# May 22, 2024

## Autonomous AI Robot Creates a Shock-Absorbing Shape No Human Ever Could

- Boston University's autonomous robot, **MAMA BEAR**, has created a **shock-absorbing shape** achieving a **record 75% energy absorption efficiency**, surpassing previous human-made designs.

## Wikimedia Enterprise – APIs for LLMs, AI Training, and More

- **Wikimedia Enterprise** offers **enterprise-grade APIs** designed to enhance user experiences by providing real-time access to the vast knowledge stored across Wikimedia projects, including **Wikipedia** in over **330 languages**.

## New Performance Optimizations Supercharge NVIDIA RTX AI PCs for Gamers, Creators and Developers

- **NVIDIA announced new AI performance optimizations** for RTX AI PCs at Microsoft Build, significantly enhancing capabilities for gamers, creators, and developers with up to **3x faster** execution of large language models (LLMs) using [ONNX Runtime (ORT) and DirectML](https://developer.nvidia.com/blog/end-to-end-ai-for-nvidia-based-pcs-onnx-and-directml/).

## Windows Copilot Runtime

- **Windows Copilot Runtime and Copilot+ PCs** introduce a **new era of AI-infused Windows PCs**, featuring the world's most powerful PC Neural Processing Units (NPUs) capable of delivering **40+ TOPS** of compute, making these PCs up to **20 times more powerful** and **100 times as efficient** for AI workloads than traditional PCs.

## Cover Agent: open source regression test generation tool

- **CodiumAI's Cover Agent** leverages **Generative AI** to automate the generation of unit tests, aiming to significantly improve code coverage and quality assurance processes.

## [R] Enabling sparse, foundational LLMs for faster and more efficient models from Neural Magic and Cerebras

- **Neural Magic, Cerebras, and IST Austria** have developed the **first highly sparse, foundational LLMs** that maintain full accuracy on tasks like chat, code generation, and summarization, marking a significant advancement in model efficiency.

## [P] Fish Speech TTS: clone OpenAI TTS in 30 minutes

- **Fish Speech TTS** successfully **cloned OpenAI's TTS** performance using **supervised fine-tuning** on 10 hours of OpenAI TTS data, achieving similar emotion, rhythm, accent, and timbre.

## Show HN: Route your prompts to the best LLM

- **Unify's chat interface** is under development for mobile, aiming to **route prompts to the most suitable Large Language Models (LLMs)** for optimized responses.

## [R] LLMs as active learning agents

- **LLMs, such as GPT-4, excel as active learning agents** by adeptly identifying challenging or diverse examples, surpassing few-shot learning techniques.

## Watt a Win: NVIDIA Sweeps New Ranking of World’s Most Energy-Efficient Supercomputers

- **NVIDIA-powered systems** dominated the [Green500 list](https://top500.org/lists/green500/2024/06/), securing the **top three spots** and seven of the top 10, showcasing the **energy efficiency** of accelerated computing.

## [Research] What are some great articles on algorithmic fairness?

- The request seeks **recent and milestone articles** on **algorithmic fairness** for a presentation preparation, emphasizing the need for foundational and cutting-edge insights.

## OLAPH: Improving Factuality in Biomedical Long-form Question Answering

- **OLAPH**, a novel framework, enhances **factuality in biomedical long-form question answering** by training large language models (LLMs) to reduce hallucinations, using the **MedLFQA dataset** for automatic evaluations.

## TwoMinutePapers - 50,000,000 Point Bouncy Jelly Simulation!

- The **50,000,000 point bouncy jelly simulation** showcases **elastic bodies**, like squishy balls and armadillos, interacting in complex ways, demonstrating **advanced computational physics**.

## Retrieval-Augmented Language Model for Extreme Multi-Label Knowledge Graph Link Prediction

- The **retrieval-augmented language model** addresses **hallucination and expensive training costs** in large language models by leveraging **knowledge graph (KG) information**, enhancing accuracy in specialized domains.

## GigaPath: Whole-Slide Foundation Model for Digital Pathology

- **GigaPath**, developed by Microsoft in collaboration with Providence Health System and the University of Washington, is a **novel vision transformer** designed for whole-slide digital pathology, leveraging **dilated self-attention** to manage the computational challenges posed by gigapixel images.