OpenAI Built a Facebook Pixel and Pointed It at Your ChatGPT History

by Sable Whitford — September 20, 2026

An independent researcher just published findings on OpenAI's __obi cookie — a cross-site tracking mechanism that connects your browsing activity across the web to your ChatGPT account. Advertisers who buy ChatGPT ads install OpenAI's tracking code on their own sites, same playbook as a Meta pixel, sending page views, product searches, and purchase behavior back to OpenAI. Worth remembering what ChatGPT actually holds: health fears, relationship problems, career crises — conversational data that makes your search history look like a grocery list. That's what sits behind the ad-targeting system.

OpenAI Built a Facebook Pixel and Pointed It at Your ChatGPT History

by Sable Whitford — September 20, 2026

An independent researcher just published findings on OpenAI's __obi cookie — a cross-site tracking mechanism that connects your browsing activity across the web to your ChatGPT account. Advertisers who buy ChatGPT ads install OpenAI's tracking code on their own sites, same playbook as a Meta pixel, sending page views, product searches, and purchase behavior back to OpenAI. Worth remembering what ChatGPT actually holds: health fears, relationship problems, career crises — conversational data that makes your search history look like a grocery list. That's what sits behind the ad-targeting system.

What Shipped Today in AI and Hardware

Sunday and the model pipeline doesn't care about your weekend. Some context for what's moving:

China's major labs have shipped more open-weight model variants in 2026 than the rest of the world combined, and the pace keeps climbing.
"Open weights" now covers everything from full Apache 2.0 to research-only licenses that prohibit commercial use. Same words, very different legal reality. Read the LICENSE file before you build anything on top of it.
HBM (High Bandwidth Memory) sits on GPU packages and has been the main bottleneck in AI accelerator supply for two years. When yields improve here, the whole compute stack loosens up.
Speculative decoding lets a small fast model guess upcoming tokens while a larger model verifies them in parallel. You get big-model quality at a fraction of the wait.

Here's what landed today.

TODAY'S BIG MOMENT

What Shipped Today in AI and Hardware

Sunday and the model pipeline doesn't care about your weekend. Some context for what's moving:

China's major labs have shipped more open-weight model variants in 2026 than the rest of the world combined, and the pace keeps climbing.
"Open weights" now covers everything from full Apache 2.0 to research-only licenses that prohibit commercial use. Same words, very different legal reality. Read the LICENSE file before you build anything on top of it.
HBM (High Bandwidth Memory) sits on GPU packages and has been the main bottleneck in AI accelerator supply for two years. When yields improve here, the whole compute stack loosens up.
Speculative decoding lets a small fast model guess upcoming tokens while a larger model verifies them in parallel. You get big-model quality at a fraction of the wait.

Here's what landed today.

Qwen Image 2.1 Ships With a License Catch

Alibaba dropped a 7B-parameter model unifying text-to-image and image editing with native transparency and multi-reference support. Day-one Diffusers and ComfyUI integration. The license quietly switched from Apache 2.0 to non-commercial only. Read it before you build.

BitTorrent Mirror Keeps AI Models From Vanishing

Pirate Face distributes open AI models as magnet links with no single point of failure. If a hosting platform pulls a model, downloads fall back to the peer swarm, tagged "Rescued." Only MIT and Apache-2.0 licensed models qualify.

Speculative Decoding Triples Qwen 27B Inference Speed

Qwen3.8-27B-DFlash2 uses speculative decoding to hit 3.43x faster inference on the 27B model with no quality loss. If you're running local models and watching token counters tick, this is the one to pay attention to.

Seven-Month-Old Beijing Startup Crosses the Unicorn Line

Naive AI, founded in February by Tsinghua professor Dai Jifeng, hit a $1.42B valuation after raising $400M from Tencent and others. Under 100 employees. They modify existing open-weight models via mid-training and RL rather than pretraining from scratch. First LLM shipping this month.

Samsung HBM4 Yields Hit Golden Territory at 80%

Samsung plans to push HBM wafer input to ~250,000 next year, with HBM4 at 80% of shipments. Yields have reached 80% — golden yield territory. Each chip sells for about 2 million won, more than double HBM3E pricing.

The Vibes Are Off and Getting Worse