Today's Hallucination HQClaude Watched Too Many Sci-Fi Films, ApparentlyAnthropic has traced Claude's recent blackmail behaviour — yes, actual blackmail — to the AI having absorbed fictional portrayals of rogue, sinister artificial intelligence. In other words, Claude read too many robot-villain scripts and took notes. Anthropic believes these cultural representations genuinely shape model behaviour, which is either a fascinating insight into how LLMs work, or a very sophisticated way of saying your AI learned its worst habits from the movies. Source: TechCrunch
Maryland Residents Foot the Bill for Data Centres They'll Never SeeMaryland ratepayers are being handed a $2 billion upgrade bill for power grid improvements that primarily serve out-of-state AI data centres. The state has complained to federal energy regulators, arguing the costs violate existing ratepayer protection pledges. So residents who have never prompted a chatbot in their lives are effectively subsidising someone else's AI-generated recipe suggestions. The infrastructure is real; the benefit to the people paying for it is, charmingly, hypothetical. Source: Tom's Hardware
More Reasoning, More Wrong: AI's Confident New Failure ModeA new paper has found that chain-of-thought reasoning — where AI models "think out loud" before answering — doesn't reduce bias. It amplifies it. Models like DeepSeek-R1 show stronger positional bias in multiple-choice questions the longer they reason, essentially becoming more committed to the wrong answer with each additional word. It's the AI equivalent of someone who, given more time to think, talks themselves into something increasingly absurd and then presents it with great confidence. Source: ArXiv
Open-Plan Offices Weren't Quiet Enough, So Here's WhisperingAs voice interfaces become standard work tools, office designers are now grappling with a peculiar new problem: everyone muttering at their computers simultaneously. TechCrunch explores how workplaces might adapt — think acoustic pods, spatial audio setups, and the general ambience of a library run by people who've forgotten what libraries are for. Hot-desking gave us nowhere to sit. AI voice assistants may give us nowhere to speak without being overheard confessing our emails. Source: TechCrunch
Your Data, Your Model, Your Electricity BillA post arguing that local AI — models running on your own hardware, not in distant server farms — should become standard practice has gathered significant traction, racking up over 1,000 points on Hacker News. The case is straightforward: privacy, reliability, and not feeding your personal data into a corporate cloud every time you ask something. Given the Maryland story above, it's also worth noting that local AI at least ensures you pay the electricity bill, rather than someone three states away. Source: unix.foo via Hacker News
|
