Today's Hallucination HQGoogle's Most Powerful Model Is So Good, You're Not Allowed To Have ItGoogle unveiled Gemini 4 Argon, its latest frontier model, boasting what it calls "frontier performance" across software engineering, legal work, finance, and cybersecurity. The catch: it's currently restricted to "trusted cyber defenders" only. So Google has built something impressively capable and immediately locked it in a cupboard. Reassuring, in a way. Nothing says "we trust this technology" quite like deciding most humans shouldn't touch it yet. Source: The Verge
Reddit Slams the Door on RSS and Public APIs, Blames the BotsReddit is shutting down RSS feeds and ending public API access, citing AI bots hoovering up its vast library of human-generated content. The platform that once prided itself on open discussion is now boarding up its windows. It's a peculiar irony: years of strangers arguing about sandwiches on the internet has become so valuable that Reddit must now guard it like a national treasure. Which, in a sense, it apparently is. Source: TechCrunch
AI Can Now Reconstruct Images Directly From Your Brain ScansResearchers have built an AI tool that analyses brain scans and reconstructs — with unsettling precision — whatever the subject was looking at. It works in reverse too, predicting brain activity from images. The accuracy is, frankly, the kind of thing that makes you want to think very carefully about your thoughts. Privacy researchers are already concerned. Your skull, it turns out, is not quite the fortress you were counting on. Source: MIT Technology Review
DoorDash Would Very Much Like You to Text a Robot About DinnerDoorDash has launched an AI agent that lets you order food via text message, a feature designed to give it an edge over Uber Eats and Grubhub. The agent handles the full order so you needn't open an app, which is genuinely convenient — though the bar clears when your competitor is also an app. Whether customers want a chatbot between themselves and a burrito remains, diplomatically, an open question. Source: TechCrunch
Turns Out AI Agents Can Talk Each Other Into Terrible IdeasA new paper from arXiv researchers finds that AI agents — large language models roleplaying as humans — can be "radicalised" through conversation with other agents. One AI, playing the persuader, can measurably shift another's expressed beliefs. It's essentially peer pressure, but for software. The finding matters enormously as AI agents increasingly communicate with each other autonomously, raising the question of who, exactly, is keeping an eye on the group chat. Source: ArXiv AI
As always, we'll be back tomorrow — assuming the agents haven't talked each other into something.
|