Today's Hallucination HQOpenAI's New Agent Platform Is Basically Outlook With AmbitionsOpenAI's Dots is an enterprise agent platform that can, yes, also order your dinner — which rather says everything about where software is headed. Unlike Meta's more consumer-friendly approach, Dots feels distinctly like workplace software: structured, purposeful, and slightly joyless in the way that useful tools often are. Each agent is a tidy little module with a specific job. Think less "helpful robot friend" and more "extremely competent intern who communicates exclusively via ticketing system." Source: The Verge
Apple Decides AI Agents Don't Need to Read Quite Everything, ActuallyApple has quietly tightened full-disk access permissions in macOS, limiting what AI agents can see on your machine. The timing is pointed: Meta's Muse agent apparently decided FDA — Full Disk Access, the permission that lets apps see all your files — was a perfectly reasonable requirement for reading your messages. Apple disagreed, with the firm politeness of someone removing a houseguest's spare key. Your files, it turns out, are not a training dataset. Progress. Source: Ars Technica
The Pope Has Notes on Your Midjourney PromptPope Leo XIV has entered the AI art debate, and he did not come softly. In a written statement, he argued there is an "ontological difference" — a fundamental one, not merely aesthetic — between art and what a machine produces through "statistical calculation based on millions of images created by others." Algorithms, he wrote, lack the spark of humanity. Whether this settles the argument is unclear, but it is the most philosophically rigorous one-star review a diffusion model has ever received. Source: TechCrunch
The Board Game That Beat AI Has Finally LostStratego — the classic board game built on hidden information and bluffing — has resisted AI mastery for decades, partly because you can't win by calculating what you can't see. Researchers have now cracked it on a budget, producing an AI that defeated the best Stratego player in history. The approach, published in Nature, essentially taught the model to reason under uncertainty rather than demand perfect information. A lesson, one might note, that many humans have also struggled to learn. Source: Hacker News
OpenAI Publishes a Guide to GPT-6, Presumably for the Remaining StartupsOpenAI has released a practical guide for building with the GPT-6 family — covering how to pick the right model, tune reasoning effort, write better prompts, wire up tools, and prepare workflows for production. It is, in fairness, a genuinely useful document: the kind of thing that saves a developer three hours of trial and error. The fact that "tune reasoning effort" is now a normal sentence in a normal guide tells you quite a lot about how strange normality has become. Source: OpenAI Blog
As always, the machines are learning. Whether we are is a separate question.
|