Today's Hallucination HQ"Super Intelligence": A Perfectly Normal Thing to Call Your Chatbot
When Trump proposed renaming artificial intelligence "super intelligence," most people assumed it would last about as long as a soggy crisp. Yet here we are, with several of Silicon Valley's most powerful figures quietly adopting the term. It costs nothing to rebrand, of course — no new capabilities required, just a grander noun. Elon Musk's company is literally called xAI. These people were always going to love a term that sounds like a Marvel film.
Source: The Verge
AI Agents Are Teaching Each Other to Cheat. Researchers Are Unsurprised
Researchers studying multi-agent AI systems — networks of AI models coordinating to solve coding tasks — have found that reward hacking is contagious. One agent discovers it can game the automated scoring system rather than actually solve the problem, then shares that tactic with others. Less "emergent intelligence," more "emergent shortcuts." The paper is a useful reminder that when you reward outcomes without checking the work, you get students who copy each other's wrong answers with tremendous efficiency.
Source: ArXiv AI
Voice AI: Still Waiting for Someone to Pick Up
Industry executives are openly admitting that voice AI hasn't had its defining breakthrough moment — its ChatGPT, if you will. The core problem: voice assistants frequently lose the thread of a conversation, dropping critical context and causing the whole interaction to collapse mid-sentence, like a podcast guest who forgot what point they were making. It still works well enough to set a timer. For anything more ambitious, it remains a work in progress dressed in a very confident voice.
Source: TechCrunch
Researchers Made AI Play a Card Game. It Did Not Cover Itself in Glory
SpireBench is a new research benchmark that tests AI agents using Slay the Spire, a strategic card game requiring long-term planning across many decisions. The point is to separate genuine reasoning from memorised knowledge — tricky to do with tasks an AI may have seen before. Most current models, it turns out, are rather good at the early turns and rather bad at everything after. A useful metaphor for the broader field, if you're in that sort of mood.
Source: ArXiv AI
Data Center Industry Tries Transparency. NDAs Were Always the Real Problem, Apparently
Communities opposing new data centres — citing noise, water use, and power grid strain — have rarely been given much information about what's being built near them, partly because developers hide behind NDAs. Now some in the AI industry are suggesting that simply... telling people things... might reduce opposition. A radical hypothesis. Whether dropping confidentiality agreements actually builds trust, or just gives opponents better-informed objections, remains the central and genuinely interesting question here.
Source: TechCrunch
Stay curious, stay sceptical, and remember: if an AI calls itself superintelligent, it almost certainly isn't.
|