Alibaba Cloud's Qwen chatbot evolved from Tongyi Qianwen to Qwen2.5 with 2.4 trillion parameters in just 3.5 years, revolutionizing AI models and licensing. Qwen became a default choice for developers in 2024, with multiple generations and sizes, including open mixture-of-experts and specialized models for math, audio, and video analysis.
Yandex's Sona AI model streamlines recommendation pipelines, combining candidate generation and ranking into one system. Sona's architecture eliminates multiple stages and hand-engineered features, improving efficiency and user experience on smart speakers.
AI startups in NVIDIA Inception program are transforming breast cancer care, from screening to treatment planning. iSono Health's AI-powered ATUSA platform offers faster, more accurate 3D breast scans, improving early detection and treatment decisions.
Amazon Quick offers flexible user management options, including role assignments and access reviews. Downgrading user roles is crucial for security and cost optimization in Quick environments, with detailed steps provided for smooth transitions.
Generative AI in Amazon Quick enables compliance teams to efficiently check 50,000 apartment leases for changing state laws. The Adjudicated Query pattern ensures provable completeness and defensibility in high-stakes compliance domains like lease regulations.
Search agents powered by large language models (LLMs) are revolutionizing information retrieval for enterprises by autonomously deciding what to search for and when to stop. Fine-tuning with Amazon SageMaker AI multi-turn reinforcement learning (MTRL) optimizes search agents for reliability, speed, and cost efficiency.
MIT's Bill Aulet's book "Disciplined Entrepreneurship" has inspired a free AI-powered youth platform called Dear Dreamer, aiming to empower 50,000 young entrepreneurs by 2030. The platform adapts MIT's proven framework into a digital, self-paced learning experience for middle and high school students, offering interactive exercises and personalized feedback.
Cathy Wu '12, MNG '13, uses machine learning to improve transportation systems at MIT. Inspired by family experiences and games like "SimCity," Wu aims to design safer, more efficient systems for all.
Tech industry workers, inspired by a call to action, organized against corporate policies in the book "Against Tech Oligarchy." The movement chronicles successful strategies and backlash faced by employees advocating for democratic values in Silicon Valley.
Amazon S3 Vectors in NVIDIA NeMo Agent Toolkit (NAT) revolutionize AI agent memory storage, offering rich metadata and elastic scale. NAT's memory module allows for custom memory providers, enhancing conversation history storage and retrieval.
Machine learning regression problem aims to predict values using various techniques like Gradient Boosting with Extra Trees, yielding surprisingly accurate results. The complex model involves weak base learners and sequential tree construction, showcasing potential for successful predictions in datasets.
Amazon Quick Apps introduce Live Data in Apps, allowing real-time querying of governed Quick Sight datasets. Users can now access current data without manual refresh, bringing a new level of efficiency and data security to app development and usage.
AI factories, like NVIDIA's, cost millions to build but maximize returns by delivering high throughput, low costs, and versatility. NVIDIA's CUDA-X libraries and standardized architecture ensure productivity, durability, and fungibility, leading to significant gains in token production and reduced costs.
AI-driven InstructMesh simplifies 3D design for practical, creative objects, from dragon mugs to shrimp robots, using Microsoft's TRELLIS and GPT-4. MIT researchers blend text and image prompts to produce unique, functional items, like a denim knee brace.
uniopen, by Uni-President Enterprises Group, uses a unique moderation policy to connect customers to ecommerce experiences. Amazon Nova 2 Lite, fine-tuned in Amazon SageMaker AI, optimizes this policy for web, tablet, and mobile channels, ensuring consistent decisions with separate correction and deployment paths.