Google has introduced a new technology called R4T-Diffusion, designed to perform background searches up to 20 times faster than traditional methods. This innovation is intended to improve the speed and relevance of responses in Google's AI Mode, a feature that has been available in France since July. AI Mode is capable of breaking down complex questions into simpler searches, performing them simultaneously, and compiling the results into a single answer. This technique, known as "query fan-out," allows Google to launch several searches on behalf of the user. For instance, instead of sending a complex request directly to its index, the AI Mode can search separately for different methods, their difficulty, necessary equipment, cost, or differences in taste before synthesizing all of this information. Google provides an example of this process: for a search on group activities in a city, its system can decide to search independently for restaurants, bars, child-friendly activities, or different places to visit. Some of these intermediate searches are never typed by the user. R4T-Diffusion aims to improve this mechanism by generating multiple relevant sub-queries more efficiently, reducing both computational demand and time. The current challenge is to generate many relevant sub-queries without launching redundant searches. R4T-Diffusion is trained to produce queries that remain linked to the initial question while exploring genuinely different leads. The major novelty of R4T-Diffusion lies in its speed. Traditional language models generally generate content progressively, whereas R4T-Diffusion works differently. Its small model with 53.9 million parameters can produce several search paths in parallel. According to measurements published by Google, this method would be 12 to 20 times faster than the autoregressive approaches used as a reference. On the most important workloads tested by researchers, traditional methods can approach 50 seconds, while R4T-Diffusion remains under a second or just a few seconds. However, these figures concern only a specific step in the process, during which the AI decides which additional searches it needs to perform. For the user, the primary benefit is that Google could explore more leads without the time required to answer a complex question exploding. With the AI Mode and AI Insights, Google is already handling questions much longer and more complex than a traditional search. R4T-Diffusion could help it explore more leads in parallel, especially to plan a trip, compare products, or answer a request with many criteria. Google presents its technology as suitable for large-scale use, without specifying whether it is already used in Search. If deployed, its effect should mainly be invisible: obtaining a complete answer faster where it previously required several searches in sequence.