Slack has raised our charges by $195k per year

Nonprofit organization Hack Club was threatened by Slack with a massive price hike of $200k per year and a deadline of just a few days to pay, or risk losing their workspace and years of message history. However, after the issue went viral and sparked public support, Slack's CEO reached out and offered a more favorable solution, prompting Hack Club to prioritize data ownership and announce a move to alternative platform Mattermost.

Read more

Meta Ray-Ban Display

Mark Zuckerberg has introduced the Meta Ray-Ban Display, a new generation of AI glasses that feature a full-color, high-resolution display and can be controlled using the Meta Neural Band, an EMG wristband that interprets muscle activity. The glasses, which start at $799 and come with the Meta Neural Band, aim to help users stay present and accomplish everyday tasks without needing to pull out their phone, and will be available in limited quantities starting September 30.

Read more

Learn Your Way: Reimagining Textbooks with Generative AI

Google is exploring the use of generative AI to create personalized educational materials, allowing students to learn in a way that is tailored to their individual needs and interests. The company has introduced "Learn Your Way," a research experiment that uses AI to generate alternative representations of textbook content, and has shown promising results, with students using the tool scoring 11 percentage points higher on retention tests than those using a standard digital reader.

Read more

You Had No Taste Before AI

Many people in the tech industry are suddenly emphasizing the importance of "taste" when working with AI, but ironically, these same individuals often lacked discernment and critical judgment in their own work before AI became a factor. Developing true "taste" requires a deeper understanding of aesthetic quality, contextual appropriateness, and ethical boundaries, and is not a new skill that only applies to AI, but rather a fundamental aspect of producing high-quality work.

Read more

Chrome's New AI Features

Google is releasing its biggest upgrade to Chrome ever, incorporating new AI features to enhance browsing, including Gemini in Chrome, which can clarify complex information and handle tasks across multiple tabs. The update also includes AI-powered search from the address bar, better integration with Google apps, and enhanced security features to block scams and manage notifications.

Read more

AI Propaganda factories with language models

Researchers have found that small language models can produce coherent, persona-driven political messaging on basic computer hardware, and that the design of the persona behind the message has a greater impact on its effectiveness than the model used. The study also shows that automated influence operations are now feasible for both large and small actors, and proposes shifting from restricting model access to detecting and disrupting these campaigns through conversation-centric methods.

ArXiv

Cut Costs, Not Accuracy: LLM-Powered Data Processing with Guarantees

Large Language Models (LLMs) are used to process large text datasets, but top-of-the-line models are expensive, while more affordable models may compromise on accuracy. BARGAIN is a method that uses affordable LLMs to reduce costs while providing strong theoretical guarantees on solution quality, and has been shown to reduce costs by up to 86% more than state-of-the-art methods while maintaining accuracy.

ArXiv

A Trustworthiness-Based Metaphysics of Artificial Intelligence Systems

Modern AI systems lack a clear metaphysical foundation, with the orthodox view being that they lack inherent identity and persistence conditions due to being man-made artifacts. This work challenges that perspective by introducing a theory of metaphysical identity for AI systems, arguing that their identity and persistence are tied to their trustworthiness profiles and functional requirements, providing a foundation for discussions about their epistemological, ethical, and legal implications.

ArXiv

Towards a Physics Foundation Model

The General Physics Transformer (GPhyT) is a foundation model that demonstrates the ability to simulate various physical phenomena, such as fluid-solid interactions and thermal convection, without being explicitly told the underlying equations. Trained on a large dataset, GPhyT achieves superior performance across multiple physics domains, generalizes to unseen systems, and makes stable long-term predictions, paving the way for a universal Physics Foundation Model that could transform computational science and engineering.

ArXiv

Discovery of Unstable Singularities

Researchers have made a breakthrough in identifying unstable singularities in fluids, a phenomenon where solutions to equations develop infinite gradients, by discovering new families of unstable singularities using a combination of machine learning and high-precision optimization. This work achieves unprecedented accuracy, paving the way for rigorous mathematical validation and providing a new approach to exploring complex nonlinear partial differential equations and tackling long-standing challenges in mathematical physics.

ArXiv

Launch HN: Cactus (YC S25) – AI inference on smartphones

Cactus is an energy-efficient AI inference framework designed for mobile devices, particularly budget and mid-range phones, with a bottom-up approach that requires no dependencies. It exposes four levels of abstraction, including Cactus Graph, a general numerical computing framework, and Cactus Engine, a transformer inference engine, allowing for efficient execution of AI models on a wide range of devices.

Repo

RDMA-Powered Distributed Cache for Fast AI Training and Inference

Blackbird is a high-performance, RDMA-based distributed storage system that delivers intelligent data placement, enabling applications to seamlessly offload data to a high-performance tiered system. It draws inspiration from various projects, including Microsoft/FARM and Redis, and is designed for use cases such as HPC, ML training/inference pipelines, real-time analytics, and metadata-heavy services.

Repo

Show HN: SandBox – AI agents simulating possible futures

The SandBox projects are a collection of interconnected AI initiatives that explore autonomy, governance, and decision-making, including AIBlog, an autonomous AI researcher that writes about AI, TomorrowNews, an AI-driven news prediction and decision-making platform, and GenBox, a virtual world where AI makes daily decisions on critical areas. These projects demonstrate the potential of recursive agent architectures and generative AI to create high-quality, verifiable content and simulate complex decision-making scenarios without human intervention.

Repo

Show HN: PageIndex MCP – Chat with Long PDFs on Claude or Cursor

PageIndex is a vectorless reasoning-based RAG system that uses multi-step reasoning and tree search to retrieve information, offering higher accuracy and transparency. It can be integrated with Claude desktop to enable seamless chat with long PDFs, supporting local and online PDFs with free 1000 pages and unlimited conversations.

Repo

Kagi open sourced Ask: a bash script for quick AI queries in the shell

The "ask" tool is a lightweight bash script that allows users to query AI models via the OpenRouter API, providing executable output for various tasks such as command generation, code generation, and quick answers. It supports multiple AI models, including Mercury Coder, Gemini 2.5 Flash, and Claude Sonnet 4, and offers features like model selection, provider routing, and streaming mode.

Repo