Demand for human radiologists is at an all-time high

Works in Progress is launching a print magazine, with the first issue available in November, and despite predictions that AI would replace human radiologists, demand for them is higher than ever, with record numbers of residency positions and high salaries. This is because AI models, while able to detect certain diseases with greater accuracy than humans, struggle to replicate this performance in real-world hospital conditions, face regulatory hurdles, and only replace a small part of a radiologist's job, which includes tasks like communicating with patients and clinicians.

Read more

Accenture to 'exit' staff that cannot be retrained for age of AI

Accenture plans to "exit" staff who cannot be retrained for the age of artificial intelligence. The company is offering a severance package to employees who are unable to adapt to new technologies, as part of its efforts to prepare its workforce for an AI-driven future.

Read more

Cloudflare Introduces NET Dollar

Cloudflare is introducing the NET Dollar, a US dollar-backed stablecoin that will enable instant and secure transactions for the "agentic web", supporting a new business model for the AI-driven internet. The NET Dollar aims to power global transactions at internet speed, rewarding originality, sustaining creativity, and enabling innovation in an AI-driven world where humans interact with the web through autonomous AI agents.

Read more

Gemini Robotics 1.5 brings AI agents into the physical world

DeepMind has introduced two new models, Gemini Robotics 1.5 and Gemini Robotics-ER 1.5, which enable robots to perceive, plan, think, use tools, and act to better solve complex tasks. These models work together to allow robots to understand their environment, think before taking action, and learn across different embodiments, making them more capable and versatile in completing complex, multi-step tasks.

Read more

Factory Raises $50M Series B

Factory has raised $50M in Series B funding at a valuation of $300M, led by prominent investors such as NEA and Nvidia, to further develop its AI-powered software development platform. The company's Droids have achieved the top ranking on Terminal Bench, a benchmark for general software development, and are now available to anyone, with any model, in any interface, providing developers with greater choice and flexibility.

Read more

Bit is all we need: binary normalized neural networks

Researchers have developed a new type of neural network layer, called binary normalized layers, which use only single-bit parameters, allowing for models that use 32 times less memory than current models without sacrificing performance. These binary normalized layers can be easily implemented on standard computers and enable the deployment of large neural network models on simple and affordable hardware, such as mobile devices or CPUs.

ArXiv

The Memory Paradox: Why Our Brains Need Knowledge in an Age of AI

Heavy reliance on AI systems and digital tools may impair human memory and cognitive development, particularly in regards to expertise, critical thinking, and long-term retention. The paper argues that over-reliance on AI can hinder the formation of strong internal models and neural connections, and instead proposes a balanced approach to human-AI interaction that fosters both technological proficiency and robust cognitive abilities.

ArXiv

Ransomware 3.0: Self-Composing and LLM-Orchestrated

Ransomware 3.0 is a new threat model that utilizes large language models to autonomously plan and execute ransomware attacks, generating polymorphic variants that adapt to the environment at runtime. This AI-enabled ransomware can perform reconnaissance, payload generation, and extortion without human involvement, and has been shown to be effective across various environments, highlighting the need for improved defenses against such attacks.

ArXiv

Multi-Modal vs. Text-Based: Benchmarking LLM Strategies for Invoice Processing

Eight multi-modal large language models were benchmarked on three invoice document datasets using zero-shot prompting, comparing direct image processing to a structured parsing approach. The results showed that native image processing generally outperformed structured approaches, with performance varying across models and document characteristics, providing insights for selecting suitable models and processing strategies for automated document systems.

ArXiv

A Software Engineering Analysis of the XZ Utils Supply Chain Attack

An estimated 90% of modern applications contain open-source components, but this has also created security vulnerabilities, particularly in under-resourced projects. A recent attack on the XZ Utils project exploited the open-source development process to inject a backdoor into a fundamental Linux compression library, demonstrating a new breed of supply chain attack that manipulates software engineering practices to establish legitimacy and control.

ArXiv

TallMountain – Stoic Virtue Ethics for an LLM Agent

TallMountain-Raku is an AI agent framework in Raku that incorporates a formal machine ethics system based on a Stoic-inspired normative calculus, ensuring that all outputs are filtered through its ethical reasoning system. The framework integrates with a Large Language Model (LLM) for natural language understanding and generation, and can be run using Docker or directly, with configuration options for ethical rules and threat scanner thresholds.

Repo

ai.robots.txt – A list of AI agents and robots to block

The ai.robots.txt project provides a list of AI-related crawlers and tools to block them from accessing websites, including a robots.txt file and configuration files for various web servers. The project encourages contributions and provides resources for implementing the list, as well as links to additional information and tools for blocking AI bots and crawlers.

Repo

A C++ library to efficiently run language models across edge platforms

LiteRT-LM is a C++ library that enables efficient execution of language models across various edge platforms, including Android, macOS, Windows, and Linux, with support for CPU, GPU, and NPU acceleration. The library is currently in early preview, with its first full release expected late summer or early fall, and offers a range of features, including a flexible API, cross-platform support, and customizable models.

Repo

Show HN: SQLite-RAG – A semantic search engine built on top of SQLite

SQLite RAG is a hybrid search engine built on SQLite that combines vector similarity search with full-text search, allowing for optimal document retrieval. It features a command-line interface, support for multiple file formats, and customizable configuration options, making it easy to use and integrate into various applications.

Repo

Verifiers: Environments for LLM Reinforcement Learning

Verifiers is a library of modular components for creating reinforcement learning (RL) environments and training large language models (LLMs), allowing for the creation of custom environments and integration with various RL frameworks. The library provides a range of tools and features, including an async GRPO implementation, support for large-scale training, and easy integration with OpenAI-compatible models, making it a reliable toolkit for building and training LLMs.

Repo