AI News

Princess Cruises now offers live cruise search and planning inside ChatGPT: 17 ships, more than 330 destinations, and pricing pulled live from Princess's own booking systems rather than a cached brochure. You tag @princesscruises, install the app, and plan a cruise in plain language. Final booking is completed on Princess.com, not in the chat. Which means an American AI platform, an American cruise brand and a European traveller now sit in the same conversation thread — and that is where the interesting questions begin.

Picture a hospital delivery drone crossing a city at 120 metres. Its detect-and-avoid software spots a rescue helicopter and swerves — in under a second, without asking anyone. Under one European rulebook that software may be part of an aviation safety-regulated product. Under another, it may be a high-risk artificial intelligence system requiring appropriate and effective human oversight.

NVIDIA is moving agent safety outside the model. OpenShell is an open-source runtime designed to keep autonomous agents inside strict sandboxes; NVIDIA Sentry is an announced hardware-monitoring design for BlueField DPUs that can isolate an agent when it crosses a defined boundary. More than 100 partners have been associated with the announcement.

Europe has spent years debating what could go wrong with artificial intelligence. On 23 September 2026, a German company that had just announced a $300 million funding round made the opposite case: that the more expensive mistake is hesitating while the rest of the world builds. Black Forest Labs, the firm behind the FLUX image models, is urging European policymakers and investors to trade caution for optimism — and its argument lands in the middle of a transatlantic fight over who gets to slow AI down.

Skylark Labs, working with researchers from Carnegie Mellon University and UC Berkeley, has published an architecture that lets a robot keep learning after it has been shipped — no cloud round-trip, no full retraining. The measured numbers are believable: action success up from 74.0% to 87.9%, forgetting down to −0.5 percentage points against −11.4 for standard LoRA fine-tuning. The awkward part for European buyers is not the maths.

Microsoft has reportedly removed the "Copilot+ PC" badge from new Surface devices, while a member of its AI division argued on X that "most people love" Copilot and that the criticism comes from a San Francisco tech bubble. Both things can be true at once: 30 million paid enterprise seats on one side, and a hardware label that Microsoft is no longer putting on some new Surface products on the other. The gap between those two facts is the actual story.

According to a SiliconValley.com report that first detailed the project, the Buck Institute for Research on Aging in Novato, California is seeking roughly $1 million to build a safety mechanism that lives inside the AI simulator instead of in a checklist sitting next to it. The prototype the institute calls OASIS is aimed at the scenario biosecurity specialists worry about: an AI that could design a pathogen nobody has seen before before human review can realistically respond.

Microsoft has rebuilt Copilot around three tabs — Home, Code and Autopilot — while keeping per-seat licensing alongside additional consumption charges. Code opens in early access and Autopilot in private preview at the end of September 2026, with broader previews for Microsoft 365 Premium and Pro subscribers later in the year. The demo looks clean. The meter is where it gets interesting for European IT departments.

DeepSeek V4 Pro logged a headline-grabbing 4,096,000 tokens per second in our AI Arena harness this week — and we're here to tell you that number is a measurement artifact. The endpoint returned the entire answer as one non-streamed response, leaving the harness with no measurable TTFT and triggering a fallback timing value for the tok/s calculation. Measured by total elapsed time, Gemini 3.6 Flash beat DeepSeek in all four tests we ran.