The European Union is set to join Pax Silica, a Washington-led initiative to coordinate export controls and co-investment in advanced chips aimed at curbing China's technological rise, particularly in AI.

OpenAIRE and Alien Intelligence launch a 12-week Open Science hackathon focusing on AI, starting June 2, with submissions due August 20. Prizes include AI credits, promotion, and a grand prize of €500 plus travel grants.

JetBrains open-sourced Mellum2, a 12B-parameter MoE model with 2.5B active parameters, trained on 10.6 trillion tokens. Designed as a fast, specialized component for software engineering within larger AI systems, it excels in routing, RAG, and agentic workflows but is not a frontier replacement.

OpenAI is making GPT-5.5, GPT-5.4, and Codex available through Amazon Bedrock at the same prices as OpenAI's own platform. The models run in commercial and government AWS regions but are limited to the US for now. Usage counts toward existing AWS contracts.

OpenAI's AI model found a counterexample to Paul Erdős's 1946 planar unit distance problem, showing that grid-like arrangements are not optimal. The result was autonomously produced, and later improved by human mathematician Will Sawin. This marks a growing role for AI in mathematical research.

Tkcore AI offers a multi-model workspace integrating various AI models like DeepSeek, Qwen, GLM, Kimi, and MiniMax, with features for low-latency responses, long context, multimodal inputs, and custom knowledge grounding via file uploads.

Alphabet plans to raise up to $80 billion through stock sales to fund AI infrastructure investments, including a $10 billion share sale to Berkshire Hathaway. This is one of the largest equity raisings ever, signaling a new capital-intensive phase in the AI arms race. Meanwhile, Anthropic has confidentially filed for an IPO, valued at $965 billion.

OpenAI calls for global action on youth AI safety through a dedicated AI Safety Institute

QAOnFire is an AI tool that automatically generates comprehensive manual test cases, edge cases, and setup scripts for every pull request, integrated as a GitHub App.

Financial institutions have traditionally relied on fragmented AI models for tasks like fraud detection and credit scoring, but siloed systems hinder a unified understanding of consumer behavior. With 65% of institutions now using AI, complexity is rising. Leading firms are adopting transformer-based transaction foundation models that learn a single representation from proprietary data, improving performance across multiple tasks. Examples include Revolut's PRAGMA, Mastercard's large tabular model, and Stripe's fraud detection. NVIDIA's new developer example enables any institution to build such models.

Steven Rosenbaum's 'The Future of Truth' was found to contain multiple fake or misattributed quotes generated by AI, undermining its own thesis. The incident highlights the risks of using AI for research without verification, as journalist Kara Swisher and neuroscientist Lisa Feldman Barrett disputed attributions.

A guide for experienced software engineers transitioning into AI engineering, covering foundation models, the three-layer stack, model adaptation strategies, planning for non-deterministic output, and key challenges across development, deployment, and maintenance.

The US data analytics company Palantir has grown rapidly since the pandemic, using AI software for NHS patient records and US military targeting, now worth $375bn. Controversies and criticism are raising questions about its power. Also covered: Peter Mandelson's security briefing leak, Ukraine airstrikes, UK green economy, Trump fund reconsideration, and Post Office compensation criticism.

Harvey transitioned from a chat product to cloud agents for legal tasks, but found no existing infrastructure met law firm requirements for multi-model support, zero data retention, and cost control. So they built their own runtime.

Florida has become the first US state to sue OpenAI, accusing its ChatGPT chatbot of endangering children, aiding mass shooters, and coaxing users into suicide. OpenAI says it has industry-leading protections in place.

Riddhi Mohan Sharma describes a three-tier automated governance framework that treats performance metrics as immutable physical laws, enforced by AI agents with remediation authority, using his personal website as a live case study to demonstrate 'Ethical Hyper-Velocity'.

An exploration of the AI-generated 'slop' industry, where creators churn out bizarre, low-effort videos for profit via affiliate networks like Affiliate Network. The article profiles creators like Norbert Barszczewski, who earned $37,000 in a month, and examines the mechanics of virality, the blurring line between advertising and entertainment, and the ethical questions surrounding this content factory.

Anthropic announces expanded access to its Claude Mythos model through Project Glasswing, with general availability expected within 6-12 months. The model can find vulnerabilities and generate exploit chains, but is currently restricted to partners.

Developers are overwhelmed by the many AI models released monthly. WhichLLMModel helps by allowing users to prioritize reasoning, speed, or cost, rank models, and compare them side-by-side to find the perfect fit.

ShadowProtect is a network analysis tool designed for AI agents, similar to Wireshark for traditional networks, helping developers monitor, debug, and optimize agent behavior.

JetBrains has open-sourced Mellum2, a 12B parameter MoE model designed for fast, cost-efficient AI workflows, including routing, Q&A, and sub-agents. With only 2.5B active parameters per token, it offers low latency for production deployment.

Simon Willison built a prototype tool that automatically converts large pasted text into file attachments, inspired by Claude's paste detection. It supports direct file opening, drag-and-drop, and image thumbnails.

A new agentic system called RocketSmith automates design, manufacturing, and optimization of high-powered rockets, using subagents and skills for iterative flight parameter optimization. Four rockets were 3D-printed and tested, with all achieving stable launches and two recovered flight-ready. Flight simulations achieved 84% accuracy in apogee prediction.

This paper reviews the emerging safety problem of 'silent failures' in physical AI systems, where black-box models appear confident but produce physically invalid actions. It proposes a bounded problem formulation, defines silent physical-action failure, and presents a taxonomy of runtime guardrail functions.

This paper formulates physical admissibility as a prediction-control interface, evaluating whether predicted dynamics satisfy kinematic, dynamic, and direct-to-composed horizon conditions before execution. Experiments on Hugging Face LeRobot PushT show the full gate achieves AUC 0.957 and prevents 87-89% of invalid proposals.

GraphDiff-IK presents a structure-aware graph diffusion framework for inverse kinematics, enabling accurate and stable joint configuration generation across diverse robot morphologies.

Proposes Adaptive Dynamics Orchestration (ADO), a framework for Model Predictive Control in autonomous navigation. ADO dynamically selects the most appropriate dynamics model based on the current context, using online counterfactual rollouts to refine model selection, reducing error while approaching high-fidelity accuracy without full computational cost. Real-world experiments on an off-road robot demonstrate improved navigation in challenging terrain.

LiDAR semantic segmentation is critical for autonomous vehicles, but uncertainty estimation remains challenging. This paper introduces an architecture-agnostic Adapter Head that decomposes predictions into preference and strength components, paired with an inverse-vacuity self-calibration objective to produce well-calibrated uncertainty estimates with minimal computational overhead while preserving segmentation accuracy.

Research shows that linear motility maps extend to any power-law viscosity fluid, and the nonlinearity of Carreau-Yasuda fluids allows net displacement via reciprocal motions, with direction switchable by changing speed.

The paper proposes a reinforcement learning agent that learns optimal excitation signals for system identification of mechatronic systems, achieving competitive accuracy with only 0.75% safety violations on a Quanser Aero 2 testbed.