Anthropic agrees to pay $1.5B to settle lawsuit with book authors

Anthropic, a leading artificial intelligence company, has agreed to pay $1.5 billion to a group of authors and publishers in a landmark settlement after a judge ruled it had illegally downloaded and stored millions of copyrighted books. The settlement, the largest payout in US copyright history, could set a precedent for other AI companies to pay rights holders for use of their works, and may lead to more court decisions and settlements or licensing fees in the ongoing battle between AI companies and copyright holders.

Read more

GLM 4.5 with Claude Code

GLM-4.5 and GLM-4.5-Air are advanced AI models designed for agent-oriented applications, leveraging a Mixture-of-Experts architecture and achieving high performance on various benchmark tests, including coding, reasoning, and agent-specific tasks. These models offer high parameter efficiency, low cost, and high speed, making them suitable for real-world applications such as tool invocation, web browsing, software engineering, and front-end development, with capabilities including intelligent code generation, real-time code completion, and automated bug fixing.

Read more

Why Everybody Is Losing Money On AI

The generative AI industry is plagued by unprofitability, with every company offering AI services, including those providing the models themselves, losing money due to high compute costs. Despite claims that costs will come down, the expense of running large language models continues to rise, with companies like OpenAI and Anthropic burning billions of dollars, and it's unclear how they can become profitable without a significant change in their business model.

Read more

Should we revisit Extreme Programming in the age of AI?

The rapid advancement of software development tools, including AI-enabled code generation, has not led to better delivery outcomes, with many projects still underdelivering and overrunning budgets. The solution lies not in further acceleration, but in implementing smarter constraints, such as those found in Extreme Programming (XP), which emphasizes human-centered practices, collaboration, and feedback to ensure that software is built with intent and meets real-world needs.

Read more

Why RDF is the natural knowledge layer for AI systems

Knowledge graphs can triple the accuracy of large language models (LLMs) on enterprise data, as they provide a structured and semantic representation of information that aligns with how LLMs process information. However, many teams that build custom knowledge graphs without using established RDF standards often end up rebuilding the same features and capabilities, leading to unnecessary complexity and costs, and ultimately converging on the same patterns and solutions as RDF.

Read more

How Developers Wield Agentic AI in Real Software Engineering Tasks

Software Engineering Agents (SWE agents) can perform development tasks, but struggle with complex real-world tasks, leading to a collaborative approach with developers to solve problems. A study of 19 developers using an in-IDE agent to resolve open issues found that incremental and interactive collaboration with the agent led to greater success, but also highlighted challenges in trusting the agent's responses and collaborative debugging and testing.

ArXiv

Fantastic pretraining optimizers and where to find them

AdamW remains the dominant optimizer in language model pretraining despite claims of alternative optimizers offering speedups, largely due to unequal hyperparameter tuning and limited evaluation setups. A systematic study of ten deep learning optimizers found that while some alternative optimizers offer speedups, they decrease with model size, and that rigorous hyperparameter tuning and evaluations are necessary to make fair comparisons, with matrix-based optimizers offering the most significant speedups, albeit decreasing with model scale.

ArXiv

Ultra Ethernet's Design Principles and Architectural Innovations

The Ultra Ethernet (UE) 1.0 specification introduces a new High-Performance Ethernet standard for AI and HPC systems, featuring the innovative Ultra Ethernet Transport (UET) protocol for fast and efficient communication. UE builds upon the established Ethernet ecosystem, offering significant improvements over previous standards like InfiniBand, and is designed to deliver a new era of high-performance networking with vastly improved computational efficiency.

ArXiv

Fleeting memory improves language learning but impairs reading time prediction

Human memory's fleeting nature may actually aid in language learning, a theory supported by experiments on transformer language models that showed improved language learning with simulated fleeting memory. However, this improvement came at the cost of worse prediction of human reading times, suggesting that memory limitations benefit language learning but not necessarily the modeling of human behavior.

ArXiv

Do Language Models Agree with Human Perceptions of Suspense in Stories?

Researchers replicated psychological studies on suspense using large language models (LMs) and found that while LMs can identify texts intended to induce suspense, they cannot accurately estimate the relative amount of suspense or capture its progression like human readers. The study concluded that LMs process suspense differently than humans, despite being able to superficially identify certain suspenseful elements, and that their understanding of suspense diverges from human perceptions.

ArXiv

Show HN: Open-sourcing our text-to-CAD app

CADAM is an open-source text-to-CAD web app that uses AI to transform natural language and images into 3D models, featuring parametric controls, multiple export formats, and a browser-based interface. The app is built with React, Three.js, OpenSCAD WebAssembly, and Supabase, and allows users to create and edit 3D models in a web browser, with features like real-time preview, parameter extraction, and custom fonts.

Repo

Show HN: I Built Open Source Claude Code Alternative, but Better at API Testing

Ani Code is an open-source alternative to Claude Code, designed for better API testing, providing features such as QA testing, security scans, API exploration, and persistent agent memory. It works with various AI providers, including OpenRouter, OpenAI, Anthropic, and Google Gemini, and offers a range of tools and commands to simplify testing, security, and development workflows.

Repo

Smart Speed – adjusts YouTube playback speed based on real-time audio using AI

YouTube Smart Speed is a Firefox extension that uses AI to adjust YouTube playback speed in real-time, speeding up during silence and slowing down during speech, with customizable settings and manual override options. The extension utilizes web audio analysis, machine learning, and neural networks to intelligently control playback speed, and allows users to customize parameters such as minimum and maximum speed, volume threshold, and analysis frequency.

Repo

LLM Evaluation via Rap Battles

The Great LLM Rap-off is a tournament where various language models compete against each other through rap battles, with the goal of evaluating their verbal intelligence and generative poetry capabilities. The competition features numerous models from different companies, such as OpenAI, Google, and Microsoft, with each model being ranked based on its performance in the battles.

Repo

Awesome AI Agent Frameworks

This repository reviews and compares various AI agent frameworks, evaluating them based on reliability, ergonomics, deployment, and ecosystem fit, with top recommendations including Pydantic AI, OpenAI Agents SDK, and CrewAI. These frameworks are assessed on their strengths, such as type safety, observability, and modularity, and are suited for different use cases, including production-grade systems, multi-agent workflows, and practical assistants.

Repo