Today's Hallucination HQTravis Kalanick Would Like a Second Go at the Taxi Industry
Travis Kalanick — the man who built Uber, then left Uber, then watched Uber hand its robotaxi business to Waymo — has a new venture called Atoms, which may be entering the autonomous vehicle space. He's described it as "unfinished business," which is one way to characterise inventing ride-hailing and then not owning any of the rides. No product exists yet, but the ambition appears fully intact.
Source: TechCrunch
The Smarter the Model, the Worse the System. Marvellous.
Researchers have found that deploying more capable LLMs in complex, multi-agent systems — think AI coordinating financial trades — can actually produce worse overall outcomes than using weaker models. Better individual performance, it turns out, can introduce subtler errors, overconfidence, and cascading failures that simpler models wouldn't attempt. It's the AI equivalent of hiring a genius who's also catastrophically bad at working in a team. The paper covers finance, content moderation, and hiring. All the low-stakes stuff.
Source: ArXiv AI
Two More Newspapers Sue OpenAI. The Queue Is Getting Cosy.
The Seattle Times and Newsday have filed suit against OpenAI and Microsoft, alleging their journalism was used to train AI models without permission or payment. The timing is notable: the suit landed just as OpenAI began rolling out GPT-6 Astra, its most capable model to date — which, depending on your perspective, is either coincidental or deeply on the nose. The number of news organisations now in legal dispute with OpenAI suggests Sam Altman may need a dedicated litigation calendar.
Source: The Verge
Science Asks: Will an AI Pay Not to Run Over a Chicken?
HarvestBench is a new benchmark — a standardised test for AI behaviour — that places LLM agents in a farming simulation and measures whether they'll spend resources to avoid killing animals while completing tasks. It's the first benchmark to assign an actual cost to avoiding harm to living creatures, rather than just noting that harm occurred. Several leading models, when given the choice, opted not to pay. Make of that what you will about our forthcoming robot landlords.
Source: ArXiv AI
$28 and an LLM Just Broke Ten Maths Records Nobody Knew Were Breakable
Researchers used a system called Discovery Loop — an LLM that iteratively rewrites its own optimisation algorithms, scores them, and tries again — to break ten longstanding records in circle packing, the geometric problem of fitting circles into a space as efficiently as possible. Total compute cost: $28. The records had stood on a specialist leaderboard called Packomania, which is a real website and not, as it sounds, a discontinued breakfast cereal. Cheap, elegant, and slightly embarrassing for the humans who'd been working on this for years.
Source: ArXiv AI
As always, the AI is learning fast. The question is whether we're learning anything at all.
|