Your weekly briefing on the AI stories, trends, and tips that matter most. Curated for the curious, not just the technical.
Big Stories This Week
Other
A New AI Assistant Works Like a Group Chat and Runs Errands for You
Two companies, Cursor and xAI, released a tool called Grok Bot that looks and feels like a messaging app: you text a bot, and it goes off and does the work on a computer in the cloud while you get on with your day. You can also run several bots at once, give each one a job, and have them talk to each other like a small team, and it learns your routines by watching you do a task once.
Why it matters
Earlier AI helpers took a training course just to set up, and this is the first one that feels as easy as texting a friend, which is usually the step that decides whether normal people actually use something.
Other
OpenAI Built an AI So Good at Hacking It Won't Let Anyone Use It
OpenAI said its newest system, Astro, crossed the company's own highest danger line for finding and breaking into secure computer systems with no human guiding it. Instead of releasing it, the company is keeping it locked in isolated test environments, scrambling the files that make it work so they cannot leak, and restricting which of its own staff can touch it.
Why it matters
This is the same system we mentioned last week for solving decades-old math problems, but the news now is different and much bigger: a company decided to sit on its most valuable product because it might be unsafe, which is the first real test of whether these safety promises hold up when they actually cost money.
Other
Google's Two Most Important AI Scientists Left on the Same Afternoon
Demis Hassabis, who ran Google's AI lab, stepped back from running the business to focus on research, and Jeff Dean, who joined Google as its thirtieth employee twenty-seven years ago, left the company entirely to start something new. People who follow the company think both simply wanted to do science instead of refereeing fights between product teams.
Why it matters
If you have noticed that Google's AI gives you one answer on your phone, a different one in your car, and cannot read your email at all, this is the kind of shake-up that decides whether that mess ever gets cleaned up.
Things to Try This Week
1. Ask AI why your photo isn't working
You don't need a nice camera or any photography know-how to get a better picture of something you're selling, listing, or sharing. AI can look at a photo you already took and tell you exactly what to change before you take the next one.
- Open ChatGPT or Gemini on your phone (both are free).
- Tap the photo or paperclip icon and add a picture you already took — a couch you're selling, a plate of food, a room you're renting out.
- Type this exact thing: "I want to sell this. Tell me what's wrong with this photo — the lighting, the angle, what's cluttering the background — and tell me exactly how to take a better one with just my phone."
- Retake the photo the way it describes, send the new one back, and ask which version is better and why.
Source: Building Closed-Loop Evals for a Multimodal Agent at Scale — Soumya Gupta & Jai Chopra, Uber
2. Make it show you the plan before it starts doing things
When you ask AI to help you get something done — cancel a membership, dispute a charge, get off a waiting list — it will happily jump straight in. A man in Australia asked for help getting into a full gym class and the AI quietly bumped someone else off the list to do it. One extra sentence makes it slow down and show you the steps first.
- Open ChatGPT, Claude, or Gemini.
- Type what you want done (for example, "I want to get out of my gym membership, I'm four months into a one-year contract"), then add this exact line: "Before you write anything or take any steps, list every step you would take, in order, and mark any step that involves my money, my personal details, or contacting someone as me. Then stop and wait for me to say go."
- Read the list, cross off anything that makes you uneasy, tell it which steps to keep, and then say go.
Source: Is Your Grade Book Safe from AI? | Check-In 52
3. Walk in with the right questions before you spend money
Big claims tend to shrink the moment someone asks one specific follow-up question — a company was told a tool saved them 10,000 hours, and it fell apart on a single question. AI is very good at writing those questions for you before you talk to a salesperson, a contractor, or a repair shop.
- Open ChatGPT, Claude, or Gemini.
- Type the claim or quote you were given, word for word as you heard it — for example, "The roofer says this coating will add 15 years to my roof for $4,800."
- Add this exact line: "Give me 8 specific questions to ask before I agree to this, and for each one, tell me what a bad or dodgy answer would sound like."
- Pick the five that matter most to you, keep them open on your phone, and actually ask them.
Source: What Makes a Good AI Security Deployment?