Tencent's New AI Squeezes Your Whole Browsing History Into a Note

Every time you scroll a shopping app or a video feed, the company behind it is trying to guess what you'll want next. To do that, its computers read through your history — several hundred items, which turn into tens of thousands of words of text for the AI to chew on. With a billion users getting refreshed every week, that adds up to an enormous electricity and chip bill. Tencent's researchers say that approach is hitting a wall, so they built something to shrink it.

Their new system, KuaFu, compresses each individual thing you did — one click, one watch, one purchase — into just two to four tiny shorthand tokens, the basic units AI models read. Think of it like turning a page-long diary entry about you into a couple of sticky notes. The storage for each item drops from about 10 kilobytes, roughly a small photo, down to 0.5 kilobytes. The authors also warn that sloppy compression makes AI quietly invent things about you — made-up interests, skipped details, wrong dates — so they built checks along the way.

The results are notable. KuaFu matched or beat Tencent's older, uncompressed models on all five of their key measures, while letting each graphics chip do 37% to 350% more work. Tencent says it now saves 190 GPUs, expensive chips that cost thousands of dollars each. On public tests, a smaller version of the model even beat a bigger one. And on Tencent's live ad and recommendation platform, ten months of running it lifted overall merchandise sales by 1.37%.

The catch: this is Tencent reporting its own numbers, and the real prize is sharper ad targeting — meaning the system knows your habits better, even if it stores less of them in full. Expect other big platforms to copy the trick.