Hey Folks!

This newsletter will discuss some cool AI research papers and AI tools. Happy learning!

👉 What is Trending in AI/ML Research? 

➡️ Large Language Models Surprise Meta AI Researchers at Compiler Optimization!

How can we harness the capabilities of Large Language Models (LLMs) for code optimization? In this study by Meta AI, the authors dive into this emerging application, developing a 7B-parameter transformer model tailored to optimize LLVM assembly code for size. Unlike conventional methods, this model analyzes unoptimized assembly to suggest optimal compiler options. During training, the model is fine-tuned to forecast instruction counts pre- and post-optimization and the optimized code itself. This training strategy enriches the model's optimization prowess. Remarkably, on tested programs, the approach trims instruction counts by 3.0% more than the compiler, surpassing two top-tier benchmarks, while showcasing a profound code understanding—yielding compilable code in 91% of instances and matching the compiler's output 70% of the time.

➡️  Meet BlindChat: An Open-Source Artificial Intelligence Project to Develop Fully in-Browser and Private Conversational AI

BlindChat, an open-source and privacy-first alternative to ChatGPT, was just launched by MithrilSecurity. BlindChat is an open-source AI initiative aiming to create the world’s first conversational AI that operates entirely within a web browser without any third-party access. Today’s prevalent everyday AI solutions typically include sharing user data with AI service providers in exchange for AI model usage. Users risk having their data stolen if they let this happen. Since data is a valuable resource for enhancing LLMs, several approaches implicitly adjust users’ data to train the model better. Users run the danger of having LLMs memorize private information in this way. By performing local inference or employing secure, isolated environments called secure enclaves, BlindChat ensures that users’ data is kept private at all times and that they retain complete control over it.

➡️  Microsoft Researchers Introduce Kosmos-2.5: A Multimodal Literate Model for Machine Reading of Text-Intensive Images

How can machines effectively read and interpret text-rich images? Microsoft researchers Introduce Kosmos-2.5, a state-of-the-art multimodal literate model adept at processing text-intensive images. Trained on a vast dataset of such images, it serves two main purposes: First, it can generate spatially-aware text blocks, pinpointing the location of each text section within the image. Secondly, it translates these texts into a structured markdown format, preserving styles and structures. This dual functionality is powered by a shared Transformer design combined with specialized prompts and dynamic text representations. Assessed on extensive text recognition and image-to-markdown tasks, Kosmos-2.5 showcases potential as a universal tool for deciphering text-heavy images and represents a leap towards more advanced multimodal language models.

➡️ How Green is AI? Comparing the Carbon Footprint of Artificial Intelligence and Human Tasks

How do the greenhouse gas emissions of AI systems compare to those of humans performing the same tasks? This study compared the CO2e emissions of various AI systems, including ChatGPT, BLOOM, DALL-E2, and Midjourney, with human counterparts. Findings revealed that an AI writing a page of text produces 130 to 1500 times less CO2e than a human, and an AI generating an image emits 310 to 2900 times less. However, this analysis doesn't factor in societal effects such as job displacement, legal concerns, and potential increased AI usage. While AI can't replace all human tasks, it offers the possibility of conducting certain activities with significantly reduced emissions.

➡️ Deci AI Unveils DeciDiffusion 1.0: A 820 Million Parameter Text-to-Image Latent Diffusion Model and 3x the Speed of Stable Diffusion

Deci AI introduces DeciDiffusion 1.0 – A New Approach To solve the text-to-image generation problem; a research team introduced DeciDiffusion 1.0, a groundbreaking model representing a significant leap forward in this domain. DeciDiffusion 1.0 builds upon the foundations of previous models but introduces several key innovations that set it apart. One of the key innovations is the substitution of the traditional U-Net architecture with the more efficient U-Net-NAS. This architectural change reduces the number of parameters while maintaining or even improving performance. The result is a model that is not only capable of generating high-quality images but also does so more efficiently in terms of computation. The model’s training process is also noteworthy. It undergoes a four-phase training procedure to optimize sample efficiency and computational speed. This approach is crucial for ensuring the model can generate images with fewer iterations, making it more practical for real-world applications.

👉 What is Trending in AI Tools?