🚀 Headlines & Launches
Microsoft’s Copilot App Lands on iOS
Microsoft has unveiled its Copilot AI chatbot app for iOS and iPadOS, bringing the power of GPT-4 to users without a subscription. The app empowers users to ask questions, draft emails, summarize text, and create images. The move signals Microsoft’s shift towards a standalone experience akin to ChatGPT, with Copilot offering a distinct web experience separate from Bing.
Insights into AI Developments in 2023
A comprehensive roundup delves into the highlights of AI development in 2023. Covering topics such as running LLMs on personal devices, fine-tuning models, challenges in gullibility, diverse LLM applications, and more, the article explores the transformative impact of LLMs on users’ quality of life.
IBM’s Perspective on AI Game-Changers
IBM’s Vice President of Software and Technology, Raj Datta, and Director of Startups, Kylie Rutherford, share insights into how AI is revolutionizing businesses across industries. The article highlights various use cases of AI products, showcasing the transformative potential for companies of all sizes.
🧠 Research & Innovation
MosaicBERT: Optimizing Bidirectional Encoder for Fast Pretraining
Mosaic introduces a bidirectional encoder, MosaicBERT, optimized for rapid pretraining. Incorporating innovations like FlashAttention and GLU, MosaicBERT significantly improves pretraining speed while matching the performance of larger traditional BERT models.
Enhancing Text Embeddings with Large Language Models
Microsoft researchers leverage synthetic data to train a decoder-only transformer based on Mistral for embeddings. This two-step prompting strategy with GPT-4 showcases the potential of large language models to generate synthetic retrieval training data effectively.
Influence of Altered Images on Human Perception
Recent research reveals that subtle alterations to digital images designed to deceive AI vision systems can also impact human perception. While humans and AI perceive images differently, controlled conditions demonstrated that humans can be systematically biased by adversarial perturbations intended to mislead AI.
🧑💻 Engineering & Resources
Comprehensive LLM Course
A detailed exploration of hot topics in the Large Language Model (LLM) space, including merging, GGUF, quantization, DPO, and more. Designed for beginners, scientists, and engineers, the course provides valuable insights to quickly get up to speed in the field.
Llama File One-Liners
Delving into the Llamafile project, which integrates model and inference code into a portable executable. The blog explores using command line output as input for the language model, showcasing practical applications of Llamafile.
Unveiling Gemini’s Potential: Multimodal Commonsense Reasoning
An in-depth analysis of Google’s Gemini, a Multimodal Large Language Model, assesses its performance in common sense reasoning across various tasks. The project compares Gemini with other models, highlighting its competitive edge in integrating knowledge across modalities.
🎁 Miscellaneous
LLMs and Programming in 2024
Large Language Models (LLMs) have become indispensable for programmers in 2023, accelerating code writing and enhancing productivity. While having limitations in complex system programming, LLMs excel in high-level Python coding and mundane tasks, serving as efficient tools for developers.
Advancements in Single Image Super-Resolution
Researchers unveil a novel method to improve single image super-resolution, focusing on the optimal centroid of potential high-resolution images and mitigating inherent noise that affects image quality.
Recap of Consumer AI in 2023
A comprehensive thread recaps AI products and trends that gained traction in 2023. Highlights include ChatGPT reaching 100 million monthly active users, viral moments like Balenciaga Pope, the rise of AI-generated covers, TikTok’s AI memes, and the launch of X’s Grok.
