AI, Wikipedia, and uncorrected machine translations of vulnerable languages

Wikipedia editions for low-resource languages are being flooded with poor-quality machine translations, often from non-speakers. Since Wikipedia is a primary data source for training LLMs on these languages, this creates a "garbage in, garbage out" cycle where models learn from and perpetuate errors. This linguistic doom loop degrades translation quality and threatens the digital viability of vulnerable languages, leading to extreme cases like the shutdown of the Greenlandic Wikipedia.

Read more

Armed police handcuff teen after AI mistakes chip bag for gun in Baltimore

An AI weapon detection system generated a false positive, mistaking a student's crisp packet for a gun and triggering an armed police response. A failure in the human-in-the-loop workflow occurred when the school principal missed the cancellation notice from the system's human reviewers and contacted police. The system's vendor stated the technology "operated as designed," highlighting the gap between system function and real-world operational procedure.

Read more

Deepagent: A powerful desktop AI assistant

DeepAgent is an AI platform that builds a wide range of solutions from prompts. Its capabilities include generating full-stack web applications with database and payment integration, creating RAG-powered chatbots from documents, and developing features directly in code repositories by raising PRs. The platform also automates complex tasks like data analysis, financial modeling, and generating reports or AI videos.

Read more

The design space of AI coding tools

A recent paper analyzes 90 AI coding assistants to define a design space with 10 key dimensions, tracing the UI evolution from autocomplete to chat and agents. The study finds industry is converging on polished, high-speed features, while academia explores novel interactions and explainability. The paper concludes by mapping these design trade-offs to six user personas, emphasizing the challenge of shaping how these tools fit real workflows while keeping humans in the loop.

Read more

Is AI's Circular Financing Inflating a Bubble? [video]

The video analyzes the potential for an AI bubble driven by circular financing, where major tech companies invest in AI startups that in turn spend the capital on their investors' compute services. This self-reinforcing loop raises questions about the true demand for hardware from companies like Nvidia and the systemic risks associated with the massive infrastructure and energy requirements. The analysis explores whether this financial structure could collapse under its own weight.

Read more

SynthID-Image: Invisibly Watermarking AI-Generated Imagery

SynthID-Image is a deep learning system for invisibly watermarking AI-generated imagery, already deployed at internet scale across billions of Google images. The paper details the threat models and practical challenges of such a large-scale media provenance system. An external variant, SynthID-O, is benchmarked against other post-hoc methods, demonstrating state-of-the-art robustness to common perturbations while maintaining high visual quality.

ArXiv

LLM-empowered knowledge graph construction: A survey

This survey provides a comprehensive overview of how LLMs are driving a paradigm shift in Knowledge Graph (KG) construction, moving from traditional rule-based pipelines to language-driven, generative frameworks. It systematically analyzes the impact of LLMs on the classical stages of ontology engineering, knowledge extraction, and knowledge fusion, categorizing new approaches into schema-based and schema-free paradigms. The paper also outlines future research directions, including KG-based reasoning for LLMs and dynamic knowledge memory for agentic systems.

ArXiv

A Framework to Support Technical Assessment in AI Regulatory Sandboxes

To address fragmented and non-standardized AI assessment within the EU's AI Regulatory Sandboxes (AIRS), a new modular, open-source framework called the Sandbox Configurator is proposed. It enables users to generate custom testing environments from a shared library of domain-relevant tests via a plug-in architecture. The framework aims to streamline compliance for AI providers and create a standardized, interoperable ecosystem for trustworthy AI governance under the supervision of Competent Authorities (CAs).

ArXiv

AI PB: A Grounded Generative Agent for Personalized Investment Insights

AI PB is a production-scale generative agent for retail finance that proactively generates grounded and personalized investment insights. Its architecture features a hybrid retrieval pipeline with a finance-domain embedding model, a multi-stage recommendation engine, and a deterministic orchestration layer that routes between internal and external LLMs based on data sensitivity. Deployed on-premises using vLLM on H100 GPUs, the system demonstrates a method for delivering trustworthy AI insights in a highly regulated environment.

ArXiv

Pico Banana: Large-Scale Dataset for Image Editing by Apple

Pico-Banana-400K is a new 400K-image dataset for instruction-based image editing, created to address the lack of large-scale, high-quality resources built from real images. Generated from the OpenImages collection using the Nano-Banana model, it is curated with a fine-grained taxonomy and MLLM-based quality scoring to ensure diversity and instruction faithfulness. The dataset includes specialized subsets for multi-turn sequential editing, preference learning for alignment and reward model training, and instruction rewriting, providing a robust resource for training and benchmarking next-generation models.

ArXiv

Show HN: Diagram as code tool with draggable customizations

oxdraw is a tool that combines Mermaid's declarative syntax with a visual GUI editor for fine-grained diagram customization. It addresses the common workflow of exporting AI-generated .mmd files to separate tools for polishing by allowing direct manipulation of node positions, connector paths, and styles. All visual edits are saved back into the source .mmd file as declarative comments, ensuring diagrams remain versionable and reproducible.

Repo

SPL – AI patterns with verifiable truth scores

Semantic Pattern Language (SPL) is a verifiable language for composing AI systems from layered patterns, each defined by a three-part structure: contract for the LLM's goal and output, execution for runtime logic, and guarantees for ethical alignment. It uses a quantum-inspired composition model to contextually select and combine patterns at runtime. This approach creates auditable systems that produce a truth score with every execution, aiming for generalization through composition instead of model retraining.

Repo

Neurosymbolic AI server combining Prolog's symbolic reasoning with MCP

Prolog-MCP is a neurosymbolic AI server that exposes Prolog's symbolic reasoning to LLMs via the Model Context Protocol (MCP). It enables stateful tool use by providing a persistent Prolog session with functions for loading programs, running queries, and managing session persistence. The server utilizes Trealla Prolog within a WASI runtime for efficient, sandboxed execution.

Repo

Show HN: Pyxis CodeCanvas a lightweight, client-side IDE for iPad and browsers

Pyxis is a client-side, zero-setup IDE designed for fast, lightweight coding on web and iPad platforms. It features a custom, non-WASM Node.js/TypeScript runtime that emulates core modules like fs and leverages IndexedDB for performance. The IDE integrates Git with GitHub PAT support and includes an AI assistant for code diff suggestions and in-context editing.

Repo

Show HN: The Slow AI Commons – here's why it might fail

The Slow AI Commons is a project to build a collectively owned "commons of intelligence," aiming to break the historical pattern of developers creating systems that enrich others and automate their own roles. It proposes using latent resources like idle compute to create a fair, distributed AI that cannot be captured by a single entity. The project is in its initial formation stage, prioritizing the definition of its vision and core principles—like forkability and contributor-led governance—before tackling the technical architecture.

Repo