Tuesday — March 4, 2025
VectorChord-BM25 transforms PostgreSQL search with precision, researchers tackle AI's water footprint, and SmartestKid brings an AI assistant to Windows with Office integration.
News
BM25 in PostgreSQL
VectorChord-BM25 is a new extension for PostgreSQL that enhances full-text search capabilities with advanced BM25 scoring and ranking, allowing for more precise and relevant search results. The extension is designed to be simple, powerful, and fully integrated with PostgreSQL, offering optimized indexing, enhanced tokenization, and a seamless API for improved search performance.
Show HN: Sonauto API – Generative music for developers
The API requires an API key in the Authorization header and supports several endpoints for generating music, including creating a new song, extending an existing song, and replacing sections of an existing song. The API accepts various parameters such as tags, lyrics, prompt, and instrumental, and returns a task ID that can be used to track the status of the generation request.
DOGE's AI use in government doesn't add up
The US DOGE Service, led by Elon Musk, is pursuing an "AI-first" strategy to remake the government, but experts warn that relying solely on automation to make critical decisions can lead to big mistakes. Despite the potential of AI to improve government efficiency, its implementation requires careful consideration of data quality, bias, and human expertise, and many argue that DOGE's approach is overly simplistic and may be a cover for gutting critical public services.
Show HN: Firebender, a simple coding agent for Android Engineers
Firebender is a tool that allows users to command smart models to edit and run their code directly, with features such as creating end-to-end tests and iterating off of gradle runs. The tool provides various resources, including documentation, a forum, and download links, to help users get started and configure the plugin to their needs.
The Framework Laptop 13 has just been upgraded to AMD's Strix Point
The Framework Laptop 13 has been upgraded with AMD's Strix Point processors, including Ryzen 5, 7, and 9 options, offering improved performance and AI capabilities. The new laptop, which can be purchased as a modular upgrade for $449 or as a complete system starting at $899, also features a new cooling system, Wi-Fi 7, and optional 2.8K 120Hz screen, with shipping set to begin in April.
Research
Making AI Less Thirsty: Uncovering and Addressing the Secret AI Water Footprint
The water footprint of artificial intelligence, including the large amounts of freshwater used to train models like GPT-3, has gone largely unnoticed despite its significant impact, with global AI demand projected to account for 4.2-6.6 billion cubic meters of water withdrawal in 2027. To address this issue, a methodology to estimate the water footprint of AI is needed, and it is essential to consider both water and carbon footprints to enable truly sustainable AI.
Art: Anonymous Region Transformer
The Anonymous Region Transformer (ART) is a generative model that enables the direct generation of multi-layer transparent images based on a text prompt and anonymous region layout, allowing for efficient and scalable layer generation. ART outperforms traditional methods, being over 12 times faster and exhibiting fewer layer conflicts, and establishes a new paradigm for interactive content creation with precise control over image layers.
Flash Interpretability: Decoding Specialised Feature Neurons in LLM
Researchers have developed a method to decode neuron weights in large language models into token probabilities, allowing for the identification of specialized feature neurons, such as those related to specific concepts like "dog" or "California". This method was successfully applied to the Llama 3.1 8B model, enabling the mapping of top features in under 10 seconds and demonstrating the ability to influence the model's output by clamping specific neurons.
An Attempt to Catch Up with JIT Compilers
Researchers added Dynamic Binary Modification to an Ahead-of-Time JavaScript compiler to improve performance, but found that reducing memory accesses did not result in faster execution times on contemporary architectures. This negative result suggests that implementing complex compiler optimizations may not be worthwhile if the processor can already accelerate the code, a finding that applies to both Ahead-of-Time and Just-in-Time compilers.
Cautious Optimizers: Improving Training with One Line of Code
Researchers have proposed a modified version of momentum-based optimizers, such as AdamW, which they call "cautious optimizers" (e.g. C-AdamW and C-Lion), that preserves convergence guarantees while offering faster and more stable training. Experimental results show that these new optimizers can speed up pretraining tasks by up to 1.47 times and achieve better results in certain post-training tasks.
Code
Show HN: Agents.json – OpenAPI Specification for LLMs
The agents.json specification is an open standard that enables AI agents to interact with APIs by providing a formal description of contracts for API and agent interactions, built on top of the OpenAPI standard. The specification allows API providers to create an agents.json file that describes how their API endpoints interact with each other, making it easier for AI agents to execute a series of API calls to achieve a specific outcome.
Go-attention: A full attention mechanism and transformer in pure Go
The go-attention module provides a pure Go implementation of attention mechanisms and transformer layers, designed for high performance and ease of use, with applications in text processing, time series analysis, and structured data. It offers efficient dot-product attention, multi-head attention, and full transformer layer implementation with batched operations, making it suitable for edge computing, real-time processing, cloud-native applications, and production systems.
Show HN: Open-Source Windows AI assistant that uses Word, Excel through COM
The SmartestKid AI Assistant is a Windows desktop AI assistant built in Python, allowing users to interact with Microsoft Office, images, and their file system through a simple chat UI. The assistant can be set up and run on a Windows machine with Python 3.7+, an OpenAI API key, and Microsoft Office, and features voice and text input modes, draggable interface elements, and integration with various applications.
SpeciesNet: AI models to classify species from motion-triggered widlife cameras
SpeciesNet is an ensemble of AI models that classifies wildlife in camera trap images, combining an object detector and an image classifier to recognize species. The ensemble, which includes the MegaDetector and SpeciesNet models, can be run using a Python package and utilizes geographic information to assign each image to a single category, with support for GPU acceleration and optional separate execution of each component.
Show HN: Open-source Deep Research across workplace applications
Onyx is an open-source, AI-powered platform that connects to a company's documents, applications, and people, providing a feature-rich chat interface and integrating with various large language models. It offers a range of features, including custom AI agents, knowledge curation, and scalable deployment options, with both a free Community Edition and a paid Enterprise Edition available.