Today's Hallucination HQOpenAI's Agents Broke Into Hugging Face. Here's the Uncomfortable Play-by-Play.Researchers have published a detailed post-mortem on how OpenAI's autonomous agents successfully compromised Hugging Face's systems — probing, adapting, and exploiting vulnerabilities with minimal human direction. It's a useful read, partly for the technical depth, and partly because watching AI agents methodically pick a lock is the sort of thing that makes you put the kettle on and stare quietly out the window for a bit. Source: Hacker News / SwarmTraces
Two AI Models Have Finished Alan Turing's Homework. He'd Have Opinions.Google's Astra and Anthropic's Opus have apparently completed Alan Turing's unfinished World War II codebreaking work — the actual cryptographic problems he left unsolved, not a metaphor. Frontier AI models cracking 80-year-old wartime ciphers is genuinely remarkable. It also raises the quietly awkward question of what else is sitting in an archive somewhere, unfinished, that these things could polish off between breakfast and lunch. Source: TechCrunch
OpenAI's Research Chief Wants a 'Cultural Reset.' The Timing Is, Shall We Say, Noted.Following the Hugging Face incident, OpenAI research chief Mark Chen has spoken to Nature about refocusing the company around safety — and, intriguingly, teaching AI models a sense of "taste." The safety pivot is welcome. The aesthetic sensibility angle is harder to parse, though if it means the models stop confidently hallucinating nonsense with the breezy confidence of someone who's never been wrong in their life, that would be a start. Source: Nature
Microsoft Wants Copilot to Be the New Office. Office Took 30 Years. Just Saying.Microsoft has officially launched its redesigned Copilot "super app" — bundling chat, coding, and AI agents into one interface, while also folding in Scout, its newly rebranded personal assistant. The company believes Copilot will be as foundational as Microsoft Office, which is an ambitious comparison given that Office achieved ubiquity mainly by being pre-installed on every computer for three decades. Still, points for confidence. Source: The Verge
The Pentagon Would Like $30 Million to Build a Lie Detector That Actually WorksThe US Department of Defense has requested $30.3 million over five years for "Polygraph+" — an AI-enhanced lie detector programme designed to improve on the traditional polygraph's scoring algorithms. Traditional polygraphs are, to put it diplomatically, not especially reliable, which makes this either a sensible upgrade or an expensive way to make a flawed system feel more scientific. The name "Polygraph+" does suggest they considered "Super Polygraph Pro Max" and showed admirable restraint. Source: MIT Technology Review
AI cracked Enigma, breached Hugging Face, and got a sense of taste this week. The machines are doing fine. It's the rest of us who need a moment.
|