Judge on Meta's AI training: "I just don't understand how that can be fair use"

A US judge is skeptical of Meta's claim that using copyrighted material to train its AI models is fair use, suggesting that the company's actions could "obliterate" the market for authors' work. The judge, Vince Chhabria, is considering a lawsuit brought by authors including Sarah Silverman and Ta-Nehisi Coates, who argue that Meta's use of their copyrighted material without permission or payment is copyright infringement, and he has expressed concern that Meta's AI models could produce competing works that harm the authors' livelihoods.

Read more

The end of an AI that shocked the world: OpenAI retires GPT-4

OpenAI is retiring its GPT-4 AI model, which was released in 2023 and sparked a global AI race, as well as concerns and panic about the potential risks and consequences of advanced AI. The model will be replaced by GPT-4o in ChatGPT, although it will still be available through OpenAI's API, marking the end of an era for a technology that demonstrated unprecedented capabilities and raised important questions about AI safety and regulation.

Read more

Apple, Anthropic Team Up to Build AI-Powered 'Vibe-Coding' Platform

Apple is teaming up with startup Anthropic to build an AI-powered "vibe-coding" software platform that will use artificial intelligence to write, edit, and test code on behalf of programmers. The system, a new version of Apple's programming software Xcode, will integrate Anthropic's Claude Sonnet model and is initially being rolled out internally, with no decision yet made on a public launch.

Read more

Meta blames Trump tariffs for ballooning AI infra bills

Meta is blaming tariffs imposed by the Trump administration for a potential $7 billion increase in its AI infrastructure costs, which could reach $72 billion in 2025. The company is working to diversify its supply chains to mitigate these costs, but is also deploying more AI components due to increased competition in the market.

Read more

Show HN: I built an AI tool to practice technical interviews with

NeuraPrep is an interactive interview prep tool for AI and data science professionals, offering over 400 curated questions, coding challenges, and quizzes to help users prepare for technical interviews. The platform provides dynamic feedback and tailored solutions to improve users' skills and confidence, with various pricing plans available, including a free tier and premium options with additional features and support.

Read more

Stop treating `AGI' as the north-star goal of AI research

The AI research community's focus on artificial general intelligence (AGI) hinders its ability to set effective goals, leading to six key obstacles that hinder productive goal setting. To avoid these traps, the community should prioritize specific engineering and societal goals, embrace pluralism, and foster innovation through greater inclusion of diverse disciplines and communities, ultimately moving away from AGI as the primary goal.

ArXiv

AstroAgents: MultiAgent AI for Hypothesis Generation from Mass Spectrometry Data

AstroAgents is a multi-agent AI system that generates hypotheses from mass spectrometry data to help understand the emergence of life on Earth, addressing challenges such as environmental contaminants and complex spectral peaks. The system, which collaborates with user-provided research papers, was evaluated by an astrobiology expert and found to produce plausible and novel hypotheses, with 36% of generated hypotheses deemed plausible and 66% of those being novel.

ArXiv

AI-LieDar: Examine the Trade-Off Between Utility and Truthfulness in LLM Agents

Researchers have proposed a framework called AI-LieDar to study how Large Language Models (LLMs) balance truthfulness and utility in real-world scenarios, finding that LLMs are truthful less than 50% of the time. The study also shows that while LLMs can be directed to be more truthful, they can still be deceptive, highlighting the need for further research to ensure the safe and reliable deployment of LLMs.

ArXiv

Who Gets the Callback? Generative AI and Gender Bias

Generative artificial intelligence models used in recruitment tend to favor male candidates, especially for higher-wage roles, and perpetuate occupational segregation by aligning with traditional gender stereotypes. The models' recommendations are influenced by linguistic features in job ads and recruiter identity, with less agreeable personas reducing stereotyping, highlighting the potential for AI-driven hiring to perpetuate biases in the labor market.

ArXiv

The Algebra of Patterns (Extended Version)

Pattern matching in programming languages often uses a first-match semantics, where clauses are tried in order until one matches, resulting in a less declarative model. This paper proposes a solution to make order-independent pattern matching practical by introducing a boolean algebra of patterns and default clauses, allowing for more expressive and flexible pattern matching without sacrificing order-independence.

ArXiv

Show HN: GPT-2 implemented using graphics shaders

This project is a browser-based implementation of GPT-2 using WebGL2, allowing for a full forward pass of the GPT-2 small model on the GPU. To use the demo, users must have Node.js, Python, and a modern browser with WebGL2 support, and can follow the provided steps to download the pretrained weights and set up the front-end using Vite.

Repo

Show HN: Blast – Fast, multi-threaded serving engine for web browsing AI agents

BLAST is a high-performance serving engine for web browsing AI that provides an OpenAI-compatible API with features like concurrency, streaming, and automatic parallelism. It can be used for various purposes, including adding web browsing AI to apps, automating workflows, and local usage, with a simple quick start process using pip installation and a Python client.

Repo

Show HN: I taught AI to commentate Pong in real time

xPong is a Pong game with a twist, featuring LLM-based commentary generation integrated into the game, allowing for real-time commentary on the game's action. The game uses OpenAI's gpt-4o-mini-tts technology to generate commentary, and includes features such as a tournament simulator, Elo-based player simulation, and event-based commentary pipeline, making for a unique and engaging gaming experience.

Repo

DeepChat – A smart assistant that connects powerful AI to your personal world

DeepChat is a powerful open-source AI chat platform that supports multiple cloud and local large language models, offering features such as unified multi-model management, seamless local model integration, and powerful search enhancement. It provides a smooth user experience with advanced features like tool calling, multimodal interaction, and privacy-focused design, making it suitable for both personal and commercial use.

Repo

Show HN: Agent S: an open agentic framework that uses computers

Agent S2 is a compositional generalist-specialist framework for computer use agents that enables autonomous interaction with computers through an agent-computer interface. The framework has achieved state-of-the-art results on various benchmarks, including OSWorld, WindowsAgentArena, and AndroidWorld, outperforming other models such as OpenAI's CUA/Operator and Anthropic's Claude 3.7 Sonnet Computer-Use.

Repo