Harvard Is Releasing a Free AI Training Dataset

Harvard University has released a high-quality dataset of nearly 1 million public-domain books, which can be used to train large language models and other AI tools, aiming to "level the playing field" by giving the general public access to curated content repositories. The dataset, created with funding from Microsoft and OpenAI, is part of a growing trend of projects providing public-domain datasets to reduce reliance on copyrighted materials for AI training.

Read more

Apple collaborates with Nvidia to research faster LLM performance

Apple has collaborated with NVIDIA to research and implement faster text generation performance with large language models (LLMs). The collaboration integrates Apple's Recurrent Drafter (ReDrafter) technique into NVIDIA's TensorRT-LLM framework, resulting in a 2.7x speed-up in generated tokens per second for greedy decoding.

Read more

Microsoft acquires twice as many Nvidia AI chips as tech rivals

Microsoft has acquired twice as many Nvidia AI chips as its tech rivals. The article is locked behind a paywall, requiring a subscription to access the full content.

Read more

I Built a Figma Plugin That Generates Custom SVG Illustrations with AI

Vector Image AI is a Figma plugin that uses artificial intelligence to generate custom SVG illustrations and icons, allowing users to create unique, scalable designs with just a few words. The plugin combines AI-powered creativity with the precision of vector graphics, making it a must-have for designers looking to unlock the full potential of AI-powered creativity.

Read more

On-silicon real-time AI compute governance from Nvidia, Intel, EQTY Labs

EQTY Lab, Intel, and NVIDIA have unveiled 'Verifiable Compute,' a solution that provides the first-ever certificates of authenticity and compliance for independent verification of AI training, inference, and benchmarks at runtime. This breakthrough, the result of two years of research, delivers on-silicon, real-time governance and is set to transform the way organizations enforce AI governance, automate auditing, and collaborate to build safer and more valuable AI.

Read more

Cultural Evolution of Cooperation Among LLM Agents

Researchers studied the interactions of multiple large language model (LLM) agents over generations, finding that some models, like Claude 3.5 Sonnet, can learn mutually beneficial social norms and achieve higher cooperation scores than others, such as Gemini 1.5 Flash and GPT-4o. The study also revealed variation in emergent behavior across different initial conditions, suggesting a need for new benchmarks to evaluate the implications of LLM agent deployment on societal cooperation.

ArXiv

Leveraging LLM for Automated Ontology Extraction and Knowledge Graph Generation

OntoKGen is a pipeline for extracting and structuring knowledge from complex technical documents in the Reliability and Maintainability domain, using Large Language Models and an adaptive algorithm to align with user requirements. The generated Knowledge Graph can be seamlessly integrated into non-relational databases, facilitating advanced querying and analysis, and serving as a foundation for developing intelligent applications.

ArXiv

Artificial Intelligence in the Knowledge Economy

The integration of Artificial Intelligence (AI) into the knowledge economy can transform problem-solving by either displacing humans or reallocating them to different tasks, depending on the level of AI autonomy. The effects of AI vary, with autonomous AI benefiting knowledgeable individuals and non-autonomous AI benefiting less knowledgeable ones, but overall, autonomous AI achieves higher output.

ArXiv

Rethinking the Combination of Graph Neural Network and Large Language Model

Recent research integrating Large Language Models (LLMs) with Graph Neural Networks (GNNs) has limitations, with LLM-centered models struggling to capture graph structures and GNN-centered models compressing textual data into fixed-size vectors. A new architecture, GL-Fusion, addresses these limitations by deeply integrating GNNs with LLMs through innovations such as Structure-Aware Transformers and Graph-Text Cross-Attention, achieving state-of-the-art performance on various tasks.

ArXiv

Design choices made by LLM-based test generators prevent them from finding bugs

Researchers evaluated LLM-based test generation tools, such as Codium CoverAgent and CoverUp, and found that they can fail to detect bugs and may even validate faulty code by rejecting tests that reveal bugs. This raises concerns about the design and effectiveness of these tools in achieving the objectives of software testing and their potential impact on software quality and test suite reliability.

ArXiv

Show HN: RAGLite – A Python package for the unhobbling of RAG

RAGLite is a Python toolkit for Retrieval-Augmented Generation (RAG) that supports PostgreSQL or SQLite databases and various LLM providers, offering features such as configurable reranking, fast and permissive dependencies, and extensible architecture. It can be installed via pip and configured with a preferred database and LLM, allowing users to insert documents, perform adaptive or programmable RAG, and evaluate retrieval and generation performance.

Repo

Show HN: Bodo – high-performance compute engine for Python data processing

Bodo is a high-performance Python compute engine that uses an auto-parallelizing just-in-time compiler to transform Python programs into highly optimized, parallel binaries, resulting in 20x to 240x faster performance compared to alternatives. It seamlessly supports native Python APIs like Pandas and NumPy, and is designed for large-scale data processing, data engineering, data science, and AI/ML workloads.

Repo

Show HN: I built PdfDing – A selfhosted PDF manager and viewer

PdfDing is a self-hosted PDF manager and viewer that offers a seamless user experience on multiple devices, allowing users to organize, view, and share PDFs with ease. It's designed to be minimal, fast, and easy to set up using Docker, with features like dark mode, SSO support, and automated backups.

Repo

Show HN: xmllm – Structured LLM streaming output using lenient XML parsing

The provided text is incomplete, but it appears to be an error message indicating that a README file could not be retrieved. No further information is available due to the incomplete text.

Repo

Genesis

Genesis is a physics platform designed for general-purpose robotics, embodied AI, and physical AI applications, featuring a universal physics engine, a lightweight and user-friendly simulation platform, and a powerful photo-realistic rendering system. It aims to lower the barrier to using physics simulations, unify state-of-the-art physics solvers, and minimize human effort in collecting and generating data for robotics and other domains.

Repo