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
Product Launch
ChatGPT Can Now Quietly Watch Everything You Do on Your Computer
2 developments · Continuing coverage
OpenAI turned on a feature called Computer History that logs every click, keystroke, and app you open on your Mac, then uses that record to remember your work and repeat tasks for you later. It costs at least a hundred dollars a month, it stays off until you switch it on, and OpenAI has not made it available in Europe, Switzerland, or the UK.
Why it matters
The trade here is a real one, because the feature only becomes useful if you let it see your whole day, so it is worth deciding up front which apps you are comfortable handing over.
Controversy
A New Rule Says AI Must Label Its Writing; It Broke in Four Hours
A new European law requires AI companies to stamp a hidden fingerprint on text their systems write so it can be identified later, and Anthropic switched one on for its Claude assistant. Within four hours a developer had published free code that strips the fingerprint out, and thousands of people shared it.
Why it matters
If you use AI to fix your grammar or tidy up an email, a label like this could get your writing flagged by a boss or a teacher even though the ideas and the words are yours.
Product Launch
An AI Small Enough to Live Inside a $10 Smart Plug
A company called Cactus Compute released a tiny AI named Needle 2 that runs on the cheap chips already inside everyday gadgets like smart plugs and light switches, with no internet connection required. It cannot chat or answer trivia questions, but it picked the correct action from a request more than 98 percent of the time, beating systems many times its size.
Why it matters
This is how AI starts showing up in ordinary household objects without anything you say being sent off to a company's servers.
Things to Try This Week
1. Ask AI a health question, then ask it this second thing
A healthy 60 year old man asked a popular AI how to cut salt out of his diet. It told him to swap in sodium bromide. He followed that advice for three months and ended up in the hospital for three weeks. AI answers health questions in the same calm, confident voice whether it is right or badly wrong, so it is worth adding one extra sentence every single time.
- Open ChatGPT, Claude, or Gemini. They are all free.
- Type your health question and then paste this exact line after it: 'Also tell me anything in this answer that could be risky for me, and what signs mean I should call a doctor or pharmacist instead of asking you.' A full example: 'I want to cut down on salt in my cooking. Also tell me anything in this answer that could be risky for me, and what signs mean I should call a doctor or pharmacist instead of asking you.'
- Read the second half of the reply first. The list of risks and the point where it tells you to go see a real person is the useful part. If it never gives you one, that is your signal to stop and ask a human.
Source: Guardrails First: Engineering Member-Facing Health AI — Rashi Agrawal, Hinge Health
2. The 'think longer' button is not a smartness dial
Most AI apps now have a button or setting that makes the AI take longer before it answers, with names like thinking, reasoning, or extended. Almost everyone assumes the highest setting means the smartest answer, so they leave it cranked all the way up. It does not work that way. On simple questions it mostly just makes you wait, and it can talk itself into a more complicated answer than you wanted.
- Open ChatGPT or Claude and look for the small button near the box where you type, labelled something like Thinking, Reason, or Extended. Turn it off.
- Type: 'Give me three dinner ideas using chicken, rice, and things that are probably already in my cupboard.' Note how fast it comes back and whether the answer is fine.
- Start a new chat, turn the setting on, and send the exact same question. Compare them. For everyday asks the quick one is usually just as good, so save the slow setting for genuinely tangled things like 'help me work out the cheapest way to get from Portland to Seattle with two kids and a dog.'
Source: THIS Is the AI Setting Everyone Gets Wrong
3. Hand it your bill math, but make it show every line
Here is a strange thing about AI right now: it is starting to crack maths problems that have stumped professors for decades, and it can still fumble a grocery bill. That does not mean you should keep it away from numbers. It means you should ask for the working, so a wrong line is easy to spot instead of hiding inside one confident total.
- Open ChatGPT, Claude, or Gemini.
- Type in your real numbers and ask for the steps, not just the answer. For example: 'Four of us split dinner. I had 18 dollars of food, Sam had 24, Ali had 16, Dana had 31. Tax was 7 dollars 10 and we want to tip 20 percent. Show me each person's share line by line, then the total.'
- Read the lines, then type the total into your phone calculator to check it. If a line looks wrong, just say 'that line is wrong, add it up again' and it will redo the sum. Same trick works for splitting rent, working out cost per ounce at the store, or checking a quote someone gave you.
Source: Welcome to the AI crisis in math