• AI Breakfast
  • Posts
  • OpenAI pauses Browsing feature to stop paywall bypass

OpenAI pauses Browsing feature to stop paywall bypass

Author

AI Breakfast
July 05, 2023

Good morning. It’s Wednesday, July 5th.

Did you know: AI has revealed what Roman emperor Marcus Aurelius may have looked like based on stable diffusion models of marble busts?

In today’s email:

  • Google to scrape public data for AI training

  • Nvidia acquires OmniML for device-based AI

  • MoonlanderAI creates 3D games from text

  • Midjourney introduces "panning" feature

  • U.S. to limit China's cloud computing access

  • AI and CRISPR combined for gene control

  • Human beats AI in Go game

  • AWS VP downplays AI threats

  • Greyparrot uses AI for waste management

  • ChatGPT pauses Browsing feature to stop paywall bypass

  • Future predicted without human programmers

  • App for AI-based skin diagnosis developed

  • Celestial AI raises $100M for tech expansion

  • 7 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.

Special note:

Hi, Jeff here, founder of AI Breakfast. I’m visiting Canada this month and would love to meet a few members of the community! Are you in Toronto and interested in meeting up for a coffee and a chat this month?

Today’s trending AI news stories

Google Says It'll Scrape Everything You Post Online for AI: Google has updated its privacy policy, stating that it can scrape and utilize public data from the entire internet to train its AI models. The new policy expands on the previous language regarding data usage for “language models” and now includes “AI models.” This raises privacy concerns as users’ public posts may be analyzed and repurposed by Google AI’s systems, such as Google Translate, Bard, and Cloud AI capabilities. The legality of this data collection method and its potential consequences are still subjects of debate and court cases.

Nvidia Acquired AI Startup That Shrinks Machine-Learning Models: Nvidia quietly acquired OmniML, an AI startup specializing in shrinking machine-learning models to run on devices instead of the cloud. The acquisition suggests Nvidia’s interest in enhancing its AI chips for autonomous vehicles, industrial robots, and drones. It could also contribute to optimizing AI software for chatbots to operate locally rather than relying on data centers.

Moonlander launches AI-based platform for 3D game development: MoonlanderAI’s platform includes a “text-2-game” feature that enables the creation of 3D games based on text descriptions. The company aims to democratize high-fidelity 3D game development and plans to integrate UnityAI’s tools into its ecosystem to enhance its capabilities and facilitate collaborative ecosystem building. Check out the website here.

MidJourney Adds "Panning" Feature: Midjourney has announced the addition of a new “panning” feature on Discord. Users can now pan their images up, down, left, and right, expanding the image in the desired direction. The feature is accompanied by some limitations, such as no support for variations on panned images and a decrease in resolution when zooming out. Our friend Matt Wolfe at FutureTools.io wrote a great breakdown of the feature.

U.S. Looks to Restrict China’s Access to Cloud Computing to Protect Advanced Technology: The move will close a loophole that allowed Chinese AI companies to bypass export controls by using cloud services. U.S. cloud providers like Amazon and Microsoft would need government permission to offer cloud-computing services with advanced AI chips to Chinese customers. The action is part of the broader Chips Act, a $53 billion effort to reduce U.S. reliance on foreign-made semiconductors, especially those used by the Pentagon.

AI and CRISPR precisely control gene expression: RNA-based predictive models developed: Researchers from New York University and Columbia Engineering have combined AI with CRISPR screens to control the expression of human genes through RNA targeting. The AI system, known as TIGER (Targeted Inhibition of Gene Expression via guide RNA design), accurately predicts on- and off-target activity of CRISPR tools that target RNA. TIGER enables precise gene controls, potentially leading to the development of new CRISPR-based therapies including RNA editing and drug screening.

A human just defeated an AI in Go. Here’s why that matters: A human amateur defeated the powerful AI, KataGo, in the complex game of Go. The victory came through exploiting the AI’s vulnerabilities rather than superior play. This underscores the importance of understanding AI robustness, as exceptional performance in controlled environments may not guarantee success in real-world situations.

AWS exec downplays existential threat of AI, calls it a 'mathematical parlor trick': Matt Wood, VP of product at AWS, downplays AI’s existential threat, calling it a “mathematical parlor trick” that helps humans make better decisions. He highlights generative AI’s potential for businesses, with over 100,000 AWS customers already using AI/ML services. Mr. Wood is wrong about the threat of AI, but we forgive him.

The AI trained to recognize waste for recycling: UK start-up Greyparrot has developed an AI system to analyze waste processing and recycling facilities. By placing cameras above conveyor belts in around 50 sites in Europe, the AI software tracks 32 billion waste objects annually. The data collected not only helps waste managers become more efficient but also allows regulators to understand material usage, influencing packaging design.

ChatGPT pauses Bing integration to stop people from bypassing paywalls: OpenAI temporarily disabled the “Browse with Bing” beta feature in its ChatGPT platform after users found a way to bypass paywalls using the chatbot. The feature, introduced exclusively to ChatGPT Plus subscribers, aimed to enhance the search experience. The company disabled the feature after we wrote about the loophole last week. Sorry!

Stability AI CEO: There Will Be No (Human) Programmers in Five Years: Emad Mostaque, CEO of Stability AI, products that there will be no human programmers in five years as AI continues to advance. Mostaque cites data from GitHub, showing that 41% of all code is currently AI-generated. Mostaque envisions AI models residing on mobile phones without internet connectivity, revolutionizing communication.

Toronto engineer develops app that can diagnose skin cancer for free using AI analysis: The app aims to provide faster access to dermatological consultations, as Canadians often wait 90 days for an initial appointment. The app’s deep learning network analyzes live video, making it inclusive for all skin types and colors. The app’s creator previously worked on prediction models for tech giants Tesla and General Motors.

Celestial AI raises $100M to expand Photonic Fabric technology platform: Celestial AI has secured $100 million in a successful Series B funding round led by IAG Capital Partners, Koch Disruptive Technologies, and Temasek’s Xora Innovation fund. The funding will be used to further develop the company’s Photonic Fabric technology platform, which offers advanced optical interconnect technology for AI computing and memory clusters. Photonic Fabric aims to address the limitations of electrical interconnects, such as restricted bandwidth and high power consumption by providing increased bandwidth and low-latency data transmission.

🎧 Did you know AI Breakfast has a podcast read by a human? Join AI Breakfast team member Luke (an actual AI researcher!) as he breaks down the week’s AI news, tools, and research: Listen here

5 new AI-powered tools from around the web

Myreader AI: An innovative tool that uses AI to read books aloud, allowing users to chat with books, navigate specific pages, and upload large amounts of content for easy reference.

VribbleAI: A speech-to-text platform that converts recordings into summaries, offering transcription search, Telegram integration, and three pricing options for different needs.

EducUp Studio: A platform that allows users to create interactive and gamified courses quickly, connecting them with a community of over 1.5 million learners and providing opportunities for monetization and audience expansion.

Body Scan by Zing: An AI-driven fitness tool that measures body fat and lean mass using full-body selfies, providing personalized fitness programs, micronutrient plans, accuracy comparable to DEXA scans, privacy at home, video workouts, and progress tracking.

Ween.ai: An AI platform that automates the analysis of customer interviews, extracting important insights to aid product decision-making and align research with business goals.

Splash Pro: An AI-driven platform that generates music based on user input, allowing users to specify themes, styles, or moods and create high-quality music without extensive musical knowledge.

TalkVisions: A mobile application that provides real-time closed captioning translations for videos, facilitating communication and learning by overcoming language barriers with its intuitive design and powerful features.

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

Researchers have been studying deep learning models called Transformers, which are used in tasks like language processing. They discovered a problem called “rank collapse” where the model’s ability to distinguish different inputs becomes poor, making it hard to train effectively. To address this, they proposed a modifier Transformer called the “shaped Transformer.” It includes changes to the attention mechanism, which helps the model focus on relevant information. The modifications stabilize the model’s behavior and improve its performance.

Researchers propose a real-time monocular full-body motion capture method that accurately captures human movements in the real world. They use a new data generation strategy, creating a proxy dataset of 2D skeleton sequences and 3D rotational motions. This dataset enables training a learning-based network for accurate full-body motion capture while addressing generalization issues. To improve predictions, they introduce a contact-aware neural motion descent module to handle foot-ground contact and motion misalignment. The system achieves real-time performance with plausible foot-ground contact in the real world.

The paper presents the challenges and solutions for deploying large diffusion models, specifically Stable Diffusion, on mobile devices with limited computational and memory resources. The authors introduce Mobile State Diffusion implemented using the TensorFlow Lite framework, achieving an inference latency of less than 7 seconds for generating 512×512 image on Android devices with mobile GPUs.

The research paper investigates the personality traits exhibited in the text generated by large language models, with a focus on their validity and reliability. The study, conducted by researchers from Google DeepMInd, presents a comprehensive methodology using psychometric tests to quantify and shape LLM-simulated personalities. The findings demonstrate the effectiveness of certain LLM configurations in simulating reliable and valid personalities. The paper also discusses the ethical implications of manipulating LLM personalities.

DISCO (Disentangled Control for Referring Human Dance Generation in Real World) addresses the challenge of generating human dance content that accurately represents real-world scenarios. The paper introduces the problem of Referring Human Dance Generation, which focuses on generating dance sequences with faithfulness, generalizability, and compositionally. DISCO proposes a novel approach that includes a model architecture with disentangled control and an effective human attribute pre-training strategy.

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.

Attending Ai4 this year? We will be!