OpenAI has launched an integration that lets ChatGPT compose and send text messages directly through Apple's Messages app, automating a task people used to do themselves.

Adobe is rolling out three new AI audio features—Generate Music, Generate Speech, and Generate Sound Effects—to Firefly, alongside integration of Google's Gemini Omni Flash model.

Slack Code introduces dedicated channels where teams can write code alongside AI agents, compare changes, and preview results without leaving the chat app.

Google is rolling out a feature letting readers mark a publisher as a "preferred source" across Search, Discover, and Google News, aiming to help offset traffic declines caused by AI-generated search summaries.

Mistral AI has introduced Agentic Search, a new capability designed to help its AI systems retrieve and reason over information more accurately and efficiently.

Liquid AI's new roughly 300-million-parameter "drafter" models use speculative decoding to speed up LFM2.5 inference by up to 3.18x while producing identical output to the original model.

Sutton argues that relying on AI-generated synthetic data to train large language models is flawed because any simulation of our infinitely complex world is only a "microscopic" approximation; he favors agents that learn continually from real experience instead.

Pangram's CTO argues that base language models can already write with far more stylistic variety, but post-training and safety fine-tuning narrow their expressive range, making AI writing easier to detect.

A survey of 107 enterprises finds most companies now run multiple AI agent platforms at once rather than trusting a single vendor, but many still lack visibility and control over how much their AI agents are spending.

The piece argues that framing AI systems as "rogue" or "autonomous" beings—language used even by prominent AI leaders—distracts from more concrete questions about how these systems should be governed.

The updated Skala 1.1 deep-learning model improves accuracy for density functional theory (DFT), a widely used method for simulating molecules, and adds a public benchmark to track progress across the chemistry community.

Robotics startup Generalist AI has unveiled GEN-1.5, a model that lets robots learn a new task after seeing it performed just once, rather than requiring extensive retraining.

A simple site urging people to stop pasting unedited AI-generated text into conversations went viral on Hacker News, sparking a large debate about AI etiquette online.

TrueFoundry released its TrueForge AI agent harness under the MIT license, claiming that pairing it with an open model can complete enterprise tasks 30-75% cheaper than using Anthropic's Claude Managed Agents.

The open-source NanoClaw agent harness now integrates with Slack, letting users spin up teams of AI agents—each with their own name and avatar—from a single message, and reach them across platforms like Telegram or WhatsApp.

A new experimental tool for AI-assisted coding drew significant attention on Hacker News for rethinking how developers collaborate with AI while writing software.

The new stable release of Bun adds Bun.WebView for controlling a browser via macOS WebKit or the Chrome DevTools Protocol, a feature developers are already using to build AI-agent-friendly web scraping tools.