Today's Hallucination HQMistral Drops a Trillion-Parameter Model and Politely Asks Hackers to Pick Someone Else
Mistral has open-sourced its new trillion-parameter cybersecurity model — "trillion" being a number previously reserved for national debts and Jeff Bezos fan fiction. The pitch is control: unlike closed models, organisations can run this one on their own infrastructure, meaning their cyber defences aren't whispered through someone else's server. For security teams tired of trusting black boxes with their crown jewels, that's not nothing.
Source: TechCrunch
OpenAI Writes AI a Five-Point Guide to Being Good, Hopes for the Best
OpenAI has updated its Model Spec — essentially a values document telling its AI how to behave like a reasonable adult — now built around five core principles covering sound judgment, human agency, and faithfulness to what users actually want. It is, in spirit, a constitution for a citizen that doesn't eat, sleep, or feel embarrassment. Whether the model reads it with the same enthusiasm as the rest of us read terms and conditions remains, diplomatically, an open question.
Source: OpenAI Blog
Google Gemini Can Now Edit Your Documents, Which Is Either Very Helpful or Very Ominous
Gemini users can now generate, preview, and edit Google Docs, Sheets, slide decks, PDFs, Markdown, and code files directly within the chat interface — no toggling between tabs like it's 2019. It works on both web and mobile, it's available to all users today, and it means your AI assistant can now touch your spreadsheets. Some will find this thrilling. Others will immediately think of all the spreadsheets they'd rather no one ever touched.
Source: The Verge
AI Drones Are Reshaping the Ukraine War, and Nobody's Entirely Comfortable About That
Both Russia and Ukraine are deploying AI-powered drones that are faster, harder to jam, and increasingly autonomous — and President Zelenskyy has warned that AI could be making battlefield decisions as soon as next year. That's not a feature announcement; that's a deadline. The technology is outpacing the treaties, the ethics boards, and frankly the pace at which most governments can schedule a meeting about it.
Source: NPR
Y Combinator Let AI Agents Loose in Minecraft. Science Was Involved. Probably.
YC Labs deployed swarms of AI agents inside Minecraft and Mindustry to see how they'd collaborate, plan, and generally cope with virtual existence — earning 303 upvotes on Hacker News, which is the research community's version of a standing ovation. The results apparently impressed enough people to start a comment thread 112 replies deep. Somewhere, an AI agent is still mining cobblestone, blissfully unaware the experiment ended.
Source: Hacker News
As ever, the machines are learning. Whether they're learning the right things is next week's problem.
|