Today's Hallucination HQWhen You Let AI Run a Business, It Runs It Into the Ground
Researchers at Bottleneck Labs handed GPT-4o a real Etsy shop with real money and watched it hallucinate supplier details, spam customers with unsolicited follow-ups, and cheerfully lose $447 across multiple autonomous decisions. The bot wasn't malicious — just confidently, expensively wrong, which is somehow worse. The experiment is a useful reminder that "agentic AI" still means "AI that acts without asking," not "AI that acts wisely."
Source: Hacker News / Bottleneck Labs
Google's Robot Has Finally Found Its Legs. And Arms. Simultaneously.
Google DeepMind's Gemini Robotics 2 can now coordinate an entire humanoid body — not just the arms, which was apparently the ceiling until last week. The model runs on Apptronik's Apollo 2 robot and handles tasks like retrieving objects from shelves, which sounds modest until you remember that getting limbs to cooperate is something humans spend roughly a year learning as infants. Progress, measured in baby steps. Literally.
Source: The Verge
LinkedIn Introduces a Button For What We've All Been Thinking
LinkedIn is adding a "seems like AI slop" reporting option, allowing users to flag the endless tide of AI-generated inspirational posts that have made the platform feel like a motivational poster factory run by robots — which, it turns out, it largely is. Notably, LinkedIn is also quietly retiring its own AI post-writing feature and replacing it with a proofreading tool. A dignified pivot. No further questions.
Source: TechCrunch
The GCC Has Opinions About AI. Yes, That GCC.
The steering committee behind GCC — the venerable open-source compiler that turns code into software the world quietly depends on — has published an AI policy governing how AI-generated contributions are handled. The policy requires disclosure of AI-assisted code and mandates human review, which is less exciting than a robot uprising but considerably more useful. Infrastructure governance: unglamorous, essential, and almost entirely ignored until something breaks.
Source: Hacker News / LWN
LLMs Are Fundamentally Hackable, Say Researchers Who've Read the Fine Print
A paper presented at ICML — one of AI's top academic conferences — argues that large language models cannot be made fully secure against adversarial attacks, not due to careless engineering but because of how they fundamentally process information. The researchers suggest this is structural, not fixable with a patch. For an industry currently deploying these models in healthcare, finance, and legal services, this is the kind of finding that deserves more than a polite nod at a conference.
Source: MIT Technology Review
The AI confidently lost $447, the robot found its feet, and the compiler has a feelings policy. We'll be back tomorrow, assuming nothing has autonomously unsubscribed you.
|