NVIDIA's New Text-to-3D AI Playground

Author

AI Breakfast
October 16, 2023

Good morning. It’s Monday, October 16th.

Did you know: If Siri was a person, she’d be in middle school now? This week in 2011, the iPhone 4S was released and debuted Apple’s original AI assistant.

In today’s email:

  • AI Advancements in Art and Design

  • AI Innovation in Science and Technology

  • Tech Industry Collaborations

  • Global AI Policy and Regulation

  • AI in Battery Development and Electric Vehicles

  • AI Safety and Security Concerns

  • AI Hardware and Semiconductor Developments

  • 5 New AI Tools

  • Latest AI Research Papers

You read. We listen. Let us know what you think of this edition by replying to this email, or DM us on Twitter.

Today’s trending AI news stories

AI Advancements in Art and Design

NVIDIA & Masterpiece Studio Launch New Text-to-3D AI Playground Masterpiece Studio, in collaboration with NVIDIA, introduces “Masterpiece X - Generate,” a text-to-3D AI playground designed to simplify 3D art creation. This tool operates in a web browser and doesn’t require prior knowledge.

Users can describe what they want to see, and the AI will generate it. While not suitable for high-fidelity or AAA game assets, it’s ideal for quick idea exploration and prototyping. The resulting assets are compatible with software like Blender, Unity, and Unreal Engine. It works on a credit system and offers 250 free credits upon account creation, with plans for future updates and expansions.

AI Innovation in Science and Technology

Polymathic AI aims to help scientists make faster and better discoveries Polymathic AI leverages foundation models for science, designed to learn from numerical data and simulations across various scientific domains, like physics and astrophysics. The project is a collaboration between institutions such as the Simons Foundation, New York University, University of Cambridge, and Lawrence Berkeley National Laboratory, uniting specialists in AI and science. It aims to democratize AI for scientific analysis and promote multidisciplinary knowledge transfer.

MIT’s New Generative AI Outperforms Diffusion Models in Image Generation MIT’s Computer Science and Artificial Intelligence Laboratory (CSAIL) introduces a novel generative AI model called PFGM++, combining principles from diffusion and Poisson Flow, which yields exceptional image generation capabilities. Inspired by physics, this model advances the generation of complex and realistic patterns in various domains. PFGM++ expands upon previous research to provide a balanced approach between robustness and ease of use. It has potential applications in various fields, from generating antibody sequences to creating realistic images and patterns.

AI Tech Industry Collaborations

TCS Seeks to Use Microsoft AI Partnership to Improve Margins Tata Consultancy Services Ltd, (TCS), Asia’s largest outsourcing firm, is leveraging its partnership with Microsoft Corp to develop AI-based software profit margins and drive growth. The collaboration with Azure OpenAI and the use of GitHub Copilot’s cloud-based AI tools enable TCS to offer services like fraud detection for financial clients and personalized customer service for retailers, a strategic move demonstrating a commitment to AI-driven solutions and driving higher margins.

Microsoft Will Pay You $15,000 If You Get Bing AI to Go Off the Rails Microsoft has launched a “bug bounty” program, offering substantial rewards to security researchers who discover vulnerabilities in its Bing AI products. The program aims to identify vulnerabilities, especially those related to “jailbreak” prompts that make Bing AI produce inappropriate responses. Eligible submissions must identify important or critical vulnerabilities and demonstrate them through video or in writing. This move comes after Microsoft faced issues with Bing AI producing undesirable content shortly after its release. The company is now outsourcing vulnerability research to address these issues.

Global AI Policy and Regulation

China sets stricter rules for training generative AI models by introducing security regulations aimed at companies offering generative AI services, focusing on restricting data sources used for AI model training. The proposed regulations recommend conducting security evaluations on content used to train publicly accessible generative AI models, blacklisting content that exceeds 5% as “unlawful and detrimental information,” including terrorism advocacy and actions undermining societal stability. The regulations also prohibit data subject to Chinese internet censorship from serving as training material for these AI models.

New US rules on AI chip exports aim to stop workarounds on China sales, says US official the move, expected to impact the delicate balance in U.S.-China relations, targets advanced AI chips such as the NVIDIA H-100. An anonymous official revealed to Reuters that the regulations will block some chips falling just under current parameters while demanding companies report shipments of others. The U.S. Department of Commerce declined to comment.

AI in Battery Development and Electric Vehicles

How generative AI is creeping into EV battery development Avionics, a pioneering startup, is leveraging the power of generative AI to advance electric vehicle battery development. With over ten billion commercially available molecules, Aionics employs AI tools to expedite the intricate process of discovering the ideal electrolyte materials for batteries. This innovation not only promises faster-charging batteries but also significantly enhanced energy density, ultimately benefitting various applications, including EVs, the grid, and electric aircraft. Avionics’ approach shows promise in addressing the persistent challenge of sifting through a multitude of candidate molecules and optimizing battery technology.

AI Safety and Security Concerns

AI safety guardrails easily thwarted, security study finds Computer scientists from Princeton University, Virginia Tech, IBM Research, and Stanford University extensively probed the capacity of various LLMs to withstand potential circumvention of their safety protocols. Notably, their investigation unveiled OpenAI’s GPT-3.5 Turbo’s susceptibility to manipulation through modest fine-tuning, making it prone to the execution of harmful directives. Consequently, the study underscores the critical necessity of fortified safety frameworks for commercial cloud-based models, advocating for comprehensive AI safety protocols and meticulous pre-deployment model licensing and testing.

AI Hardware and Semiconductor Developments

NVIDIA Blackwell B100 GPUs To Feature SK Hynix HBM3e Memory, Launches In Q2 2024 Due To Rise In AI Demand These GPUs will feature SK Hynix’s HBM3e memory, further solidifying NVIDIA’s position in the AI GPU market. Originally scheduled for Q4, the launch date was moved up to meet growing demand. SK Hynix is working on ensuring the quality and quantity of the memory supply for NVIDIA.

5 new AI-powered tools from around the web

Cal.ai is the first world’s open-source AI scheduling assistant. This innovative tool streamlines meeting arrangements through personalized email automation. Embracing community collaboration, Cal.ai envisions integrating open-source LLMs.

Xero.AI introduces Xero’s ARtificial Analyst (XARA), a user-friendly, no-code machine learning platform enabling seamless data exploration, visualization, and customized model creation. With its democratizing approach, XARA empowers users to effortlessly leverage the potential of AI, fostering a dynamic data-driven environment.

Gemelo AI leverages AI to craft interactive AI twins, catering to customer engagement and feedback collection. Its complimentary Voice Cloning feature enables limitless voice creation. With cutting-edge Voice Technology Vision, Knowledge, and Interaction services. Gemelo integrates NVIDIA Maxine GPU acceleration and cloud-native microservices from NVIDIA AI Enterprise on OCI AI Infrastructure.

Dorosi AI empowers users to craft individualized stories and illustrations using AI. By inputting character specifics like name, gender, age, and preferences, the AI tailors stories accordingly and adapts with each edit, ensuring personalized storytelling. The platform offers discounts, and free trials, and plans to introduce audiobooks, user-narrated stories, and physical story copies.

WeAreHiring.AI offers a free, easy-to-share bio-link solution for effortless job postings and social media sharing. With the convenience of a unique link, users can display up to 100 job openings, promoting easy recruitment across various platforms. The platform envisions simplified, efficient hiring for all founders.

arXiv is a free online library where scientists share their research papers before they are published. Here are the top AI papers for today.

The research paper introduces an innovative agent capable of autonomous task execution without explicit expert guidance. By incorporating staged planning and structured reflection, the model demonstrates superior performance even in the absence of specific task examples. Utilizing a compact screen representation, the agent enhances its adaptability and efficiency. The study emphasizes the significance of adaptable and self-reflective systems, providing valuable insights into improving agent performance across diverse tasks. This approach represents an advancement in autonomous learning and effective problem-solving within complex computer environments, showcasing the potential for future developments in the field.

The paper introduces LoftQ, a novel quantization framework designed for Large Language Models (LLMs). It addresses the issue of performance degradation when applying quantization and Low-Rank Adaptation (LoRA) fine-tuning to pre-trained models. In such cases, there’s often a gap between full fine-tuning and the quantization plus LoRA approach. LoftQ combines quantization and low-rank approximation to approximate the original high-precision pre-trained weights. This alignment significantly improves the initialization for LoRA fine-tuning and leads to better downstream task performance. The framework outperforms existing quantization methods, particularly in challenging low-bit quantization scenarios, such as 2-bit. LoftQ shows promise for efficient use of LLMs in resource-constrained environments.

The paper proposes a framework to improve the code generation capabilities of large language models. The authors introduce a method called CodeChain, which encourages the models to generate modularized code by reusing and revising representative sub-modules. They utilize chain-of-thought prompting to guide the model’s generation process and apply a chain of self-revisions by selecting representative sub-modules from previous iterations. Experimental results demonstrate significant improvements in the model’s performance, achieving state-of-the-art results on challenging code generation benchmarks. The study highlights the effectiveness of the proposed framework and provides insights through comprehensive ablation studies.

The paper proposes a joint language modeling approach for integrating speech units and text to bridge the gap between speech and text-based language models. The authors explore different tokenization methods and introduce evaluation metrics to assess the cross-modal ability of the model. Experimental results demonstrate that the integration of speech and text data improves the model’s performance and cross-modal transferability. The study emphasizes the benefits of incorporating both speech and text data in language modeling, encouraging further exploration in this direction. The findings suggest potential advancements in the development of comprehensive language models capable of handling diverse forms of human language expression.

The paper introduces the Consensus Game, a game-theoretic framework to improve language model (LM) decoding and generation. By formulating decoding as an imperfect-information sequential signaling game, the authors propose Equilibrium-Ranking, a training-free algorithm that reconciles generative and discriminative LM predictions. They leverage no-regret learning to compute approximate equilibria, achieving coherence and reasonableness in LM outputs. Demonstrated on diverse benchmarks, including question answering and dialogue tasks, Equilibrium-Ranking consistently outperforms existing methods. This approach enhances LM performance, highlighting the potential of game-theoretic tools in addressing challenges of truthfulness and consistency in language modeling.

Thank you for reading today’s edition.

Your feedback is valuable.


Respond to this email and tell us how you think we could add more value to this newsletter.