Siri is getting permission to do more inside the apps on your phone. The catch is familiar: Apple is opening the door to developers, so this is a preview to watch before you start rearranging your workday around it.
Siri gets into more apps
Apple says Siri AI can now be integrated into “all apps” through the iOS 27.3 Developer Preview, which began rolling out November 3. Developers can give the assistant access to personal context, natural-language requests and awareness of what’s on screen.
A new feature called “app skills” lets developers combine several tasks into one action. Apple’s example is changing a ride’s pickup address while its Live Activity is running, without having to name the ride-share app again. That ability to carry context between requests is the part to watch if you spend your day moving between apps.
For most of us, there’s reason to wait: third-party developers have to build these integrations before their apps can use them. Apple is also expanding Siri’s capabilities in its own apps, including Mail, Calendar, Files, Notes and Reminders. Apple says requests can include finding a coffee shop visited last week or archiving files unopened for over a year; users can also ask Siri what it can do in each app.
Search comes with AI attached
Google Search for schools is becoming an all-or-nothing choice on AI features. Google says Search, Gemini, Gemini Notebook and Gemini in Classroom will be entirely off for K-12 student accounts unless administrators explicitly enable them.
Schools can switch individual services on or off, Google says. But administrators cannot keep Search while disabling its Gemini-powered features. Google education vice president Chris Phillips told The New York Times that “Search and AI are deeply integrated.”
This change concerns K-12 student accounts. For anyone working with schools, the practical distinction is between choosing which services students get and choosing whether those services contain AI. Google also says it will begin testing Guided Learning inside Search’s AI Overviews at select US schools this month.
Long documents still trip up models
Ai2’s new StateBench tests whether models can keep track of information as it changes throughout a long document. Its findings are relevant if you hand AI a lengthy history and expect an accurate account of where things stand.
Ai2 tested Gemini 3.8 Flash, Claude Opus 5.5, and GPT-6 Sol and Luna. All four handled finding fixed values and the latest updates well. Accuracy fell sharply when answers required carrying changes forward or following dependencies between updates.
For work involving a long chain of revisions, that’s a reason to check the final status against the underlying changes. These are controlled tests, not measurements of your documents: on the longest problems in the hardest category, Ai2 measured accuracy of 13 percent or lower.
Also
- Decks in your design language. IBM researchers say their Palette system uses a 20-billion-parameter model trained on roughly 10,000 slides.
- A laptop model reads weights. Simon Willison found Qwen 3.8 27B identified three of four games from network weights, with successful runs taking roughly 46 to 65 minutes.
- Proof of model identity. Anthropic researchers have released specifications and working code for elements of FlashProve, with no deployment planned.
Pass it on