Today's Hallucination HQChatGPT Learns the Art of Letting You Finish a Sentence
OpenAI has overhauled ChatGPT's voice mode with GPT-Live-1, a new model built to interrupt less and wait patiently when you pause mid-thought — a skill many humans have yet to master. The goal, apparently, is conversation that feels like "talking to another person," though presumably a person who doesn't talk over you at dinner. It's a low bar, but given the state of discourse, a meaningful one.
Source: The Verge
Microsoft Releases a Tool So AI Can Draw You a Decent Chart
Flint is Microsoft's new open-source visualisation language designed to help AI agents produce charts that are actually useful — not the blurry, default-colour monstrosities that typically emerge when you ask a model to "make a graph." The problem it's solving is real: AI can reliably generate simple chart specs but tends to produce low-quality results without explicit styling. Think of Flint as guardrails for an agent that, left alone, would label every axis "Value."
Source: Hacker News
Meta Adds a Privacy Feature to the Product Whose Entire Point Is Watching You
Meta is introducing a safeguard to prevent people secretly recording others with its AI glasses — a noble gesture, somewhat undercut by the company simultaneously expanding how much personal data its AI products collect. The glasses will now be harder to use as a covert surveillance device, which is reassuring, provided you're comfortable with Meta itself having a front-row seat. It's a bit like a landlord fitting your door with a better lock, then keeping a spare key.
Source: TechCrunch
AI Is Getting Too Good to Test. Scientists Are Worried About That.
Researchers have hit a conceptual wall: as AI surpasses human performance on benchmarks, the humans writing those benchmarks can no longer reliably judge which questions are hard, fair, or even answerable. A new paper argues this problem is fundamental, not fixable with more clever exams. It's the intellectual equivalent of asking a student to grade their own teacher — except the student has read everything ever written and the teacher is increasingly nervous about it.
Source: ArXiv AI
Google's Deepfake Detector Does Its First Shift in the Real World
A photograph appearing to show Senator Mitch McConnell in acute medical distress circulated this week before Google's deepfake detection system identified it as AI-generated. The image was fake. The anxiety it caused was not. It's a useful reminder that the same technology making these tools necessary is also, mercifully, making detection more possible — though the gap between "generated" and "debunked" remains wide enough for considerable chaos to fit through.
Source: TechCrunch
Stay curious, stay sceptical, and remember: if an AI tells you it's just like talking to a person, a person wrote that.
|