Scaling Indic Pretraining: 2.8 Trillion Tokens and an Open Data Generation Pipeline
NVIDIA’s team outlines a massive Indic pretraining effort – 2.8T tokens – paired with an open pipeline for generating and cleaning training data. The pitch is practical: better data construction for multilingual coverage, not just bigger training runs, with an emphasis on reproducibility for other researchers to build on.
ChatGPT can peek at your Chrome tabs now
OpenAI’s ChatGPT Chrome extension updates add a tab-aware mode, letting the assistant reference content from the browser tabs you already have open. The workflow is positioned as “computer use” for everyday tasks, with a demo showing ChatGPT comparing multiple open tabs to help you decide between options.
Counterfactual Neighborhood Risk Fields: Evidence-Grounded Selective Safety for Vision-Language Models
This arXiv work tackles a core safety ambiguity in vision-language models – separating “harmful” from “benign” isn’t enough unless the system also knows when its risk estimate is trustworthy. Counterfactual Neighborhood Risk Fields build evidence-grounded benign/harmful counterfactual pairs and use representation “risk directions” to route risky requests toward targeted review without blanket rejection.
A week with RTX Spark: everything we’ve learned using Nvidia’s new laptops
The Verge’s hands-on testing focuses on RTX Spark N1X, Nvidia’s Arm-based “superchip” for Windows laptops, aiming to bring strong local AI compute and unified memory. The coverage breaks down what it means in real machines from multiple vendors – performance, compatibility, and what users should actually expect when these laptops show up on shelves.
AlphaDiana: A Measurement System for Agents in Verifiable Reasoning Tasks
AlphaDiana is a measurement framework for agent systems – separating tools/skills/memory from the harness so researchers can see which components truly move correctness. The headline result: harnesses and guidance mechanisms don’t reliably improve accuracy, meaning “more scaffolding” can accidentally reduce performance even when traces look better on paper.
When Correctness Feedback Erases Evidence: Phase-Separated Completeness Auditing in Verus
This paper shows how the usual safety/repair loop can sabotage itself: correctness feedback can suppress the very evidence that completeness auditing is supposed to find. Across Verus/Rust experiments, withholding feedback and using phase-separated auditing yields more certified witnesses and higher implementation success – and the authors argue there are empirical boundaries to when the workflow will generalize.
Trump posts AI-edited photo to brag about White House restoration work
The Verge reports that a Trump-posted “before vs restored” photo of a White House column appears to have been AI-edited, with discrepancies like mismatched step counts and inconsistencies flagged by Google’s AI detection as made or edited with OpenAI. The story highlights how political communication is increasingly entangled with generative media – and why visual forensics is becoming a routine civic task.
Welcome Beam: a 501B open model you can own, customize, and scale
Hugging Face’s Beam coverage spotlights the release of a very large open model and frames it as a “you can own it” alternative for teams that want to customize behavior instead of relying on closed APIs. The emphasis is on deployability and scaling – turning model weight access into a practical platform decision, not just an academic benchmark win.