Sponsored by
What would you delegate if you had a whole team?
With Skydive, you can build a team of AI agents to take work off your plate.
Create agents for customer support, sales, marketing, engineering, ops, and more. Give each one a role, connect the tools they need, and hand off the work.
Your agents can work on their own or together, sharing context and handing off tasks to get bigger jobs done.
Start with one agent. Build a whole team around you.
Beginners in AI
Good morning and thank you for joining us again!
Last week ChatGPT started showing ads. This week a researcher found the code behind those ads showing up on shopping sites and reporting back to OpenAI. Today's lead walks through what it collects and the one browser setting worth changing. As always, every story here is human picked and edited.
THE FRONT PAGE
ChatGPT's Ad Tracker Follows You to Other Websites
TLDR: A security researcher says OpenAI's new ad tracker links what you do on sites like Wayfair and Chewy back to your ChatGPT account, using a cookie that lasts a full year.
The Story:
Businesses that buy ChatGPT ads put a small piece of OpenAI code on their own websites to track results, the same way stores already use Meta and Google trackers. On Sunday, a researcher who writes as Buchodi showed that this code sends an ID tied to your ChatGPT account back to OpenAI from ordinary shopping sites. On his phone, one ID went out from 12 of them, including Chewy, Wayfair, HelloFresh, Coursera and SeatGeek.
The code also grabs emails and phone numbers from the page, scrambled first, and sends your city and zip code as plain text. It works when you're logged out. It's filed under "analytics," so saying no to marketing doesn't stop it. He says OpenAI's support team got his questions on September 14 and didn't answer them.
Its Significance:
Meta has run this playbook for years, and the researcher says so himself. What's different is what people type into a chatbot: health worries, money trouble, fights with the boss. He found page addresses in the data pointing to a medical condition, a debt-help signup and a legal intake form. His testing only covered Chrome on Android. iPhone browsers block it, and he never saw OpenAI match the data to accounts on its own servers. If you use ChatGPT on Android, block third-party cookies in your browser settings today.
QUICK TAKES
The story: After about eight hours of talks in New York on Sunday, Treasury Secretary Scott Bessent said the US proposed a US-China AI dialogue. It would include a system for warning each other about AI incidents serious enough to be a national security issue. Trade chief Jamieson Greer said chip export limits aren't part of it, and Trump and Xi meet in Washington on September 24.
Your takeaway: Trump and Xi talked about an AI forum in May, and it never got off the ground. A warning system is a small, specific step, which makes it more likely to survive. Watch Thursday's summit for whether it becomes real.
The story: On CBS Sunday Morning, Jensen Huang said there's a 0% chance AI ends the world by 2030. He called scaring people irresponsible, and added that companies should never ship unsafe products. He was answering Evan Hubinger, a former Anthropic researcher who put the risk above 10% within a decade.
Your takeaway: Huang sells the chips everyone is racing to buy, so zero is a handy number for him. Neither figure comes from anything you can measure. Treat both as opinions from people with a stake in the answer.
The story: A Financial Times investigation found tech giants are backing roughly $300 billion in AI data center and chip debt. They do it through guarantees instead of borrowing the money directly. Alphabet's guarantees jumped from $16.9 billion to $43.8 billion in six months, and less than 2% of that shows up on its balance sheet.
Your takeaway: It's legal, standard accounting. But it means the debt numbers you see in earnings reports don't show how much is riding on AI demand staying high. If you own an index fund, you own a piece of this bet.
TOOLS ON OUR RADAR
🎙️ Jamie Freemium: Get AI meeting notes, transcripts and action items without a bot joining your call, free for 10 meetings a month. After today's lead, one detail stands out: Jamie says it deletes the audio once your notes are ready and never trains AI on your data.
📚 Readwise Paid: Pull your Kindle, article and PDF highlights into one place, then get a daily email that brings them back so you remember what you read.
⏱️ Clockify Freemium: Track hours across projects with a one-click timer and weekly timesheets, free for up to 5 people.
📅 Cal.com Free and Open Source: Share a booking link so people pick a time that works, with unlimited calendars and event types free for one person, alternative to Calendly.
TRENDING
Doctors tripled their AI use and drew a hard line at the same time. The FT reports clinicians are pushing back on AI beyond reading scans, citing thin real-world evidence. In a survey of 355 US doctors and nurses, 74% worried about losing their own skills and 74% about made-up answers. Another 72% worried ads would bend medical advice, which is worth remembering after today's lead.
A Swedish drone found, ranked and bombed a target with no human input. Scaleout Systems ran the demo for BAE Systems Bofors on a small Nvidia board that hobbyists can buy. The mission took under 320 seconds with no outside signal, and a pilot stayed on only as a backup. A UN expert group says human judgment is required under the laws of war, and it meets in Geneva in November.
Apple's Siri-powered home screen could arrive next month. Bloomberg's Mark Gurman says the device, code-named J490, is being tested in employees' homes. It's a small square display that recognizes whoever is looking at it and shows their calendar and notes. Apple has had it planned since 2024 and waited on Siri to catch up.
GPT-6 Astra cracked a German radio message from 1918. A developer named Prinz pulled it from a public list of 50 unsolved ciphers. The decoded message warns of a British cruiser at Sevastopol, and it matches the HMS Canterbury's log for November 24, 1918. Not bad for a message nobody had read in 108 years.
A startup raised about £15 million to send AI copies of musicians on tour. Unit1, run by the former CEO of Andrew Lloyd Webber's company, builds lifelike digital versions of retired and dead artists. Balderton Capital led the round, and U2's former manager Paul McGuinness joined. The pitch is a cheaper version of ABBA Voyage that can play in several countries at once.
TRY THIS PROMPT (copy and paste into Claude, Grok, ChatGPT, or Gemini)
🏭 Type in your town and find out which AI data centers are going up nearby, and how much power and water they use.
Build a single-file HTML app with vanilla HTML/CSS/JS. Data Center Next Door: enter a town, get real nearby AI and cloud data centers with their reported power and water use, found via live web search. No API key needed.
Aesthetic: deep slate background (#0f1720), electric amber (#f2b63c) for power with a glow top-left, cooling-water cyan (#4fc3d9) for water with a glow bottom-right. Archivo for headings and facility names, Source Serif 4 (italic for notes) for body, JetBrains Mono for small data labels.
Form: one text input for a town and state, and a "Find data centers" button.
Technical: first geocode the town with the free Open-Meteo geocoding API (https://geocoding-api.open-meteo.com/v1/search?name=...) to confirm the place and get coordinates. Then call the Anthropic Messages API WITH tools:[{type:'web_search_20250305', name:'web_search'}] (works automatically). The system prompt is a local infrastructure researcher who searches news, utility and county filings, and company announcements for REAL data centers (operating, under construction or proposed) within about 50 miles. It must never invent a facility or number. It includes power and water figures only when a source states them, and says whose figure it is. It returns found:false if nothing credible turns up. Return raw JSON: found, summary (2-3 sentences), facilities (max 5: name, operator, status, where, power, water, note), local_context (1-2 sentences on reported effects on bills, grid, water, taxes or jobs).
Render: the confirmed place name and summary. Then one block per facility: a status badge (gray = operating, amber = under construction, cyan = proposed), the operator, and two side-by-side gauges for Power (amber border) and Water (cyan border) that show "Not reported" when empty. Then an italic note, a bordered "What it means locally" box, and a footer saying the figures come from the named sources and can change. Include a loading state with a pulsing ⚡ and an inline error box.What this does: Enter a town or county. The app confirms the location, then searches the live web for real data centers within about 50 miles, whether they're running, under construction or only proposed. Each one shows its operator, status, and power and water figures, but only when a company, utility or news report has published them. Otherwise it says "Not reported." It ends with what's been reported locally about electric bills, the grid, water or jobs, and it says so plainly when nothing turns up.
What this looks like:
One idea shouldn't take six rewrites to post.
Posting everywhere means rewriting one idea six times, so you post to one, or none. SureThing turns one idea into native posts for every platform.
WHERE WE STAND(based on today’s news)
✅ AI Can Now: Break a 108-year-old military cipher and check its answer against real ship logs.
❌ Still Can't: Explain every loose end. The key it used wasn't issued until December 9, 1918, two weeks after the message was sent, and that gap still needs a human historian.
✅ AI Can Now: Spot, rank and fly to a target on a small hobby-grade chip with no connection to anyone.
❌ Still Can't: Set its own mission. People chose the target type and rules ahead of time, and a pilot stayed on as a backup.
RECOMMENDED LISTENING/READING/WATCHING
If today's lead made you think twice about what you type into ChatGPT, start here. Reporter Joseph Cox got hold of documents about the contractors who review real users' prompts, and he saw some of those prompts himself. The episode came out September 16 and pairs well with the ad tracker story: one is about who sees your chats, the other is about who sees your shopping.
How was today's edition? Tap one below. Pick "Could be better" and tell me what to fix; I read every answer.
How was today's edition?
Thank you for reading. We’re all beginners in something. With that in mind, 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.
