As Google I/O approaches, a little-known experimental project called Illuminate is subtly exposing what may turn out to be one of the most significant AI media improvements to date. A more comprehensive version of Illuminate, which has been providing audio summaries of academic articles for a while, was just released, revealing a site with these AI-generated descriptions. Nevertheless, feature flags continue to obscure the majority of its new features.
In the past, testers found that Illuminate enables the construction of personalized audio overviews, enabling users to change prompts, choose hosts, or even ignore the entire chat. Something far more significant now appears to be taking shape. The interface suggests capability for audio summaries of research papers as well as great novels like The Great Gatsby, Alice in Wonderland, and Frankenstein, using the same generation format, however it is still concealed. Additionally, there are experimental options that are not available to the general public, such as an Edit button, toggles for the captions, and even the ability to generate images for cover photographs.
The page’s most intriguing find is a brand-new section called Sparks that is listed as Early Preview. “Imagine any question could be instantly transformed into a short video, 100% AI-generated,” the description says. Here are some examples of vertical videos that cover a variety of subjects and are usually one to three minutes long. The term “100% AI-generated” implies that these movies are created by a single model that can output synchronized video and audio from input, doing away with the need for separate pipelines, even if the production tool isn’t publically accessible and appears to be limited to internal Google accounts.
The excellent quality of the data suggests a tie to Veo 3 or a multi-modal Gemini (Ultra?) form, even if we are unable to confirm the precise model underlying it. Furthermore, the NotebookLM video overview function is likely powered by the same tech stack because it is confirmed to include two AI hosts and has a similar format. In that case, Video Overviews derived from uploaded sources and presented as fully produced dialog snippets may soon be supported by NotebookLM.
The preview of Sparks provides a clear indication of Google’s direction, which is toward seamless, multi-modal content generation from a single prompt, even though the majority of this is still theoretical and concealed behind feature flags.
