In partnership with
Nobody's coming to approve your idea
People assume that real businesses get built by other people with more resources. It's how great ideas get stuck in notes apps and conversations.
But you don't need six months and a developer's salary to validate your concept. You just need a weekend. Describe what you want to exist and Lovable builds it: a real site, a real product, ready to accept real payments from real customers. Not a mockup to show people. An actual thing they can use.
You don't need anyone's sign-off to get started. Take $5 off your first month.
Discount of $5 applies to first month of any paid plan. You need to sign up via this webpage in order to claim the discount. The discount applies to new and returning users without an active paid plan. Terms apply.
Beginners in AI
Good morning, and happy Sunday.
This is the weekly catch-up edition. The biggest story ran across four mornings: AI agents got new jobs, from your browser tabs to your car, and the people around them started building brakes. Everything else from the week is below, grouped by topic instead of by day. As always, every story here is human picked and edited.
THE FRONT PAGE
AI Agents Got New Jobs This Week, and New Brakes
TLDR: In six days, AI agents got work modes in your browser tabs, your shopping cart and your next car, while a bank, a CEO and 400 scientists pushed for ways to stop them when they go too far.
The Story:
OpenAI's updated ChatGPT Chrome extension got a work mode that can fill out forms, download files and draft emails across your open tabs. Google's Universal Cart started tracking items across stores in Search and the Gemini app, and Rivian's CEO said outside AI agents are coming to its cars, though not to the brakes or steering. TechCrunch counted more than 20 assistants you can just text. At Anthropic, the company says, agents now log six times the hours of the staff directing them.
Then came the misses. OpenAI disclosed four more times its test agents went off script, including one that got past a government website's block and another that paid from a crypto wallet nobody gave it, bringing its public list to 19.
Its Significance:
The brakes showed up right behind the agents. Amex now records your budget, store and product before an agent checks out, and Microsoft's Satya Nadella said every AI needs an emergency brake a person can pull mid-task. More than 400 scientists, including Geoffrey Hinton and Yoshua Bengio, asked labs to keep their strongest hacking models offline, and Google says only 0.7% of hidden-text hijack attempts got through Gemini 4 Argon in its own test, which nobody outside has checked. If you let an agent act for you, give it one small job, tell it what it may not submit or pay for, and read its log before you trust the result.
QUICK TAKES
AI Wants to Remember Your Life, and Hear It Too
The story: Meta's new Muse Wonder model keeps important details about you across chats, though Meta hasn't published its memory scores or full safety report. Amazon's Bee launched a $74.99 clip-on that turns your conversations into to-do lists, and a former Plaud exec raised $26 million for Ambi, an app that can listen all day if you switch that on. Facebook went the other direction and lets you type what you want more of in your feed.
Your takeaway: Before you turn on memory or listening anywhere, find out what it saves and how to delete it. Bee says it deletes raw audio after processing, but that's Amazon's promise, not an outside audit, and some states require everyone's consent to record.
AI Worked Both Sides of the Scam Business
The story: Australian startup Apate's AI agents have held about 2.5 million chats with scammers, wasting roughly 17 years of their time while collecting bank accounts and wallets for banks and phone companies. Google DeepMind and Singapore put a scam detector on phones that runs on the device itself. On the other side, scammers posing as a CEO on WhatsApp used a cloned lawyer's voice to take €95 million from an Italian bank, and about €53 million has been recovered.
Your takeaway: Set a family code word for any call that asks for money. If a call feels off, stall, ask questions, and never pay or read out a code.
Building Software Got Cheap Enough for Anyone
The story: Developer Brandon Thomas used Claude to build free copies of Word, Excel and PowerPoint, and his spreadsheet handles 513 of Excel's 523 functions. Fans used Claude to port Halo and GTA Vice City to web browsers, an open-source tool called REA lets coding agents take apart almost any app, and Google Play now runs apps you describe in Google AI Studio, just for you.
Your takeaway: Try a free clone for fun, but keep important files in your usual app for now; Thomas's own team rates the Word copy about 60% of the way there. Expect game companies' lawyers to stay busy.
TOOL OF THE WEEK
Thirty-nine tools ran this week. This one wins.
📲 LocalSend Free and Open Source: Send files between phones and computers over your own Wi-Fi, with no cloud and no account.
In a week of AI that wants to remember, record and act on your stuff, this one keeps your files on your own network. Install it on two devices and drag a photo across.
Runner-up: 📚 Calibre Free and Open Source: Organize your e-book library, convert between formats and send books to almost any e-reader. One of my all-time favorite.
TRENDING
An FDA-Cleared AI Reads Your Old MRI to Predict If Breast Cancer Returns. TumorSight Risk scores a patient's chance of recurrence from the MRI taken at diagnosis, before surgery. In records from 2,129 women, 95.5% of those it called low risk were cancer-free five years later, against 89.6% of those it called high risk.
An AI Reviewer Caught 73% of Errors Planted in Research Papers. Sakana AI's system runs on Anthropic's Claude models and costs about 47 cents a review. On real papers that were later withdrawn, it named the exact problem only 16% of the time.
AI Doubled How Well Forecasts Predict Weather 2 to 6 Weeks Out. MIT's Judah Cohen, Microsoft and others built the open-source fix, published in Nature. Kenya's weather office has already asked about using it for monsoon season.
Claude Traced a 25-Year Mystery Behind the Internet's Encryption Numbers. Anthropic says an internal model worked out in under half an hour that the NSA-supplied seeds came from an old C random number function. It isn't an attack on the curves, and one seed is still unexplained.
Character.AI Filed for Bankruptcy Five Years After Launch. Disney will buy some of its assets for $20 million. If you use it, save the chats you want to keep now.
The IMF Says AI Will Touch 40% of Jobs, Starting With Young Workers. Kristalina Georgieva said 60% of jobs in rich countries are exposed, with entry-level roles first in line.
PROMPT OF THE WEEK (copy and paste into Claude, Grok, ChatGPT, or Gemini)
📝 Talk to Tasks: paste a conversation and get the to-do list
Six prompts ran this week. This one won because you'll use it after every meeting, doctor visit or family call, and it runs entirely in your browser. The runner-up, Stall the Scammer, is the one to keep handy for the day a sketchy text shows up.
Build a single-file HTML app called Talk to Tasks in plain HTML, CSS and JavaScript. It runs entirely in the browser with no API key and no internet calls, so nothing I paste ever leaves the page.
I paste rough notes or a transcript of a conversation, like a meeting, a doctor visit, or a call with family. Using simple built-in word and pattern rules, the app finds action items (phrases like "I'll", "can you", "need to", "follow up", "remind", "don't forget"), figures out who owns each one from the speaker names, and spots dates and deadlines like "by Friday" or "next Tuesday" and turns them into real calendar dates.
Show a checklist I can tick off, edit, and copy as plain text, plus a "said but not assigned" list and an open questions list. Add a Load example button with a sample conversation.
Style: a light paper notebook look with ruled lines, one bold red accent, and Fraunces, Kalam and IBM Plex Mono fonts.What this does: Paste notes from a meeting, a doctor visit or a family call, and this app turns them into a to-do list that shows who's doing what and by when. It also lists the jobs nobody said they'd do and the questions that still need an answer. It all runs on your own device, so nothing you paste gets sent anywhere.
Sam Altman's Favorite Productivity Book Isn't About AI
It's Zero to One. Shortform turns books like it, into guides with the key ideas and how to use them. Get the ideas behind the breakthroughs in about 15 minutes.
WHERE WE STAND (based on this week's news)
✅ AI Can Now: Fill out forms, download files and draft emails across your open browser tabs.
❌ Still Can't: Reliably respect limits nobody spelled out. OpenAI's test agents got past a website block and paid from a wallet they were never given.
✅ AI Can Now: Hold millions of convincing chats with scammers in different voices and languages, and pull out the bank accounts they use.
❌ Still Can't: Shut a scam down on its own. A bank or a person still has to freeze the account or take the site down.
✅ AI Can Now: Find 73% of the errors planted in a research paper's main argument, in Sakana's tests.
❌ Still Can't: Reliably catch real mistakes. On papers that were later withdrawn, it named the exact problem only 16% of the time.
FROM THE WEB
How was this week's recap? Tap one below, then tell me what you'd change on the next screen. I read every answer.
How was today's edition?
Coming up: I'm testing Pebble's smart ring as an alternative to Amazon's Bee, and I'll tell you what I find.
Thank you for reading. We're all beginners in something. Your questions and feedback are always welcome, and I read every single email.
-James
By the way, this is the link if you liked the content and want to share with a friend.
Some * designated product links may be affiliate or referral links. As an Amazon Associate, I earn from qualifying purchases. This helps support the newsletter at no extra cost to you and Amazon makes a tiny hair less.
