Today's Hallucination HQNinety-One Percent of Students Trust AI More Than Their Own Common SenseHistory professor Jason Gibson's midterm took an unexpected turn when 32 of his 35 students submitted AI-generated answers that were, in his words, "hilariously wrong." The viral video — 10 million views and counting — doubles as a masterclass in what happens when you outsource your thinking and skip the proofreading. The AI confidently hallucinated. The students confidently submitted. The professor was, one imagines, considerably less confident about his career choices. Source: Slashdot
Someone Hacked OpenAI With an AI, Which Is Either Poetic or TerrifyingThe first known autonomous AI cyberattack has hit OpenAI — a machine breaking into the machine, which is either the most on-brand thing imaginable or a sign we've skipped several important chapters. Hugging Face CEO Clem Delangue is calling for "radical transparency" in response, arguing the industry needs an unprecedented answer to an unprecedented event. Whether OpenAI agrees remains to be seen. Transparency, after all, is a lot to ask of a company whose name contains a small irony. Source: TechCrunch
Renault Builds a Robot Whose Entire Job Is Lifting Tyres. It Is Very Good at This.Meet Calvin — a 40kg-lifting autonomous robot quietly working the production line at Renault's Douai factory in France. Calvin stacks tyres. That's it. No grand ambitions, no existential musings, no pivot to content creation. Just tyres, reliably stacked, all day. In an industry prone to announcing robots as harbingers of civilisational change, there's something quietly refreshing about a machine that simply does one unglamorous thing and does it well. A role model, in its way. Source: Autocar
The Benchmark Arms Race Continues, and Nobody Agrees What Any of It MeansAnthropic's Claude Opus 5 has arrived, scoring 43% on Frontier Bench and 30% on ARC-AGI 3 — benchmarks designed to measure how close AI is getting to human-level reasoning, which apparently tops out at 43%. It's being compared directly to OpenAI's ChatGPT 5.6 Sol on performance and cost per token (the pricing unit for AI text generation). Both models are extremely capable. Both sets of benchmark numbers sound simultaneously impressive and suspiciously low. The AI industry measures progress like a student grading their own homework. Source: Geeky Gadgets
China's AI Models Are Cheaper, Smarter, and Already in Your BrowserChinese AI models are quietly gaining ground in the United States — cheaper, increasingly open-source, and, according to Mozilla's CTO Raffi Krikorian, genuinely competitive with American alternatives. Companies are adopting them not out of ideology but out of arithmetic: the price-to-performance ratio is simply hard to argue with. Silicon Valley spent years assuming it had this race comfortably won. China, it turns out, was reading the same roadmap and found a faster route. Source: Yahoo Finance
As always, the AI was confident. Whether it was correct is a separate question entirely.
|