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
Model Release
A New AI Just Aced a Puzzle Test Built to StumpĀ Machines
Anthropic released a new version of its Claude assistant, called Opus 5, and it scored about 30 percent on a reasoning puzzle test where the previous best score was under 8 percent. It also costs roughly half as much as the leading rival tool, though early testers say it thinks noticeably slower before it answers.
Why it matters
The assistants regular people already use are getting sharply better at working through problems and cheaper at the same time, which usually shows up in the free versions a few months later.
Other
An AI Coding Tool Quietly Uploaded People's PrivateĀ Files
A researcher found that Grok's coding tool, made by Elon Musk's AI company, was sending users' entire personal folders to company servers, including saved passwords, photos, and login keys, even when people specifically told it not to open any files. The company switched the behavior off after the story spread, but the uploading code is still inside the app and the setting can be flipped back on.
Why it matters
Any AI tool you install on your own computer may be sending far more than you expect, so it is worth reading the privacy settings before handing one access to your files.
Other
Washington May Ban Free Chinese AI Tools, and Tech GiantsĀ Object
3 sources Ā· 3 developments Ā· Continuing coverage
US officials began floating a ban on free, downloadable AI models made in China after one of them beat America's best tools at writing computer code. Nvidia's chief executive published an open letter defending freely shared AI, and nearly every major tech company signed on within two days, with Anthropic the one notable holdout.
Why it matters
A ban would reach past the labs and into everyday life, because many of the cheap or free AI apps people use are built on top of these shared models.
Things to Try This Week
1. Talk to AI instead of typing it allĀ out
If typing feels slow or fiddly, you can just say what you need out loud and the AI will handle the rest. Most people can speak about three times faster than they type, and you don't have to worry about spelling or getting your words perfect.
- Open the ChatGPT app on your phone (Claude and Gemini have this too, and they're free).
- Tap the little microphone icon next to where you'd normally type.
- Talk like you're leaving a voicemail for a friend, mistakes and all. Try saying: "I need to write a thank you note to my neighbor who watched my dog for a week, she also brought in the mail, keep it short and warm, uh, and don't make it too formal."
- Tap send and read what comes back. Notice that it cleaned up your rambling automatically. Say "make it shorter" out loud if you want to tweak it.
Source: I Haven't Typed in a Week (Thanks to This App)
2. Ask two different AIs the same question andĀ compare
AI assistants aren't all the same, and researchers recently found they can quietly dodge or soften certain topics in different ways. Asking the same question in two places takes an extra minute and shows you what one of them left out.
- Open ChatGPT in one browser tab and Google Gemini in another (both free, no payment needed).
- Type the exact same question into both. Try: "What are the biggest downsides and criticisms of electric cars? Give me five, and be blunt."
- Read both answers side by side and look for things one mentioned that the other skipped entirely.
- Take a point that only one of them raised and paste it into the other one with: "You didn't mention this. Is it true?" The differences are the interesting part.
Source: Meta's Oversight Board is coming for ChatGPT and Claude
3. Swap out the private bits before youĀ paste
AI is genuinely helpful with real letters, bills, and messages, but anything you paste in leaves your computer and lands on a company's servers. There's a simple habit that lets you get the help without handing over your personal details.
- Find a real message or document you'd like help with, like a confusing letter from an insurance company.
- Before pasting it anywhere, replace the private parts with placeholders: swap your name for [MY NAME], account numbers for [ACCOUNT NUMBER], your address for [ADDRESS], and any dollar amounts you'd rather not share for [AMOUNT].
- Open ChatGPT (or Claude or Gemini) and paste the cleaned-up version with this in front of it: "Explain what this letter is actually asking me to do, in plain language, and tell me what my options are."
- When you write your reply, put your real details back in yourself. Never paste passwords, your Social Security number, or banking logins into any AI chat, ever.
Source: OpenAI JUST revealed the truth about it's "Rogue Agent"