- The Digital Buff
- Posts
- The Anti-Rot Weekly Buff Agenda 🧠: A 1-year ChatGPT cookie, Meta's latest research + more
The Anti-Rot Weekly Buff Agenda 🧠: A 1-year ChatGPT cookie, Meta's latest research + more
Plus: bring your ideas to life with me & Metricool, free, this Wednesday!
The Digital Buff
September 28, 2026
“While we teach, we learn.” - Seneca
This Just In From The Buff Desk ☕
Your tabs multiplied. I narrowed them down, and here’s what actually mattered.
Pinterest expanded its NVIDIA partnership to rebuild the AI behind search and shopping on Blackwell GPUs, and the upgrade is dramatic: Pinterest Assistant can now take in 25x more visual context per request, with responses starting about 85x faster.
96% of Pinterest searches are unbranded, which makes it one of the last places people shop before they've picked a brand. The AI deciding who wins that moment just got much smarter, and it's worth a spot in your 2027 discovery plan.
Made On YouTube gave creators text-prompt editing for Shorts, Shorts organized into seasons and episodes (microdramas have officially arrived), and "Ask Studio," an AI assistant that drafts brand pitches and media kits. The stat to remember: over half of YouTube's top 100 creators now get most of their views on TV. Short-form is going serialized, and the living room is the new feed.
An independent researcher traced how ChatGPT's ad system can connect your ChatGPT account to the brand websites you visit afterward.
Here's how it works. ChatGPT sets a cookie called __obi that lasts a full year and is configured to travel with you onto other sites. Brands advertising on ChatGPT install OpenAI's tracking pixel, which sends that cookie back to OpenAI along with what you bought or signed up for.
The part that's getting attention: OpenAI files this cookie under "Analytics," not "Marketing." So if you accept analytics cookies but refuse marketing ones, it still gets set and sent.
Must-Read Research Paper: MaD-RL: Matching Distributions for Calibrating LLMs with Reinforcement Learning
Ever asked AI for ten campaign ideas and gotten one idea wearing ten different outfits? Same. Turns out there's a reason.
Most models are trained by rewarding the "right" answer. So over time, they quietly pick a favorite and keep coming back to it. The weird part: train the same model twice, and it can land on a different favorite each time. You don't get a say in which one. This paper gives you the knife. With MaD-RL, you decide the mix: "give me 40% of this, 20% of that," and the model learns to hit those proportions while still getting the answer right.
Why I'm obsessed as an AI Marketer: synthetic personas, simulated audiences and AI-generated survey data are only as good as their range. If your AI keeps defaulting to the loudest average customer, you're not modeling your audience. You're modeling a stereotype. 👀
Challenge Set :
Three things to actually try this week!
Gallup x Microsoft's global AI survey - 37 countries were asked how they feel about AI & 34 feel positive. The US? Only 36% think it will make life better. The Plot twist: the people using AI the most worry about it the least.  | Jensen Huang on The Ezra Klein Show - Ezra went to NVIDIA HQ to ask Jensen Huang if the AI doom talk has gone too far. Hear how the CEO whose chips fuel almost every AI tool you use, sees what comes next.  | Whimsical AI and The Art of Vibe Coding I'm joining Mya Shell for a free masterclass hosted by Metricool on vibe coding: building real, working tools just by describing what you want. Come build your ideas for FREE.  |
 | Buff Picks: This year’s Marketing Brew Summit is getting into a pretty accurate list of everything marketers are trying to figure out right now. Featuring speakers from Dove, Reddit, IBM, and Profound. |
Looking for rooms worth showing up for?! Head to thedigitalbuff.com and check the events tab before you miss another one. 👀