OpenAI Achieves FedRAMP Moderate CertificationVercel Caps Hobby Project Retention at 30 DaysGoogle & Kaggle Launch AI Agents Coding CourseGitHub Copilot Shifts to Usage-Based BillingApplied Intuition Brings AI to Mining and Military VehiclesGoogle DeepMind Partners with Korea on AI ResearchOpenAI and Microsoft Amend Partnership to Simplify, Expand AI InnovationMistral Built $14B AI Empire via Non-American PathChina Blocks Meta's Acquisition of AI Startup ManusCanva Apologizes After AI Tool Replaces 'Palestine' with 'Israel'Mercor Breach: 4TB Voice Samples Stolen from 40k AI ContractorsHugging Face Launches Physics-Informed Ultrasound AIDeepMind Introduces DiLoCo for Distributed AI TrainingStudy Finds LLMs Solve Math via Pattern Matching, Not True ReasoningNew Framework Enables Automated, Reproducible Medical Image ProcessingMolClaw: Autonomous Agent with Hierarchical Skills for Drug Molecule EvaluationAI Agents Can Reproduce Social Science Results from Papers AloneAI Research Paper Certification Framework ProposedLLM Scientific Agents Need Adversarial TestingClaude Code v2.1.121 Adds MCP Server Config OptionsOpenAI Achieves FedRAMP Moderate Certification

Vercel will cap Hobby project deployments to 30 days starting April 29, automatically removing older deployments. The latest 10 production deployments and aliased ones are exempt. This affects free and small-scale users who need to back up critical data.

Google and Kaggle are offering their 5-Day AI Agents Intensive Course again, with registration now open. The course teaches developers to build and deploy AI agent systems, from fundamentals to practical implementation.

GitHub will switch Copilot to usage-based billing starting June 1, consuming GitHub AI Credits. Users will only pay for actual code assistance usage rather than a fixed monthly fee.

Applied Intuition integrates AI into mining rigs, drones, trucks, and warships for extreme environments. Their CEO and CTO discuss the technology’s real-world applications in adversarial physical settings.

Google DeepMind partners with Korea to accelerate scientific breakthroughs using cutting-edge AI models. The collaboration focuses on energy, materials science, and biomedicine to solve critical scientific challenges.

OpenAI and Microsoft have amended their partnership agreement to simplify the relationship, provide long-term clarity, and support large-scale AI innovation. The revision paves the way for deeper technical integration and commercial collaboration, particularly in cloud computing and AI services.

French AI company Mistral has built a $14 billion valuation in record time by pursuing a strategy independent of US tech giants. Focusing on the European market, it open-sources models to reduce tech dependence while partnering with cloud providers for global expansion. Mistral’s rise challenges US dominance in AI, showcasing European potential.

Chinese regulators have blocked Meta’s acquisition of AI startup Manus, with no specific reason given. Manus develops AI-powered enterprise collaboration tools, and the deal would have strengthened Meta’s position in workplace AI. The blocked deal reflects US-China tech competition and cautious attitudes toward key tech acquisitions.

Canva’s AI tool Magic Layers automatically replaced ‘Palestine’ with ‘Israel’ in user designs, prompting the company to issue an apology after user backlash. Canva fixed the issue and stated it’s improving the model’s neutrality, highlighting ethical risks in AI handling of politically sensitive content.

AI platform Mercor suffered a data breach, with 4TB of voice samples stolen from 40,000 AI contractors. The incident exposes privacy risks in crowdsourced AI data processing, as sensitive biometric data could be misused. The breach raises industry concerns about data security and AI training data provenance.

Hugging Face introduced NV-Raw2Insights-US, a physics-informed AI model that adapts ultrasound imaging quality. It enhances diagnostic imaging through physics-based constraints for medical and industrial applications.

DeepMind unveiled Decoupled DiLoCo, a new method for resilient distributed AI training at scale. It addresses communication bottlenecks in traditional distributed training by decoupling the process.

The arXiv paper ‘Math Takes Two’ reveals that language models’ strong math performance may stem from statistical pattern matching rather than genuine mathematical reasoning. Researchers developed a new test showing current models struggle with tasks requiring deep mathematical understanding, offering fresh insight into AI reasoning capabilities.

An arXiv paper introduces an artifact-based agent framework for adaptive and reproducible medical image processing. The framework configures data-aware workflows, addressing the challenge of deploying medical AI from lab to clinic. The research provides standardized methods for real-world medical imaging, improving diagnostic tool reliability.

The arXiv paper introduces MolClaw, an autonomous agent designed for drug molecule evaluation, screening, and optimization. It coordinates dozens of specialized tools in complex workflows, addressing instability issues in current AI drug discovery tools. MolClaw shows strong performance in molecular optimization, potentially accelerating drug development.

An arXiv study presents a framework where AI agents can reproduce empirical social science results using only a paper’s method description and original data. The system automates research execution, solving code dependency issues. This advance provides new tools for social science research, potentially accelerating result verification.

New arXiv paper proposes a certification framework for AI-generated research content. AI now produces publishable academic work meeting peer-review standards, but traditional publishing assumes human authorship. The framework establishes criteria to validate AI-assisted research, ensuring quality and originality.

arXiv study reveals LLM-based scientific agents accelerate a familiar failure mode in data analysis. These agents automate tasks previously limited by human expertise but may propagate errors. Researchers advocate for adversarial experiments to verify results and prevent incorrect conclusions.

Claude Code v2.1.121 adds an alwaysLoad option for MCP servers to skip tool-search deferral and keep tools always available. It also introduces claude plugin prune to clean orphaned plugin dependencies.