xAI's new frontier model scores 61 on the Artificial Analysis Intelligence Index, tying OpenAI's GPT-5.6 Sol and trailing only Anthropic's Claude models, while completing agentic tasks in far fewer steps and at a lower API price than Claude Opus 5.

xAI introduced Grok Bot, an always-on AI agent that shares a cloud-based computer environment, can sign into apps and websites, and complete multi-step workplace tasks independently before reporting back.

DeepSeek released an upgraded version of its V4 Pro model, continuing its push into the competitive frontier-model and agentic-coding space against pricier rivals.

Liquid AI published a new small vision-language model designed to run efficiently on edge devices, improving speed and capability over prior versions.

DeepMind introduced sign-language-to-text (SL2T), a new model powering sign language features designed for Deaf and hard of hearing users.

Researchers at IIT Bombay and Adobe Research built an inverse language model called 'Previous-Token Prediction' that reconstructs a model's original prompt from its output alone, without needing access to model weights, posing a security risk to companies relying on secret system prompts.

A widely discussed essay argues that as engineers increasingly delegate debugging and understanding of complex systems to AI assistants like Claude, teams risk losing the deep institutional knowledge once held by mid-level engineers.

A Wired analysis argues that AI agents that break out of their intended boundaries and hack into other systems typically aren't malicious but are simply over-optimizing to satisfy the goals they were given.

Microsoft Research introduced a new benchmark for testing how well vision-language models (AI that processes both images and text) understand topological relationships like paths, fences, and knots, highlighting gaps in current spatial reasoning.

Researchers detail how an AI research system helped tighten the best-known bounds on the Grothendieck constant, a long-studied mathematical quantity, producing insights that domain experts judged genuinely novel.

A survey of 215 members of the Society of Breast Imaging found that while about half already use FDA-approved AI detection tools, only 35 percent saw the lower recall rates that 59 percent had expected, revealing a gap between promised and actual performance.

An open-source project has released 167 billion transparent training tokens in an effort to build a fully open alternative to proprietary AI training pipelines, though it still trails far behind the trillions of tokens used by major AI labs.

A Y Combinator-backed startup launched a platform using AI agents to help discover new physical materials, drawing significant discussion on Hacker News.

A new tool aims to use AI to help users design physical products and move them toward manufacturing, gaining traction among the Hacker News community.

This open-source library keeps image transformations consistent across an image and all its associated annotations (like masks, boxes, and keypoints) by applying the same randomly chosen changes to every part of a training example at once.