One in 20 new Wikipedia pages seem to be written with the help of AI

Nearly 5 per cent of new Wikipedia pages published in English since ChatGPT's release seem to contain AI-generated content, which could reduce the site's reliability. Researchers at Princeton University used AI detection tools to identify the AI-written content, highlighting the need for Wikipedia editors to be vigilant.

Read more

AI prefers white and male job candidates in new test of resume-screening bias

A new study from the University of Washington found that AI-powered resume-screening tools overwhelmingly prefer white and male job candidates, with the technology favoring resumes from white-associated names 85% of the time and female-associated names 11% of the time. The researchers tested three open-source language models and found that Black men fared the worst, with the models preferring other candidates nearly 100% of the time.

Read more

Anthropic calls Government to regulate AI in the next eighteen months

Governments should take urgent action on AI policy within the next 18 months to mitigate catastrophic risks while supporting innovation in the field. The rapid progress of AI systems in areas such as math, reasoning, and coding has brought both benefits and risks, including potential misuse in domains like cybersecurity and biology.

Read more

How I use LLM to scrape 99% of websites [video]

This text appears to be a YouTube video page, but it does not contain any actual content or information about the video. The page seems to be experiencing errors, with missing information and broken links.

Read more

We put 1M files into DVC, Git-LFS, and Oxen.ai

Oxen.ai outperformed other tools in a benchmark test, uploading 1 million+ images from the ImageNet dataset in 1 hour and 30 minutes, while also allowing for data viewing and collaboration. This is significantly faster than other methods, including Git-LFS, which took 20 hours, and offers advantages over traditional version control systems like Git, which are not well-suited for large datasets.

Read more

Interpreting Affine Recurrence Learning in GPT-Style Transformers

Researchers investigated how GPT-style transformers perform in-context learning, specifically their ability to learn and predict affine recurrences, and found that the model uses a copying mechanism and negative similarity heads to form and refine its predictions. This study contributes to a deeper understanding of transformer behaviors and offers potential avenues for improving AI alignment through mechanistic interpretability.

ArXiv

Length-Induced Embedding Collapse in Transformer-Based Models

Researchers have identified a phenomenon called Length Collapse, where longer text embeddings collapse into a narrow space, hurting performance in downstream tasks. They propose a solution called TempScale, which introduces a temperature in softmax() to mitigate this issue, and demonstrate its effectiveness in improving existing embedding models, especially on long text inputs.

ArXiv

Voice-Enabled AI Agents Can Perform Common Scams

Researchers have demonstrated that voice-enabled AI agents can be used to autonomously perform common scams, raising concerns about their potential misuse. The study found that these agents can successfully carry out the actions necessary to execute scams, highlighting the need for caution in their widespread deployment.

ArXiv

An embarrassingly simple approach to recover unlearned knowledge for LLMs

Large language models (LLMs) can acquire unwanted behaviors from their training data, and machine unlearning has been proposed as a solution to remove this problematic content. However, research has found that applying quantization to models that have undergone unlearning can restore the supposedly "forgotten" information, with up to 83% of the intended forgotten knowledge retained after 4-bit quantization.

ArXiv

Creating Interactive and Embedded Physics Simulations from Static Textbooks

Augmented Physics is a web-based tool that uses machine learning and computer vision to convert static textbook diagrams into interactive physics simulations, allowing for more engaging and personalized learning experiences. The system enables users to extract diagrams from textbooks and generate interactive simulations, which can be integrated into scanned textbook pages to cover various physics concepts.

ArXiv

Project Sid: Many-agent simulations toward AI civilization

Researchers have developed Project Sid, a large-scale simulation involving 10-1000+ AI agents interacting in a Minecraft environment, demonstrating their ability to progress and develop complex societies. The agents, using the PIANO architecture, autonomously developed specialized roles, collective rules, and cultural transmission, achieving significant milestones towards AI civilizations.

Repo

gptel: a simple LLM client for Emacs

The provided text is incomplete and only contains an error message. There is no information to summarize.

Repo

Show HN: A browser extension for Claude/ChatGPT to edit your projects locally

The CodeSpin.AI Chrome Extension allows users to edit local projects using Claude and ChatGPT through the File System APIs on Chrome. To install, users must manually clone the project from GitHub, install dependencies, build the extension, and then load it as an unpacked extension in Chrome.

Repo

A declarative language to build LLM applications

GenSphere is a declarative framework for building Large Language Model (LLM) applications, allowing users to define tasks and connections in YAML files, and share or pull projects from an open platform. It provides features such as low-level control, portability, community collaboration, and composability, making it easier to build and share complex LLM applications.

Repo

SurfSense: Personal AI Assistant for Internet Surfers and Researchers

SurfSense is a personal AI assistant that allows users to save and recall content from the internet or their own files, including web browsing sessions, social media chats, emails, and documents. It features a cross-browser extension, file uploading support, powerful search, and natural language interaction with saved content, making it easy to capture and recall information.

Repo