News

This paper introduces a method to create a better large-scale dataset for training large language models (LLMs) by combining techniques like classifier ensembling, synthetic data rephrasing, and reduced reliance on traditional filters. The approach results in a 6.3T token dataset with more unique, high-quality tokens compared to previous datasets like DCLM. This improves model performance… Continue reading Nemotron-CC: Transforming Common Crawl into a Refined Long-Horizon Pretraining Dataset

Wake Vision is a new dataset designed to advance TinyML, enabling machine learning on low-power devices like microcontrollers. It is 100 times larger than the current benchmark dataset, Visual Wake Words (VWW), with 6 million images. Wake Vision provides two training sets: one focusing on size and the other on data quality, allowing researchers to… Continue reading Introducing Wake Vision: A High-Quality, Large-Scale Dataset for TinyML Computer Vision Applications

Microsoft Research has developed innovative tools, the Large Market Model (LMM) and Financial Market Simulation Engine (MarS), to harness the power of generative AI in the financial industry. These tools use advanced models trained on order flow data—the detailed records of market transactions—to simulate market behaviors and analyze trends. Order flow data is ideal for… Continue reading MarS: A unified financial market simulation engine in the era of generative foundation models

MaskedMimic is an innovative system designed to improve humanoid robot movement and control. It uses motion inpainting, a method inspired by generative AI, to create natural, full-body motion from incomplete instructions like partial body movements, text commands, or object interactions. Unlike traditional task-specific controllers, MaskedMimic can handle diverse tasks without retraining, such as following paths,… Continue reading Unified Whole-Body Control for Physically Simulated Humanoids

Google Cloud is revolutionizing content creation with advanced AI models, Veo and Imagen 3, now available on its Vertex AI platform. Veo generates high-quality videos from simple text or image prompts, enabling faster and cost-effective video production for tasks like marketing and social media. Imagen 3 produces lifelike, photorealistic images from text prompts, offering businesses… Continue reading Veo and Imagen 3: Announcing new video and image generation models on Vertex AI

Antibodies, known for their precision in targeting diseases like cancer and autoimmune disorders, are challenging to model due to their flexible and diverse structures. To address this, A-Alpha Bio developed AlphaBind, an AI model designed to predict and optimize antibody-antigen binding affinity. Leveraging NVIDIA and AWS technology, AlphaBind uses experimental data and advanced machine learning… Continue reading In-Silico Antibody Development with AlphaBind Using NVIDIA BioNeMo and AWS HealthOmics

AI shopping agents are emerging tools that can navigate retail websites, find products, and complete purchases on behalf of users through simple prompts. Companies like Perplexity, Google, and OpenAI are developing such agents, with Perplexity already offering one that can browse and purchase items using a system powered by Stripe’s single-use debit cards. However, challenges… Continue reading The race is on to make AI agents do your online shopping for you

NYU Langone Health achieved a world-first by performing a fully robotic double lung transplant on a 57-year-old woman with severe COPD. Using the Da Vinci Xi robotic system, the surgical team, led by Dr. Stephanie Chang, replaced both lungs through small incisions, offering a minimally invasive approach with reduced pain and quicker recovery. The patient,… Continue reading World’s First Fully Robotic Double Lung Transplant Performed by NYU Langone Health

This study examines how large language models (LLMs) use their training data to solve reasoning tasks, such as math problems, compared to answering factual questions. It found that while factual answers often come directly from specific training documents, reasoning tasks rely on general strategies learned from documents demonstrating how to solve similar problems. These strategies… Continue reading Procedural Knowledge in Pretraining Drives Reasoning in Large Language Models