You click an ad inside ChatGPT, land on a shop, and leave. Here's the part you didn't see: according to an independent researcher, a cookie on OpenAI's domain may have followed you onto that shop's site, tied to your ChatGPT account. That's the story I can't stop thinking about this week, because ChatGPT is quietly turning into an ad network, and ad networks run on tracking.
Also on the list: Tencent shipped an image model that eats 20 reference photos at once, xAI put out a cheaper Grok for serious work, Alibaba announced a chip and a model plan the size of a small country's power grid, and a Canadian province took OpenAI to court. I read all of it so you don't have to. Here's what matters if you sell online.
ChatGPT's ad cookie reportedly follows people onto advertisers' sites
On September 21, an independent researcher published an analysis of a cookie called __obi. It lives on openai.com and, per the report, links a ChatGPT user's browser to websites that have installed OpenAI's advertising pixel. Security outlets that picked it up say the behaviour was reproduced on a phone and seen across 936 advertiser pixels on more than 1,000 hostnames.
The awkward detail: OpenAI's cookie policy lists __obi as an analytics cookie, not a marketing one. The researcher says the sync tokens showed up even when marketing consent had been denied, and that logged-out visitors got a device identifier that stayed stable for at least 27 days. Safari and other WebKit browsers block this kind of third-party cookie, so iPhone users are mostly out of the picture.
Why you should care: this is what a grown-up ad business looks like. Yesterday I wrote about ChatGPT's new conversational ads. A pixel that connects a chat to a purchase on your site is the other half of that machine. If you ever advertise on ChatGPT, you'll be offered this pixel, and your customers' trust will ride on how it behaves. My take: read the consent settings before you paste anyone's tracking code into your store, and expect regulators in Europe to ask the same questions loudly.
Tencent's new image model takes 20 reference photos
Tencent launched Hy Image 3.5 Preview on September 22. The headline numbers: up to 20 reference images per job (the old version took 3), native output up to 4096×4096, and multi-round editing, so you can keep refining the same picture instead of starting over. Tencent says it's about 30% better overall than the last version. The market liked it; Tencent shares rose more than 7% in Hong Kong.
For sellers, the number that matters is 20. More references means the model can see your product from every angle, plus your model, plus your background, and keep them consistent. That's exactly where product-photo AI usually breaks: the bag's stitching changes, the label drifts, the colour shifts. It's free to try on Tencent's Miora design agent until October 7.
Honest caveat: it's a preview, and "better than the competition" claims from a launch day are marketing until real users test them. But the direction is clear. Consistent product shots are getting cheap. If you still pay per photo shoot for every catalogue update, it's worth running a side-by-side test this month.
Grok 4.7: more brain, same price
xAI released Grok 4.7 on September 21. It's built for coding, long reasoning and professional work, and xAI kept the price where Grok 4.6 was: $2 per million input tokens and $6 per million output. It went live in Cursor and the xAI API right away and is rolling out in GitHub Copilot. Independent coverage points out it still trails the top models from OpenAI and Anthropic on several benchmarks.
So why is it in a sellers' digest? Because of the pattern, not the model. Every few weeks, somebody ships "almost frontier" capability at a lower price. That pressure is what makes AI tools for small shops affordable at all. You won't use Grok 4.7 directly to answer customers. You will feel it in what AI tools cost you next year.
Alibaba goes big: a new chip and a 10-trillion-parameter plan
At its Apsara conference in Hangzhou, Alibaba CEO Eddie Wu unveiled the Zhenwu V900 AI chip, which Alibaba says delivers three times the performance of its predecessor, with mass production planned for early 2027. He also said Alibaba plans to train a model with 5 to 10 trillion parameters and wants more than 20 gigawatts of data-center capacity by 2032.
That's big-picture news, and I'll keep the take short. The AI race is no longer two or three American labs. Chinese companies are building their own chips, their own models and, as Tencent showed the same day, their own creative tools. More serious competitors means faster prices falling, and more choice for the tools you build a business on. It also means the market will stay noisy. Don't switch tools every time a new chart goes viral.
British Columbia sues OpenAI over a school shooting
This one is serious, and I'll report it plainly. The Canadian province of British Columbia has sued OpenAI in a US federal court over February's Tumbler Ridge school shooting, in which eight people were killed. The lawsuit alleges OpenAI's safety team flagged the shooter's ChatGPT activity months earlier and recommended contacting police, but leadership overruled them. OpenAI had earlier apologised for not alerting law enforcement. More than 30 other lawsuits have been filed in the same court.
The province wants compensation and a court order forcing OpenAI to change how it handles conversations that threaten violence. Whatever the court decides, the question is now on the table for every company that runs AI conversations: when does a machine have to hand things to a human?
You're not running a chatbot for millions of strangers, but the principle scales down to a shop. An AI that talks to your customers needs clear rules for when to stop and bring in a person: an angry customer, a payment dispute, anything that sounds like harm. I wrote about the escalation rules that decide when AI should hand off to a human; this week is a good week to check yours.
Quick takes
- Labels vs. Suno, round two. Universal and Sony filed a new lawsuit on September 18 saying Suno's v6 music model still builds on 60,202 of their recordings, even though Suno signed licences with Warner, BMG and Believe. If you put AI-generated music under your reels, keep an eye on this.
- OpenAI DevDay is September 29 in San Francisco. Expect new developer tools; the useful bits for sellers usually arrive in apps a few weeks later.
- Amazon Accelerate runs September 22 to 24 in Seattle, and AI tools for sellers are expected to be a big theme. I'll report what actually ships, not what the preview blogs promise.
If you're trying to decide which of these tools deserves a place in your shop, start with our breakdown of AI tools for Instagram sellers and what to adopt first.
The thread running through this week: AI is getting cheaper and more capable, and the rules about trust (who tracks whom, who owns the training data, when a human steps in) are being written in real time. Your customers feel that too. If you want an assistant that answers them on Instagram, WhatsApp and Telegram in their own language, knows when to call you in, and never tracks them around the web, see how Vardast works.