May 13, 2024

Daily

• Hazy Research introduces ThunderKittens, an embedded DSL for CUDA, designed to maximize GPU efficiency by simplifying the writing of high-performance kernels, particularly for AI applications.

• Page dewarping is achieved by transforming curled page images into flat versions, using a parametric model that simulates the page's 3D shape and orientation.

• GitHub Copilot's magic fades when faced with unique coding projects like picom, revealing limitations in handling atypical codebases.

• Scientists have discovered a "sperm whale phonetic alphabet" in the click patterns of sperm whales, suggesting a complex communication system akin to human language.

• The Technology Innovation Institute launched Falcon 2 11B, a multilingual and multimodal AI model with unique vision-to-language capabilities, trained on 5.5 trillion tokens with 11 billion parameters.

• Researchers have developed methods to automatically detect under-trained tokens, like the notorious SolidGoldMagikarp, in Large Language Models (LLMs), addressing a significant gap in tokenizer and model training alignment.

• Mohammed Jamal explores the concept of homoiconicity in Python by emulating Lisp's unique characteristic where code and data are interchangeable, providing a fresh perspective on programming language design.

• The solar storm has prompted NOAA to issue warnings about potential damage to transformers and grid systems, which could indirectly affect the operation of models like Llama3 8B.

• Yi-1.5 is an enhanced AI model over its predecessor, boasting improvements in coding, math, reasoning, and instruction-following, while retaining its prowess in language understanding and commonsense reasoning.

• Pipecat is a versatile framework designed for creating voice and multimodal conversational agents, including personal coaches, meeting assistants, and customer support bots, with examples available on GitHub.

• LinearBoost, a new machine learning classification algorithm, outperforms traditional GBDT algorithms like CatBoost, XGBoost, and LightGBM in accuracy and response time across five benchmark datasets.

• AlphaFold3 (AF3), developed by Google DeepMind, represents a significant advancement in bioinformatics, capable of predicting unseen protein structures solely from their amino acid sequences and extending its predictive capabilities to complexes like nucleic acids and small molecules.

• Discord has been utilizing machine learning to classify users by gender and age since August 2022, revealing a sophisticated approach to user analytics.

• Pi-C.A.R.D is an AI-powered voice assistant designed to run on a Raspberry Pi, capable of conversational interactions and, with a camera, visual recognition tasks.

• Deblur-GS introduces a method to reconstruct a sharp radiance field from camera motion blurred images, addressing the limitations of 3D Gaussian Splatting (3DGS) in handling blurred inputs.