ChatGPT Ads expands across Europe – and more advertisers get targeting access
OpenAI says ChatGPT Ads are rolling out to 31 European markets, expanding how advertisers can reach people as they explore, compare options, and make decisions. The move signals a broader shift from experimentation to scaled ad inventory inside generative AI experiences.
Google adds a Gemini student hub with study notebooks, flashcards, and calendar sync
The Verge reports Google is launching a dedicated Gemini student hub built around “study notebooks” that can generate flashcards, quizzes, and practice materials. It also adds richer notebook features like graphs and images, plus the ability to populate study dates and deadlines into Google Calendar from a syllabus.
DiSCO – a black-box defense that fixes unsafe text-to-image prompts without model access
DiSCO (from arXiv) targets a tricky safety gap called “benign adversarial” prompts – inputs that look linguistically safe but still trigger harmful outputs based on a model’s internal learned distribution. The method is plug-and-play at the prompt level: it uses distribution-guided suffix expansion and contrastive scoring against safe/unsafe image pools generated by the target model, reducing attack success rates while preserving semantic fidelity.
Meta launches a Mac app for Meta AI – share your screen, get productivity help
Meta is bringing its AI chatbot to macOS with a new app that can share your active window so it can answer questions, suggest actions, and create content based on what’s on-screen. The launch also aligns with Meta’s push to make Meta AI feel more like a desktop productivity assistant to better compete with rivals that already offer desktop experiences.
Replit expands GPT-5.6 Luna access with Free Mode for turning ideas into working software
Replit introduces Free Mode, powered by GPT-5.6 Luna, so more users can generate and iterate on software without worrying about token costs. For builders, it’s a push toward frictionless prototyping – turning natural-language intent into runnable code inside Replit.
GxP-Agent – reliably generating CDISC-compliant clinical trial datasets using a DAG of agents
This arXiv paper tackles why LLM code generation fails in clinical trial programming – it’s not just accuracy, it’s producing structurally valid, analysis-ready datasets that meet CDISC requirements. GxP-Agent encodes regulatory process ordering as a directed acyclic graph, decomposes the work into domain-specific nodes with validation gates and conditional retries, and reports 100% structural match across multiple runs on CDISC-Bench (including a challenging ADAE setup).
KernelArc – multi-agent framework for autonomous GPU kernel optimization that targets real leaderboards
KernelArc (arXiv) describes a multi-agent system that optimizes GPU kernels across heterogeneous workloads by coordinating specialist agents through shared memory and benchmarking guardrails. The paper claims top ranking across multiple categories on a SOL-ExecBench leaderboard snapshot, including kernels for BF16 GEMM configuration, MoE backward fusion, attention variants, and quantization-related tasks.
Checked agent memory and reasoning workflows in LiquidAI’s QAD experiments
LiquidAI’s post explores how quantitative guardrails can improve question-answering reliability by checking intermediate steps instead of trusting a single generation path. The focus is practical: building systems that can self-validate outputs and reduce silent failure modes when models drift from expected reasoning patterns.