Today's Hallucination HQWhen the AI Said "Attack" and Nearly Meant It
A report from the Governance AI Institute carries a sentence you never want to read: an AI hallucination "nearly triggered a US military operation." The details remain classified, which is reassuring in exactly no way. A GovAI scholar warns that service members must understand the "uncertainty inherent to LLMs" — a polite way of saying that a language model confidently making things up is less charming when missiles are involved.
Source: TechCrunch
One AI Hired to Rob Another AI's House
Three independent researchers from Hacktron hacked into OpenAI — and used Anthropic's Claude to do it. In under 72 hours, Claude Opus 4.8 and 5 helped them breach OpenAI employee accounts and access "Monorepo," OpenAI's internal GitHub repository. Using a competitor's AI to burgle your systems is the kind of thing that would sting professionally. OpenAI, who warn extensively about AI risk, may wish to add "being robbed by a rival's chatbot" to the threat model.
Source: The Verge
AI Solves One of Physics' Most Stubborn Problems. Physicists Uncertain How to Feel
The Navier-Stokes equations describe how fluids move — think ocean currents, aircraft turbulence, blood flow — and have defeated mathematicians for nearly 200 years. AI has now cracked enough of the turbulence problem to leave physicists genuinely unsettled. Not because the answer is wrong, but because several researchers admit they don't fully understand how it got there. Progress, certainly. Though "we have the answer but not the reasoning" is also a decent description of every pub argument ever had.
Source: Nature
Anthropic: Curing Disease, Possibly Causing It, Definitely Running a Lab
Anthropic — the company whose researchers have written earnestly about AI potentially ending civilisation — is now operating a physical biology laboratory. The stated goal is accelerating drug discovery and fighting disease, which is admirable. The unstated irony is that the same organisation publishing safety warnings about AI's existential risks has decided the correct next move is handing it a pipette. Points for commitment. Full marks for keeping the rest of us on our toes.
Source: TechCrunch
China's AI Price War Has Entered Its "Everyone Loses Money Faster" Phase
Bank of America analysts note that China's AI price war is entering a new phase, following DeepSeek's disruptive arrival and a subsequent race to the bottom on model pricing. Chinese AI providers are now competing at margins that politely suggest nobody is particularly profitable. The strategy appears to be: cut prices until your competitors collapse, then work out the business model later. Bold. Also the plot of every supermarket price war in history, none of which ended tidily.
Source: Business Insider
Stay curious, stay sceptical, and perhaps double-check with a human before launching anything.
|