Meta Made Its Ad AI Faster Using One-Third the Code

Meta has rebuilt one small but crucial engine inside the AI that powers its advertising. That engine handles "attention" — the part of an AI that decides which pieces of information matter most. Meta's ad AI reads user activity in uneven chunks, like a stack of notes of different lengths, so the engine has to work efficiently without wasting effort on blank space. The old version worked, but it was slow and fussy to improve.

Why should you care? Because this is the kind of unglamorous engineering that quietly decides how expensive AI is to run. Meta says the new version is about 13% faster at making predictions and roughly 50% faster at the learning step behind the scenes. It also needs only about 3,200 lines of code, down from 10,000 — meaning engineers can tweak and improve it far more easily. When AI runs cheaper, the company can train bigger models or serve more users for the same money.

There is a real catch. This is internal research code tested on one specific type of expensive chip (Nvidia's Blackwell), on Meta's own ad model. It is not a tool you can download and use, and the impressive numbers come from carefully chosen situations that favor Meta's setup. Competitors may see smaller gains, or none, on different hardware.

The bigger story is a pattern worth noticing: AI companies are now competing not just on who has the smartest model, but on who can run it most cheaply and change it most quickly. Meta is essentially saying it found a way to get top performance without the painful, hard-to-edit coding style that usually requires. If that holds up, it means faster experimentation — and faster experimentation is how AI products improve. For everyday users, the practical effect shows up as ads that respond more quickly to what you do, and as AI features that cost companies less to keep online.