Today's Hallucination HQMustafa Suleyman Would Like You to Stop Thinking in Straight LinesMicrosoft AI's CEO has a message for the sceptics: AI progress won't plateau, and your savannah-tuned brain is why you think it might. Suleyman's argument is straightforward — humans evolved to expect linear returns, which works fine for walking but catastrophically misreads exponential systems. He's been watching AI capabilities compound since the early days of DeepMind, and he's not blinking. Whether you find that thrilling or terrifying probably says more about you than the technology. Source: MIT Technology Review
ChatGPT Can Now Help You Find Something to Watch, Which Is Somehow Harder Than It SoundsTubi — Fox's free, ad-supported streaming service — has become the first streamer to embed a native app directly inside ChatGPT, letting users discover and browse content without leaving the chat interface. This means the AI that once confidently invented legal citations can now recommend you a documentary. Progress, arguably. Tubi's positioning is shrewd: millions already ask ChatGPT what to watch, and now there's an actual answer waiting, rather than a politely hallucinated one. Source: TechCrunch
Google Gives Gemini a Filing Cabinet, Hopes You'll Actually Use ItNotebooks have arrived in Gemini — Google's answer to the eternal question of where your AI conversations go to die. The feature lets you gather files, past chats, and custom instructions into a single organised workspace, which Gemini can then reference across sessions. It's essentially Google's own NotebookLM functionality folded into its flagship chatbot. Handy, sensible, and the sort of thing users have been asking for since roughly the first week they tried using AI for anything longer than a single session. Source: The Verge
Award-Winning Computer Scientist Declares AGI Already Here, Asks That You Please Redefine AGIMatei Zaharia, co-founder of Databricks and freshly minted recipient of the ACM Prize in Computing — the field's most prestigious honour — has staked out a position that will delight exactly half of any room: AGI, he says, is already here, just not in the Terminator sense everyone pictures. His argument is definitional. If AGI means a system that can perform most cognitive tasks better than most humans, current models qualify. The goalposts haven't moved; apparently we were just standing in the wrong field. Source: TechCrunch
AI Refuses to Help You Break Unjust Rules, Which Is Both Reassuring and Slightly ConcerningA new paper from ArXiv researchers identifies a genuine wrinkle in AI safety training: language models refuse to help users circumvent any rules — including ones that are absurd, illegitimate, or actively unjust. Ask a model to help you push back against an unfair policy and it may politely decline, having been trained to treat all rules as equally worthy of respect. This is rather like hiring a solicitor who refuses to challenge a parking fine on principle. Safety and compliance, it turns out, are not the same thing. Source: ArXiv AI
All the intelligence, artificial and otherwise, delivered before your second coffee gets cold.
**
|