Self-hosting your own media considered harmful according to YouTube

Jeff Geerling's YouTube video demonstrating the use of LibreELEC on a Raspberry Pi 5 for 4K video playback was removed due to a community guidelines violation, with YouTube claiming it promoted "dangerous or harmful content" by describing how to access audiovisual content without payment. The video was later reinstated after a human review process, and Geerling has re-uploaded it to alternative platforms such as Internet Archive and Floatplane, highlighting the challenges of self-hosting media and navigating YouTube's content guidelines.

Read more

Meta: Shut down your invasive AI Discover feed

The Mozilla Foundation is calling on Meta to shut down its invasive AI Discover Feed, which is quietly turning private AI chats into public content without users' knowledge or consent. The foundation, a global nonprofit dedicated to keeping the internet open and accessible, is demanding that Meta prioritize user privacy and transparency in its AI practices.

Read more

Doge Developed Error-Prone AI Tool to "Munch" Veterans Affairs Contracts

The Trump administration used a software engineer with no healthcare experience to develop an AI tool to identify non-essential contracts at the Department of Veteran Affairs, which resulted in flawed and inaccurate recommendations to cancel over 2,000 contracts. The tool, which was built quickly using outdated AI models, made significant errors, including misreading contract values and flagging essential services such as cancer treatment and nurse care for cancellation.

Read more

Show HN: AI game animation sprite generator

God Mode AI is an online platform that uses artificial intelligence to generate professional game animation sprites from uploaded character designs or text descriptions. The platform offers a range of features, including diverse animation types, production-ready sprites, and customizable styles, making it suitable for indie developers, game studios, and artists.

Read more

What “working” means in the era of AI apps

Startups in the era of AI apps are growing faster than ever, with median enterprise companies reaching over $2 million in annual recurring revenue in their first year and median consumer companies reaching $4.2 million. The gap between "good" and "exceptional" companies is growing, with top performers achieving rapid growth and pulling away from the rest, and consumer companies are now generating significant revenue, outpacing their B2B counterparts.

Read more

Algebra Unveils Deep Learning – An Invitation to Neuroalgebraic Geometry

This position paper proposes studying function spaces in machine learning models through the lens of algebraic geometry, focusing on algebraic models like neural networks with polynomial activations. It outlines a connection between algebro-geometric invariants and fundamental machine learning aspects, laying the groundwork for a new research direction called neuroalgebraic geometry that bridges algebraic geometry and deep learning.

ArXiv

LLM-Explorer: Efficient and Affordable LLM-Based Exploration for Mobile Apps

Large language models (LLMs) have improved automated mobile app exploration, but existing approaches rely too heavily on LLMs, leading to high costs and resource usage. LLM-Explorer, a new exploration agent, uses LLMs more efficiently by leveraging them for knowledge maintenance rather than action generation, resulting in significantly faster and more cost-effective exploration with higher coverage.

ArXiv

LongCodeBench: Evaluating Coding LLMs at 1M Context Windows

Context lengths for models have rapidly increased, making it challenging to create realistic long-context benchmarks, but code comprehension and repair tasks have been identified as a suitable testbed. The LongCodeBench benchmark tests long-context language models' coding abilities, revealing that even top models struggle with long contexts, experiencing significant performance drops as context size increases.

ArXiv

Efficient Streaming Language Models with Attention Sinks

Deploying Large Language Models in streaming applications poses challenges due to memory consumption and limited generalization to longer texts. The proposed StreamingLLM framework addresses these issues, enabling models to generalize to infinite sequence lengths without fine-tuning, and achieves significant speedup over baseline methods in streaming settings.

ArXiv

Holo1: Cost-Efficient Web Agent Powered by Open Weights

Surfer-H is a cost-efficient web agent that uses Vision-Language Models (VLM) to perform tasks on the web, and when paired with the Holo1 model, it achieves 92.2% state-of-the-art performance on the WebVoyager benchmark. The Holo1 model, which was trained on a curated dataset, tops generalist UI benchmarks and a new web UI localization benchmark called WebClick, and both the model and the WebClick dataset are being open-sourced.

ArXiv

Microsoft Trace: End-to-End Generative Optimization for AI Agents

Trace is a Python library that enables end-to-end generative optimization for AI agents, allowing users to train AI systems with general feedback such as numerical rewards or losses, natural language text, and compiler errors. It provides a PyTorch-like interface, with primitives like node and bundle, to define and optimize computation graphs, and can be used to build self-improving agents that react to environment feedback.

Repo

Show HN: Resumable Web Streams

The resumable-stream library allows clients to resume streams of strings after a lost connection or for a second client to follow along, using a pubsub mechanism with Redis to minimize latency impact. It works by creating a producer that completes the stream and listens for additional consumers, which can then receive the stream content via pubsub after requesting it.

Repo

I built a knowledge system that gives AI perfect codebase memory

Octocode is a powerful code indexer and semantic search engine that builds intelligent knowledge graphs of your codebase, combining advanced AI capabilities with local-first design to provide deep code understanding and intelligent assistance for developers. It offers features such as semantic code search, knowledge graph building, multi-language support, AI-powered features, and integration with AI assistants, and can be installed and used through various commands and configurations.

Repo

Show HN: Cpdown – Copy any webpage/YouTube subtitle as clean Markdown(LLM-ready)

Cpdown is a browser extension that allows users to copy the content of any webpage as clean, formatted markdown with one click or keyboard shortcut, also supporting YouTube subtitle copying. The extension uses tools like Defuddle and Mozilla's Readability to extract the main content, remove unnecessary HTML elements, and provide features like token counting and keyboard shortcut support.

Repo

The WeightWatcher tool for predicting the accuracy of Deep Neural Networks

WeightWatcher is an open-source diagnostic tool for analyzing Deep Neural Networks (DNNs) without needing access to training or test data, using ideas from Random Matrix Theory, Statistical Mechanics, and Strongly Correlated Systems. It can analyze pre-trained models, monitor over-training or over-parameterization, predict test accuracies, and detect potential problems when compressing or fine-tuning pre-trained models.

Repo