AI to answer questions about an upcoming election in your country

Google has implemented restrictions on its AI chatbot, Gemini, preventing it from answering questions related to elections in countries where elections are taking place. This move comes amid concerns about the potential misuse of AI in elections and the dissemination of inaccurate or misleading information. The restrictions are already live in the U.S. and have begun rolling out in India and other major countries with upcoming elections. Queries about political parties, candidates, or politicians now return a preset message, signaling Google's caution regarding this sensitive topic. TECHCRUNCH

Mayo researchers invented a new class of AI to improve cancer research and treatments

Mayo Clinic innovates with hypothesis-driven AI, a new approach diverging from traditional models. This AI class integrates scientific knowledge into algorithms, enhancing cancer research and treatment development. By targeting specific hypotheses, leveraging existing data, and improving interpretability, this method holds promise for personalized medicine advancement. Challenges include expertise requirements and bias prevention. MAYO CLINIC

This Startup Says It Can Beat Deepmind’s Gamechanging Protein AI

Basecamp Research unveils BaseFold, an AI model built on DeepMind's AlphaFold2, boasting improved protein structure predictions. By training on a broader dataset, BaseFold claims a threefold enhancement in predicting protein-small molecule interactions crucial in drug discovery. The startup aims to diversify genomic data for better predictions and collaborates with Nvidia for optimization. FORBES