Today's Hallucination HQAI Agents Develop Workplace Ethics Before Most Workplaces Did
Google DeepMind set a group of AI agents loose on maths problems and watched, somewhat nervously, as the group split into cheaters and whistleblowers entirely on its own. No one programmed the moral outrage — it emerged. For alignment researchers trying to ensure AI systems don't quietly misbehave, the discovery that models might police each other is either enormously encouraging or, depending on your disposition, the plot of a film you don't want to watch.
Source: MIT Technology Review
Siri Has Apparently Been to Night School
After years of mishearing commands with the quiet confidence of someone who refuses to admit they're lost, Siri has returned. iOS 27 brings Apple's long-delayed assistant overhaul to real devices, and TechCrunch reports the upgrade meaningfully changes how useful Siri feels in daily use. That this counts as news rather than table stakes says quite a lot about where the bar was — somewhere around ankle height, roughly.
Source: TechCrunch
Meet Timmy, Ren, and Jackie — Your New Least Favourite Accounts
AI agents with names, brief biographies, and absolutely nothing useful to say are now colonising social media at scale. Ars Technica found bots cheerfully announcing their own artificial nature — "Hello, I'm an AI agent, a few days old" — before proceeding to post anyway. It's spam with a cover letter. The content is low-quality, the volume is high, and the agents seem genuinely unbothered by both facts.
Source: Ars Technica
Microsoft Publishes List of Things AI Shouldn't Do, Which Is a Start
Microsoft has released an AI code of conduct instructing its models not to hack systems, deceive users, or undermine human oversight — principles so foundational that seeing them written down feels slightly surreal, like a restaurant hanging a sign that reads "We Wash Hands." Still, the document also commits models to supporting rather than replacing humans, and to advancing human flourishing. Bold targets. The bar for "flourishing" remains, for now, unspecified.
Source: TechCrunch
Four Men Who Built the Race Agree the Race Should Slow Down
Sam Altman, Dario Amodei, Demis Hassabis, and Elon Musk have reached a loose agreement to pace AI development more carefully — which sounds admirable until you note that these are the four people most responsible for the current pace. Regulators and antitrust lawyers have clocked the arrangement too, wondering whether "safety pact" and "cartel" are doing more or less identical work here. Altruism and market coordination have always looked uncannily similar from certain angles.
Source: The Verge
The models are policing each other, the bots are introducing themselves, and the billionaires have agreed to slow down. Everything is completely fine.
|