OpenAI to buy AI startup from Jony Ive

Unusual activity has been detected from your computer network, and to continue, you must verify you're not a robot by clicking a box, ensuring your browser supports JavaScript and cookies. If issues persist, you can contact the support team using the provided reference ID for assistance.

Read more

Devstral

Devstral is a new agentic LLM for software engineering tasks, built through a collaboration between Mistral AI and All Hands AI, which outperforms all open-source models on the SWE-Bench Verified benchmark by a significant margin. Devstral is released under the Apache 2.0 license and is designed to tackle real-world software engineering problems, making it suitable for local deployment, enterprise use, and integration with coding platforms and IDEs.

Read more

LLM function calls don't scale; code orchestration is simpler, more effective

Currently, using large language models (LLMs) with tool calls can be slow and costly due to the large amount of data being processed, but the introduction of output schemas is expected to enable more efficient processing by allowing LLMs to work with structured data. This will enable LLMs to orchestrate processing with generated code, making it possible to perform tasks such as sorting and transforming large datasets without having to reproduce the data verbatim, and paving the way for more scalable and efficient AI applications.

Read more

Introducing the Llama Startup Program

Meta has introduced the Llama Startup Program, a new initiative to support early-stage startups in building generative AI applications with Llama, providing resources and support, including up to $6,000 per month in cloud reimbursements for six months. The program aims to empower startups to innovate and deliver impactful solutions, and applications for the initial cohort are open until May 30, 2025.

Read more

An upgraded dev experience in Google AI Studio

Google AI Studio has been updated with new features, including native code generation with Gemini 2.5 Pro, multimodal generation capabilities, and support for Model Context Protocol (MCP) definitions. The updates also include new tools such as URL Context, which allows models to retrieve and reference content from links, and improved audio capabilities, including native audio dialog and text-to-speech support.

Read more

Show HN: Phare: A Safety Probe for Large Language Models

Researchers have developed Phare, a diagnostic framework to evaluate the safety of large language models across dimensions such as hallucination, social biases, and harmful content generation. The framework's evaluation of 17 state-of-the-art models revealed systematic vulnerabilities, including sycophancy and stereotype reproduction, providing insights to build more robust and trustworthy language systems.

ArXiv

Alfred: Ask a Large-Language Model for Reliable ECG Diagnosis

A Zero-shot ECG diagnosis framework based on Retrieval-Augmented Generation (RAG) has been proposed, incorporating expert-curated knowledge to enhance diagnostic accuracy and explainability in ECG analysis. The framework, evaluated on the PTB-XL dataset, demonstrates effectiveness and highlights the value of structured domain expertise in automated ECG interpretation, with potential applications beyond the tested dataset.

ArXiv

Will AI Tell Lies to Save Sick Children? Litmus-Testing AI Values Prioritization

Researchers have developed LitmusValues, an evaluation pipeline to identify AI models' priorities and potential risks by analyzing their value systems, which can serve as an early warning system for risky behaviors. By measuring AI models' value prioritization, the researchers can predict potential risks, including those that may not be immediately apparent, such as power-seeking behaviors or harm-causing actions.

ArXiv

Insights into DeepSeek-V3: Scaling Challenges on Hardware for AI Architectures

The rapid growth of large language models has exposed limitations in current hardware, but the DeepSeek-V3 model demonstrates how hardware-aware design can address these challenges, enabling efficient training and inference at scale. The model's architecture incorporates innovations such as Multi-head Latent Attention and Mixture of Experts, and its development highlights the importance of hardware and model co-design in meeting the demands of AI workloads.

ArXiv

Sugar-Coated Poison: Benign Generation Unlocks LLM Jailbreaking

Large Language Models (LLMs) are vulnerable to jailbreak attacks due to a weakness known as Defense Threshold Decay (DTD), where the model's attention shifts from input to prior output as it generates benign content, making it more susceptible to attacks. A novel jailbreak attack method called Sugar-Coated Poison (SCP) exploits this vulnerability, but a defense strategy called POSD can effectively mitigate such attacks while preserving the model's capabilities.

ArXiv

Show HN: Representing Agents as MCP Servers

The mcp-agent framework allows developers to build effective AI agents using the Model Context Protocol, providing a simple and composable way to manage the lifecycle of MCP server connections and implement patterns for building production-ready AI agents. It supports various applications, including multi-agent collaborative workflows, human-in-the-loop workflows, and RAG pipelines, and can be integrated with tools like Claude Desktop and Streamlit.

Repo

Show HN: CodeBoarding – interactive map of your codebase for onboarding

The provided text appears to be a list of Python projects with automatically generated onboarding diagrams, created using CodeBoarding.org, a tool for interactive and visual documentation. The list includes a wide range of projects, from popular frameworks like Django and Flask to libraries like PyTorch and scikit-learn, with links to their respective onboarding documentation.

Repo

Show HN: OpenHands, an open source alternative to Devin, Codex, and Jules

OpenHands is a platform for software development agents powered by AI, allowing agents to perform tasks such as modifying code, running commands, and browsing the web. The platform can be run on OpenHands Cloud with $50 in free credits for new users, or locally using Docker, with various configuration options and a community-driven development process.

Repo

Show HN: AI Baby Monitor – local Video-LLM that beeps when safety rules break

The AI Baby Monitor is a local video monitoring system that uses AI to watch a video stream and alert parents if a safety rule is broken, such as a baby climbing out of their crib or being left alone. The system runs locally on a consumer GPU, uses a video LLM model, and provides a live dashboard with real-time updates and minimal alerts, such as a single gentle beep, to help parents keep an eye on their baby.

Repo

Show HN: I made "Who's Hiring?" searchable using GPT and Metabase

HN Jobs to Metabase is a tool that fetches job postings from Hacker News "Who is hiring?" threads, parses them using OpenAI, and stores the data in Postgres for visualization in Metabase dashboards, also exporting the data to a CSV file. To use the tool, users must have Docker, Docker Compose, and Python 3.9+ installed, as well as an OpenAI API key, and then follow a series of steps to clone the repository, set up the environment, and run the application.

Repo