Modelli, agenti, applicazioni: l’intelligenza artificiale messa al lavoro su problemi reali.
Qui l’intelligenza artificiale si giudica sulle prove: che cosa fa, per chi, a quale costo. La rubrica segue i modelli e i loro progressi misurabili, gli agenti che eseguono compiti reali e soprattutto le adozioni sul campo — un ospedale che esamina più pazienti, una rete elettrica gestita meglio, un’amministrazione che risponde più in fretta. L’IA spettacolo, i round di finanziamento senza prodotto e le profezie non trovano posto.
Preferiamo i casi d’uso documentati, numeri alla mano: quanti esami, quale tasso d’errore, quale risparmio. Gli attori non occidentali — laboratori cinesi, indiani, emiratini — sono trattati come quelli americani, perché buona parte dell’adozione reale avviene lì. Ogni articolo cita le sue fonti e distingue il prototipo dal servizio in produzione.
“Clues for Choosing a Nursery” displays information on more than 55,000 childcare facilities nationwide in one place.
The Claude-developed controller restored the lock in 695 of 700 trials on a QuEra neutral-atom testbed.
Janelia neuroscientist Arco Bast built the Model Hardware Standard with Anthropic and Claude Code.
The stack joins Evon v3.3, trained across 11 Indian languages, with Plexus for workflow-specific AI agents.
Science Tokyo and NEC link camera geometry with visual clues to match people across separate camera views.
The new version eliminates sessions, simplifying the operation of MCP servers.
A new setting is designed to prevent posts from unintentionally going viral in the algorithmic Discover feed.
- The shopping assistant is a voluntary Early Preview for selected users of the Apple Store app.
Version 1.1 now uses 10 seconds of video context instead of 1 second for extensions.
AI coding agents can reload running Compose apps, take screenshots and simulate input.
Qoder hides code generation and execution behind a chat workflow for product, design, operations and creator roles.
The N60 completed the Guiyang qualifier with zero interventions, scoring 117.81 overall and setting a record since the competition’s launch 30 stops ago.
The desktop agent ranked second in July, behind Tencent’s WorkBuddy at 11.15 million monthly active users.
WorkBuddy now covers nine Shanghai Jahwa departments, from supply chain to finance and R&D.
Meta’s fix closes a loophole that let wearers cover the capture LED after recording had already begun.
City University of Hong Kong and Chinese University of Hong Kong researchers studied mice performing two tasks at once.
LL-Refiner handles lighting and scene structure at lower resolution before restoring fine details.
Researchers at Pusan National University developed two systems that combine specialized models for different kinds of movement.