I gave the AI arms and legs then it rejected me

The author of the open-source library "enigo" discovered that it is being used by Anthropic, a leading AI company, in their "Claude Desktop" application, which allows an AI to control a computer. Despite being proud that his library was chosen for the project, the author's application to work at Anthropic was rejected, leaving him to wonder if the rejection letter was written by a human or the very AI his library helped to equip with new capabilities.

Read more

Kitten TTS: 25MB CPU-Only, Open-Source Voice Model

The AI industry is shifting away from large, resource-intensive models towards smaller, more efficient ones, with Kitten TTS being a prime example, boasting just 15 million parameters and a size of under 25MB. This tiny text-to-speech model can run without a GPU, has multiple high-quality voices, and is ultra-fast, making it a game-changer for on-device AI and a significant step towards a more distributed, community-driven ecosystem.

Read more

LLM Inflation

Data compression has long been a crucial aspect of computing, allowing us to reduce the size of data while retaining its information, but now we're seeing a trend where large language models (LLMs) are being used to do the opposite, inflating simple content into lengthy, complex text. This "LLM inflation" can implicitly reward obfuscation and time wasting, and may be covering up a lack of clear thinking, prompting us to consider why we're using LLMs in this way and potentially encouraging us to change our approach.

Read more

Running GPT-OSS-120B at 500 tokens per second on Nvidia GPUs

As a launch partner for OpenAI's new open-source LLM, GPT-oss-120b, the company worked to optimize the model's performance, achieving the best latency and throughput on NVIDIA GPUs by the end of the launch day. Through a series of engineering efforts, including testing and benchmarking, fixing compatibility bugs, and optimizing model configuration, the company was able to improve the model's performance, with plans to continue optimizing and adding new features such as speculative decoding.

Read more

AI in Search is driving more queries and higher quality clicks

Google's integration of AI in Search is driving more queries and higher quality clicks, with users searching more than ever and discovering new capabilities. Despite some concerns about traffic to websites, Google's data shows that total organic click volume has remained relatively stable year-over-year, with average click quality actually increasing and slightly more quality clicks being sent to websites than a year ago.

Read more

A candidate giant planet imaged in the habitable zone of α Cen A

The James Webb Space Telescope's MIRI instrument was used to observe the star $\alpha$ Cen A, achieving high sensitivity to detect planets and exozodiacal dust emission, and setting a new limit on the latter. A point source, potentially a planet, was detected in August 2024, but not in subsequent observations, although its possible orbital motion could explain its non-detection, and if confirmed, the planet candidate would have a temperature of around 225 K and a mass between 90-150 $M_{\rm Earth}$.

ArXiv

Solving the compute crisis with physics-based ASICs

The increasing demands of artificial intelligence are creating a "compute crisis" due to high energy consumption and training costs, but Physics-based Application-Specific Integrated Circuits (ASICs) offer a solution by harnessing physical dynamics for computation. These devices have the potential to greatly improve energy efficiency and computational power, accelerating AI applications and traditional workloads, and could lead to the development of heterogeneous, specialized computing platforms that overcome current scaling limitations.

ArXiv

The possibility of a giant impact on Venus

A giant impact on Venus could have altered its rotation rate and thermal evolution, and researchers used simulations to explore various impact scenarios, considering factors such as impactor mass, velocity, and geometry. The results suggest that a wide range of impact scenarios can reproduce Venus' current rotation characteristics, and that such an impact could also explain the planet's lack of a moon, as debris from the collision would likely reaccrete onto the planet.

ArXiv

The wall confronting large language models

Large language models' ability to improve the uncertainty of their predictions is severely limited by scaling laws, making it difficult to raise their reliability to meet scientific standards. The models' learning mechanism, which generates complex output distributions, may also contribute to errors and degenerative behavior, and avoiding these issues will require a deeper understanding of the problems being investigated and the structural characteristics of the data.

ArXiv

Recyclable 3D-Printed Electronics with Liquid Metal Conductors and PVA Substrate

DissolvPCB is a technique for creating fully recyclable printed circuit board assemblies using 3D printing and water-soluble materials, allowing for easy recycling by dissolving the substrate and reusing the components. The technique has been demonstrated through the fabrication and recycling of several functional prototypes, and its environmental impact has been evaluated through a lifecycle assessment comparing it to conventional methods.

ArXiv

Show HN: An open-source e-book reader for conversational reading with an LLM

BookWith is a next-generation conversational reading platform that uses AI to transform the reading experience, allowing users to converse with an AI that understands the book's content in real-time, answering questions, deepening comprehension, and offering new perspectives. The platform features various tools, including an AI reading assistant, AI podcast generation, multi-layer memory system, smart annotation, and semantic search, making it perfect for researchers, book lovers, students, and business professionals.

Repo

Bitfrost – LLM gateway 90x faster than Litellm at p99

Bifrost is a high-performance AI gateway that connects to 10+ providers, including OpenAI and Anthropic, through a single API, offering features like automatic failover, load balancing, and zero-downtime deployments. It can be set up in under 30 seconds and adds only 11µs latency while handling 5,000+ requests per second, making it a reliable and efficient solution for building AI applications.

Repo

An Open-Source Asynchronous Coding Agent

Open SWE is an open-source, cloud-based asynchronous coding agent that autonomously understands codebases, plans solutions, and executes code changes across entire repositories. It offers features such as planning, human-in-the-loop interaction, parallel execution, and end-to-end task management, and can be used through a web application or directly from GitHub issues.

Repo

Run LLM's Locally on iPhone

Q2 Edge Chat is a privacy-focused chat application that runs large language models locally on your iPhone, allowing for 100% local processing and no data collection. The app features a modern chat experience, model management, and a beautiful interface, and is designed with privacy as its core principle, with all conversations staying on the device and no analytics or telemetry.

Repo

Show HN: XRAY MCP – AST-grep wrapped in a tiny server for code-aware AI

XRAY is a code intelligence tool that helps AI assistants navigate and understand codebases by providing capabilities such as code mapping, symbol searching, and impact analysis. It uses ast-grep, a tree-sitter powered structural search tool, to accurately parse code syntax and provide reliable results, making it a practical solution for AI assistants to gain deep code intelligence without information overload.

Repo