OpenAI is adding new watermarks to DALL-E 3 Images

In today’s email:

📚 Tutorial: Running Open-Source AI Models Locally With Ruby
🔥 DeepSeekMath: Approaching Mathematical Reasoning Capability of GPT-4
🇺🇸 Two Texas companies were behind the AI Joe Biden robocalls
🧰 10 new AI-powered tools and resources. Make sure to check the online version for the full list of tools.

OpenAI is adding new watermarks to DALL-E 3

OpenAI is implementing watermarks in the image metadata of DALL-E 3 generated images to improve trust in digital information. This move comes as part of the adoption of standards by the Coalition for Content Provenance and Authenticity (C2PA), including major companies like Adobe and Microsoft.

The watermarks will be visible as a CR symbol in the top left corner and an invisible metadata component, indicating the AI tool used for creation. This feature aims to allow users to verify the origin of images through platforms like Content Credentials Verify. However, OpenAI acknowledges that this method isn't foolproof since metadata can be easily removed, intentionally or accidentally, and social media platforms often strip metadata from uploads. The initiative is in line with efforts to identify AI-generated content, as emphasized by the Biden administration's executive order on AI, and is intended to bolster the credibility of digital content despite potential limitations in combating misinformation.

Related: Meta to expand labeling of AI-generated imagery in election-packed year

DeepSeekMath: Approaching Mathematical Reasoning Capability of GPT-4 with a 7B Model.

DeepSeekMath is an advanced AI model developed on the DeepSeek-Coder-v1.5 7B, enhanced with an additional 500B tokens focused on mathematics, natural language, and code. It achieves a 51.7% score on the MATH benchmark, showcasing high proficiency in math problem solving without external aids, rivaling top models like Gemini-Ultra and GPT-4.

Key points include:

Evaluation Results: DeepSeekMath-Base 7B excels in mathematical problem solving, tool use, and shows comparable abilities in reasoning and coding to leading models.
Specialized Versions: DeepSeekMath-Instruct and -RL 7B models demonstrate strong quantitative reasoning skills, with the RL version reaching up to 60% accuracy on MATH with tool use, surpassing all existing open-source models.
Data Collection: The model's training involved a rigorous selection process of mathematical web pages, resulting in a comprehensive corpus of 35.5M pages.
Availability: DeepSeekMath 7B models are publicly available for research, subject to licensing terms for both academic and commercial use.

This condensed summary highlights DeepSeekMath's significant contributions to AI in mathematics, its robust evaluation results, and its availability for broader research applications.

Google’s Lookout App released a new feature called ‘Image Question and Answer'.

"Lookout," an app developed by Google aimed at assisting blind and partially sighted individuals by leveraging artificial intelligence (AI) to interpret and describe the world around them. The app features a new tool that allows users to upload images and ask specific questions about them, such as identifying landmarks (e.g., Mount Fuji) or describing elements within a photo (e.g., the color of a temple or the expression on a person's face).

The app has the ability to focus on and elucidate different aspects of an image, ranging from landscapes and objects to personal photos, thereby enhancing the user's understanding of their environment. Through community member testimonials, the video illustrates how Lookout can enrich the lives of visually impaired individuals by providing detailed descriptions of their surroundings, making daily tasks more manageable, and fostering inclusion in visual conversations.

Meet ‘Smaug-72B’: The new king of open-source AI

The recently released "Smaug-72B" language model by Abacus AI has taken the top spot in the Hugging Face Open LLM leaderboard, marking a significant achievement in the world of open-source AI. This model, a fine-tuned version of "Qwen-72B" developed by Alibaba Group's Qwen team, surpasses other leading models including GPT-3.5 and Mistral Medium in various benchmarks, achieving an average score of over 80 across all major LLM evaluations.

Smaug-72B's strength lies particularly in reasoning and math tasks, thanks to specific fine-tuning techniques that will be detailed in an upcoming research paper. This development reflects a broader trend towards the democratization of AI, with open-source models like Smaug-72B and Qwen 1.5 (which also includes a suite of models and a new large vision language model) challenging the dominance of proprietary models from big tech companies.

Their availability on Hugging Face allows for widespread access and use, contributing to rapid advancements and applications in AI. This marks a significant moment in open-source AI, highlighting its potential to rival and even surpass the capabilities of models developed by well-funded private entities.