• AI Breakfast
  • Posts
  • Stanford Cracks the AI Code: The Groundbreaking Law of Equi-Separation

Stanford Cracks the AI Code: The Groundbreaking Law of Equi-Separation

Author

AI Breakfast
September 04, 2023

Good morning. It’s Monday, September 4th.

Did you know: On this day in 1995, eBay was founded as AuctionWeb.

In today’s email:

  • Film and Entertainment

  • Supercomputing and Hardware

  • AI Research and Innovation

  • Supply Chain and Commerce

  • Security and Defense

  • Social Media and Data Monetization

  • Digital Artistry and Music

  • AI Policy and Regulation

  • 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

Venice Film Festival 2023 Is Cinematic AI’s Big Debut At the 90th Venice Film Festival, Hollywood stars like Tom Hanks, Robert Zemeckis, Joe Russo, and Darren Aronofsky are endorsing AI’s cinematic potential, countering Hollywood’s recent hesitancy towards AI. Hanks envisions an AI version of himself acting after his death, while Russo foresees a future where people request AI-generated movies starring themselves and iconic figures. Director Harmony Korine premiered an AI-assisted film called Aggro Dr1ft," suggesting AI is an exciting creative tool. Three AI film studios also launched at the festival focusing on the convergence of AI, game engines, XR, and cinema to create innovative content.

Russian 400 AI PetaFLOPS Supercomputer for AI Comes Online: Lomonosov Moscow State University (MSU) has unveiled its MSU0270 supercomputer with a peak computation power of 400 AI PetaFLOPS. The machine will be utilized for various AI and high-performance computing (HPC) applications, including training large AI models. While the exact graphics accelerators used are undisclosed, given past affiliations with Nvidia, it’s possible they are using GPUs from companies like AMD or Nvidia. The supercomputer will contribute to AI research, the development of AI tools, and the training of AI specialists.

Stanford Cracks the AI Code: The Groundbreaking Law of Equi-Separation: Stanford University’s groundbreaking “Law of EquiSeparation” is set to revolutionize the field of AI and deep learning. The empirical law, developed by Hangfeng He and Weijie J. Su, promises to unveil the inner workings of neural networks, enabling precise ethical governance, model optimization, and improved robustness. The implications are far-reaching, spanning diverse sectors, including medicine, finance, and autonomous vehicles. The law could pave the way for standardized AI practices, trust-building in AI applications, and educational advancements, while challenging our understanding of intelligence and learning in an AI-driven world.

Global Trade Sector Taps AI to Smooth Supply Chain Issues: Artificial Intelligence is being increasingly used in the global trade sector to address supply chain issues worth $32 trillion. Generative AI and large-language models are assisting governments and businesses in identifying problems and opportunities in international trade. This shift towards AI in supply chains promises more efficient cross-border commerce, with companies like ImportGenius and Nestle SA using machine learning tools to simplify trade data analysis and enhance product quality control. AI-driven solutions may lead to fully resilient and sustainable supply chains, automating decision-making and mitigating risk exposure.

CLIPN: New computer vision method teaches AI to say 'no': Researchers at the Hong Kong University of Science and Technology have developed CLIPN, a new technique for improving out-of-distribution (OOD) detection in computer vision models. CLIPN teaches the CLIP model to reject unknown inputs by adding learnable “no” prompts and “no” text encoders, allowing the model to recognize when an image falls outside its known classes. In experiments, CLIPN improved OOD detection in nine reference datasets by up to nearly 12 percent compared to existing methods. However, its suitability for specialized datasets and applications like medical imaging remains unclear.

Hyderabad Firm Unveils India's First AI-Powered Anti-Drone System: A Hyderabad-based robotics firm, Grene Robotics, has unveiled India’s first AI-powered anti-drone system named Indrajaal. This advanced system can protect vital installations, including nuclear facilities and oil rigs, as well as entire cities from multiple drones. Indrajaal utilizes AI and offers 360-degree protection, detecting, identifying, classifying, tracking, and neutralizing threats in real time. The system is designed to defend against various types of autonomous drones over an area of 4,000 square kilometers and aims to address the increasing drone-related security challenges faced by India.

Elon Musk's xAI could train AI models on your Twitter data: Elon Musk’s recent move to update Twitter’s terms of service, allowing for the use of user data in AI training, could significantly benefit his AI startup, xAI. While this change ensures that only publicly available information will be used, it could provide valuable data for xAI’s AI model development. Musk’s strategic focus on data monetization is evident as he aims to improve data quality on the platform and charge for API access, creating incentives for content creators to stay within the platform.

New AI technology gives robot recognition skills a big lift: Researchers at the University of Texas at Dallas have developed a new AI system that enhances robots’ object recognition skills through repeated interactions. Unlike previous methods that relied on a single push or grasp by the robot to identify an object, this new approach has the robot push each item 15 to 20 times, allowing it to capture more detailed images with its RGB-D camera. The system reduces potential errors and contributes to robots’ ability to recognize and remember objects, a crucial function for performing tasks in real-world environments.

Model who never ages: Noonoouri becomes the first digital artist to be signed by Warner Music: Noonoouri, the digital avatar and model, has made history as the first digital artist signed by Warner Music. Despite her debut occurring five years ago, Noonoouri remains 18 years old, as she is not a real person but a CGI creation. She has amassed over 400,000 Instagram followers and has been featured in campaigns for top fashion brands. Now, Warner Music Central Europe has given her a singing voice, thanks to AI-assisted technology, releasing her debut single, “Dominoes,” featuring German DJ Alle Farven.

Google DeepMind co-founder argues US should set AI global standards: Mustafa Suleyman, co-founder of Google DeepMind and CEO of Inflection AI, advocates for the US to establish global AI standards. He proposes that the US leverage its chip leadership to enforce minimum global AI standards and suggests that tech companies, particularly those using Nvidia chips, commit to these standards. This call comes as the Biden administration expands restrictions on Nvidia and Advanced Micro Devices AI chips to regions beyond China. Suleyman’s recommendation aligns with the commitments made by leading AI firms to the White House earlier in the year, emphasizing the importance of regulating AI technologies.

5 new AI-powered tools from around the web’

Match AI by color.io is a free web app for creative color grading, utilizing AI to extract and apply colors from reference images. Users can fine-tune color grades and even create custom 3D LUTs with the Pro version for consistent looks in various media applications.

RiskAssessmentAI is an AI-powered platform that streamlines vendor risk and cybersecurity questionnaires. It automates the process based on internal policies and previous assessments, reducing the time and effort required. Ideal for B2B companies, it aims to simplify risk assessments and seeks user feedback for further improvements.

Seona V2, an intelligent AI-powered SEO tool by Style AI, is here with significant improvements. The new version offers enhanced features like improved onboarding, a redesigned user interface, a comprehensive dashboard to monitor site changes and rankings, and an upgraded blog writing system. This SEO tool aims to simplify SEO for businesses and incorporates user feedback for continuous enhancement.

Marketsy.ai simplifies e-commerce with a single prompt, instantly generating fully functional online stores or marketplaces. Say goodbye to the complexities of website building and let AI create your e-store effortlessly.

Connect Google Sheets to the OpenAI API seamlessly without relying on third-party tools like Zapier or Make. The script simplifies the process, allowing instant access to responses from GPT-4 and GPT 3.5 models.

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 Large Content and Behavior Models (LCMBs) developed by Adobe and IIIT-Delhi bridge the gap in communication theory by addressing the effectiveness problem. These models, including LCBM-13B, introduce behavior tokens alongside content tokens in the language space, enabling them to predict and optimize receiver behavior. LCBMs excel in content understanding, behavior simulation, content simulation, and behavior understanding, with applications in content recommendation, A/B testing, and more. The research introduces the Content Behavior Corpus (CBC) dataset and benchmarks to prompt further exploration of large content and behavior models. LCBMs showcase promise in transforming communication by considering both content and behavior.

VideoGen is a groundbreaking text-to-video generation method introduced by researchers from Baidu, Inc. This approach uses reference-guided latent diffusion to create high-definition videos with impressive frame quality and temporal consistency. It utilizes a pre-trained text-to-image generation model to produce a high-quality reference image from a text prompt, guiding the video generation process. VideoGen employs an efficient cascaded latent diffusion module conditioned on both the reference image and the text prompt, followed by temporal upsampling to enhance temporal resolution. During training, it uses the first frame of a ground-truth video as the reference image. VideoGen sets a new state-of-the-art in text-to-video generation, achieving remarkable qualitative and quantitative results.

This paper by Google Research explores the feasibility of using AI-generated feedback as an alternative to human feedback for reinforcement learning. They compare Reinforcement Learning from Human Feedback with Reinforcement Learning from AI Feedback and find that both result in similar improvements. For summarization tasks, human evaluators prefer both RLAIF and RLHF over a supervised fine-tuned model in approximately 70% of cases. The study suggests that RLAIF can achieve human-level performance, addressing scalability challenges in RLHF due to the need for human annotation. They also investigate techniques to enhance AI-generated preferences for better alignment with human preferences.

The research paper introduces CityDreamer, a novel generative model designed for creating realistic 3D city layouts. Unlike previous methods that struggled with the complexity of 3D city generation, CityDreamer separates building generation from background objects like roads and green areas. It also utilizes two datasets, OSM and Google Earth, to enhance realism. CityDreamer is shown to outperform existing methods in generating diverse and lifelike 3D cities. This research is significant in the context of the metaverse and urban planning, offering a practical approach to generating complex 3D urban environments.

The paper introduces a novel benchmark named FACET (FAirness in Computer Vision EvaluaTion), designed for evaluating computer vision models in terms of fairness across demographic attributes. The benchmark comprises 32k images covering common vision tasks like image classification, object detection, and segmentation. Expert reviewers manually annotate attributes such as perceived skin tone, hair type, and fine-grained person-related classes for images. FACET is used to assess state-of-the-art vision models, revealing performance disparities and biases across demographic attributes and intersections. The benchmark aims to contribute to the development of fairer and more robust vision models and is publicly available.

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.