• AI Emergence
  • Posts
  • Meta, OpenAI add labels to AI-generated images

Meta, OpenAI add labels to AI-generated images

Along with: Google taking lead and Apple showing its muscles!

Hey there, 

With the election season intensifying, OpenAI and Meta are taking steps to fight Deepfakes. OpenAI has introduced watermarks to its Dall-E 3 images, while Meta is implementing labels on AI-generated content throughout its platforms.

Google began February with a surge of new features, such as adding image generation capabilities to Bard, making Gemini Pro available in more languages and locations, and significantly updating Google Maps. Rumors are going around that Gemini Ultra lands this week! I look forward to seeing how the most powerful LLM from Google Stable comes out.

Let’s dive into the most significant developments this week!

What would be the format? Every week, we will break the newsletter into the following sections:

  • The Input - All about recent developments in AI

  • The Tools - Interesting finds and launches

  • The Algorithm - Resources for learning

  • The Output - Our reflection 

  • Question to ponder before we meet next!

Table of Contents

OpenAI is updating its DALL-E 3 image generator to add watermarks to images, following the C2PA standards. Starting with images from the ChatGPT website and the DALL-E 3 API, and extending to mobile users by February 12th, each image will feature a visible CR symbol in the top left corner and an invisible metadata watermark.  (source)

Meta Implements AI Content Labels and Penalties Amidst the Election Season

Meta is set to label AI-generated photos on Facebook, Instagram, and Threads in the next few months, especially as election season approaches globally. The platform will also penalize users who fail to disclose the AI origin of realistic videos or audio. Besides adding an “Imagined with AI” watermark to its AI-generated images, Meta will extend this practice to photos created using tools from major tech entities like Google, OpenAI, and others. (source)

In the latest Google updates, there's plenty to unpack:

  • Google Bard to be renamed Gemini?

    According to a leaked changelog dated 7 Feb by the developer Dylan Roussel, it claimed that Google is rebranding its chatbot Bard to Gemini altogether. It also mentions that Google is launching its most advanced model Ultra 1.0 to power Gemini Advanced which will be far more powerful to solve complex problems and directly compete with ChatGPT 4. Google will also launch the Gemini app that can be integrated with other Google apps to super-boost your productivity. (source)

  • Gemini Pro now available in multiple languages

    Google launched the Gemini Pro for English in December and this week it rolled out the Gemini Pro for over 40 languages and more than 230 countries and territories. Additionally, the option to double-check responses is also unlocked for these languages and places. Finally, Google is bringing the ability to generate images into Bard powered by the updated image gen 2 model. (source)

  • GenAI makeover to Google Maps

    Google Maps is transforming how we explore locally with Generative AI, offering tailored recommendations that are more intuitive. In a U.S. early access experiment, this feature utilizes generative AI to match users with spots that fit their unique tastes, like "places to visit on a rainy day in Bengaluru" by sifting through Maps' extensive data. (source)

Apple's team has just unveiled an exciting tool called MGIE, short for MLLM-Guided Image Editing. Imagine tweaking your photos - cropping, resizing, flipping, or adding filters, all by simply telling the tool what you want, no editing software needed! Whether it's basic adjustments or more intricate edits like changing an object's shape or brightness in your picture, MGIE's got you covered. (source)

Meta is gearing up to launch its own custom AI chip, "Artemis," in its data centers this year. This strategic move is designed to bolster Meta's AI projects and lessen dependence on Nvidia chips, leading to cost savings in AI operations. 
Mark Zuckerberg aims to combine 350,000 Nvidia processors with Meta's custom chip, potentially reducing annual energy expenses significantly. Despite setbacks with the initial version, the new chip prioritizes inference tasks, offering improved efficiency for Meta's recommendation models compared to power-hungry Nvidia processors. (source)

AI helps read charred text buried for over 2000 years

A group of student researchers, winners of the $1 Million Vesuvius Challenge, used AI to decipher a 2,000-year-old charred scroll from Herculaneum. This scroll, part of the Herculaneum scrolls, was too delicate to open before.

 With AI, they uncovered a previously unknown philosophical text discussing senses and pleasure. By training machine-learning algorithms on scans of the papyrus, they made this breakthrough, which could revolutionize our understanding of the ancient world. (source)

Amazon has recently unveiled Rufus, an AI shopping assistant, now available in the Amazon mobile app.

Rufus is trained on Amazon's vast product catalog and web data to provide personalized assistance to shoppers. With over 25 years of collected data, Amazon has equipped Rufus with features like tailored recommendations, review highlights, and fit guidance, aiming to enhance the shopping experience. Shoppers can now easily seek guidance, compare items, receive recommendations, and ask specific questions about products. (source)

The DeepSeek-Math a tiny 7B model has achieved over "50% accuracy" on the challenging MATH benchmark. This performance is comparable to that of leading models like GPT-4, marking a notable advancement in the field of mathematical reasoning.
It enables the tackling of complex quantitative challenges in financial analysis, scientific computing, engineering design, and mathematical research, signifying a step forward in AI-assisted problem-solving across these crucial areas. (source)

In 2023, AI lobbying surged by 185%, involving over 450 organizations. Notable newcomers include ByteDance, Tesla, Palantir, and OpenAI. These efforts span industries from Big Tech to academia, with a total investment of $957 million. President Biden's executive order on AI has sparked discussions on fairness and real-world impacts, with ongoing public input sought by NIST. (source)

Microsoft has teamed up with media platform Semafor and other news outlets to pioneer AI-assisted news content creation, amidst a legal battle with The New York Times. Semafor's upcoming global news feed, Signals, will empower journalists with diverse perspectives on major stories, using OpenAI and Microsoft tools. (source)

  • Hugging Chat Assistants is the free, open-source alternative to OpenAI's GPT Store, simplifying personalized chatbot creation with a "two-click" process. Unlike OpenAI's paid model, it offers an accessible way to engage with AI, though it lacks some features like web search.

  • Google's Lumiere, a video generation AI, introduces STUNet (Space-Time-U-Net), a novel diffusion model that understands objects' placement (space) and their motion and transformation over time. Unlike other models that piece together videos from pre-generated keyframes, STUNet enables Lumiere to directly generate content by focusing on how and where things move within the video timeline.

  • Meta has released an interactive Jupyter notebook titled Prompt Engineering with Llama 2 which explores techniques and best practices in prompt engineering, along with demonstrating different methods of prompting.

  • In the recent episode of 'Leading With Data', I engaged in a fascinating discussion with Sebastian Raschka, a renowned author and advocate for AI and machine learning education. He shared insights on his journey into AI research, his role at Lightning AI in optimizing LLM models, and his vision for the future of open-source LLMs and Generative AI.

  • Here is an interesting GitHub Repository for people handling Engineering behind training LLMs (source)

  • Analytics Vidhya along with Fractal has released Fractal Data Science Professional Certificate on Coursera. (source) The Certificate not only includes all the technical skills you need to start your career in Data Science but also brings a problem-solving first approach to solving Data Science problems. I don’t think there is any other course bringing the practical aspects of the program. Do check them out.

I appreciate the move from OpenAI, Meta, and previously by Samsung on tagging AI content - a significant step in the much-needed direction. Is it enough - probably not!

Content generation is reaching a point where it is impossible to distinguish AI-generated content and actual content for most people.

It is also not clear how can we enforce this on people using open-source libraries and using it to generate content! We don’t have an answer yet.

How do you rate this issue of AI Emergence?

Would love to hear your thoughts

Login or Subscribe to participate in polls.

Reply

or to participate.